1 /* 2 * Copyright (c) 1997, 2014, 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 "libadt/vectset.hpp" 27 #include "memory/allocation.inline.hpp" 28 #include "compiler/compilerDirectives.hpp" 29 #include "opto/block.hpp" 30 #include "opto/cfgnode.hpp" 31 #include "opto/chaitin.hpp" 32 #include "opto/loopnode.hpp" 33 #include "opto/machnode.hpp" 34 #include "opto/matcher.hpp" 35 #include "opto/opcodes.hpp" 36 #include "opto/rootnode.hpp" 37 #include "utilities/copy.hpp" 38 39 void Block_Array::grow( uint i ) { 40 assert(i >= Max(), "must be an overflow"); 41 debug_only(_limit = i+1); 42 if( i < _size ) return; 43 if( !_size ) { 44 _size = 1; 45 _blocks = (Block**)_arena->Amalloc( _size * sizeof(Block*) ); 46 _blocks[0] = NULL; 47 } 48 uint old = _size; 49 while( i >= _size ) _size <<= 1; // Double to fit 50 _blocks = (Block**)_arena->Arealloc( _blocks, old*sizeof(Block*),_size*sizeof(Block*)); 51 Copy::zero_to_bytes( &_blocks[old], (_size-old)*sizeof(Block*) ); 52 } 53 54 void Block_List::remove(uint i) { 55 assert(i < _cnt, "index out of bounds"); 56 Copy::conjoint_words_to_lower((HeapWord*)&_blocks[i+1], (HeapWord*)&_blocks[i], ((_cnt-i-1)*sizeof(Block*))); 57 pop(); // shrink list by one block 58 } 59 60 void Block_List::insert(uint i, Block *b) { 61 push(b); // grow list by one block 62 Copy::conjoint_words_to_higher((HeapWord*)&_blocks[i], (HeapWord*)&_blocks[i+1], ((_cnt-i-1)*sizeof(Block*))); 63 _blocks[i] = b; 64 } 65 66 #ifndef PRODUCT 67 void Block_List::print() { 68 for (uint i=0; i < size(); i++) { 69 tty->print("B%d ", _blocks[i]->_pre_order); 70 } 71 tty->print("size = %d\n", size()); 72 } 73 #endif 74 75 uint Block::code_alignment() { 76 // Check for Root block 77 if (_pre_order == 0) return CodeEntryAlignment; 78 // Check for Start block 79 if (_pre_order == 1) return InteriorEntryAlignment; 80 // Check for loop alignment 81 if (has_loop_alignment()) return loop_alignment(); 82 83 return relocInfo::addr_unit(); // no particular alignment 84 } 85 86 uint Block::compute_loop_alignment() { 87 Node *h = head(); 88 int unit_sz = relocInfo::addr_unit(); 89 if (h->is_Loop() && h->as_Loop()->is_inner_loop()) { 90 // Pre- and post-loops have low trip count so do not bother with 91 // NOPs for align loop head. The constants are hidden from tuning 92 // but only because my "divide by 4" heuristic surely gets nearly 93 // all possible gain (a "do not align at all" heuristic has a 94 // chance of getting a really tiny gain). 95 if (h->is_CountedLoop() && (h->as_CountedLoop()->is_pre_loop() || 96 h->as_CountedLoop()->is_post_loop())) { 97 return (OptoLoopAlignment > 4*unit_sz) ? (OptoLoopAlignment>>2) : unit_sz; 98 } 99 // Loops with low backedge frequency should not be aligned. 100 Node *n = h->in(LoopNode::LoopBackControl)->in(0); 101 if (n->is_MachIf() && n->as_MachIf()->_prob < 0.01) { 102 return unit_sz; // Loop does not loop, more often than not! 103 } 104 return OptoLoopAlignment; // Otherwise align loop head 105 } 106 107 return unit_sz; // no particular alignment 108 } 109 110 // Compute the size of first 'inst_cnt' instructions in this block. 111 // Return the number of instructions left to compute if the block has 112 // less then 'inst_cnt' instructions. Stop, and return 0 if sum_size 113 // exceeds OptoLoopAlignment. 114 uint Block::compute_first_inst_size(uint& sum_size, uint inst_cnt, 115 PhaseRegAlloc* ra) { 116 uint last_inst = number_of_nodes(); 117 for( uint j = 0; j < last_inst && inst_cnt > 0; j++ ) { 118 uint inst_size = get_node(j)->size(ra); 119 if( inst_size > 0 ) { 120 inst_cnt--; 121 uint sz = sum_size + inst_size; 122 if( sz <= (uint)OptoLoopAlignment ) { 123 // Compute size of instructions which fit into fetch buffer only 124 // since all inst_cnt instructions will not fit even if we align them. 125 sum_size = sz; 126 } else { 127 return 0; 128 } 129 } 130 } 131 return inst_cnt; 132 } 133 134 uint Block::find_node( const Node *n ) const { 135 for( uint i = 0; i < number_of_nodes(); i++ ) { 136 if( get_node(i) == n ) 137 return i; 138 } 139 ShouldNotReachHere(); 140 return 0; 141 } 142 143 // Find and remove n from block list 144 void Block::find_remove( const Node *n ) { 145 remove_node(find_node(n)); 146 } 147 148 bool Block::contains(const Node *n) const { 149 return _nodes.contains(n); 150 } 151 152 // Return empty status of a block. Empty blocks contain only the head, other 153 // ideal nodes, and an optional trailing goto. 154 int Block::is_Empty() const { 155 156 // Root or start block is not considered empty 157 if (head()->is_Root() || head()->is_Start()) { 158 return not_empty; 159 } 160 161 int success_result = completely_empty; 162 int end_idx = number_of_nodes() - 1; 163 164 // Check for ending goto 165 if ((end_idx > 0) && (get_node(end_idx)->is_MachGoto())) { 166 success_result = empty_with_goto; 167 end_idx--; 168 } 169 170 // Unreachable blocks are considered empty 171 if (num_preds() <= 1) { 172 return success_result; 173 } 174 175 // Ideal nodes are allowable in empty blocks: skip them Only MachNodes 176 // turn directly into code, because only MachNodes have non-trivial 177 // emit() functions. 178 while ((end_idx > 0) && !get_node(end_idx)->is_Mach()) { 179 end_idx--; 180 } 181 182 // No room for any interesting instructions? 183 if (end_idx == 0) { 184 return success_result; 185 } 186 187 return not_empty; 188 } 189 190 // Return true if the block's code implies that it is likely to be 191 // executed infrequently. Check to see if the block ends in a Halt or 192 // a low probability call. 193 bool Block::has_uncommon_code() const { 194 Node* en = end(); 195 196 if (en->is_MachGoto()) 197 en = en->in(0); 198 if (en->is_Catch()) 199 en = en->in(0); 200 if (en->is_MachProj() && en->in(0)->is_MachCall()) { 201 MachCallNode* call = en->in(0)->as_MachCall(); 202 if (call->cnt() != COUNT_UNKNOWN && call->cnt() <= PROB_UNLIKELY_MAG(4)) { 203 // This is true for slow-path stubs like new_{instance,array}, 204 // slow_arraycopy, complete_monitor_locking, uncommon_trap. 205 // The magic number corresponds to the probability of an uncommon_trap, 206 // even though it is a count not a probability. 207 return true; 208 } 209 } 210 211 int op = en->is_Mach() ? en->as_Mach()->ideal_Opcode() : en->Opcode(); 212 return op == Op_Halt; 213 } 214 215 // True if block is low enough frequency or guarded by a test which 216 // mostly does not go here. 217 bool PhaseCFG::is_uncommon(const Block* block) { 218 // Initial blocks must never be moved, so are never uncommon. 219 if (block->head()->is_Root() || block->head()->is_Start()) return false; 220 221 // Check for way-low freq 222 if(block->_freq < BLOCK_FREQUENCY(0.00001f) ) return true; 223 224 // Look for code shape indicating uncommon_trap or slow path 225 if (block->has_uncommon_code()) return true; 226 227 const float epsilon = 0.05f; 228 const float guard_factor = PROB_UNLIKELY_MAG(4) / (1.f - epsilon); 229 uint uncommon_preds = 0; 230 uint freq_preds = 0; 231 uint uncommon_for_freq_preds = 0; 232 233 for( uint i=1; i< block->num_preds(); i++ ) { 234 Block* guard = get_block_for_node(block->pred(i)); 235 // Check to see if this block follows its guard 1 time out of 10000 236 // or less. 237 // 238 // See list of magnitude-4 unlikely probabilities in cfgnode.hpp which 239 // we intend to be "uncommon", such as slow-path TLE allocation, 240 // predicted call failure, and uncommon trap triggers. 241 // 242 // Use an epsilon value of 5% to allow for variability in frequency 243 // predictions and floating point calculations. The net effect is 244 // that guard_factor is set to 9500. 245 // 246 // Ignore low-frequency blocks. 247 // The next check is (guard->_freq < 1.e-5 * 9500.). 248 if(guard->_freq*BLOCK_FREQUENCY(guard_factor) < BLOCK_FREQUENCY(0.00001f)) { 249 uncommon_preds++; 250 } else { 251 freq_preds++; 252 if(block->_freq < guard->_freq * guard_factor ) { 253 uncommon_for_freq_preds++; 254 } 255 } 256 } 257 if( block->num_preds() > 1 && 258 // The block is uncommon if all preds are uncommon or 259 (uncommon_preds == (block->num_preds()-1) || 260 // it is uncommon for all frequent preds. 261 uncommon_for_freq_preds == freq_preds) ) { 262 return true; 263 } 264 return false; 265 } 266 267 #ifndef PRODUCT 268 void Block::dump_bidx(const Block* orig, outputStream* st) const { 269 if (_pre_order) st->print("B%d",_pre_order); 270 else st->print("N%d", head()->_idx); 271 272 if (Verbose && orig != this) { 273 // Dump the original block's idx 274 st->print(" ("); 275 orig->dump_bidx(orig, st); 276 st->print(")"); 277 } 278 } 279 280 void Block::dump_pred(const PhaseCFG* cfg, Block* orig, outputStream* st) const { 281 if (is_connector()) { 282 for (uint i=1; i<num_preds(); i++) { 283 Block *p = cfg->get_block_for_node(pred(i)); 284 p->dump_pred(cfg, orig, st); 285 } 286 } else { 287 dump_bidx(orig, st); 288 st->print(" "); 289 } 290 } 291 292 void Block::dump_head(const PhaseCFG* cfg, outputStream* st) const { 293 // Print the basic block 294 dump_bidx(this, st); 295 st->print(": #\t"); 296 297 // Print the incoming CFG edges and the outgoing CFG edges 298 for( uint i=0; i<_num_succs; i++ ) { 299 non_connector_successor(i)->dump_bidx(_succs[i], st); 300 st->print(" "); 301 } 302 st->print("<- "); 303 if( head()->is_block_start() ) { 304 for (uint i=1; i<num_preds(); i++) { 305 Node *s = pred(i); 306 if (cfg != NULL) { 307 Block *p = cfg->get_block_for_node(s); 308 p->dump_pred(cfg, p, st); 309 } else { 310 while (!s->is_block_start()) 311 s = s->in(0); 312 st->print("N%d ", s->_idx ); 313 } 314 } 315 } else { 316 st->print("BLOCK HEAD IS JUNK "); 317 } 318 319 // Print loop, if any 320 const Block *bhead = this; // Head of self-loop 321 Node *bh = bhead->head(); 322 323 if ((cfg != NULL) && bh->is_Loop() && !head()->is_Root()) { 324 LoopNode *loop = bh->as_Loop(); 325 const Block *bx = cfg->get_block_for_node(loop->in(LoopNode::LoopBackControl)); 326 while (bx->is_connector()) { 327 bx = cfg->get_block_for_node(bx->pred(1)); 328 } 329 st->print("\tLoop: B%d-B%d ", bhead->_pre_order, bx->_pre_order); 330 // Dump any loop-specific bits, especially for CountedLoops. 331 loop->dump_spec(st); 332 } else if (has_loop_alignment()) { 333 st->print(" top-of-loop"); 334 } 335 st->print(" Freq: %g",_freq); 336 if( Verbose || WizardMode ) { 337 st->print(" IDom: %d/#%d", _idom ? _idom->_pre_order : 0, _dom_depth); 338 st->print(" RegPressure: %d",_reg_pressure); 339 st->print(" IHRP Index: %d",_ihrp_index); 340 st->print(" FRegPressure: %d",_freg_pressure); 341 st->print(" FHRP Index: %d",_fhrp_index); 342 } 343 st->cr(); 344 } 345 346 void Block::dump() const { 347 dump(NULL); 348 } 349 350 void Block::dump(const PhaseCFG* cfg) const { 351 dump_head(cfg); 352 for (uint i=0; i< number_of_nodes(); i++) { 353 get_node(i)->dump(); 354 } 355 tty->print("\n"); 356 } 357 #endif 358 359 PhaseCFG::PhaseCFG(Arena* arena, RootNode* root, Matcher& matcher) 360 : Phase(CFG) 361 , _block_arena(arena) 362 , _regalloc(NULL) 363 , _scheduling_for_pressure(false) 364 , _root(root) 365 , _matcher(matcher) 366 , _node_to_block_mapping(arena) 367 , _node_latency(NULL) 368 #ifndef PRODUCT 369 , _trace_opto_pipelining(C->directive()->TraceOptoPipeliningOption) 370 #endif 371 #ifdef ASSERT 372 , _raw_oops(arena) 373 #endif 374 { 375 ResourceMark rm; 376 // I'll need a few machine-specific GotoNodes. Make an Ideal GotoNode, 377 // then Match it into a machine-specific Node. Then clone the machine 378 // Node on demand. 379 Node *x = new GotoNode(NULL); 380 x->init_req(0, x); 381 _goto = matcher.match_tree(x); 382 assert(_goto != NULL, ""); 383 _goto->set_req(0,_goto); 384 385 // Build the CFG in Reverse Post Order 386 _number_of_blocks = build_cfg(); 387 _root_block = get_block_for_node(_root); 388 } 389 390 // Build a proper looking CFG. Make every block begin with either a StartNode 391 // or a RegionNode. Make every block end with either a Goto, If or Return. 392 // The RootNode both starts and ends it's own block. Do this with a recursive 393 // backwards walk over the control edges. 394 uint PhaseCFG::build_cfg() { 395 Arena *a = Thread::current()->resource_area(); 396 VectorSet visited(a); 397 398 // Allocate stack with enough space to avoid frequent realloc 399 Node_Stack nstack(a, C->live_nodes() >> 1); 400 nstack.push(_root, 0); 401 uint sum = 0; // Counter for blocks 402 403 while (nstack.is_nonempty()) { 404 // node and in's index from stack's top 405 // 'np' is _root (see above) or RegionNode, StartNode: we push on stack 406 // only nodes which point to the start of basic block (see below). 407 Node *np = nstack.node(); 408 // idx > 0, except for the first node (_root) pushed on stack 409 // at the beginning when idx == 0. 410 // We will use the condition (idx == 0) later to end the build. 411 uint idx = nstack.index(); 412 Node *proj = np->in(idx); 413 const Node *x = proj->is_block_proj(); 414 // Does the block end with a proper block-ending Node? One of Return, 415 // If or Goto? (This check should be done for visited nodes also). 416 if (x == NULL) { // Does not end right... 417 Node *g = _goto->clone(); // Force it to end in a Goto 418 g->set_req(0, proj); 419 np->set_req(idx, g); 420 x = proj = g; 421 } 422 if (!visited.test_set(x->_idx)) { // Visit this block once 423 // Skip any control-pinned middle'in stuff 424 Node *p = proj; 425 do { 426 proj = p; // Update pointer to last Control 427 p = p->in(0); // Move control forward 428 } while( !p->is_block_proj() && 429 !p->is_block_start() ); 430 // Make the block begin with one of Region or StartNode. 431 if( !p->is_block_start() ) { 432 RegionNode *r = new RegionNode( 2 ); 433 r->init_req(1, p); // Insert RegionNode in the way 434 proj->set_req(0, r); // Insert RegionNode in the way 435 p = r; 436 } 437 // 'p' now points to the start of this basic block 438 439 // Put self in array of basic blocks 440 Block *bb = new (_block_arena) Block(_block_arena, p); 441 map_node_to_block(p, bb); 442 map_node_to_block(x, bb); 443 if( x != p ) { // Only for root is x == p 444 bb->push_node((Node*)x); 445 } 446 // Now handle predecessors 447 ++sum; // Count 1 for self block 448 uint cnt = bb->num_preds(); 449 for (int i = (cnt - 1); i > 0; i-- ) { // For all predecessors 450 Node *prevproj = p->in(i); // Get prior input 451 assert( !prevproj->is_Con(), "dead input not removed" ); 452 // Check to see if p->in(i) is a "control-dependent" CFG edge - 453 // i.e., it splits at the source (via an IF or SWITCH) and merges 454 // at the destination (via a many-input Region). 455 // This breaks critical edges. The RegionNode to start the block 456 // will be added when <p,i> is pulled off the node stack 457 if ( cnt > 2 ) { // Merging many things? 458 assert( prevproj== bb->pred(i),""); 459 if(prevproj->is_block_proj() != prevproj) { // Control-dependent edge? 460 // Force a block on the control-dependent edge 461 Node *g = _goto->clone(); // Force it to end in a Goto 462 g->set_req(0,prevproj); 463 p->set_req(i,g); 464 } 465 } 466 nstack.push(p, i); // 'p' is RegionNode or StartNode 467 } 468 } else { // Post-processing visited nodes 469 nstack.pop(); // remove node from stack 470 // Check if it the fist node pushed on stack at the beginning. 471 if (idx == 0) break; // end of the build 472 // Find predecessor basic block 473 Block *pb = get_block_for_node(x); 474 // Insert into nodes array, if not already there 475 if (!has_block(proj)) { 476 assert( x != proj, "" ); 477 // Map basic block of projection 478 map_node_to_block(proj, pb); 479 pb->push_node(proj); 480 } 481 // Insert self as a child of my predecessor block 482 pb->_succs.map(pb->_num_succs++, get_block_for_node(np)); 483 assert( pb->get_node(pb->number_of_nodes() - pb->_num_succs)->is_block_proj(), 484 "too many control users, not a CFG?" ); 485 } 486 } 487 // Return number of basic blocks for all children and self 488 return sum; 489 } 490 491 // Inserts a goto & corresponding basic block between 492 // block[block_no] and its succ_no'th successor block 493 void PhaseCFG::insert_goto_at(uint block_no, uint succ_no) { 494 // get block with block_no 495 assert(block_no < number_of_blocks(), "illegal block number"); 496 Block* in = get_block(block_no); 497 // get successor block succ_no 498 assert(succ_no < in->_num_succs, "illegal successor number"); 499 Block* out = in->_succs[succ_no]; 500 // Compute frequency of the new block. Do this before inserting 501 // new block in case succ_prob() needs to infer the probability from 502 // surrounding blocks. 503 float freq = in->_freq * in->succ_prob(succ_no); 504 // get ProjNode corresponding to the succ_no'th successor of the in block 505 ProjNode* proj = in->get_node(in->number_of_nodes() - in->_num_succs + succ_no)->as_Proj(); 506 // create region for basic block 507 RegionNode* region = new RegionNode(2); 508 region->init_req(1, proj); 509 // setup corresponding basic block 510 Block* block = new (_block_arena) Block(_block_arena, region); 511 map_node_to_block(region, block); 512 C->regalloc()->set_bad(region->_idx); 513 // add a goto node 514 Node* gto = _goto->clone(); // get a new goto node 515 gto->set_req(0, region); 516 // add it to the basic block 517 block->push_node(gto); 518 map_node_to_block(gto, block); 519 C->regalloc()->set_bad(gto->_idx); 520 // hook up successor block 521 block->_succs.map(block->_num_succs++, out); 522 // remap successor's predecessors if necessary 523 for (uint i = 1; i < out->num_preds(); i++) { 524 if (out->pred(i) == proj) out->head()->set_req(i, gto); 525 } 526 // remap predecessor's successor to new block 527 in->_succs.map(succ_no, block); 528 // Set the frequency of the new block 529 block->_freq = freq; 530 // add new basic block to basic block list 531 add_block_at(block_no + 1, block); 532 } 533 534 // Does this block end in a multiway branch that cannot have the default case 535 // flipped for another case? 536 static bool no_flip_branch(Block *b) { 537 int branch_idx = b->number_of_nodes() - b->_num_succs-1; 538 if (branch_idx < 1) { 539 return false; 540 } 541 Node *branch = b->get_node(branch_idx); 542 if (branch->is_Catch()) { 543 return true; 544 } 545 if (branch->is_Mach()) { 546 if (branch->is_MachNullCheck()) { 547 return true; 548 } 549 int iop = branch->as_Mach()->ideal_Opcode(); 550 if (iop == Op_FastLock || iop == Op_FastUnlock) { 551 return true; 552 } 553 // Don't flip if branch has an implicit check. 554 if (branch->as_Mach()->is_TrapBasedCheckNode()) { 555 return true; 556 } 557 } 558 return false; 559 } 560 561 // Check for NeverBranch at block end. This needs to become a GOTO to the 562 // true target. NeverBranch are treated as a conditional branch that always 563 // goes the same direction for most of the optimizer and are used to give a 564 // fake exit path to infinite loops. At this late stage they need to turn 565 // into Goto's so that when you enter the infinite loop you indeed hang. 566 void PhaseCFG::convert_NeverBranch_to_Goto(Block *b) { 567 // Find true target 568 int end_idx = b->end_idx(); 569 int idx = b->get_node(end_idx+1)->as_Proj()->_con; 570 Block *succ = b->_succs[idx]; 571 Node* gto = _goto->clone(); // get a new goto node 572 gto->set_req(0, b->head()); 573 Node *bp = b->get_node(end_idx); 574 b->map_node(gto, end_idx); // Slam over NeverBranch 575 map_node_to_block(gto, b); 576 C->regalloc()->set_bad(gto->_idx); 577 b->pop_node(); // Yank projections 578 b->pop_node(); // Yank projections 579 b->_succs.map(0,succ); // Map only successor 580 b->_num_succs = 1; 581 // remap successor's predecessors if necessary 582 uint j; 583 for( j = 1; j < succ->num_preds(); j++) 584 if( succ->pred(j)->in(0) == bp ) 585 succ->head()->set_req(j, gto); 586 // Kill alternate exit path 587 Block *dead = b->_succs[1-idx]; 588 for( j = 1; j < dead->num_preds(); j++) 589 if( dead->pred(j)->in(0) == bp ) 590 break; 591 // Scan through block, yanking dead path from 592 // all regions and phis. 593 dead->head()->del_req(j); 594 for( int k = 1; dead->get_node(k)->is_Phi(); k++ ) 595 dead->get_node(k)->del_req(j); 596 } 597 598 // Helper function to move block bx to the slot following b_index. Return 599 // true if the move is successful, otherwise false 600 bool PhaseCFG::move_to_next(Block* bx, uint b_index) { 601 if (bx == NULL) return false; 602 603 // Return false if bx is already scheduled. 604 uint bx_index = bx->_pre_order; 605 if ((bx_index <= b_index) && (get_block(bx_index) == bx)) { 606 return false; 607 } 608 609 // Find the current index of block bx on the block list 610 bx_index = b_index + 1; 611 while (bx_index < number_of_blocks() && get_block(bx_index) != bx) { 612 bx_index++; 613 } 614 assert(get_block(bx_index) == bx, "block not found"); 615 616 // If the previous block conditionally falls into bx, return false, 617 // because moving bx will create an extra jump. 618 for(uint k = 1; k < bx->num_preds(); k++ ) { 619 Block* pred = get_block_for_node(bx->pred(k)); 620 if (pred == get_block(bx_index - 1)) { 621 if (pred->_num_succs != 1) { 622 return false; 623 } 624 } 625 } 626 627 // Reinsert bx just past block 'b' 628 _blocks.remove(bx_index); 629 _blocks.insert(b_index + 1, bx); 630 return true; 631 } 632 633 // Move empty and uncommon blocks to the end. 634 void PhaseCFG::move_to_end(Block *b, uint i) { 635 int e = b->is_Empty(); 636 if (e != Block::not_empty) { 637 if (e == Block::empty_with_goto) { 638 // Remove the goto, but leave the block. 639 b->pop_node(); 640 } 641 // Mark this block as a connector block, which will cause it to be 642 // ignored in certain functions such as non_connector_successor(). 643 b->set_connector(); 644 } 645 // Move the empty block to the end, and don't recheck. 646 _blocks.remove(i); 647 _blocks.push(b); 648 } 649 650 // Set loop alignment for every block 651 void PhaseCFG::set_loop_alignment() { 652 uint last = number_of_blocks(); 653 assert(get_block(0) == get_root_block(), ""); 654 655 for (uint i = 1; i < last; i++) { 656 Block* block = get_block(i); 657 if (block->head()->is_Loop()) { 658 block->set_loop_alignment(block); 659 } 660 } 661 } 662 663 // Make empty basic blocks to be "connector" blocks, Move uncommon blocks 664 // to the end. 665 void PhaseCFG::remove_empty_blocks() { 666 // Move uncommon blocks to the end 667 uint last = number_of_blocks(); 668 assert(get_block(0) == get_root_block(), ""); 669 670 for (uint i = 1; i < last; i++) { 671 Block* block = get_block(i); 672 if (block->is_connector()) { 673 break; 674 } 675 676 // Check for NeverBranch at block end. This needs to become a GOTO to the 677 // true target. NeverBranch are treated as a conditional branch that 678 // always goes the same direction for most of the optimizer and are used 679 // to give a fake exit path to infinite loops. At this late stage they 680 // need to turn into Goto's so that when you enter the infinite loop you 681 // indeed hang. 682 if (block->get_node(block->end_idx())->Opcode() == Op_NeverBranch) { 683 convert_NeverBranch_to_Goto(block); 684 } 685 686 // Look for uncommon blocks and move to end. 687 if (!C->do_freq_based_layout()) { 688 if (is_uncommon(block)) { 689 move_to_end(block, i); 690 last--; // No longer check for being uncommon! 691 if (no_flip_branch(block)) { // Fall-thru case must follow? 692 // Find the fall-thru block 693 block = get_block(i); 694 move_to_end(block, i); 695 last--; 696 } 697 // backup block counter post-increment 698 i--; 699 } 700 } 701 } 702 703 // Move empty blocks to the end 704 last = number_of_blocks(); 705 for (uint i = 1; i < last; i++) { 706 Block* block = get_block(i); 707 if (block->is_Empty() != Block::not_empty) { 708 move_to_end(block, i); 709 last--; 710 i--; 711 } 712 } // End of for all blocks 713 } 714 715 Block *PhaseCFG::fixup_trap_based_check(Node *branch, Block *block, int block_pos, Block *bnext) { 716 // Trap based checks must fall through to the successor with 717 // PROB_ALWAYS. 718 // They should be an If with 2 successors. 719 assert(branch->is_MachIf(), "must be If"); 720 assert(block->_num_succs == 2, "must have 2 successors"); 721 722 // Get the If node and the projection for the first successor. 723 MachIfNode *iff = block->get_node(block->number_of_nodes()-3)->as_MachIf(); 724 ProjNode *proj0 = block->get_node(block->number_of_nodes()-2)->as_Proj(); 725 ProjNode *proj1 = block->get_node(block->number_of_nodes()-1)->as_Proj(); 726 ProjNode *projt = (proj0->Opcode() == Op_IfTrue) ? proj0 : proj1; 727 ProjNode *projf = (proj0->Opcode() == Op_IfFalse) ? proj0 : proj1; 728 729 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1]. 730 assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0"); 731 assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1"); 732 733 ProjNode *proj_always; 734 ProjNode *proj_never; 735 // We must negate the branch if the implicit check doesn't follow 736 // the branch's TRUE path. Then, the new TRUE branch target will 737 // be the old FALSE branch target. 738 if (iff->_prob <= 2*PROB_NEVER) { // There are small rounding errors. 739 proj_never = projt; 740 proj_always = projf; 741 } else { 742 // We must negate the branch if the trap doesn't follow the 743 // branch's TRUE path. Then, the new TRUE branch target will 744 // be the old FALSE branch target. 745 proj_never = projf; 746 proj_always = projt; 747 iff->negate(); 748 } 749 assert(iff->_prob <= 2*PROB_NEVER, "Trap based checks are expected to trap never!"); 750 // Map the successors properly 751 block->_succs.map(0, get_block_for_node(proj_never ->raw_out(0))); // The target of the trap. 752 block->_succs.map(1, get_block_for_node(proj_always->raw_out(0))); // The fall through target. 753 754 if (block->get_node(block->number_of_nodes() - block->_num_succs + 1) != proj_always) { 755 block->map_node(proj_never, block->number_of_nodes() - block->_num_succs + 0); 756 block->map_node(proj_always, block->number_of_nodes() - block->_num_succs + 1); 757 } 758 759 // Place the fall through block after this block. 760 Block *bs1 = block->non_connector_successor(1); 761 if (bs1 != bnext && move_to_next(bs1, block_pos)) { 762 bnext = bs1; 763 } 764 // If the fall through block still is not the next block, insert a goto. 765 if (bs1 != bnext) { 766 insert_goto_at(block_pos, 1); 767 } 768 return bnext; 769 } 770 771 // Fix up the final control flow for basic blocks. 772 void PhaseCFG::fixup_flow() { 773 // Fixup final control flow for the blocks. Remove jump-to-next 774 // block. If neither arm of an IF follows the conditional branch, we 775 // have to add a second jump after the conditional. We place the 776 // TRUE branch target in succs[0] for both GOTOs and IFs. 777 bool found_fixup_loops = false; 778 for (uint i = 0; i < number_of_blocks(); i++) { 779 Block* block = get_block(i); 780 block->_pre_order = i; // turn pre-order into block-index 781 782 Node *bh = block->head(); 783 if (bh->is_Loop()) { 784 LoopNode *loop = bh->as_Loop(); 785 if (loop->is_inner_loop() && loop->is_multiversioned() && loop->is_vectorized_loop() && !loop->range_checks_present()) { 786 found_fixup_loops = true; 787 } 788 } 789 790 // Connector blocks need no further processing. 791 if (block->is_connector()) { 792 assert((i+1) == number_of_blocks() || get_block(i + 1)->is_connector(), "All connector blocks should sink to the end"); 793 continue; 794 } 795 assert(block->is_Empty() != Block::completely_empty, "Empty blocks should be connectors"); 796 797 Block* bnext = (i < number_of_blocks() - 1) ? get_block(i + 1) : NULL; 798 Block* bs0 = block->non_connector_successor(0); 799 800 // Check for multi-way branches where I cannot negate the test to 801 // exchange the true and false targets. 802 if (no_flip_branch(block)) { 803 // Find fall through case - if must fall into its target. 804 // Get the index of the branch's first successor. 805 int branch_idx = block->number_of_nodes() - block->_num_succs; 806 807 // The branch is 1 before the branch's first successor. 808 Node *branch = block->get_node(branch_idx-1); 809 810 // Handle no-flip branches which have implicit checks and which require 811 // special block ordering and individual semantics of the 'fall through 812 // case'. 813 if ((TrapBasedNullChecks || TrapBasedRangeChecks) && 814 branch->is_Mach() && branch->as_Mach()->is_TrapBasedCheckNode()) { 815 bnext = fixup_trap_based_check(branch, block, i, bnext); 816 } else { 817 // Else, default handling for no-flip branches 818 for (uint j2 = 0; j2 < block->_num_succs; j2++) { 819 const ProjNode* p = block->get_node(branch_idx + j2)->as_Proj(); 820 if (p->_con == 0) { 821 // successor j2 is fall through case 822 if (block->non_connector_successor(j2) != bnext) { 823 // but it is not the next block => insert a goto 824 insert_goto_at(i, j2); 825 } 826 // Put taken branch in slot 0 827 if (j2 == 0 && block->_num_succs == 2) { 828 // Flip targets in succs map 829 Block *tbs0 = block->_succs[0]; 830 Block *tbs1 = block->_succs[1]; 831 block->_succs.map(0, tbs1); 832 block->_succs.map(1, tbs0); 833 } 834 break; 835 } 836 } 837 } 838 839 // Remove all CatchProjs 840 for (uint j = 0; j < block->_num_succs; j++) { 841 block->pop_node(); 842 } 843 844 } else if (block->_num_succs == 1) { 845 // Block ends in a Goto? 846 if (bnext == bs0) { 847 // We fall into next block; remove the Goto 848 block->pop_node(); 849 } 850 851 } else if(block->_num_succs == 2) { // Block ends in a If? 852 // Get opcode of 1st projection (matches _succs[0]) 853 // Note: Since this basic block has 2 exits, the last 2 nodes must 854 // be projections (in any order), the 3rd last node must be 855 // the IfNode (we have excluded other 2-way exits such as 856 // CatchNodes already). 857 MachNode* iff = block->get_node(block->number_of_nodes() - 3)->as_Mach(); 858 ProjNode* proj0 = block->get_node(block->number_of_nodes() - 2)->as_Proj(); 859 ProjNode* proj1 = block->get_node(block->number_of_nodes() - 1)->as_Proj(); 860 861 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1]. 862 assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0"); 863 assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1"); 864 865 Block* bs1 = block->non_connector_successor(1); 866 867 // Check for neither successor block following the current 868 // block ending in a conditional. If so, move one of the 869 // successors after the current one, provided that the 870 // successor was previously unscheduled, but moveable 871 // (i.e., all paths to it involve a branch). 872 if (!C->do_freq_based_layout() && bnext != bs0 && bnext != bs1) { 873 // Choose the more common successor based on the probability 874 // of the conditional branch. 875 Block* bx = bs0; 876 Block* by = bs1; 877 878 // _prob is the probability of taking the true path. Make 879 // p the probability of taking successor #1. 880 float p = iff->as_MachIf()->_prob; 881 if (proj0->Opcode() == Op_IfTrue) { 882 p = 1.0 - p; 883 } 884 885 // Prefer successor #1 if p > 0.5 886 if (p > PROB_FAIR) { 887 bx = bs1; 888 by = bs0; 889 } 890 891 // Attempt the more common successor first 892 if (move_to_next(bx, i)) { 893 bnext = bx; 894 } else if (move_to_next(by, i)) { 895 bnext = by; 896 } 897 } 898 899 // Check for conditional branching the wrong way. Negate 900 // conditional, if needed, so it falls into the following block 901 // and branches to the not-following block. 902 903 // Check for the next block being in succs[0]. We are going to branch 904 // to succs[0], so we want the fall-thru case as the next block in 905 // succs[1]. 906 if (bnext == bs0) { 907 // Fall-thru case in succs[0], so flip targets in succs map 908 Block* tbs0 = block->_succs[0]; 909 Block* tbs1 = block->_succs[1]; 910 block->_succs.map(0, tbs1); 911 block->_succs.map(1, tbs0); 912 // Flip projection for each target 913 ProjNode* tmp = proj0; 914 proj0 = proj1; 915 proj1 = tmp; 916 917 } else if(bnext != bs1) { 918 // Need a double-branch 919 // The existing conditional branch need not change. 920 // Add a unconditional branch to the false target. 921 // Alas, it must appear in its own block and adding a 922 // block this late in the game is complicated. Sigh. 923 insert_goto_at(i, 1); 924 } 925 926 // Make sure we TRUE branch to the target 927 if (proj0->Opcode() == Op_IfFalse) { 928 iff->as_MachIf()->negate(); 929 } 930 931 block->pop_node(); // Remove IfFalse & IfTrue projections 932 block->pop_node(); 933 934 } else { 935 // Multi-exit block, e.g. a switch statement 936 // But we don't need to do anything here 937 } 938 } // End of for all blocks 939 940 if (found_fixup_loops) { 941 // find all fixup-loops and process them 942 for (uint i = 0; i < number_of_blocks(); i++) { 943 Block* block = get_block(i); 944 Node *bh = block->head(); 945 if (bh->is_Loop()) { 946 LoopNode *loop = bh->as_Loop(); 947 // fixup loops are only marked for processing when they are predicated and 948 // vectorized else they are just post loops. 949 if (Matcher::has_predicated_vectors()) { 950 if (loop->is_inner_loop() && loop->is_multiversioned() && loop->is_vectorized_loop() && !loop->range_checks_present()) { 951 CFGLoop *cur_loop = block->_loop; 952 // fixup loops can have multiple exits, so we need to find the backedge 953 Block *back_edge = cur_loop->backedge_block(); 954 if (back_edge) { 955 // fetch the region of the back edge 956 Node *backedge_region = back_edge->get_node(0); 957 Block *idom = back_edge->_idom; 958 if (backedge_region->is_Region()) { 959 Node *if_true = backedge_region->in(1); 960 if (if_true->Opcode() == Op_IfTrue) { 961 Node *backedge_iff = if_true->in(0); 962 if (backedge_iff->is_MachIf() && idom) { 963 for (uint j = 0; j < idom->number_of_nodes(); j++) { 964 Node *n = idom->get_node(j); 965 if (n == backedge_iff) { 966 MachMskNode *mask = new MachMskNode(true); 967 if (mask) { 968 idom->insert_node(mask, j); 969 map_node_to_block(mask, idom); 970 break; 971 } 972 } 973 } 974 } 975 } 976 } 977 } 978 } 979 } 980 } 981 } 982 } 983 } 984 985 986 // postalloc_expand: Expand nodes after register allocation. 987 // 988 // postalloc_expand has to be called after register allocation, just 989 // before output (i.e. scheduling). It only gets called if 990 // Matcher::require_postalloc_expand is true. 991 // 992 // Background: 993 // 994 // Nodes that are expandend (one compound node requiring several 995 // assembler instructions to be implemented split into two or more 996 // non-compound nodes) after register allocation are not as nice as 997 // the ones expanded before register allocation - they don't 998 // participate in optimizations as global code motion. But after 999 // register allocation we can expand nodes that use registers which 1000 // are not spillable or registers that are not allocated, because the 1001 // old compound node is simply replaced (in its location in the basic 1002 // block) by a new subgraph which does not contain compound nodes any 1003 // more. The scheduler called during output can later on process these 1004 // non-compound nodes. 1005 // 1006 // Implementation: 1007 // 1008 // Nodes requiring postalloc expand are specified in the ad file by using 1009 // a postalloc_expand statement instead of ins_encode. A postalloc_expand 1010 // contains a single call to an encoding, as does an ins_encode 1011 // statement. Instead of an emit() function a postalloc_expand() function 1012 // is generated that doesn't emit assembler but creates a new 1013 // subgraph. The code below calls this postalloc_expand function for each 1014 // node with the appropriate attribute. This function returns the new 1015 // nodes generated in an array passed in the call. The old node, 1016 // potential MachTemps before and potential Projs after it then get 1017 // disconnected and replaced by the new nodes. The instruction 1018 // generating the result has to be the last one in the array. In 1019 // general it is assumed that Projs after the node expanded are 1020 // kills. These kills are not required any more after expanding as 1021 // there are now explicitly visible def-use chains and the Projs are 1022 // removed. This does not hold for calls: They do not only have 1023 // kill-Projs but also Projs defining values. Therefore Projs after 1024 // the node expanded are removed for all but for calls. If a node is 1025 // to be reused, it must be added to the nodes list returned, and it 1026 // will be added again. 1027 // 1028 // Implementing the postalloc_expand function for a node in an enc_class 1029 // is rather tedious. It requires knowledge about many node details, as 1030 // the nodes and the subgraph must be hand crafted. To simplify this, 1031 // adlc generates some utility variables into the postalloc_expand function, 1032 // e.g., holding the operands as specified by the postalloc_expand encoding 1033 // specification, e.g.: 1034 // * unsigned idx_<par_name> holding the index of the node in the ins 1035 // * Node *n_<par_name> holding the node loaded from the ins 1036 // * MachOpnd *op_<par_name> holding the corresponding operand 1037 // 1038 // The ordering of operands can not be determined by looking at a 1039 // rule. Especially if a match rule matches several different trees, 1040 // several nodes are generated from one instruct specification with 1041 // different operand orderings. In this case the adlc generated 1042 // variables are the only way to access the ins and operands 1043 // deterministically. 1044 // 1045 // If assigning a register to a node that contains an oop, don't 1046 // forget to call ra_->set_oop() for the node. 1047 void PhaseCFG::postalloc_expand(PhaseRegAlloc* _ra) { 1048 GrowableArray <Node *> new_nodes(32); // Array with new nodes filled by postalloc_expand function of node. 1049 GrowableArray <Node *> remove(32); 1050 GrowableArray <Node *> succs(32); 1051 unsigned int max_idx = C->unique(); // Remember to distinguish new from old nodes. 1052 DEBUG_ONLY(bool foundNode = false); 1053 1054 // for all blocks 1055 for (uint i = 0; i < number_of_blocks(); i++) { 1056 Block *b = _blocks[i]; 1057 // For all instructions in the current block. 1058 for (uint j = 0; j < b->number_of_nodes(); j++) { 1059 Node *n = b->get_node(j); 1060 if (n->is_Mach() && n->as_Mach()->requires_postalloc_expand()) { 1061 #ifdef ASSERT 1062 if (TracePostallocExpand) { 1063 if (!foundNode) { 1064 foundNode = true; 1065 tty->print("POSTALLOC EXPANDING %d %s\n", C->compile_id(), 1066 C->method() ? C->method()->name()->as_utf8() : C->stub_name()); 1067 } 1068 tty->print(" postalloc expanding "); n->dump(); 1069 if (Verbose) { 1070 tty->print(" with ins:\n"); 1071 for (uint k = 0; k < n->len(); ++k) { 1072 if (n->in(k)) { tty->print(" "); n->in(k)->dump(); } 1073 } 1074 } 1075 } 1076 #endif 1077 new_nodes.clear(); 1078 // Collect nodes that have to be removed from the block later on. 1079 uint req = n->req(); 1080 remove.clear(); 1081 for (uint k = 0; k < req; ++k) { 1082 if (n->in(k) && n->in(k)->is_MachTemp()) { 1083 remove.push(n->in(k)); // MachTemps which are inputs to the old node have to be removed. 1084 n->in(k)->del_req(0); 1085 j--; 1086 } 1087 } 1088 1089 // Check whether we can allocate enough nodes. We set a fix limit for 1090 // the size of postalloc expands with this. 1091 uint unique_limit = C->unique() + 40; 1092 if (unique_limit >= _ra->node_regs_max_index()) { 1093 Compile::current()->record_failure("out of nodes in postalloc expand"); 1094 return; 1095 } 1096 1097 // Emit (i.e. generate new nodes). 1098 n->as_Mach()->postalloc_expand(&new_nodes, _ra); 1099 1100 assert(C->unique() < unique_limit, "You allocated too many nodes in your postalloc expand."); 1101 1102 // Disconnect the inputs of the old node. 1103 // 1104 // We reuse MachSpillCopy nodes. If we need to expand them, there 1105 // are many, so reusing pays off. If reused, the node already 1106 // has the new ins. n must be the last node on new_nodes list. 1107 if (!n->is_MachSpillCopy()) { 1108 for (int k = req - 1; k >= 0; --k) { 1109 n->del_req(k); 1110 } 1111 } 1112 1113 #ifdef ASSERT 1114 // Check that all nodes have proper operands. 1115 for (int k = 0; k < new_nodes.length(); ++k) { 1116 if (new_nodes.at(k)->_idx < max_idx || !new_nodes.at(k)->is_Mach()) continue; // old node, Proj ... 1117 MachNode *m = new_nodes.at(k)->as_Mach(); 1118 for (unsigned int l = 0; l < m->num_opnds(); ++l) { 1119 if (MachOper::notAnOper(m->_opnds[l])) { 1120 outputStream *os = tty; 1121 os->print("Node %s ", m->Name()); 1122 os->print("has invalid opnd %d: %p\n", l, m->_opnds[l]); 1123 assert(0, "Invalid operands, see inline trace in hs_err_pid file."); 1124 } 1125 } 1126 } 1127 #endif 1128 1129 // Collect succs of old node in remove (for projections) and in succs (for 1130 // all other nodes) do _not_ collect projections in remove (but in succs) 1131 // in case the node is a call. We need the projections for calls as they are 1132 // associated with registes (i.e. they are defs). 1133 succs.clear(); 1134 for (DUIterator k = n->outs(); n->has_out(k); k++) { 1135 if (n->out(k)->is_Proj() && !n->is_MachCall() && !n->is_MachBranch()) { 1136 remove.push(n->out(k)); 1137 } else { 1138 succs.push(n->out(k)); 1139 } 1140 } 1141 // Replace old node n as input of its succs by last of the new nodes. 1142 for (int k = 0; k < succs.length(); ++k) { 1143 Node *succ = succs.at(k); 1144 for (uint l = 0; l < succ->req(); ++l) { 1145 if (succ->in(l) == n) { 1146 succ->set_req(l, new_nodes.at(new_nodes.length() - 1)); 1147 } 1148 } 1149 for (uint l = succ->req(); l < succ->len(); ++l) { 1150 if (succ->in(l) == n) { 1151 succ->set_prec(l, new_nodes.at(new_nodes.length() - 1)); 1152 } 1153 } 1154 } 1155 1156 // Index of old node in block. 1157 uint index = b->find_node(n); 1158 // Insert new nodes into block and map them in nodes->blocks array 1159 // and remember last node in n2. 1160 Node *n2 = NULL; 1161 for (int k = 0; k < new_nodes.length(); ++k) { 1162 n2 = new_nodes.at(k); 1163 b->insert_node(n2, ++index); 1164 map_node_to_block(n2, b); 1165 } 1166 1167 // Add old node n to remove and remove them all from block. 1168 remove.push(n); 1169 j--; 1170 #ifdef ASSERT 1171 if (TracePostallocExpand && Verbose) { 1172 tty->print(" removing:\n"); 1173 for (int k = 0; k < remove.length(); ++k) { 1174 tty->print(" "); remove.at(k)->dump(); 1175 } 1176 tty->print(" inserting:\n"); 1177 for (int k = 0; k < new_nodes.length(); ++k) { 1178 tty->print(" "); new_nodes.at(k)->dump(); 1179 } 1180 } 1181 #endif 1182 for (int k = 0; k < remove.length(); ++k) { 1183 if (b->contains(remove.at(k))) { 1184 b->find_remove(remove.at(k)); 1185 } else { 1186 assert(remove.at(k)->is_Proj() && (remove.at(k)->in(0)->is_MachBranch()), ""); 1187 } 1188 } 1189 // If anything has been inserted (n2 != NULL), continue after last node inserted. 1190 // This does not always work. Some postalloc expands don't insert any nodes, if they 1191 // do optimizations (e.g., max(x,x)). In this case we decrement j accordingly. 1192 j = n2 ? b->find_node(n2) : j; 1193 } 1194 } 1195 } 1196 1197 #ifdef ASSERT 1198 if (foundNode) { 1199 tty->print("FINISHED %d %s\n", C->compile_id(), 1200 C->method() ? C->method()->name()->as_utf8() : C->stub_name()); 1201 tty->flush(); 1202 } 1203 #endif 1204 } 1205 1206 1207 //------------------------------dump------------------------------------------- 1208 #ifndef PRODUCT 1209 void PhaseCFG::_dump_cfg( const Node *end, VectorSet &visited ) const { 1210 const Node *x = end->is_block_proj(); 1211 assert( x, "not a CFG" ); 1212 1213 // Do not visit this block again 1214 if( visited.test_set(x->_idx) ) return; 1215 1216 // Skip through this block 1217 const Node *p = x; 1218 do { 1219 p = p->in(0); // Move control forward 1220 assert( !p->is_block_proj() || p->is_Root(), "not a CFG" ); 1221 } while( !p->is_block_start() ); 1222 1223 // Recursively visit 1224 for (uint i = 1; i < p->req(); i++) { 1225 _dump_cfg(p->in(i), visited); 1226 } 1227 1228 // Dump the block 1229 get_block_for_node(p)->dump(this); 1230 } 1231 1232 void PhaseCFG::dump( ) const { 1233 tty->print("\n--- CFG --- %d BBs\n", number_of_blocks()); 1234 if (_blocks.size()) { // Did we do basic-block layout? 1235 for (uint i = 0; i < number_of_blocks(); i++) { 1236 const Block* block = get_block(i); 1237 block->dump(this); 1238 } 1239 } else { // Else do it with a DFS 1240 VectorSet visited(_block_arena); 1241 _dump_cfg(_root,visited); 1242 } 1243 } 1244 1245 void PhaseCFG::dump_headers() { 1246 for (uint i = 0; i < number_of_blocks(); i++) { 1247 Block* block = get_block(i); 1248 if (block != NULL) { 1249 block->dump_head(this); 1250 } 1251 } 1252 } 1253 1254 void PhaseCFG::verify() const { 1255 #ifdef ASSERT 1256 // Verify sane CFG 1257 for (uint i = 0; i < number_of_blocks(); i++) { 1258 Block* block = get_block(i); 1259 uint cnt = block->number_of_nodes(); 1260 uint j; 1261 for (j = 0; j < cnt; j++) { 1262 Node *n = block->get_node(j); 1263 assert(get_block_for_node(n) == block, ""); 1264 if (j >= 1 && n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CreateEx) { 1265 assert(j == 1 || block->get_node(j-1)->is_Phi(), "CreateEx must be first instruction in block"); 1266 } 1267 for (uint k = 0; k < n->req(); k++) { 1268 Node *def = n->in(k); 1269 if (def && def != n) { 1270 assert(get_block_for_node(def) || def->is_Con(), "must have block; constants for debug info ok"); 1271 // Verify that instructions in the block is in correct order. 1272 // Uses must follow their definition if they are at the same block. 1273 // Mostly done to check that MachSpillCopy nodes are placed correctly 1274 // when CreateEx node is moved in build_ifg_physical(). 1275 if (get_block_for_node(def) == block && !(block->head()->is_Loop() && n->is_Phi()) && 1276 // See (+++) comment in reg_split.cpp 1277 !(n->jvms() != NULL && n->jvms()->is_monitor_use(k))) { 1278 bool is_loop = false; 1279 if (n->is_Phi()) { 1280 for (uint l = 1; l < def->req(); l++) { 1281 if (n == def->in(l)) { 1282 is_loop = true; 1283 break; // Some kind of loop 1284 } 1285 } 1286 } 1287 assert(is_loop || block->find_node(def) < j, "uses must follow definitions"); 1288 } 1289 } 1290 } 1291 } 1292 1293 j = block->end_idx(); 1294 Node* bp = (Node*)block->get_node(block->number_of_nodes() - 1)->is_block_proj(); 1295 assert(bp, "last instruction must be a block proj"); 1296 assert(bp == block->get_node(j), "wrong number of successors for this block"); 1297 if (bp->is_Catch()) { 1298 while (block->get_node(--j)->is_MachProj()) { 1299 ; 1300 } 1301 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call"); 1302 } else if (bp->is_Mach() && bp->as_Mach()->ideal_Opcode() == Op_If) { 1303 assert(block->_num_succs == 2, "Conditional branch must have two targets"); 1304 } 1305 } 1306 #endif 1307 } 1308 #endif 1309 1310 UnionFind::UnionFind( uint max ) : _cnt(max), _max(max), _indices(NEW_RESOURCE_ARRAY(uint,max)) { 1311 Copy::zero_to_bytes( _indices, sizeof(uint)*max ); 1312 } 1313 1314 void UnionFind::extend( uint from_idx, uint to_idx ) { 1315 _nesting.check(); 1316 if( from_idx >= _max ) { 1317 uint size = 16; 1318 while( size <= from_idx ) size <<=1; 1319 _indices = REALLOC_RESOURCE_ARRAY( uint, _indices, _max, size ); 1320 _max = size; 1321 } 1322 while( _cnt <= from_idx ) _indices[_cnt++] = 0; 1323 _indices[from_idx] = to_idx; 1324 } 1325 1326 void UnionFind::reset( uint max ) { 1327 // Force the Union-Find mapping to be at least this large 1328 extend(max,0); 1329 // Initialize to be the ID mapping. 1330 for( uint i=0; i<max; i++ ) map(i,i); 1331 } 1332 1333 // Straight out of Tarjan's union-find algorithm 1334 uint UnionFind::Find_compress( uint idx ) { 1335 uint cur = idx; 1336 uint next = lookup(cur); 1337 while( next != cur ) { // Scan chain of equivalences 1338 assert( next < cur, "always union smaller" ); 1339 cur = next; // until find a fixed-point 1340 next = lookup(cur); 1341 } 1342 // Core of union-find algorithm: update chain of 1343 // equivalences to be equal to the root. 1344 while( idx != next ) { 1345 uint tmp = lookup(idx); 1346 map(idx, next); 1347 idx = tmp; 1348 } 1349 return idx; 1350 } 1351 1352 // Like Find above, but no path compress, so bad asymptotic behavior 1353 uint UnionFind::Find_const( uint idx ) const { 1354 if( idx == 0 ) return idx; // Ignore the zero idx 1355 // Off the end? This can happen during debugging dumps 1356 // when data structures have not finished being updated. 1357 if( idx >= _max ) return idx; 1358 uint next = lookup(idx); 1359 while( next != idx ) { // Scan chain of equivalences 1360 idx = next; // until find a fixed-point 1361 next = lookup(idx); 1362 } 1363 return next; 1364 } 1365 1366 // union 2 sets together. 1367 void UnionFind::Union( uint idx1, uint idx2 ) { 1368 uint src = Find(idx1); 1369 uint dst = Find(idx2); 1370 assert( src, "" ); 1371 assert( dst, "" ); 1372 assert( src < _max, "oob" ); 1373 assert( dst < _max, "oob" ); 1374 assert( src < dst, "always union smaller" ); 1375 map(dst,src); 1376 } 1377 1378 #ifndef PRODUCT 1379 void Trace::dump( ) const { 1380 tty->print_cr("Trace (freq %f)", first_block()->_freq); 1381 for (Block *b = first_block(); b != NULL; b = next(b)) { 1382 tty->print(" B%d", b->_pre_order); 1383 if (b->head()->is_Loop()) { 1384 tty->print(" (L%d)", b->compute_loop_alignment()); 1385 } 1386 if (b->has_loop_alignment()) { 1387 tty->print(" (T%d)", b->code_alignment()); 1388 } 1389 } 1390 tty->cr(); 1391 } 1392 1393 void CFGEdge::dump( ) const { 1394 tty->print(" B%d --> B%d Freq: %f out:%3d%% in:%3d%% State: ", 1395 from()->_pre_order, to()->_pre_order, freq(), _from_pct, _to_pct); 1396 switch(state()) { 1397 case connected: 1398 tty->print("connected"); 1399 break; 1400 case open: 1401 tty->print("open"); 1402 break; 1403 case interior: 1404 tty->print("interior"); 1405 break; 1406 } 1407 if (infrequent()) { 1408 tty->print(" infrequent"); 1409 } 1410 tty->cr(); 1411 } 1412 #endif 1413 1414 // Comparison function for edges 1415 static int edge_order(CFGEdge **e0, CFGEdge **e1) { 1416 float freq0 = (*e0)->freq(); 1417 float freq1 = (*e1)->freq(); 1418 if (freq0 != freq1) { 1419 return freq0 > freq1 ? -1 : 1; 1420 } 1421 1422 int dist0 = (*e0)->to()->_rpo - (*e0)->from()->_rpo; 1423 int dist1 = (*e1)->to()->_rpo - (*e1)->from()->_rpo; 1424 1425 return dist1 - dist0; 1426 } 1427 1428 // Comparison function for edges 1429 extern "C" int trace_frequency_order(const void *p0, const void *p1) { 1430 Trace *tr0 = *(Trace **) p0; 1431 Trace *tr1 = *(Trace **) p1; 1432 Block *b0 = tr0->first_block(); 1433 Block *b1 = tr1->first_block(); 1434 1435 // The trace of connector blocks goes at the end; 1436 // we only expect one such trace 1437 if (b0->is_connector() != b1->is_connector()) { 1438 return b1->is_connector() ? -1 : 1; 1439 } 1440 1441 // Pull more frequently executed blocks to the beginning 1442 float freq0 = b0->_freq; 1443 float freq1 = b1->_freq; 1444 if (freq0 != freq1) { 1445 return freq0 > freq1 ? -1 : 1; 1446 } 1447 1448 int diff = tr0->first_block()->_rpo - tr1->first_block()->_rpo; 1449 1450 return diff; 1451 } 1452 1453 // Find edges of interest, i.e, those which can fall through. Presumes that 1454 // edges which don't fall through are of low frequency and can be generally 1455 // ignored. Initialize the list of traces. 1456 void PhaseBlockLayout::find_edges() { 1457 // Walk the blocks, creating edges and Traces 1458 uint i; 1459 Trace *tr = NULL; 1460 for (i = 0; i < _cfg.number_of_blocks(); i++) { 1461 Block* b = _cfg.get_block(i); 1462 tr = new Trace(b, next, prev); 1463 traces[tr->id()] = tr; 1464 1465 // All connector blocks should be at the end of the list 1466 if (b->is_connector()) break; 1467 1468 // If this block and the next one have a one-to-one successor 1469 // predecessor relationship, simply append the next block 1470 int nfallthru = b->num_fall_throughs(); 1471 while (nfallthru == 1 && 1472 b->succ_fall_through(0)) { 1473 Block *n = b->_succs[0]; 1474 1475 // Skip over single-entry connector blocks, we don't want to 1476 // add them to the trace. 1477 while (n->is_connector() && n->num_preds() == 1) { 1478 n = n->_succs[0]; 1479 } 1480 1481 // We see a merge point, so stop search for the next block 1482 if (n->num_preds() != 1) break; 1483 1484 i++; 1485 assert(n = _cfg.get_block(i), "expecting next block"); 1486 tr->append(n); 1487 uf->map(n->_pre_order, tr->id()); 1488 traces[n->_pre_order] = NULL; 1489 nfallthru = b->num_fall_throughs(); 1490 b = n; 1491 } 1492 1493 if (nfallthru > 0) { 1494 // Create a CFGEdge for each outgoing 1495 // edge that could be a fall-through. 1496 for (uint j = 0; j < b->_num_succs; j++ ) { 1497 if (b->succ_fall_through(j)) { 1498 Block *target = b->non_connector_successor(j); 1499 float freq = b->_freq * b->succ_prob(j); 1500 int from_pct = (int) ((100 * freq) / b->_freq); 1501 int to_pct = (int) ((100 * freq) / target->_freq); 1502 edges->append(new CFGEdge(b, target, freq, from_pct, to_pct)); 1503 } 1504 } 1505 } 1506 } 1507 1508 // Group connector blocks into one trace 1509 for (i++; i < _cfg.number_of_blocks(); i++) { 1510 Block *b = _cfg.get_block(i); 1511 assert(b->is_connector(), "connector blocks at the end"); 1512 tr->append(b); 1513 uf->map(b->_pre_order, tr->id()); 1514 traces[b->_pre_order] = NULL; 1515 } 1516 } 1517 1518 // Union two traces together in uf, and null out the trace in the list 1519 void PhaseBlockLayout::union_traces(Trace* updated_trace, Trace* old_trace) { 1520 uint old_id = old_trace->id(); 1521 uint updated_id = updated_trace->id(); 1522 1523 uint lo_id = updated_id; 1524 uint hi_id = old_id; 1525 1526 // If from is greater than to, swap values to meet 1527 // UnionFind guarantee. 1528 if (updated_id > old_id) { 1529 lo_id = old_id; 1530 hi_id = updated_id; 1531 1532 // Fix up the trace ids 1533 traces[lo_id] = traces[updated_id]; 1534 updated_trace->set_id(lo_id); 1535 } 1536 1537 // Union the lower with the higher and remove the pointer 1538 // to the higher. 1539 uf->Union(lo_id, hi_id); 1540 traces[hi_id] = NULL; 1541 } 1542 1543 // Append traces together via the most frequently executed edges 1544 void PhaseBlockLayout::grow_traces() { 1545 // Order the edges, and drive the growth of Traces via the most 1546 // frequently executed edges. 1547 edges->sort(edge_order); 1548 for (int i = 0; i < edges->length(); i++) { 1549 CFGEdge *e = edges->at(i); 1550 1551 if (e->state() != CFGEdge::open) continue; 1552 1553 Block *src_block = e->from(); 1554 Block *targ_block = e->to(); 1555 1556 // Don't grow traces along backedges? 1557 if (!BlockLayoutRotateLoops) { 1558 if (targ_block->_rpo <= src_block->_rpo) { 1559 targ_block->set_loop_alignment(targ_block); 1560 continue; 1561 } 1562 } 1563 1564 Trace *src_trace = trace(src_block); 1565 Trace *targ_trace = trace(targ_block); 1566 1567 // If the edge in question can join two traces at their ends, 1568 // append one trace to the other. 1569 if (src_trace->last_block() == src_block) { 1570 if (src_trace == targ_trace) { 1571 e->set_state(CFGEdge::interior); 1572 if (targ_trace->backedge(e)) { 1573 // Reset i to catch any newly eligible edge 1574 // (Or we could remember the first "open" edge, and reset there) 1575 i = 0; 1576 } 1577 } else if (targ_trace->first_block() == targ_block) { 1578 e->set_state(CFGEdge::connected); 1579 src_trace->append(targ_trace); 1580 union_traces(src_trace, targ_trace); 1581 } 1582 } 1583 } 1584 } 1585 1586 // Embed one trace into another, if the fork or join points are sufficiently 1587 // balanced. 1588 void PhaseBlockLayout::merge_traces(bool fall_thru_only) { 1589 // Walk the edge list a another time, looking at unprocessed edges. 1590 // Fold in diamonds 1591 for (int i = 0; i < edges->length(); i++) { 1592 CFGEdge *e = edges->at(i); 1593 1594 if (e->state() != CFGEdge::open) continue; 1595 if (fall_thru_only) { 1596 if (e->infrequent()) continue; 1597 } 1598 1599 Block *src_block = e->from(); 1600 Trace *src_trace = trace(src_block); 1601 bool src_at_tail = src_trace->last_block() == src_block; 1602 1603 Block *targ_block = e->to(); 1604 Trace *targ_trace = trace(targ_block); 1605 bool targ_at_start = targ_trace->first_block() == targ_block; 1606 1607 if (src_trace == targ_trace) { 1608 // This may be a loop, but we can't do much about it. 1609 e->set_state(CFGEdge::interior); 1610 continue; 1611 } 1612 1613 if (fall_thru_only) { 1614 // If the edge links the middle of two traces, we can't do anything. 1615 // Mark the edge and continue. 1616 if (!src_at_tail & !targ_at_start) { 1617 continue; 1618 } 1619 1620 // Don't grow traces along backedges? 1621 if (!BlockLayoutRotateLoops && (targ_block->_rpo <= src_block->_rpo)) { 1622 continue; 1623 } 1624 1625 // If both ends of the edge are available, why didn't we handle it earlier? 1626 assert(src_at_tail ^ targ_at_start, "Should have caught this edge earlier."); 1627 1628 if (targ_at_start) { 1629 // Insert the "targ" trace in the "src" trace if the insertion point 1630 // is a two way branch. 1631 // Better profitability check possible, but may not be worth it. 1632 // Someday, see if the this "fork" has an associated "join"; 1633 // then make a policy on merging this trace at the fork or join. 1634 // For example, other things being equal, it may be better to place this 1635 // trace at the join point if the "src" trace ends in a two-way, but 1636 // the insertion point is one-way. 1637 assert(src_block->num_fall_throughs() == 2, "unexpected diamond"); 1638 e->set_state(CFGEdge::connected); 1639 src_trace->insert_after(src_block, targ_trace); 1640 union_traces(src_trace, targ_trace); 1641 } else if (src_at_tail) { 1642 if (src_trace != trace(_cfg.get_root_block())) { 1643 e->set_state(CFGEdge::connected); 1644 targ_trace->insert_before(targ_block, src_trace); 1645 union_traces(targ_trace, src_trace); 1646 } 1647 } 1648 } else if (e->state() == CFGEdge::open) { 1649 // Append traces, even without a fall-thru connection. 1650 // But leave root entry at the beginning of the block list. 1651 if (targ_trace != trace(_cfg.get_root_block())) { 1652 e->set_state(CFGEdge::connected); 1653 src_trace->append(targ_trace); 1654 union_traces(src_trace, targ_trace); 1655 } 1656 } 1657 } 1658 } 1659 1660 // Order the sequence of the traces in some desirable way, and fixup the 1661 // jumps at the end of each block. 1662 void PhaseBlockLayout::reorder_traces(int count) { 1663 ResourceArea *area = Thread::current()->resource_area(); 1664 Trace ** new_traces = NEW_ARENA_ARRAY(area, Trace *, count); 1665 Block_List worklist; 1666 int new_count = 0; 1667 1668 // Compact the traces. 1669 for (int i = 0; i < count; i++) { 1670 Trace *tr = traces[i]; 1671 if (tr != NULL) { 1672 new_traces[new_count++] = tr; 1673 } 1674 } 1675 1676 // The entry block should be first on the new trace list. 1677 Trace *tr = trace(_cfg.get_root_block()); 1678 assert(tr == new_traces[0], "entry trace misplaced"); 1679 1680 // Sort the new trace list by frequency 1681 qsort(new_traces + 1, new_count - 1, sizeof(new_traces[0]), trace_frequency_order); 1682 1683 // Patch up the successor blocks 1684 _cfg.clear_blocks(); 1685 for (int i = 0; i < new_count; i++) { 1686 Trace *tr = new_traces[i]; 1687 if (tr != NULL) { 1688 tr->fixup_blocks(_cfg); 1689 } 1690 } 1691 } 1692 1693 // Order basic blocks based on frequency 1694 PhaseBlockLayout::PhaseBlockLayout(PhaseCFG &cfg) 1695 : Phase(BlockLayout) 1696 , _cfg(cfg) { 1697 ResourceMark rm; 1698 ResourceArea *area = Thread::current()->resource_area(); 1699 1700 // List of traces 1701 int size = _cfg.number_of_blocks() + 1; 1702 traces = NEW_ARENA_ARRAY(area, Trace *, size); 1703 memset(traces, 0, size*sizeof(Trace*)); 1704 next = NEW_ARENA_ARRAY(area, Block *, size); 1705 memset(next, 0, size*sizeof(Block *)); 1706 prev = NEW_ARENA_ARRAY(area, Block *, size); 1707 memset(prev , 0, size*sizeof(Block *)); 1708 1709 // List of edges 1710 edges = new GrowableArray<CFGEdge*>; 1711 1712 // Mapping block index --> block_trace 1713 uf = new UnionFind(size); 1714 uf->reset(size); 1715 1716 // Find edges and create traces. 1717 find_edges(); 1718 1719 // Grow traces at their ends via most frequent edges. 1720 grow_traces(); 1721 1722 // Merge one trace into another, but only at fall-through points. 1723 // This may make diamonds and other related shapes in a trace. 1724 merge_traces(true); 1725 1726 // Run merge again, allowing two traces to be catenated, even if 1727 // one does not fall through into the other. This appends loosely 1728 // related traces to be near each other. 1729 merge_traces(false); 1730 1731 // Re-order all the remaining traces by frequency 1732 reorder_traces(size); 1733 1734 assert(_cfg.number_of_blocks() >= (uint) (size - 1), "number of blocks can not shrink"); 1735 } 1736 1737 1738 // Edge e completes a loop in a trace. If the target block is head of the 1739 // loop, rotate the loop block so that the loop ends in a conditional branch. 1740 bool Trace::backedge(CFGEdge *e) { 1741 bool loop_rotated = false; 1742 Block *src_block = e->from(); 1743 Block *targ_block = e->to(); 1744 1745 assert(last_block() == src_block, "loop discovery at back branch"); 1746 if (first_block() == targ_block) { 1747 if (BlockLayoutRotateLoops && last_block()->num_fall_throughs() < 2) { 1748 // Find the last block in the trace that has a conditional 1749 // branch. 1750 Block *b; 1751 for (b = last_block(); b != NULL; b = prev(b)) { 1752 if (b->num_fall_throughs() == 2) { 1753 break; 1754 } 1755 } 1756 1757 if (b != last_block() && b != NULL) { 1758 loop_rotated = true; 1759 1760 // Rotate the loop by doing two-part linked-list surgery. 1761 append(first_block()); 1762 break_loop_after(b); 1763 } 1764 } 1765 1766 // Backbranch to the top of a trace 1767 // Scroll forward through the trace from the targ_block. If we find 1768 // a loop head before another loop top, use the the loop head alignment. 1769 for (Block *b = targ_block; b != NULL; b = next(b)) { 1770 if (b->has_loop_alignment()) { 1771 break; 1772 } 1773 if (b->head()->is_Loop()) { 1774 targ_block = b; 1775 break; 1776 } 1777 } 1778 1779 first_block()->set_loop_alignment(targ_block); 1780 1781 } else { 1782 // Backbranch into the middle of a trace 1783 targ_block->set_loop_alignment(targ_block); 1784 } 1785 1786 return loop_rotated; 1787 } 1788 1789 // push blocks onto the CFG list 1790 // ensure that blocks have the correct two-way branch sense 1791 void Trace::fixup_blocks(PhaseCFG &cfg) { 1792 Block *last = last_block(); 1793 for (Block *b = first_block(); b != NULL; b = next(b)) { 1794 cfg.add_block(b); 1795 if (!b->is_connector()) { 1796 int nfallthru = b->num_fall_throughs(); 1797 if (b != last) { 1798 if (nfallthru == 2) { 1799 // Ensure that the sense of the branch is correct 1800 Block *bnext = next(b); 1801 Block *bs0 = b->non_connector_successor(0); 1802 1803 MachNode *iff = b->get_node(b->number_of_nodes() - 3)->as_Mach(); 1804 ProjNode *proj0 = b->get_node(b->number_of_nodes() - 2)->as_Proj(); 1805 ProjNode *proj1 = b->get_node(b->number_of_nodes() - 1)->as_Proj(); 1806 1807 if (bnext == bs0) { 1808 // Fall-thru case in succs[0], should be in succs[1] 1809 1810 // Flip targets in _succs map 1811 Block *tbs0 = b->_succs[0]; 1812 Block *tbs1 = b->_succs[1]; 1813 b->_succs.map( 0, tbs1 ); 1814 b->_succs.map( 1, tbs0 ); 1815 1816 // Flip projections to match targets 1817 b->map_node(proj1, b->number_of_nodes() - 2); 1818 b->map_node(proj0, b->number_of_nodes() - 1); 1819 } 1820 } 1821 } 1822 } 1823 } 1824 }