1 /* 2 * Copyright 1997-2008 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 // Optimization - Graph Style 26 27 #include "incls/_precompiled.incl" 28 #include "incls/_block.cpp.incl" 29 30 31 //----------------------------------------------------------------------------- 32 void Block_Array::grow( uint i ) { 33 assert(i >= Max(), "must be an overflow"); 34 debug_only(_limit = i+1); 35 if( i < _size ) return; 36 if( !_size ) { 37 _size = 1; 38 _blocks = (Block**)_arena->Amalloc( _size * sizeof(Block*) ); 39 _blocks[0] = NULL; 40 } 41 uint old = _size; 42 while( i >= _size ) _size <<= 1; // Double to fit 43 _blocks = (Block**)_arena->Arealloc( _blocks, old*sizeof(Block*),_size*sizeof(Block*)); 44 Copy::zero_to_bytes( &_blocks[old], (_size-old)*sizeof(Block*) ); 45 } 46 47 //============================================================================= 48 void Block_List::remove(uint i) { 49 assert(i < _cnt, "index out of bounds"); 50 Copy::conjoint_words_to_lower((HeapWord*)&_blocks[i+1], (HeapWord*)&_blocks[i], ((_cnt-i-1)*sizeof(Block*))); 51 pop(); // shrink list by one block 52 } 53 54 void Block_List::insert(uint i, Block *b) { 55 push(b); // grow list by one block 56 Copy::conjoint_words_to_higher((HeapWord*)&_blocks[i], (HeapWord*)&_blocks[i+1], ((_cnt-i-1)*sizeof(Block*))); 57 _blocks[i] = b; 58 } 59 60 #ifndef PRODUCT 61 void Block_List::print() { 62 for (uint i=0; i < size(); i++) { 63 tty->print("B%d ", _blocks[i]->_pre_order); 64 } 65 tty->print("size = %d\n", size()); 66 } 67 #endif 68 69 //============================================================================= 70 71 uint Block::code_alignment() { 72 // Check for Root block 73 if( _pre_order == 0 ) return CodeEntryAlignment; 74 // Check for Start block 75 if( _pre_order == 1 ) return InteriorEntryAlignment; 76 // Check for loop alignment 77 if (has_loop_alignment()) return loop_alignment(); 78 79 return 1; // no particular alignment 80 } 81 82 uint Block::compute_loop_alignment() { 83 Node *h = head(); 84 if( h->is_Loop() && h->as_Loop()->is_inner_loop() ) { 85 // Pre- and post-loops have low trip count so do not bother with 86 // NOPs for align loop head. The constants are hidden from tuning 87 // but only because my "divide by 4" heuristic surely gets nearly 88 // all possible gain (a "do not align at all" heuristic has a 89 // chance of getting a really tiny gain). 90 if( h->is_CountedLoop() && (h->as_CountedLoop()->is_pre_loop() || 91 h->as_CountedLoop()->is_post_loop()) ) 92 return (OptoLoopAlignment > 4) ? (OptoLoopAlignment>>2) : 1; 93 // Loops with low backedge frequency should not be aligned. 94 Node *n = h->in(LoopNode::LoopBackControl)->in(0); 95 if( n->is_MachIf() && n->as_MachIf()->_prob < 0.01 ) { 96 return 1; // Loop does not loop, more often than not! 97 } 98 return OptoLoopAlignment; // Otherwise align loop head 99 } 100 101 return 1; // no particular alignment 102 } 103 104 //----------------------------------------------------------------------------- 105 // Compute the size of first 'inst_cnt' instructions in this block. 106 // Return the number of instructions left to compute if the block has 107 // less then 'inst_cnt' instructions. Stop, and return 0 if sum_size 108 // exceeds OptoLoopAlignment. 109 uint Block::compute_first_inst_size(uint& sum_size, uint inst_cnt, 110 PhaseRegAlloc* ra) { 111 uint last_inst = _nodes.size(); 112 for( uint j = 0; j < last_inst && inst_cnt > 0; j++ ) { 113 uint inst_size = _nodes[j]->size(ra); 114 if( inst_size > 0 ) { 115 inst_cnt--; 116 uint sz = sum_size + inst_size; 117 if( sz <= (uint)OptoLoopAlignment ) { 118 // Compute size of instructions which fit into fetch buffer only 119 // since all inst_cnt instructions will not fit even if we align them. 120 sum_size = sz; 121 } else { 122 return 0; 123 } 124 } 125 } 126 return inst_cnt; 127 } 128 129 //----------------------------------------------------------------------------- 130 uint Block::find_node( const Node *n ) const { 131 for( uint i = 0; i < _nodes.size(); i++ ) { 132 if( _nodes[i] == n ) 133 return i; 134 } 135 ShouldNotReachHere(); 136 return 0; 137 } 138 139 // Find and remove n from block list 140 void Block::find_remove( const Node *n ) { 141 _nodes.remove(find_node(n)); 142 } 143 144 //------------------------------is_Empty--------------------------------------- 145 // Return empty status of a block. Empty blocks contain only the head, other 146 // ideal nodes, and an optional trailing goto. 147 int Block::is_Empty() const { 148 149 // Root or start block is not considered empty 150 if (head()->is_Root() || head()->is_Start()) { 151 return not_empty; 152 } 153 154 int success_result = completely_empty; 155 int end_idx = _nodes.size()-1; 156 157 // Check for ending goto 158 if ((end_idx > 0) && (_nodes[end_idx]->is_Goto())) { 159 success_result = empty_with_goto; 160 end_idx--; 161 } 162 163 // Unreachable blocks are considered empty 164 if (num_preds() <= 1) { 165 return success_result; 166 } 167 168 // Ideal nodes are allowable in empty blocks: skip them Only MachNodes 169 // turn directly into code, because only MachNodes have non-trivial 170 // emit() functions. 171 while ((end_idx > 0) && !_nodes[end_idx]->is_Mach()) { 172 end_idx--; 173 } 174 175 // No room for any interesting instructions? 176 if (end_idx == 0) { 177 return success_result; 178 } 179 180 return not_empty; 181 } 182 183 //------------------------------has_uncommon_code------------------------------ 184 // Return true if the block's code implies that it is not likely to be 185 // executed infrequently. Check to see if the block ends in a Halt or 186 // a low probability call. 187 bool Block::has_uncommon_code() const { 188 Node* en = end(); 189 190 if (en->is_Goto()) 191 en = en->in(0); 192 if (en->is_Catch()) 193 en = en->in(0); 194 if (en->is_Proj() && en->in(0)->is_MachCall()) { 195 MachCallNode* call = en->in(0)->as_MachCall(); 196 if (call->cnt() != COUNT_UNKNOWN && call->cnt() <= PROB_UNLIKELY_MAG(4)) { 197 // This is true for slow-path stubs like new_{instance,array}, 198 // slow_arraycopy, complete_monitor_locking, uncommon_trap. 199 // The magic number corresponds to the probability of an uncommon_trap, 200 // even though it is a count not a probability. 201 return true; 202 } 203 } 204 205 int op = en->is_Mach() ? en->as_Mach()->ideal_Opcode() : en->Opcode(); 206 return op == Op_Halt; 207 } 208 209 //------------------------------is_uncommon------------------------------------ 210 // True if block is low enough frequency or guarded by a test which 211 // mostly does not go here. 212 bool Block::is_uncommon( Block_Array &bbs ) const { 213 // Initial blocks must never be moved, so are never uncommon. 214 if (head()->is_Root() || head()->is_Start()) return false; 215 216 // Check for way-low freq 217 if( _freq < BLOCK_FREQUENCY(0.00001f) ) return true; 218 219 // Look for code shape indicating uncommon_trap or slow path 220 if (has_uncommon_code()) return true; 221 222 const float epsilon = 0.05f; 223 const float guard_factor = PROB_UNLIKELY_MAG(4) / (1.f - epsilon); 224 uint uncommon_preds = 0; 225 uint freq_preds = 0; 226 uint uncommon_for_freq_preds = 0; 227 228 for( uint i=1; i<num_preds(); i++ ) { 229 Block* guard = bbs[pred(i)->_idx]; 230 // Check to see if this block follows its guard 1 time out of 10000 231 // or less. 232 // 233 // See list of magnitude-4 unlikely probabilities in cfgnode.hpp which 234 // we intend to be "uncommon", such as slow-path TLE allocation, 235 // predicted call failure, and uncommon trap triggers. 236 // 237 // Use an epsilon value of 5% to allow for variability in frequency 238 // predictions and floating point calculations. The net effect is 239 // that guard_factor is set to 9500. 240 // 241 // Ignore low-frequency blocks. 242 // The next check is (guard->_freq < 1.e-5 * 9500.). 243 if(guard->_freq*BLOCK_FREQUENCY(guard_factor) < BLOCK_FREQUENCY(0.00001f)) { 244 uncommon_preds++; 245 } else { 246 freq_preds++; 247 if( _freq < guard->_freq * guard_factor ) { 248 uncommon_for_freq_preds++; 249 } 250 } 251 } 252 if( num_preds() > 1 && 253 // The block is uncommon if all preds are uncommon or 254 (uncommon_preds == (num_preds()-1) || 255 // it is uncommon for all frequent preds. 256 uncommon_for_freq_preds == freq_preds) ) { 257 return true; 258 } 259 return false; 260 } 261 262 //------------------------------dump------------------------------------------- 263 #ifndef PRODUCT 264 void Block::dump_bidx(const Block* orig) const { 265 if (_pre_order) tty->print("B%d",_pre_order); 266 else tty->print("N%d", head()->_idx); 267 268 if (Verbose && orig != this) { 269 // Dump the original block's idx 270 tty->print(" ("); 271 orig->dump_bidx(orig); 272 tty->print(")"); 273 } 274 } 275 276 void Block::dump_pred(const Block_Array *bbs, Block* orig) const { 277 if (is_connector()) { 278 for (uint i=1; i<num_preds(); i++) { 279 Block *p = ((*bbs)[pred(i)->_idx]); 280 p->dump_pred(bbs, orig); 281 } 282 } else { 283 dump_bidx(orig); 284 tty->print(" "); 285 } 286 } 287 288 void Block::dump_head( const Block_Array *bbs ) const { 289 // Print the basic block 290 dump_bidx(this); 291 tty->print(": #\t"); 292 293 // Print the incoming CFG edges and the outgoing CFG edges 294 for( uint i=0; i<_num_succs; i++ ) { 295 non_connector_successor(i)->dump_bidx(_succs[i]); 296 tty->print(" "); 297 } 298 tty->print("<- "); 299 if( head()->is_block_start() ) { 300 for (uint i=1; i<num_preds(); i++) { 301 Node *s = pred(i); 302 if (bbs) { 303 Block *p = (*bbs)[s->_idx]; 304 p->dump_pred(bbs, p); 305 } else { 306 while (!s->is_block_start()) 307 s = s->in(0); 308 tty->print("N%d ", s->_idx ); 309 } 310 } 311 } else 312 tty->print("BLOCK HEAD IS JUNK "); 313 314 // Print loop, if any 315 const Block *bhead = this; // Head of self-loop 316 Node *bh = bhead->head(); 317 if( bbs && bh->is_Loop() && !head()->is_Root() ) { 318 LoopNode *loop = bh->as_Loop(); 319 const Block *bx = (*bbs)[loop->in(LoopNode::LoopBackControl)->_idx]; 320 while (bx->is_connector()) { 321 bx = (*bbs)[bx->pred(1)->_idx]; 322 } 323 tty->print("\tLoop: B%d-B%d ", bhead->_pre_order, bx->_pre_order); 324 // Dump any loop-specific bits, especially for CountedLoops. 325 loop->dump_spec(tty); 326 } else if (has_loop_alignment()) { 327 tty->print(" top-of-loop"); 328 } 329 tty->print(" Freq: %g",_freq); 330 if( Verbose || WizardMode ) { 331 tty->print(" IDom: %d/#%d", _idom ? _idom->_pre_order : 0, _dom_depth); 332 tty->print(" RegPressure: %d",_reg_pressure); 333 tty->print(" IHRP Index: %d",_ihrp_index); 334 tty->print(" FRegPressure: %d",_freg_pressure); 335 tty->print(" FHRP Index: %d",_fhrp_index); 336 } 337 tty->print_cr(""); 338 } 339 340 void Block::dump() const { dump(0); } 341 342 void Block::dump( const Block_Array *bbs ) const { 343 dump_head(bbs); 344 uint cnt = _nodes.size(); 345 for( uint i=0; i<cnt; i++ ) 346 _nodes[i]->dump(); 347 tty->print("\n"); 348 } 349 #endif 350 351 //============================================================================= 352 //------------------------------PhaseCFG--------------------------------------- 353 PhaseCFG::PhaseCFG( Arena *a, RootNode *r, Matcher &m ) : 354 Phase(CFG), 355 _bbs(a), 356 _root(r) 357 #ifndef PRODUCT 358 , _trace_opto_pipelining(TraceOptoPipelining || C->method_has_option("TraceOptoPipelining")) 359 #endif 360 { 361 ResourceMark rm; 362 // I'll need a few machine-specific GotoNodes. Make an Ideal GotoNode, 363 // then Match it into a machine-specific Node. Then clone the machine 364 // Node on demand. 365 Node *x = new (C, 1) GotoNode(NULL); 366 x->init_req(0, x); 367 _goto = m.match_tree(x); 368 assert(_goto != NULL, ""); 369 _goto->set_req(0,_goto); 370 371 // Build the CFG in Reverse Post Order 372 _num_blocks = build_cfg(); 373 _broot = _bbs[_root->_idx]; 374 } 375 376 //------------------------------build_cfg-------------------------------------- 377 // Build a proper looking CFG. Make every block begin with either a StartNode 378 // or a RegionNode. Make every block end with either a Goto, If or Return. 379 // The RootNode both starts and ends it's own block. Do this with a recursive 380 // backwards walk over the control edges. 381 uint PhaseCFG::build_cfg() { 382 Arena *a = Thread::current()->resource_area(); 383 VectorSet visited(a); 384 385 // Allocate stack with enough space to avoid frequent realloc 386 Node_Stack nstack(a, C->unique() >> 1); 387 nstack.push(_root, 0); 388 uint sum = 0; // Counter for blocks 389 390 while (nstack.is_nonempty()) { 391 // node and in's index from stack's top 392 // 'np' is _root (see above) or RegionNode, StartNode: we push on stack 393 // only nodes which point to the start of basic block (see below). 394 Node *np = nstack.node(); 395 // idx > 0, except for the first node (_root) pushed on stack 396 // at the beginning when idx == 0. 397 // We will use the condition (idx == 0) later to end the build. 398 uint idx = nstack.index(); 399 Node *proj = np->in(idx); 400 const Node *x = proj->is_block_proj(); 401 // Does the block end with a proper block-ending Node? One of Return, 402 // If or Goto? (This check should be done for visited nodes also). 403 if (x == NULL) { // Does not end right... 404 Node *g = _goto->clone(); // Force it to end in a Goto 405 g->set_req(0, proj); 406 np->set_req(idx, g); 407 x = proj = g; 408 } 409 if (!visited.test_set(x->_idx)) { // Visit this block once 410 // Skip any control-pinned middle'in stuff 411 Node *p = proj; 412 do { 413 proj = p; // Update pointer to last Control 414 p = p->in(0); // Move control forward 415 } while( !p->is_block_proj() && 416 !p->is_block_start() ); 417 // Make the block begin with one of Region or StartNode. 418 if( !p->is_block_start() ) { 419 RegionNode *r = new (C, 2) RegionNode( 2 ); 420 r->init_req(1, p); // Insert RegionNode in the way 421 proj->set_req(0, r); // Insert RegionNode in the way 422 p = r; 423 } 424 // 'p' now points to the start of this basic block 425 426 // Put self in array of basic blocks 427 Block *bb = new (_bbs._arena) Block(_bbs._arena,p); 428 _bbs.map(p->_idx,bb); 429 _bbs.map(x->_idx,bb); 430 if( x != p ) // Only for root is x == p 431 bb->_nodes.push((Node*)x); 432 433 // Now handle predecessors 434 ++sum; // Count 1 for self block 435 uint cnt = bb->num_preds(); 436 for (int i = (cnt - 1); i > 0; i-- ) { // For all predecessors 437 Node *prevproj = p->in(i); // Get prior input 438 assert( !prevproj->is_Con(), "dead input not removed" ); 439 // Check to see if p->in(i) is a "control-dependent" CFG edge - 440 // i.e., it splits at the source (via an IF or SWITCH) and merges 441 // at the destination (via a many-input Region). 442 // This breaks critical edges. The RegionNode to start the block 443 // will be added when <p,i> is pulled off the node stack 444 if ( cnt > 2 ) { // Merging many things? 445 assert( prevproj== bb->pred(i),""); 446 if(prevproj->is_block_proj() != prevproj) { // Control-dependent edge? 447 // Force a block on the control-dependent edge 448 Node *g = _goto->clone(); // Force it to end in a Goto 449 g->set_req(0,prevproj); 450 p->set_req(i,g); 451 } 452 } 453 nstack.push(p, i); // 'p' is RegionNode or StartNode 454 } 455 } else { // Post-processing visited nodes 456 nstack.pop(); // remove node from stack 457 // Check if it the fist node pushed on stack at the beginning. 458 if (idx == 0) break; // end of the build 459 // Find predecessor basic block 460 Block *pb = _bbs[x->_idx]; 461 // Insert into nodes array, if not already there 462 if( !_bbs.lookup(proj->_idx) ) { 463 assert( x != proj, "" ); 464 // Map basic block of projection 465 _bbs.map(proj->_idx,pb); 466 pb->_nodes.push(proj); 467 } 468 // Insert self as a child of my predecessor block 469 pb->_succs.map(pb->_num_succs++, _bbs[np->_idx]); 470 assert( pb->_nodes[ pb->_nodes.size() - pb->_num_succs ]->is_block_proj(), 471 "too many control users, not a CFG?" ); 472 } 473 } 474 // Return number of basic blocks for all children and self 475 return sum; 476 } 477 478 //------------------------------insert_goto_at--------------------------------- 479 // Inserts a goto & corresponding basic block between 480 // block[block_no] and its succ_no'th successor block 481 void PhaseCFG::insert_goto_at(uint block_no, uint succ_no) { 482 // get block with block_no 483 assert(block_no < _num_blocks, "illegal block number"); 484 Block* in = _blocks[block_no]; 485 // get successor block succ_no 486 assert(succ_no < in->_num_succs, "illegal successor number"); 487 Block* out = in->_succs[succ_no]; 488 // Compute frequency of the new block. Do this before inserting 489 // new block in case succ_prob() needs to infer the probability from 490 // surrounding blocks. 491 float freq = in->_freq * in->succ_prob(succ_no); 492 // get ProjNode corresponding to the succ_no'th successor of the in block 493 ProjNode* proj = in->_nodes[in->_nodes.size() - in->_num_succs + succ_no]->as_Proj(); 494 // create region for basic block 495 RegionNode* region = new (C, 2) RegionNode(2); 496 region->init_req(1, proj); 497 // setup corresponding basic block 498 Block* block = new (_bbs._arena) Block(_bbs._arena, region); 499 _bbs.map(region->_idx, block); 500 C->regalloc()->set_bad(region->_idx); 501 // add a goto node 502 Node* gto = _goto->clone(); // get a new goto node 503 gto->set_req(0, region); 504 // add it to the basic block 505 block->_nodes.push(gto); 506 _bbs.map(gto->_idx, block); 507 C->regalloc()->set_bad(gto->_idx); 508 // hook up successor block 509 block->_succs.map(block->_num_succs++, out); 510 // remap successor's predecessors if necessary 511 for (uint i = 1; i < out->num_preds(); i++) { 512 if (out->pred(i) == proj) out->head()->set_req(i, gto); 513 } 514 // remap predecessor's successor to new block 515 in->_succs.map(succ_no, block); 516 // Set the frequency of the new block 517 block->_freq = freq; 518 // add new basic block to basic block list 519 _blocks.insert(block_no + 1, block); 520 _num_blocks++; 521 } 522 523 //------------------------------no_flip_branch--------------------------------- 524 // Does this block end in a multiway branch that cannot have the default case 525 // flipped for another case? 526 static bool no_flip_branch( Block *b ) { 527 int branch_idx = b->_nodes.size() - b->_num_succs-1; 528 if( branch_idx < 1 ) return false; 529 Node *bra = b->_nodes[branch_idx]; 530 if( bra->is_Catch() ) 531 return true; 532 if( bra->is_Mach() ) { 533 if( bra->is_MachNullCheck() ) 534 return true; 535 int iop = bra->as_Mach()->ideal_Opcode(); 536 if( iop == Op_FastLock || iop == Op_FastUnlock ) 537 return true; 538 } 539 return false; 540 } 541 542 //------------------------------convert_NeverBranch_to_Goto-------------------- 543 // Check for NeverBranch at block end. This needs to become a GOTO to the 544 // true target. NeverBranch are treated as a conditional branch that always 545 // goes the same direction for most of the optimizer and are used to give a 546 // fake exit path to infinite loops. At this late stage they need to turn 547 // into Goto's so that when you enter the infinite loop you indeed hang. 548 void PhaseCFG::convert_NeverBranch_to_Goto(Block *b) { 549 // Find true target 550 int end_idx = b->end_idx(); 551 int idx = b->_nodes[end_idx+1]->as_Proj()->_con; 552 Block *succ = b->_succs[idx]; 553 Node* gto = _goto->clone(); // get a new goto node 554 gto->set_req(0, b->head()); 555 Node *bp = b->_nodes[end_idx]; 556 b->_nodes.map(end_idx,gto); // Slam over NeverBranch 557 _bbs.map(gto->_idx, b); 558 C->regalloc()->set_bad(gto->_idx); 559 b->_nodes.pop(); // Yank projections 560 b->_nodes.pop(); // Yank projections 561 b->_succs.map(0,succ); // Map only successor 562 b->_num_succs = 1; 563 // remap successor's predecessors if necessary 564 uint j; 565 for( j = 1; j < succ->num_preds(); j++) 566 if( succ->pred(j)->in(0) == bp ) 567 succ->head()->set_req(j, gto); 568 // Kill alternate exit path 569 Block *dead = b->_succs[1-idx]; 570 for( j = 1; j < dead->num_preds(); j++) 571 if( dead->pred(j)->in(0) == bp ) 572 break; 573 // Scan through block, yanking dead path from 574 // all regions and phis. 575 dead->head()->del_req(j); 576 for( int k = 1; dead->_nodes[k]->is_Phi(); k++ ) 577 dead->_nodes[k]->del_req(j); 578 } 579 580 //------------------------------move_to_next----------------------------------- 581 // Helper function to move block bx to the slot following b_index. Return 582 // true if the move is successful, otherwise false 583 bool PhaseCFG::move_to_next(Block* bx, uint b_index) { 584 if (bx == NULL) return false; 585 586 // Return false if bx is already scheduled. 587 uint bx_index = bx->_pre_order; 588 if ((bx_index <= b_index) && (_blocks[bx_index] == bx)) { 589 return false; 590 } 591 592 // Find the current index of block bx on the block list 593 bx_index = b_index + 1; 594 while( bx_index < _num_blocks && _blocks[bx_index] != bx ) bx_index++; 595 assert(_blocks[bx_index] == bx, "block not found"); 596 597 // If the previous block conditionally falls into bx, return false, 598 // because moving bx will create an extra jump. 599 for(uint k = 1; k < bx->num_preds(); k++ ) { 600 Block* pred = _bbs[bx->pred(k)->_idx]; 601 if (pred == _blocks[bx_index-1]) { 602 if (pred->_num_succs != 1) { 603 return false; 604 } 605 } 606 } 607 608 // Reinsert bx just past block 'b' 609 _blocks.remove(bx_index); 610 _blocks.insert(b_index + 1, bx); 611 return true; 612 } 613 614 //------------------------------move_to_end------------------------------------ 615 // Move empty and uncommon blocks to the end. 616 void PhaseCFG::move_to_end(Block *b, uint i) { 617 int e = b->is_Empty(); 618 if (e != Block::not_empty) { 619 if (e == Block::empty_with_goto) { 620 // Remove the goto, but leave the block. 621 b->_nodes.pop(); 622 } 623 // Mark this block as a connector block, which will cause it to be 624 // ignored in certain functions such as non_connector_successor(). 625 b->set_connector(); 626 } 627 // Move the empty block to the end, and don't recheck. 628 _blocks.remove(i); 629 _blocks.push(b); 630 } 631 632 //---------------------------set_loop_alignment-------------------------------- 633 // Set loop alignment for every block 634 void PhaseCFG::set_loop_alignment() { 635 uint last = _num_blocks; 636 assert( _blocks[0] == _broot, "" ); 637 638 for (uint i = 1; i < last; i++ ) { 639 Block *b = _blocks[i]; 640 if (b->head()->is_Loop()) { 641 b->set_loop_alignment(b); 642 } 643 } 644 } 645 646 //-----------------------------remove_empty------------------------------------ 647 // Make empty basic blocks to be "connector" blocks, Move uncommon blocks 648 // to the end. 649 void PhaseCFG::remove_empty() { 650 // Move uncommon blocks to the end 651 uint last = _num_blocks; 652 assert( _blocks[0] == _broot, "" ); 653 654 for (uint i = 1; i < last; i++) { 655 Block *b = _blocks[i]; 656 if (b->is_connector()) break; 657 658 // Check for NeverBranch at block end. This needs to become a GOTO to the 659 // true target. NeverBranch are treated as a conditional branch that 660 // always goes the same direction for most of the optimizer and are used 661 // to give a fake exit path to infinite loops. At this late stage they 662 // need to turn into Goto's so that when you enter the infinite loop you 663 // indeed hang. 664 if( b->_nodes[b->end_idx()]->Opcode() == Op_NeverBranch ) 665 convert_NeverBranch_to_Goto(b); 666 667 // Look for uncommon blocks and move to end. 668 if (!C->do_freq_based_layout()) { 669 if( b->is_uncommon(_bbs) ) { 670 move_to_end(b, i); 671 last--; // No longer check for being uncommon! 672 if( no_flip_branch(b) ) { // Fall-thru case must follow? 673 b = _blocks[i]; // Find the fall-thru block 674 move_to_end(b, i); 675 last--; 676 } 677 i--; // backup block counter post-increment 678 } 679 } 680 } 681 682 // Move empty blocks to the end 683 last = _num_blocks; 684 for (uint i = 1; i < last; i++) { 685 Block *b = _blocks[i]; 686 if (b->is_Empty() != Block::not_empty) { 687 move_to_end(b, i); 688 last--; 689 i--; 690 } 691 } // End of for all blocks 692 } 693 694 //-----------------------------fixup_flow-------------------------------------- 695 // Fix up the final control flow for basic blocks. 696 void PhaseCFG::fixup_flow() { 697 // Fixup final control flow for the blocks. Remove jump-to-next 698 // block. If neither arm of a IF follows the conditional branch, we 699 // have to add a second jump after the conditional. We place the 700 // TRUE branch target in succs[0] for both GOTOs and IFs. 701 for (uint i=0; i < _num_blocks; i++) { 702 Block *b = _blocks[i]; 703 b->_pre_order = i; // turn pre-order into block-index 704 705 // Connector blocks need no further processing. 706 if (b->is_connector()) { 707 assert((i+1) == _num_blocks || _blocks[i+1]->is_connector(), 708 "All connector blocks should sink to the end"); 709 continue; 710 } 711 assert(b->is_Empty() != Block::completely_empty, 712 "Empty blocks should be connectors"); 713 714 Block *bnext = (i < _num_blocks-1) ? _blocks[i+1] : NULL; 715 Block *bs0 = b->non_connector_successor(0); 716 717 // Check for multi-way branches where I cannot negate the test to 718 // exchange the true and false targets. 719 if( no_flip_branch( b ) ) { 720 // Find fall through case - if must fall into its target 721 int branch_idx = b->_nodes.size() - b->_num_succs; 722 for (uint j2 = 0; j2 < b->_num_succs; j2++) { 723 const ProjNode* p = b->_nodes[branch_idx + j2]->as_Proj(); 724 if (p->_con == 0) { 725 // successor j2 is fall through case 726 if (b->non_connector_successor(j2) != bnext) { 727 // but it is not the next block => insert a goto 728 insert_goto_at(i, j2); 729 } 730 // Put taken branch in slot 0 731 if( j2 == 0 && b->_num_succs == 2) { 732 // Flip targets in succs map 733 Block *tbs0 = b->_succs[0]; 734 Block *tbs1 = b->_succs[1]; 735 b->_succs.map( 0, tbs1 ); 736 b->_succs.map( 1, tbs0 ); 737 } 738 break; 739 } 740 } 741 // Remove all CatchProjs 742 for (uint j1 = 0; j1 < b->_num_succs; j1++) b->_nodes.pop(); 743 744 } else if (b->_num_succs == 1) { 745 // Block ends in a Goto? 746 if (bnext == bs0) { 747 // We fall into next block; remove the Goto 748 b->_nodes.pop(); 749 } 750 751 } else if( b->_num_succs == 2 ) { // Block ends in a If? 752 // Get opcode of 1st projection (matches _succs[0]) 753 // Note: Since this basic block has 2 exits, the last 2 nodes must 754 // be projections (in any order), the 3rd last node must be 755 // the IfNode (we have excluded other 2-way exits such as 756 // CatchNodes already). 757 MachNode *iff = b->_nodes[b->_nodes.size()-3]->as_Mach(); 758 ProjNode *proj0 = b->_nodes[b->_nodes.size()-2]->as_Proj(); 759 ProjNode *proj1 = b->_nodes[b->_nodes.size()-1]->as_Proj(); 760 761 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1]. 762 assert(proj0->raw_out(0) == b->_succs[0]->head(), "Mismatch successor 0"); 763 assert(proj1->raw_out(0) == b->_succs[1]->head(), "Mismatch successor 1"); 764 765 Block *bs1 = b->non_connector_successor(1); 766 767 // Check for neither successor block following the current 768 // block ending in a conditional. If so, move one of the 769 // successors after the current one, provided that the 770 // successor was previously unscheduled, but moveable 771 // (i.e., all paths to it involve a branch). 772 if( !C->do_freq_based_layout() && bnext != bs0 && bnext != bs1 ) { 773 // Choose the more common successor based on the probability 774 // of the conditional branch. 775 Block *bx = bs0; 776 Block *by = bs1; 777 778 // _prob is the probability of taking the true path. Make 779 // p the probability of taking successor #1. 780 float p = iff->as_MachIf()->_prob; 781 if( proj0->Opcode() == Op_IfTrue ) { 782 p = 1.0 - p; 783 } 784 785 // Prefer successor #1 if p > 0.5 786 if (p > PROB_FAIR) { 787 bx = bs1; 788 by = bs0; 789 } 790 791 // Attempt the more common successor first 792 if (move_to_next(bx, i)) { 793 bnext = bx; 794 } else if (move_to_next(by, i)) { 795 bnext = by; 796 } 797 } 798 799 // Check for conditional branching the wrong way. Negate 800 // conditional, if needed, so it falls into the following block 801 // and branches to the not-following block. 802 803 // Check for the next block being in succs[0]. We are going to branch 804 // to succs[0], so we want the fall-thru case as the next block in 805 // succs[1]. 806 if (bnext == bs0) { 807 // Fall-thru case in succs[0], so flip targets in succs map 808 Block *tbs0 = b->_succs[0]; 809 Block *tbs1 = b->_succs[1]; 810 b->_succs.map( 0, tbs1 ); 811 b->_succs.map( 1, tbs0 ); 812 // Flip projection for each target 813 { ProjNode *tmp = proj0; proj0 = proj1; proj1 = tmp; } 814 815 } else if( bnext != bs1 ) { 816 // Need a double-branch 817 // The existing conditional branch need not change. 818 // Add a unconditional branch to the false target. 819 // Alas, it must appear in its own block and adding a 820 // block this late in the game is complicated. Sigh. 821 insert_goto_at(i, 1); 822 } 823 824 // Make sure we TRUE branch to the target 825 if( proj0->Opcode() == Op_IfFalse ) { 826 iff->negate(); 827 } 828 829 b->_nodes.pop(); // Remove IfFalse & IfTrue projections 830 b->_nodes.pop(); 831 832 } else { 833 // Multi-exit block, e.g. a switch statement 834 // But we don't need to do anything here 835 } 836 } // End of for all blocks 837 } 838 839 840 //------------------------------dump------------------------------------------- 841 #ifndef PRODUCT 842 void PhaseCFG::_dump_cfg( const Node *end, VectorSet &visited ) const { 843 const Node *x = end->is_block_proj(); 844 assert( x, "not a CFG" ); 845 846 // Do not visit this block again 847 if( visited.test_set(x->_idx) ) return; 848 849 // Skip through this block 850 const Node *p = x; 851 do { 852 p = p->in(0); // Move control forward 853 assert( !p->is_block_proj() || p->is_Root(), "not a CFG" ); 854 } while( !p->is_block_start() ); 855 856 // Recursively visit 857 for( uint i=1; i<p->req(); i++ ) 858 _dump_cfg(p->in(i),visited); 859 860 // Dump the block 861 _bbs[p->_idx]->dump(&_bbs); 862 } 863 864 void PhaseCFG::dump( ) const { 865 tty->print("\n--- CFG --- %d BBs\n",_num_blocks); 866 if( _blocks.size() ) { // Did we do basic-block layout? 867 for( uint i=0; i<_num_blocks; i++ ) 868 _blocks[i]->dump(&_bbs); 869 } else { // Else do it with a DFS 870 VectorSet visited(_bbs._arena); 871 _dump_cfg(_root,visited); 872 } 873 } 874 875 void PhaseCFG::dump_headers() { 876 for( uint i = 0; i < _num_blocks; i++ ) { 877 if( _blocks[i] == NULL ) continue; 878 _blocks[i]->dump_head(&_bbs); 879 } 880 } 881 882 void PhaseCFG::verify( ) const { 883 #ifdef ASSERT 884 // Verify sane CFG 885 for( uint i = 0; i < _num_blocks; i++ ) { 886 Block *b = _blocks[i]; 887 uint cnt = b->_nodes.size(); 888 uint j; 889 for( j = 0; j < cnt; j++ ) { 890 Node *n = b->_nodes[j]; 891 assert( _bbs[n->_idx] == b, "" ); 892 if( j >= 1 && n->is_Mach() && 893 n->as_Mach()->ideal_Opcode() == Op_CreateEx ) { 894 assert( j == 1 || b->_nodes[j-1]->is_Phi(), 895 "CreateEx must be first instruction in block" ); 896 } 897 for( uint k = 0; k < n->req(); k++ ) { 898 Node *def = n->in(k); 899 if( def && def != n ) { 900 assert( _bbs[def->_idx] || def->is_Con(), 901 "must have block; constants for debug info ok" ); 902 // Verify that instructions in the block is in correct order. 903 // Uses must follow their definition if they are at the same block. 904 // Mostly done to check that MachSpillCopy nodes are placed correctly 905 // when CreateEx node is moved in build_ifg_physical(). 906 if( _bbs[def->_idx] == b && 907 !(b->head()->is_Loop() && n->is_Phi()) && 908 // See (+++) comment in reg_split.cpp 909 !(n->jvms() != NULL && n->jvms()->is_monitor_use(k)) ) { 910 assert( b->find_node(def) < j, "uses must follow definitions" ); 911 } 912 if( def->is_SafePointScalarObject() ) { 913 assert(_bbs[def->_idx] == b, "SafePointScalarObject Node should be at the same block as its SafePoint node"); 914 assert(_bbs[def->_idx] == _bbs[def->in(0)->_idx], "SafePointScalarObject Node should be at the same block as its control edge"); 915 } 916 } 917 } 918 } 919 920 j = b->end_idx(); 921 Node *bp = (Node*)b->_nodes[b->_nodes.size()-1]->is_block_proj(); 922 assert( bp, "last instruction must be a block proj" ); 923 assert( bp == b->_nodes[j], "wrong number of successors for this block" ); 924 if( bp->is_Catch() ) { 925 while( b->_nodes[--j]->Opcode() == Op_MachProj ) ; 926 assert( b->_nodes[j]->is_Call(), "CatchProj must follow call" ); 927 } 928 else if( bp->is_Mach() && bp->as_Mach()->ideal_Opcode() == Op_If ) { 929 assert( b->_num_succs == 2, "Conditional branch must have two targets"); 930 } 931 } 932 #endif 933 } 934 #endif 935 936 //============================================================================= 937 //------------------------------UnionFind-------------------------------------- 938 UnionFind::UnionFind( uint max ) : _cnt(max), _max(max), _indices(NEW_RESOURCE_ARRAY(uint,max)) { 939 Copy::zero_to_bytes( _indices, sizeof(uint)*max ); 940 } 941 942 void UnionFind::extend( uint from_idx, uint to_idx ) { 943 _nesting.check(); 944 if( from_idx >= _max ) { 945 uint size = 16; 946 while( size <= from_idx ) size <<=1; 947 _indices = REALLOC_RESOURCE_ARRAY( uint, _indices, _max, size ); 948 _max = size; 949 } 950 while( _cnt <= from_idx ) _indices[_cnt++] = 0; 951 _indices[from_idx] = to_idx; 952 } 953 954 void UnionFind::reset( uint max ) { 955 assert( max <= max_uint, "Must fit within uint" ); 956 // Force the Union-Find mapping to be at least this large 957 extend(max,0); 958 // Initialize to be the ID mapping. 959 for( uint i=0; i<max; i++ ) map(i,i); 960 } 961 962 //------------------------------Find_compress---------------------------------- 963 // Straight out of Tarjan's union-find algorithm 964 uint UnionFind::Find_compress( uint idx ) { 965 uint cur = idx; 966 uint next = lookup(cur); 967 while( next != cur ) { // Scan chain of equivalences 968 assert( next < cur, "always union smaller" ); 969 cur = next; // until find a fixed-point 970 next = lookup(cur); 971 } 972 // Core of union-find algorithm: update chain of 973 // equivalences to be equal to the root. 974 while( idx != next ) { 975 uint tmp = lookup(idx); 976 map(idx, next); 977 idx = tmp; 978 } 979 return idx; 980 } 981 982 //------------------------------Find_const------------------------------------- 983 // Like Find above, but no path compress, so bad asymptotic behavior 984 uint UnionFind::Find_const( uint idx ) const { 985 if( idx == 0 ) return idx; // Ignore the zero idx 986 // Off the end? This can happen during debugging dumps 987 // when data structures have not finished being updated. 988 if( idx >= _max ) return idx; 989 uint next = lookup(idx); 990 while( next != idx ) { // Scan chain of equivalences 991 idx = next; // until find a fixed-point 992 next = lookup(idx); 993 } 994 return next; 995 } 996 997 //------------------------------Union------------------------------------------ 998 // union 2 sets together. 999 void UnionFind::Union( uint idx1, uint idx2 ) { 1000 uint src = Find(idx1); 1001 uint dst = Find(idx2); 1002 assert( src, "" ); 1003 assert( dst, "" ); 1004 assert( src < _max, "oob" ); 1005 assert( dst < _max, "oob" ); 1006 assert( src < dst, "always union smaller" ); 1007 map(dst,src); 1008 } 1009 1010 #ifndef PRODUCT 1011 static void edge_dump(GrowableArray<CFGEdge *> *edges) { 1012 tty->print_cr("---- Edges ----"); 1013 for (int i = 0; i < edges->length(); i++) { 1014 CFGEdge *e = edges->at(i); 1015 if (e != NULL) { 1016 edges->at(i)->dump(); 1017 } 1018 } 1019 } 1020 1021 static void trace_dump(Trace *traces[], int count) { 1022 tty->print_cr("---- Traces ----"); 1023 for (int i = 0; i < count; i++) { 1024 Trace *tr = traces[i]; 1025 if (tr != NULL) { 1026 tr->dump(); 1027 } 1028 } 1029 } 1030 1031 void Trace::dump( ) const { 1032 tty->print_cr("Trace (freq %f)", first_block()->_freq); 1033 for (Block *b = first_block(); b != NULL; b = next(b)) { 1034 tty->print(" B%d", b->_pre_order); 1035 if (b->head()->is_Loop()) { 1036 tty->print(" (L%d)", b->compute_loop_alignment()); 1037 } 1038 if (b->has_loop_alignment()) { 1039 tty->print(" (T%d)", b->code_alignment()); 1040 } 1041 } 1042 tty->cr(); 1043 } 1044 1045 void CFGEdge::dump( ) const { 1046 tty->print(" B%d --> B%d Freq: %f out:%3d%% in:%3d%% State: ", 1047 from()->_pre_order, to()->_pre_order, freq(), _from_pct, _to_pct); 1048 switch(state()) { 1049 case connected: 1050 tty->print("connected"); 1051 break; 1052 case open: 1053 tty->print("open"); 1054 break; 1055 case interior: 1056 tty->print("interior"); 1057 break; 1058 } 1059 if (infrequent()) { 1060 tty->print(" infrequent"); 1061 } 1062 tty->cr(); 1063 } 1064 #endif 1065 1066 //============================================================================= 1067 1068 //------------------------------edge_order------------------------------------- 1069 // Comparison function for edges 1070 static int edge_order(CFGEdge **e0, CFGEdge **e1) { 1071 float freq0 = (*e0)->freq(); 1072 float freq1 = (*e1)->freq(); 1073 if (freq0 != freq1) { 1074 return freq0 > freq1 ? -1 : 1; 1075 } 1076 1077 int dist0 = (*e0)->to()->_rpo - (*e0)->from()->_rpo; 1078 int dist1 = (*e1)->to()->_rpo - (*e1)->from()->_rpo; 1079 1080 return dist1 - dist0; 1081 } 1082 1083 //------------------------------trace_frequency_order-------------------------- 1084 // Comparison function for edges 1085 static int trace_frequency_order(const void *p0, const void *p1) { 1086 Trace *tr0 = *(Trace **) p0; 1087 Trace *tr1 = *(Trace **) p1; 1088 Block *b0 = tr0->first_block(); 1089 Block *b1 = tr1->first_block(); 1090 1091 // The trace of connector blocks goes at the end; 1092 // we only expect one such trace 1093 if (b0->is_connector() != b1->is_connector()) { 1094 return b1->is_connector() ? -1 : 1; 1095 } 1096 1097 // Pull more frequently executed blocks to the beginning 1098 float freq0 = b0->_freq; 1099 float freq1 = b1->_freq; 1100 if (freq0 != freq1) { 1101 return freq0 > freq1 ? -1 : 1; 1102 } 1103 1104 int diff = tr0->first_block()->_rpo - tr1->first_block()->_rpo; 1105 1106 return diff; 1107 } 1108 1109 //------------------------------find_edges------------------------------------- 1110 // Find edges of interest, i.e, those which can fall through. Presumes that 1111 // edges which don't fall through are of low frequency and can be generally 1112 // ignored. Initialize the list of traces. 1113 void PhaseBlockLayout::find_edges() 1114 { 1115 // Walk the blocks, creating edges and Traces 1116 uint i; 1117 Trace *tr = NULL; 1118 for (i = 0; i < _cfg._num_blocks; i++) { 1119 Block *b = _cfg._blocks[i]; 1120 tr = new Trace(b, next, prev); 1121 traces[tr->id()] = tr; 1122 1123 // All connector blocks should be at the end of the list 1124 if (b->is_connector()) break; 1125 1126 // If this block and the next one have a one-to-one successor 1127 // predecessor relationship, simply append the next block 1128 int nfallthru = b->num_fall_throughs(); 1129 while (nfallthru == 1 && 1130 b->succ_fall_through(0)) { 1131 Block *n = b->_succs[0]; 1132 1133 // Skip over single-entry connector blocks, we don't want to 1134 // add them to the trace. 1135 while (n->is_connector() && n->num_preds() == 1) { 1136 n = n->_succs[0]; 1137 } 1138 1139 // We see a merge point, so stop search for the next block 1140 if (n->num_preds() != 1) break; 1141 1142 i++; 1143 assert(n = _cfg._blocks[i], "expecting next block"); 1144 tr->append(n); 1145 uf->map(n->_pre_order, tr->id()); 1146 traces[n->_pre_order] = NULL; 1147 nfallthru = b->num_fall_throughs(); 1148 b = n; 1149 } 1150 1151 if (nfallthru > 0) { 1152 // Create a CFGEdge for each outgoing 1153 // edge that could be a fall-through. 1154 for (uint j = 0; j < b->_num_succs; j++ ) { 1155 if (b->succ_fall_through(j)) { 1156 Block *target = b->non_connector_successor(j); 1157 float freq = b->_freq * b->succ_prob(j); 1158 int from_pct = (int) ((100 * freq) / b->_freq); 1159 int to_pct = (int) ((100 * freq) / target->_freq); 1160 edges->append(new CFGEdge(b, target, freq, from_pct, to_pct)); 1161 } 1162 } 1163 } 1164 } 1165 1166 // Group connector blocks into one trace 1167 for (i++; i < _cfg._num_blocks; i++) { 1168 Block *b = _cfg._blocks[i]; 1169 assert(b->is_connector(), "connector blocks at the end"); 1170 tr->append(b); 1171 uf->map(b->_pre_order, tr->id()); 1172 traces[b->_pre_order] = NULL; 1173 } 1174 } 1175 1176 //------------------------------union_traces---------------------------------- 1177 // Union two traces together in uf, and null out the trace in the list 1178 void PhaseBlockLayout::union_traces(Trace* updated_trace, Trace* old_trace) 1179 { 1180 uint old_id = old_trace->id(); 1181 uint updated_id = updated_trace->id(); 1182 1183 uint lo_id = updated_id; 1184 uint hi_id = old_id; 1185 1186 // If from is greater than to, swap values to meet 1187 // UnionFind guarantee. 1188 if (updated_id > old_id) { 1189 lo_id = old_id; 1190 hi_id = updated_id; 1191 1192 // Fix up the trace ids 1193 traces[lo_id] = traces[updated_id]; 1194 updated_trace->set_id(lo_id); 1195 } 1196 1197 // Union the lower with the higher and remove the pointer 1198 // to the higher. 1199 uf->Union(lo_id, hi_id); 1200 traces[hi_id] = NULL; 1201 } 1202 1203 //------------------------------grow_traces------------------------------------- 1204 // Append traces together via the most frequently executed edges 1205 void PhaseBlockLayout::grow_traces() 1206 { 1207 // Order the edges, and drive the growth of Traces via the most 1208 // frequently executed edges. 1209 edges->sort(edge_order); 1210 for (int i = 0; i < edges->length(); i++) { 1211 CFGEdge *e = edges->at(i); 1212 1213 if (e->state() != CFGEdge::open) continue; 1214 1215 Block *src_block = e->from(); 1216 Block *targ_block = e->to(); 1217 1218 // Don't grow traces along backedges? 1219 if (!BlockLayoutRotateLoops) { 1220 if (targ_block->_rpo <= src_block->_rpo) { 1221 targ_block->set_loop_alignment(targ_block); 1222 continue; 1223 } 1224 } 1225 1226 Trace *src_trace = trace(src_block); 1227 Trace *targ_trace = trace(targ_block); 1228 1229 // If the edge in question can join two traces at their ends, 1230 // append one trace to the other. 1231 if (src_trace->last_block() == src_block) { 1232 if (src_trace == targ_trace) { 1233 e->set_state(CFGEdge::interior); 1234 if (targ_trace->backedge(e)) { 1235 // Reset i to catch any newly eligible edge 1236 // (Or we could remember the first "open" edge, and reset there) 1237 i = 0; 1238 } 1239 } else if (targ_trace->first_block() == targ_block) { 1240 e->set_state(CFGEdge::connected); 1241 src_trace->append(targ_trace); 1242 union_traces(src_trace, targ_trace); 1243 } 1244 } 1245 } 1246 } 1247 1248 //------------------------------merge_traces----------------------------------- 1249 // Embed one trace into another, if the fork or join points are sufficiently 1250 // balanced. 1251 void PhaseBlockLayout::merge_traces(bool fall_thru_only) 1252 { 1253 // Walk the edge list a another time, looking at unprocessed edges. 1254 // Fold in diamonds 1255 for (int i = 0; i < edges->length(); i++) { 1256 CFGEdge *e = edges->at(i); 1257 1258 if (e->state() != CFGEdge::open) continue; 1259 if (fall_thru_only) { 1260 if (e->infrequent()) continue; 1261 } 1262 1263 Block *src_block = e->from(); 1264 Trace *src_trace = trace(src_block); 1265 bool src_at_tail = src_trace->last_block() == src_block; 1266 1267 Block *targ_block = e->to(); 1268 Trace *targ_trace = trace(targ_block); 1269 bool targ_at_start = targ_trace->first_block() == targ_block; 1270 1271 if (src_trace == targ_trace) { 1272 // This may be a loop, but we can't do much about it. 1273 e->set_state(CFGEdge::interior); 1274 continue; 1275 } 1276 1277 if (fall_thru_only) { 1278 // If the edge links the middle of two traces, we can't do anything. 1279 // Mark the edge and continue. 1280 if (!src_at_tail & !targ_at_start) { 1281 continue; 1282 } 1283 1284 // Don't grow traces along backedges? 1285 if (!BlockLayoutRotateLoops && (targ_block->_rpo <= src_block->_rpo)) { 1286 continue; 1287 } 1288 1289 // If both ends of the edge are available, why didn't we handle it earlier? 1290 assert(src_at_tail ^ targ_at_start, "Should have caught this edge earlier."); 1291 1292 if (targ_at_start) { 1293 // Insert the "targ" trace in the "src" trace if the insertion point 1294 // is a two way branch. 1295 // Better profitability check possible, but may not be worth it. 1296 // Someday, see if the this "fork" has an associated "join"; 1297 // then make a policy on merging this trace at the fork or join. 1298 // For example, other things being equal, it may be better to place this 1299 // trace at the join point if the "src" trace ends in a two-way, but 1300 // the insertion point is one-way. 1301 assert(src_block->num_fall_throughs() == 2, "unexpected diamond"); 1302 e->set_state(CFGEdge::connected); 1303 src_trace->insert_after(src_block, targ_trace); 1304 union_traces(src_trace, targ_trace); 1305 } else if (src_at_tail) { 1306 if (src_trace != trace(_cfg._broot)) { 1307 e->set_state(CFGEdge::connected); 1308 targ_trace->insert_before(targ_block, src_trace); 1309 union_traces(targ_trace, src_trace); 1310 } 1311 } 1312 } else if (e->state() == CFGEdge::open) { 1313 // Append traces, even without a fall-thru connection. 1314 // But leave root entry at the begining of the block list. 1315 if (targ_trace != trace(_cfg._broot)) { 1316 e->set_state(CFGEdge::connected); 1317 src_trace->append(targ_trace); 1318 union_traces(src_trace, targ_trace); 1319 } 1320 } 1321 } 1322 } 1323 1324 //----------------------------reorder_traces----------------------------------- 1325 // Order the sequence of the traces in some desirable way, and fixup the 1326 // jumps at the end of each block. 1327 void PhaseBlockLayout::reorder_traces(int count) 1328 { 1329 ResourceArea *area = Thread::current()->resource_area(); 1330 Trace ** new_traces = NEW_ARENA_ARRAY(area, Trace *, count); 1331 Block_List worklist; 1332 int new_count = 0; 1333 1334 // Compact the traces. 1335 for (int i = 0; i < count; i++) { 1336 Trace *tr = traces[i]; 1337 if (tr != NULL) { 1338 new_traces[new_count++] = tr; 1339 } 1340 } 1341 1342 // The entry block should be first on the new trace list. 1343 Trace *tr = trace(_cfg._broot); 1344 assert(tr == new_traces[0], "entry trace misplaced"); 1345 1346 // Sort the new trace list by frequency 1347 qsort(new_traces + 1, new_count - 1, sizeof(new_traces[0]), trace_frequency_order); 1348 1349 // Patch up the successor blocks 1350 _cfg._blocks.reset(); 1351 _cfg._num_blocks = 0; 1352 for (int i = 0; i < new_count; i++) { 1353 Trace *tr = new_traces[i]; 1354 if (tr != NULL) { 1355 tr->fixup_blocks(_cfg); 1356 } 1357 } 1358 } 1359 1360 //------------------------------PhaseBlockLayout------------------------------- 1361 // Order basic blocks based on frequency 1362 PhaseBlockLayout::PhaseBlockLayout(PhaseCFG &cfg) : 1363 Phase(BlockLayout), 1364 _cfg(cfg) 1365 { 1366 ResourceMark rm; 1367 ResourceArea *area = Thread::current()->resource_area(); 1368 1369 // List of traces 1370 int size = _cfg._num_blocks + 1; 1371 traces = NEW_ARENA_ARRAY(area, Trace *, size); 1372 memset(traces, 0, size*sizeof(Trace*)); 1373 next = NEW_ARENA_ARRAY(area, Block *, size); 1374 memset(next, 0, size*sizeof(Block *)); 1375 prev = NEW_ARENA_ARRAY(area, Block *, size); 1376 memset(prev , 0, size*sizeof(Block *)); 1377 1378 // List of edges 1379 edges = new GrowableArray<CFGEdge*>; 1380 1381 // Mapping block index --> block_trace 1382 uf = new UnionFind(size); 1383 uf->reset(size); 1384 1385 // Find edges and create traces. 1386 find_edges(); 1387 1388 // Grow traces at their ends via most frequent edges. 1389 grow_traces(); 1390 1391 // Merge one trace into another, but only at fall-through points. 1392 // This may make diamonds and other related shapes in a trace. 1393 merge_traces(true); 1394 1395 // Run merge again, allowing two traces to be catenated, even if 1396 // one does not fall through into the other. This appends loosely 1397 // related traces to be near each other. 1398 merge_traces(false); 1399 1400 // Re-order all the remaining traces by frequency 1401 reorder_traces(size); 1402 1403 assert(_cfg._num_blocks >= (uint) (size - 1), "number of blocks can not shrink"); 1404 } 1405 1406 1407 //------------------------------backedge--------------------------------------- 1408 // Edge e completes a loop in a trace. If the target block is head of the 1409 // loop, rotate the loop block so that the loop ends in a conditional branch. 1410 bool Trace::backedge(CFGEdge *e) { 1411 bool loop_rotated = false; 1412 Block *src_block = e->from(); 1413 Block *targ_block = e->to(); 1414 1415 assert(last_block() == src_block, "loop discovery at back branch"); 1416 if (first_block() == targ_block) { 1417 if (BlockLayoutRotateLoops && last_block()->num_fall_throughs() < 2) { 1418 // Find the last block in the trace that has a conditional 1419 // branch. 1420 Block *b; 1421 for (b = last_block(); b != NULL; b = prev(b)) { 1422 if (b->num_fall_throughs() == 2) { 1423 break; 1424 } 1425 } 1426 1427 if (b != last_block() && b != NULL) { 1428 loop_rotated = true; 1429 1430 // Rotate the loop by doing two-part linked-list surgery. 1431 append(first_block()); 1432 break_loop_after(b); 1433 } 1434 } 1435 1436 // Backbranch to the top of a trace 1437 // Scroll foward through the trace from the targ_block. If we find 1438 // a loop head before another loop top, use the the loop head alignment. 1439 for (Block *b = targ_block; b != NULL; b = next(b)) { 1440 if (b->has_loop_alignment()) { 1441 break; 1442 } 1443 if (b->head()->is_Loop()) { 1444 targ_block = b; 1445 break; 1446 } 1447 } 1448 1449 first_block()->set_loop_alignment(targ_block); 1450 1451 } else { 1452 // Backbranch into the middle of a trace 1453 targ_block->set_loop_alignment(targ_block); 1454 } 1455 1456 return loop_rotated; 1457 } 1458 1459 //------------------------------fixup_blocks----------------------------------- 1460 // push blocks onto the CFG list 1461 // ensure that blocks have the correct two-way branch sense 1462 void Trace::fixup_blocks(PhaseCFG &cfg) { 1463 Block *last = last_block(); 1464 for (Block *b = first_block(); b != NULL; b = next(b)) { 1465 cfg._blocks.push(b); 1466 cfg._num_blocks++; 1467 if (!b->is_connector()) { 1468 int nfallthru = b->num_fall_throughs(); 1469 if (b != last) { 1470 if (nfallthru == 2) { 1471 // Ensure that the sense of the branch is correct 1472 Block *bnext = next(b); 1473 Block *bs0 = b->non_connector_successor(0); 1474 1475 MachNode *iff = b->_nodes[b->_nodes.size()-3]->as_Mach(); 1476 ProjNode *proj0 = b->_nodes[b->_nodes.size()-2]->as_Proj(); 1477 ProjNode *proj1 = b->_nodes[b->_nodes.size()-1]->as_Proj(); 1478 1479 if (bnext == bs0) { 1480 // Fall-thru case in succs[0], should be in succs[1] 1481 1482 // Flip targets in _succs map 1483 Block *tbs0 = b->_succs[0]; 1484 Block *tbs1 = b->_succs[1]; 1485 b->_succs.map( 0, tbs1 ); 1486 b->_succs.map( 1, tbs0 ); 1487 1488 // Flip projections to match targets 1489 b->_nodes.map(b->_nodes.size()-2, proj1); 1490 b->_nodes.map(b->_nodes.size()-1, proj0); 1491 } 1492 } 1493 } 1494 } 1495 } 1496 }