1 /* 2 * Copyright (c) 1997, 2010, 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 "opto/block.hpp" 29 #include "opto/cfgnode.hpp" 30 #include "opto/chaitin.hpp" 31 #include "opto/loopnode.hpp" 32 #include "opto/machnode.hpp" 33 #include "opto/matcher.hpp" 34 #include "opto/opcodes.hpp" 35 #include "opto/rootnode.hpp" 36 #include "utilities/copy.hpp" 37 38 // Optimization - Graph Style 39 40 41 //----------------------------------------------------------------------------- 42 void Block_Array::grow( uint i ) { 43 assert(i >= Max(), "must be an overflow"); 44 debug_only(_limit = i+1); 45 if( i < _size ) return; 46 if( !_size ) { 47 _size = 1; 48 _blocks = (Block**)_arena->Amalloc( _size * sizeof(Block*) ); 49 _blocks[0] = NULL; 50 } 51 uint old = _size; 52 while( i >= _size ) _size <<= 1; // Double to fit 53 _blocks = (Block**)_arena->Arealloc( _blocks, old*sizeof(Block*),_size*sizeof(Block*)); 54 Copy::zero_to_bytes( &_blocks[old], (_size-old)*sizeof(Block*) ); 55 } 56 57 //============================================================================= 58 void Block_List::remove(uint i) { 59 assert(i < _cnt, "index out of bounds"); 60 Copy::conjoint_words_to_lower((HeapWord*)&_blocks[i+1], (HeapWord*)&_blocks[i], ((_cnt-i-1)*sizeof(Block*))); 61 pop(); // shrink list by one block 62 } 63 64 void Block_List::insert(uint i, Block *b) { 65 push(b); // grow list by one block 66 Copy::conjoint_words_to_higher((HeapWord*)&_blocks[i], (HeapWord*)&_blocks[i+1], ((_cnt-i-1)*sizeof(Block*))); 67 _blocks[i] = b; 68 } 69 70 #ifndef PRODUCT 71 void Block_List::print() { 72 for (uint i=0; i < size(); i++) { 73 tty->print("B%d ", _blocks[i]->_pre_order); 74 } 75 tty->print("size = %d\n", size()); 76 } 77 #endif 78 79 //============================================================================= 80 81 uint Block::code_alignment() { 82 // Check for Root block 83 if( _pre_order == 0 ) return CodeEntryAlignment; 84 // Check for Start block 85 if( _pre_order == 1 ) return InteriorEntryAlignment; 86 // Check for loop alignment 87 if (has_loop_alignment()) return loop_alignment(); 88 89 return 1; // no particular alignment 90 } 91 92 uint Block::compute_loop_alignment() { 93 Node *h = head(); 94 if( h->is_Loop() && h->as_Loop()->is_inner_loop() ) { 95 // Pre- and post-loops have low trip count so do not bother with 96 // NOPs for align loop head. The constants are hidden from tuning 97 // but only because my "divide by 4" heuristic surely gets nearly 98 // all possible gain (a "do not align at all" heuristic has a 99 // chance of getting a really tiny gain). 100 if( h->is_CountedLoop() && (h->as_CountedLoop()->is_pre_loop() || 101 h->as_CountedLoop()->is_post_loop()) ) 102 return (OptoLoopAlignment > 4) ? (OptoLoopAlignment>>2) : 1; 103 // Loops with low backedge frequency should not be aligned. 104 Node *n = h->in(LoopNode::LoopBackControl)->in(0); 105 if( n->is_MachIf() && n->as_MachIf()->_prob < 0.01 ) { 106 return 1; // Loop does not loop, more often than not! 107 } 108 return OptoLoopAlignment; // Otherwise align loop head 109 } 110 111 return 1; // no particular alignment 112 } 113 114 //----------------------------------------------------------------------------- 115 // Compute the size of first 'inst_cnt' instructions in this block. 116 // Return the number of instructions left to compute if the block has 117 // less then 'inst_cnt' instructions. Stop, and return 0 if sum_size 118 // exceeds OptoLoopAlignment. 119 uint Block::compute_first_inst_size(uint& sum_size, uint inst_cnt, 120 PhaseRegAlloc* ra) { 121 uint last_inst = _nodes.size(); 122 for( uint j = 0; j < last_inst && inst_cnt > 0; j++ ) { 123 uint inst_size = _nodes[j]->size(ra); 124 if( inst_size > 0 ) { 125 inst_cnt--; 126 uint sz = sum_size + inst_size; 127 if( sz <= (uint)OptoLoopAlignment ) { 128 // Compute size of instructions which fit into fetch buffer only 129 // since all inst_cnt instructions will not fit even if we align them. 130 sum_size = sz; 131 } else { 132 return 0; 133 } 134 } 135 } 136 return inst_cnt; 137 } 138 139 //----------------------------------------------------------------------------- 140 uint Block::find_node( const Node *n ) const { 141 for( uint i = 0; i < _nodes.size(); i++ ) { 142 if( _nodes[i] == n ) 143 return i; 144 } 145 ShouldNotReachHere(); 146 return 0; 147 } 148 149 // Find and remove n from block list 150 void Block::find_remove( const Node *n ) { 151 _nodes.remove(find_node(n)); 152 } 153 154 //------------------------------is_Empty--------------------------------------- 155 // Return empty status of a block. Empty blocks contain only the head, other 156 // ideal nodes, and an optional trailing goto. 157 int Block::is_Empty() const { 158 159 // Root or start block is not considered empty 160 if (head()->is_Root() || head()->is_Start()) { 161 return not_empty; 162 } 163 164 int success_result = completely_empty; 165 int end_idx = _nodes.size()-1; 166 167 // Check for ending goto 168 if ((end_idx > 0) && (_nodes[end_idx]->is_Goto())) { 169 success_result = empty_with_goto; 170 end_idx--; 171 } 172 173 // Unreachable blocks are considered empty 174 if (num_preds() <= 1) { 175 return success_result; 176 } 177 178 // Ideal nodes are allowable in empty blocks: skip them Only MachNodes 179 // turn directly into code, because only MachNodes have non-trivial 180 // emit() functions. 181 while ((end_idx > 0) && !_nodes[end_idx]->is_Mach()) { 182 end_idx--; 183 } 184 185 // No room for any interesting instructions? 186 if (end_idx == 0) { 187 return success_result; 188 } 189 190 return not_empty; 191 } 192 193 //------------------------------has_uncommon_code------------------------------ 194 // Return true if the block's code implies that it is likely to be 195 // executed infrequently. Check to see if the block ends in a Halt or 196 // a low probability call. 197 bool Block::has_uncommon_code() const { 198 Node* en = end(); 199 200 if (en->is_Goto()) 201 en = en->in(0); 202 if (en->is_Catch()) 203 en = en->in(0); 204 if (en->is_Proj() && en->in(0)->is_MachCall()) { 205 MachCallNode* call = en->in(0)->as_MachCall(); 206 if (call->cnt() != COUNT_UNKNOWN && call->cnt() <= PROB_UNLIKELY_MAG(4)) { 207 // This is true for slow-path stubs like new_{instance,array}, 208 // slow_arraycopy, complete_monitor_locking, uncommon_trap. 209 // The magic number corresponds to the probability of an uncommon_trap, 210 // even though it is a count not a probability. 211 return true; 212 } 213 } 214 215 int op = en->is_Mach() ? en->as_Mach()->ideal_Opcode() : en->Opcode(); 216 return op == Op_Halt; 217 } 218 219 //------------------------------is_uncommon------------------------------------ 220 // True if block is low enough frequency or guarded by a test which 221 // mostly does not go here. 222 bool Block::is_uncommon( Block_Array &bbs ) const { 223 // Initial blocks must never be moved, so are never uncommon. 224 if (head()->is_Root() || head()->is_Start()) return false; 225 226 // Check for way-low freq 227 if( _freq < BLOCK_FREQUENCY(0.00001f) ) return true; 228 229 // Look for code shape indicating uncommon_trap or slow path 230 if (has_uncommon_code()) return true; 231 232 const float epsilon = 0.05f; 233 const float guard_factor = PROB_UNLIKELY_MAG(4) / (1.f - epsilon); 234 uint uncommon_preds = 0; 235 uint freq_preds = 0; 236 uint uncommon_for_freq_preds = 0; 237 238 for( uint i=1; i<num_preds(); i++ ) { 239 Block* guard = bbs[pred(i)->_idx]; 240 // Check to see if this block follows its guard 1 time out of 10000 241 // or less. 242 // 243 // See list of magnitude-4 unlikely probabilities in cfgnode.hpp which 244 // we intend to be "uncommon", such as slow-path TLE allocation, 245 // predicted call failure, and uncommon trap triggers. 246 // 247 // Use an epsilon value of 5% to allow for variability in frequency 248 // predictions and floating point calculations. The net effect is 249 // that guard_factor is set to 9500. 250 // 251 // Ignore low-frequency blocks. 252 // The next check is (guard->_freq < 1.e-5 * 9500.). 253 if(guard->_freq*BLOCK_FREQUENCY(guard_factor) < BLOCK_FREQUENCY(0.00001f)) { 254 uncommon_preds++; 255 } else { 256 freq_preds++; 257 if( _freq < guard->_freq * guard_factor ) { 258 uncommon_for_freq_preds++; 259 } 260 } 261 } 262 if( num_preds() > 1 && 263 // The block is uncommon if all preds are uncommon or 264 (uncommon_preds == (num_preds()-1) || 265 // it is uncommon for all frequent preds. 266 uncommon_for_freq_preds == freq_preds) ) { 267 return true; 268 } 269 return false; 270 } 271 272 //------------------------------dump------------------------------------------- 273 #ifndef PRODUCT 274 void Block::dump_bidx(const Block* orig) const { 275 if (_pre_order) tty->print("B%d",_pre_order); 276 else tty->print("N%d", head()->_idx); 277 278 if (Verbose && orig != this) { 279 // Dump the original block's idx 280 tty->print(" ("); 281 orig->dump_bidx(orig); 282 tty->print(")"); 283 } 284 } 285 286 void Block::dump_pred(const Block_Array *bbs, Block* orig) const { 287 if (is_connector()) { 288 for (uint i=1; i<num_preds(); i++) { 289 Block *p = ((*bbs)[pred(i)->_idx]); 290 p->dump_pred(bbs, orig); 291 } 292 } else { 293 dump_bidx(orig); 294 tty->print(" "); 295 } 296 } 297 298 void Block::dump_head( const Block_Array *bbs ) const { 299 // Print the basic block 300 dump_bidx(this); 301 tty->print(": #\t"); 302 303 // Print the incoming CFG edges and the outgoing CFG edges 304 for( uint i=0; i<_num_succs; i++ ) { 305 non_connector_successor(i)->dump_bidx(_succs[i]); 306 tty->print(" "); 307 } 308 tty->print("<- "); 309 if( head()->is_block_start() ) { 310 for (uint i=1; i<num_preds(); i++) { 311 Node *s = pred(i); 312 if (bbs) { 313 Block *p = (*bbs)[s->_idx]; 314 p->dump_pred(bbs, p); 315 } else { 316 while (!s->is_block_start()) 317 s = s->in(0); 318 tty->print("N%d ", s->_idx ); 319 } 320 } 321 } else 322 tty->print("BLOCK HEAD IS JUNK "); 323 324 // Print loop, if any 325 const Block *bhead = this; // Head of self-loop 326 Node *bh = bhead->head(); 327 if( bbs && bh->is_Loop() && !head()->is_Root() ) { 328 LoopNode *loop = bh->as_Loop(); 329 const Block *bx = (*bbs)[loop->in(LoopNode::LoopBackControl)->_idx]; 330 while (bx->is_connector()) { 331 bx = (*bbs)[bx->pred(1)->_idx]; 332 } 333 tty->print("\tLoop: B%d-B%d ", bhead->_pre_order, bx->_pre_order); 334 // Dump any loop-specific bits, especially for CountedLoops. 335 loop->dump_spec(tty); 336 } else if (has_loop_alignment()) { 337 tty->print(" top-of-loop"); 338 } 339 tty->print(" Freq: %g",_freq); 340 if( Verbose || WizardMode ) { 341 tty->print(" IDom: %d/#%d", _idom ? _idom->_pre_order : 0, _dom_depth); 342 tty->print(" RegPressure: %d",_reg_pressure); 343 tty->print(" IHRP Index: %d",_ihrp_index); 344 tty->print(" FRegPressure: %d",_freg_pressure); 345 tty->print(" FHRP Index: %d",_fhrp_index); 346 } 347 tty->print_cr(""); 348 } 349 350 void Block::dump() const { dump(0); } 351 352 void Block::dump( const Block_Array *bbs ) const { 353 dump_head(bbs); 354 uint cnt = _nodes.size(); 355 for( uint i=0; i<cnt; i++ ) 356 _nodes[i]->dump(); 357 tty->print("\n"); 358 } 359 #endif 360 361 //============================================================================= 362 //------------------------------PhaseCFG--------------------------------------- 363 PhaseCFG::PhaseCFG( Arena *a, RootNode *r, Matcher &m ) : 364 Phase(CFG), 365 _bbs(a), 366 _root(r), 367 _node_latency(NULL) 368 #ifndef PRODUCT 369 , _trace_opto_pipelining(TraceOptoPipelining || C->method_has_option("TraceOptoPipelining")) 370 #endif 371 #ifdef ASSERT 372 , _raw_oops(a) 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 (C, 1) GotoNode(NULL); 380 x->init_req(0, x); 381 _goto = m.match_tree(x); 382 assert(_goto != NULL, ""); 383 _goto->set_req(0,_goto); 384 385 // Build the CFG in Reverse Post Order 386 _num_blocks = build_cfg(); 387 _broot = _bbs[_root->_idx]; 388 } 389 390 //------------------------------build_cfg-------------------------------------- 391 // Build a proper looking CFG. Make every block begin with either a StartNode 392 // or a RegionNode. Make every block end with either a Goto, If or Return. 393 // The RootNode both starts and ends it's own block. Do this with a recursive 394 // backwards walk over the control edges. 395 uint PhaseCFG::build_cfg() { 396 Arena *a = Thread::current()->resource_area(); 397 VectorSet visited(a); 398 399 // Allocate stack with enough space to avoid frequent realloc 400 Node_Stack nstack(a, C->unique() >> 1); 401 nstack.push(_root, 0); 402 uint sum = 0; // Counter for blocks 403 404 while (nstack.is_nonempty()) { 405 // node and in's index from stack's top 406 // 'np' is _root (see above) or RegionNode, StartNode: we push on stack 407 // only nodes which point to the start of basic block (see below). 408 Node *np = nstack.node(); 409 // idx > 0, except for the first node (_root) pushed on stack 410 // at the beginning when idx == 0. 411 // We will use the condition (idx == 0) later to end the build. 412 uint idx = nstack.index(); 413 Node *proj = np->in(idx); 414 const Node *x = proj->is_block_proj(); 415 // Does the block end with a proper block-ending Node? One of Return, 416 // If or Goto? (This check should be done for visited nodes also). 417 if (x == NULL) { // Does not end right... 418 Node *g = _goto->clone(); // Force it to end in a Goto 419 g->set_req(0, proj); 420 np->set_req(idx, g); 421 x = proj = g; 422 } 423 if (!visited.test_set(x->_idx)) { // Visit this block once 424 // Skip any control-pinned middle'in stuff 425 Node *p = proj; 426 do { 427 proj = p; // Update pointer to last Control 428 p = p->in(0); // Move control forward 429 } while( !p->is_block_proj() && 430 !p->is_block_start() ); 431 // Make the block begin with one of Region or StartNode. 432 if( !p->is_block_start() ) { 433 RegionNode *r = new (C, 2) RegionNode( 2 ); 434 r->init_req(1, p); // Insert RegionNode in the way 435 proj->set_req(0, r); // Insert RegionNode in the way 436 p = r; 437 } 438 // 'p' now points to the start of this basic block 439 440 // Put self in array of basic blocks 441 Block *bb = new (_bbs._arena) Block(_bbs._arena,p); 442 _bbs.map(p->_idx,bb); 443 _bbs.map(x->_idx,bb); 444 if( x != p ) // Only for root is x == p 445 bb->_nodes.push((Node*)x); 446 447 // Now handle predecessors 448 ++sum; // Count 1 for self block 449 uint cnt = bb->num_preds(); 450 for (int i = (cnt - 1); i > 0; i-- ) { // For all predecessors 451 Node *prevproj = p->in(i); // Get prior input 452 assert( !prevproj->is_Con(), "dead input not removed" ); 453 // Check to see if p->in(i) is a "control-dependent" CFG edge - 454 // i.e., it splits at the source (via an IF or SWITCH) and merges 455 // at the destination (via a many-input Region). 456 // This breaks critical edges. The RegionNode to start the block 457 // will be added when <p,i> is pulled off the node stack 458 if ( cnt > 2 ) { // Merging many things? 459 assert( prevproj== bb->pred(i),""); 460 if(prevproj->is_block_proj() != prevproj) { // Control-dependent edge? 461 // Force a block on the control-dependent edge 462 Node *g = _goto->clone(); // Force it to end in a Goto 463 g->set_req(0,prevproj); 464 p->set_req(i,g); 465 } 466 } 467 nstack.push(p, i); // 'p' is RegionNode or StartNode 468 } 469 } else { // Post-processing visited nodes 470 nstack.pop(); // remove node from stack 471 // Check if it the fist node pushed on stack at the beginning. 472 if (idx == 0) break; // end of the build 473 // Find predecessor basic block 474 Block *pb = _bbs[x->_idx]; 475 // Insert into nodes array, if not already there 476 if( !_bbs.lookup(proj->_idx) ) { 477 assert( x != proj, "" ); 478 // Map basic block of projection 479 _bbs.map(proj->_idx,pb); 480 pb->_nodes.push(proj); 481 } 482 // Insert self as a child of my predecessor block 483 pb->_succs.map(pb->_num_succs++, _bbs[np->_idx]); 484 assert( pb->_nodes[ pb->_nodes.size() - pb->_num_succs ]->is_block_proj(), 485 "too many control users, not a CFG?" ); 486 } 487 } 488 // Return number of basic blocks for all children and self 489 return sum; 490 } 491 492 //------------------------------insert_goto_at--------------------------------- 493 // Inserts a goto & corresponding basic block between 494 // block[block_no] and its succ_no'th successor block 495 void PhaseCFG::insert_goto_at(uint block_no, uint succ_no) { 496 // get block with block_no 497 assert(block_no < _num_blocks, "illegal block number"); 498 Block* in = _blocks[block_no]; 499 // get successor block succ_no 500 assert(succ_no < in->_num_succs, "illegal successor number"); 501 Block* out = in->_succs[succ_no]; 502 // Compute frequency of the new block. Do this before inserting 503 // new block in case succ_prob() needs to infer the probability from 504 // surrounding blocks. 505 float freq = in->_freq * in->succ_prob(succ_no); 506 // get ProjNode corresponding to the succ_no'th successor of the in block 507 ProjNode* proj = in->_nodes[in->_nodes.size() - in->_num_succs + succ_no]->as_Proj(); 508 // create region for basic block 509 RegionNode* region = new (C, 2) RegionNode(2); 510 region->init_req(1, proj); 511 // setup corresponding basic block 512 Block* block = new (_bbs._arena) Block(_bbs._arena, region); 513 _bbs.map(region->_idx, block); 514 C->regalloc()->set_bad(region->_idx); 515 // add a goto node 516 Node* gto = _goto->clone(); // get a new goto node 517 gto->set_req(0, region); 518 // add it to the basic block 519 block->_nodes.push(gto); 520 _bbs.map(gto->_idx, block); 521 C->regalloc()->set_bad(gto->_idx); 522 // hook up successor block 523 block->_succs.map(block->_num_succs++, out); 524 // remap successor's predecessors if necessary 525 for (uint i = 1; i < out->num_preds(); i++) { 526 if (out->pred(i) == proj) out->head()->set_req(i, gto); 527 } 528 // remap predecessor's successor to new block 529 in->_succs.map(succ_no, block); 530 // Set the frequency of the new block 531 block->_freq = freq; 532 // add new basic block to basic block list 533 _blocks.insert(block_no + 1, block); 534 _num_blocks++; 535 } 536 537 //------------------------------no_flip_branch--------------------------------- 538 // Does this block end in a multiway branch that cannot have the default case 539 // flipped for another case? 540 static bool no_flip_branch( Block *b ) { 541 int branch_idx = b->_nodes.size() - b->_num_succs-1; 542 if( branch_idx < 1 ) return false; 543 Node *bra = b->_nodes[branch_idx]; 544 if( bra->is_Catch() ) 545 return true; 546 if( bra->is_Mach() ) { 547 if( bra->is_MachNullCheck() ) 548 return true; 549 int iop = bra->as_Mach()->ideal_Opcode(); 550 if( iop == Op_FastLock || iop == Op_FastUnlock ) 551 return true; 552 } 553 return false; 554 } 555 556 //------------------------------convert_NeverBranch_to_Goto-------------------- 557 // Check for NeverBranch at block end. This needs to become a GOTO to the 558 // true target. NeverBranch are treated as a conditional branch that always 559 // goes the same direction for most of the optimizer and are used to give a 560 // fake exit path to infinite loops. At this late stage they need to turn 561 // into Goto's so that when you enter the infinite loop you indeed hang. 562 void PhaseCFG::convert_NeverBranch_to_Goto(Block *b) { 563 // Find true target 564 int end_idx = b->end_idx(); 565 int idx = b->_nodes[end_idx+1]->as_Proj()->_con; 566 Block *succ = b->_succs[idx]; 567 Node* gto = _goto->clone(); // get a new goto node 568 gto->set_req(0, b->head()); 569 Node *bp = b->_nodes[end_idx]; 570 b->_nodes.map(end_idx,gto); // Slam over NeverBranch 571 _bbs.map(gto->_idx, b); 572 C->regalloc()->set_bad(gto->_idx); 573 b->_nodes.pop(); // Yank projections 574 b->_nodes.pop(); // Yank projections 575 b->_succs.map(0,succ); // Map only successor 576 b->_num_succs = 1; 577 // remap successor's predecessors if necessary 578 uint j; 579 for( j = 1; j < succ->num_preds(); j++) 580 if( succ->pred(j)->in(0) == bp ) 581 succ->head()->set_req(j, gto); 582 // Kill alternate exit path 583 Block *dead = b->_succs[1-idx]; 584 for( j = 1; j < dead->num_preds(); j++) 585 if( dead->pred(j)->in(0) == bp ) 586 break; 587 // Scan through block, yanking dead path from 588 // all regions and phis. 589 dead->head()->del_req(j); 590 for( int k = 1; dead->_nodes[k]->is_Phi(); k++ ) 591 dead->_nodes[k]->del_req(j); 592 } 593 594 //------------------------------move_to_next----------------------------------- 595 // Helper function to move block bx to the slot following b_index. Return 596 // true if the move is successful, otherwise false 597 bool PhaseCFG::move_to_next(Block* bx, uint b_index) { 598 if (bx == NULL) return false; 599 600 // Return false if bx is already scheduled. 601 uint bx_index = bx->_pre_order; 602 if ((bx_index <= b_index) && (_blocks[bx_index] == bx)) { 603 return false; 604 } 605 606 // Find the current index of block bx on the block list 607 bx_index = b_index + 1; 608 while( bx_index < _num_blocks && _blocks[bx_index] != bx ) bx_index++; 609 assert(_blocks[bx_index] == bx, "block not found"); 610 611 // If the previous block conditionally falls into bx, return false, 612 // because moving bx will create an extra jump. 613 for(uint k = 1; k < bx->num_preds(); k++ ) { 614 Block* pred = _bbs[bx->pred(k)->_idx]; 615 if (pred == _blocks[bx_index-1]) { 616 if (pred->_num_succs != 1) { 617 return false; 618 } 619 } 620 } 621 622 // Reinsert bx just past block 'b' 623 _blocks.remove(bx_index); 624 _blocks.insert(b_index + 1, bx); 625 return true; 626 } 627 628 //------------------------------move_to_end------------------------------------ 629 // Move empty and uncommon blocks to the end. 630 void PhaseCFG::move_to_end(Block *b, uint i) { 631 int e = b->is_Empty(); 632 if (e != Block::not_empty) { 633 if (e == Block::empty_with_goto) { 634 // Remove the goto, but leave the block. 635 b->_nodes.pop(); 636 } 637 // Mark this block as a connector block, which will cause it to be 638 // ignored in certain functions such as non_connector_successor(). 639 b->set_connector(); 640 } 641 // Move the empty block to the end, and don't recheck. 642 _blocks.remove(i); 643 _blocks.push(b); 644 } 645 646 //---------------------------set_loop_alignment-------------------------------- 647 // Set loop alignment for every block 648 void PhaseCFG::set_loop_alignment() { 649 uint last = _num_blocks; 650 assert( _blocks[0] == _broot, "" ); 651 652 for (uint i = 1; i < last; i++ ) { 653 Block *b = _blocks[i]; 654 if (b->head()->is_Loop()) { 655 b->set_loop_alignment(b); 656 } 657 } 658 } 659 660 //-----------------------------remove_empty------------------------------------ 661 // Make empty basic blocks to be "connector" blocks, Move uncommon blocks 662 // to the end. 663 void PhaseCFG::remove_empty() { 664 // Move uncommon blocks to the end 665 uint last = _num_blocks; 666 assert( _blocks[0] == _broot, "" ); 667 668 for (uint i = 1; i < last; i++) { 669 Block *b = _blocks[i]; 670 if (b->is_connector()) break; 671 672 // Check for NeverBranch at block end. This needs to become a GOTO to the 673 // true target. NeverBranch are treated as a conditional branch that 674 // always goes the same direction for most of the optimizer and are used 675 // to give a fake exit path to infinite loops. At this late stage they 676 // need to turn into Goto's so that when you enter the infinite loop you 677 // indeed hang. 678 if( b->_nodes[b->end_idx()]->Opcode() == Op_NeverBranch ) 679 convert_NeverBranch_to_Goto(b); 680 681 // Look for uncommon blocks and move to end. 682 if (!C->do_freq_based_layout()) { 683 if( b->is_uncommon(_bbs) ) { 684 move_to_end(b, i); 685 last--; // No longer check for being uncommon! 686 if( no_flip_branch(b) ) { // Fall-thru case must follow? 687 b = _blocks[i]; // Find the fall-thru block 688 move_to_end(b, i); 689 last--; 690 } 691 i--; // backup block counter post-increment 692 } 693 } 694 } 695 696 // Move empty blocks to the end 697 last = _num_blocks; 698 for (uint i = 1; i < last; i++) { 699 Block *b = _blocks[i]; 700 if (b->is_Empty() != Block::not_empty) { 701 move_to_end(b, i); 702 last--; 703 i--; 704 } 705 } // End of for all blocks 706 } 707 708 //-----------------------------fixup_flow-------------------------------------- 709 // Fix up the final control flow for basic blocks. 710 void PhaseCFG::fixup_flow() { 711 // Fixup final control flow for the blocks. Remove jump-to-next 712 // block. If neither arm of a IF follows the conditional branch, we 713 // have to add a second jump after the conditional. We place the 714 // TRUE branch target in succs[0] for both GOTOs and IFs. 715 for (uint i=0; i < _num_blocks; i++) { 716 Block *b = _blocks[i]; 717 b->_pre_order = i; // turn pre-order into block-index 718 719 // Connector blocks need no further processing. 720 if (b->is_connector()) { 721 assert((i+1) == _num_blocks || _blocks[i+1]->is_connector(), 722 "All connector blocks should sink to the end"); 723 continue; 724 } 725 assert(b->is_Empty() != Block::completely_empty, 726 "Empty blocks should be connectors"); 727 728 Block *bnext = (i < _num_blocks-1) ? _blocks[i+1] : NULL; 729 Block *bs0 = b->non_connector_successor(0); 730 731 // Check for multi-way branches where I cannot negate the test to 732 // exchange the true and false targets. 733 if( no_flip_branch( b ) ) { 734 // Find fall through case - if must fall into its target 735 int branch_idx = b->_nodes.size() - b->_num_succs; 736 for (uint j2 = 0; j2 < b->_num_succs; j2++) { 737 const ProjNode* p = b->_nodes[branch_idx + j2]->as_Proj(); 738 if (p->_con == 0) { 739 // successor j2 is fall through case 740 if (b->non_connector_successor(j2) != bnext) { 741 // but it is not the next block => insert a goto 742 insert_goto_at(i, j2); 743 } 744 // Put taken branch in slot 0 745 if( j2 == 0 && b->_num_succs == 2) { 746 // Flip targets in succs map 747 Block *tbs0 = b->_succs[0]; 748 Block *tbs1 = b->_succs[1]; 749 b->_succs.map( 0, tbs1 ); 750 b->_succs.map( 1, tbs0 ); 751 } 752 break; 753 } 754 } 755 // Remove all CatchProjs 756 for (uint j1 = 0; j1 < b->_num_succs; j1++) b->_nodes.pop(); 757 758 } else if (b->_num_succs == 1) { 759 // Block ends in a Goto? 760 if (bnext == bs0) { 761 // We fall into next block; remove the Goto 762 b->_nodes.pop(); 763 } 764 765 } else if( b->_num_succs == 2 ) { // Block ends in a If? 766 // Get opcode of 1st projection (matches _succs[0]) 767 // Note: Since this basic block has 2 exits, the last 2 nodes must 768 // be projections (in any order), the 3rd last node must be 769 // the IfNode (we have excluded other 2-way exits such as 770 // CatchNodes already). 771 MachNode *iff = b->_nodes[b->_nodes.size()-3]->as_Mach(); 772 ProjNode *proj0 = b->_nodes[b->_nodes.size()-2]->as_Proj(); 773 ProjNode *proj1 = b->_nodes[b->_nodes.size()-1]->as_Proj(); 774 775 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1]. 776 assert(proj0->raw_out(0) == b->_succs[0]->head(), "Mismatch successor 0"); 777 assert(proj1->raw_out(0) == b->_succs[1]->head(), "Mismatch successor 1"); 778 779 Block *bs1 = b->non_connector_successor(1); 780 781 // Check for neither successor block following the current 782 // block ending in a conditional. If so, move one of the 783 // successors after the current one, provided that the 784 // successor was previously unscheduled, but moveable 785 // (i.e., all paths to it involve a branch). 786 if( !C->do_freq_based_layout() && bnext != bs0 && bnext != bs1 ) { 787 // Choose the more common successor based on the probability 788 // of the conditional branch. 789 Block *bx = bs0; 790 Block *by = bs1; 791 792 // _prob is the probability of taking the true path. Make 793 // p the probability of taking successor #1. 794 float p = iff->as_MachIf()->_prob; 795 if( proj0->Opcode() == Op_IfTrue ) { 796 p = 1.0 - p; 797 } 798 799 // Prefer successor #1 if p > 0.5 800 if (p > PROB_FAIR) { 801 bx = bs1; 802 by = bs0; 803 } 804 805 // Attempt the more common successor first 806 if (move_to_next(bx, i)) { 807 bnext = bx; 808 } else if (move_to_next(by, i)) { 809 bnext = by; 810 } 811 } 812 813 // Check for conditional branching the wrong way. Negate 814 // conditional, if needed, so it falls into the following block 815 // and branches to the not-following block. 816 817 // Check for the next block being in succs[0]. We are going to branch 818 // to succs[0], so we want the fall-thru case as the next block in 819 // succs[1]. 820 if (bnext == bs0) { 821 // Fall-thru case in succs[0], so flip targets in succs map 822 Block *tbs0 = b->_succs[0]; 823 Block *tbs1 = b->_succs[1]; 824 b->_succs.map( 0, tbs1 ); 825 b->_succs.map( 1, tbs0 ); 826 // Flip projection for each target 827 { ProjNode *tmp = proj0; proj0 = proj1; proj1 = tmp; } 828 829 } else if( bnext != bs1 ) { 830 // Need a double-branch 831 // The existing conditional branch need not change. 832 // Add a unconditional branch to the false target. 833 // Alas, it must appear in its own block and adding a 834 // block this late in the game is complicated. Sigh. 835 insert_goto_at(i, 1); 836 } 837 838 // Make sure we TRUE branch to the target 839 if( proj0->Opcode() == Op_IfFalse ) { 840 iff->negate(); 841 } 842 843 b->_nodes.pop(); // Remove IfFalse & IfTrue projections 844 b->_nodes.pop(); 845 846 } else { 847 // Multi-exit block, e.g. a switch statement 848 // But we don't need to do anything here 849 } 850 } // End of for all blocks 851 } 852 853 854 //------------------------------dump------------------------------------------- 855 #ifndef PRODUCT 856 void PhaseCFG::_dump_cfg( const Node *end, VectorSet &visited ) const { 857 const Node *x = end->is_block_proj(); 858 assert( x, "not a CFG" ); 859 860 // Do not visit this block again 861 if( visited.test_set(x->_idx) ) return; 862 863 // Skip through this block 864 const Node *p = x; 865 do { 866 p = p->in(0); // Move control forward 867 assert( !p->is_block_proj() || p->is_Root(), "not a CFG" ); 868 } while( !p->is_block_start() ); 869 870 // Recursively visit 871 for( uint i=1; i<p->req(); i++ ) 872 _dump_cfg(p->in(i),visited); 873 874 // Dump the block 875 _bbs[p->_idx]->dump(&_bbs); 876 } 877 878 void PhaseCFG::dump( ) const { 879 tty->print("\n--- CFG --- %d BBs\n",_num_blocks); 880 if( _blocks.size() ) { // Did we do basic-block layout? 881 for( uint i=0; i<_num_blocks; i++ ) 882 _blocks[i]->dump(&_bbs); 883 } else { // Else do it with a DFS 884 VectorSet visited(_bbs._arena); 885 _dump_cfg(_root,visited); 886 } 887 } 888 889 void PhaseCFG::dump_headers() { 890 for( uint i = 0; i < _num_blocks; i++ ) { 891 if( _blocks[i] == NULL ) continue; 892 _blocks[i]->dump_head(&_bbs); 893 } 894 } 895 896 void PhaseCFG::verify( ) const { 897 #ifdef ASSERT 898 // Verify sane CFG 899 for( uint i = 0; i < _num_blocks; i++ ) { 900 Block *b = _blocks[i]; 901 uint cnt = b->_nodes.size(); 902 uint j; 903 for( j = 0; j < cnt; j++ ) { 904 Node *n = b->_nodes[j]; 905 assert( _bbs[n->_idx] == b, "" ); 906 if( j >= 1 && n->is_Mach() && 907 n->as_Mach()->ideal_Opcode() == Op_CreateEx ) { 908 assert( j == 1 || b->_nodes[j-1]->is_Phi(), 909 "CreateEx must be first instruction in block" ); 910 } 911 for( uint k = 0; k < n->req(); k++ ) { 912 Node *def = n->in(k); 913 if( def && def != n ) { 914 assert( _bbs[def->_idx] || def->is_Con(), 915 "must have block; constants for debug info ok" ); 916 // Verify that instructions in the block is in correct order. 917 // Uses must follow their definition if they are at the same block. 918 // Mostly done to check that MachSpillCopy nodes are placed correctly 919 // when CreateEx node is moved in build_ifg_physical(). 920 if( _bbs[def->_idx] == b && 921 !(b->head()->is_Loop() && n->is_Phi()) && 922 // See (+++) comment in reg_split.cpp 923 !(n->jvms() != NULL && n->jvms()->is_monitor_use(k)) ) { 924 bool is_loop = false; 925 if (n->is_Phi()) { 926 for( uint l = 1; l < def->req(); l++ ) { 927 if (n == def->in(l)) { 928 is_loop = true; 929 break; // Some kind of loop 930 } 931 } 932 } 933 assert( is_loop || b->find_node(def) < j, "uses must follow definitions" ); 934 } 935 if( def->is_SafePointScalarObject() ) { 936 assert(_bbs[def->_idx] == b, "SafePointScalarObject Node should be at the same block as its SafePoint node"); 937 assert(_bbs[def->_idx] == _bbs[def->in(0)->_idx], "SafePointScalarObject Node should be at the same block as its control edge"); 938 } 939 } 940 } 941 } 942 943 j = b->end_idx(); 944 Node *bp = (Node*)b->_nodes[b->_nodes.size()-1]->is_block_proj(); 945 assert( bp, "last instruction must be a block proj" ); 946 assert( bp == b->_nodes[j], "wrong number of successors for this block" ); 947 if( bp->is_Catch() ) { 948 while( b->_nodes[--j]->Opcode() == Op_MachProj ) ; 949 assert( b->_nodes[j]->is_Call(), "CatchProj must follow call" ); 950 } 951 else if( bp->is_Mach() && bp->as_Mach()->ideal_Opcode() == Op_If ) { 952 assert( b->_num_succs == 2, "Conditional branch must have two targets"); 953 } 954 } 955 #endif 956 } 957 #endif 958 959 //============================================================================= 960 //------------------------------UnionFind-------------------------------------- 961 UnionFind::UnionFind( uint max ) : _cnt(max), _max(max), _indices(NEW_RESOURCE_ARRAY(uint,max)) { 962 Copy::zero_to_bytes( _indices, sizeof(uint)*max ); 963 } 964 965 void UnionFind::extend( uint from_idx, uint to_idx ) { 966 _nesting.check(); 967 if( from_idx >= _max ) { 968 uint size = 16; 969 while( size <= from_idx ) size <<=1; 970 _indices = REALLOC_RESOURCE_ARRAY( uint, _indices, _max, size ); 971 _max = size; 972 } 973 while( _cnt <= from_idx ) _indices[_cnt++] = 0; 974 _indices[from_idx] = to_idx; 975 } 976 977 void UnionFind::reset( uint max ) { 978 assert( max <= max_uint, "Must fit within uint" ); 979 // Force the Union-Find mapping to be at least this large 980 extend(max,0); 981 // Initialize to be the ID mapping. 982 for( uint i=0; i<max; i++ ) map(i,i); 983 } 984 985 //------------------------------Find_compress---------------------------------- 986 // Straight out of Tarjan's union-find algorithm 987 uint UnionFind::Find_compress( uint idx ) { 988 uint cur = idx; 989 uint next = lookup(cur); 990 while( next != cur ) { // Scan chain of equivalences 991 assert( next < cur, "always union smaller" ); 992 cur = next; // until find a fixed-point 993 next = lookup(cur); 994 } 995 // Core of union-find algorithm: update chain of 996 // equivalences to be equal to the root. 997 while( idx != next ) { 998 uint tmp = lookup(idx); 999 map(idx, next); 1000 idx = tmp; 1001 } 1002 return idx; 1003 } 1004 1005 //------------------------------Find_const------------------------------------- 1006 // Like Find above, but no path compress, so bad asymptotic behavior 1007 uint UnionFind::Find_const( uint idx ) const { 1008 if( idx == 0 ) return idx; // Ignore the zero idx 1009 // Off the end? This can happen during debugging dumps 1010 // when data structures have not finished being updated. 1011 if( idx >= _max ) return idx; 1012 uint next = lookup(idx); 1013 while( next != idx ) { // Scan chain of equivalences 1014 idx = next; // until find a fixed-point 1015 next = lookup(idx); 1016 } 1017 return next; 1018 } 1019 1020 //------------------------------Union------------------------------------------ 1021 // union 2 sets together. 1022 void UnionFind::Union( uint idx1, uint idx2 ) { 1023 uint src = Find(idx1); 1024 uint dst = Find(idx2); 1025 assert( src, "" ); 1026 assert( dst, "" ); 1027 assert( src < _max, "oob" ); 1028 assert( dst < _max, "oob" ); 1029 assert( src < dst, "always union smaller" ); 1030 map(dst,src); 1031 } 1032 1033 #ifndef PRODUCT 1034 static void edge_dump(GrowableArray<CFGEdge *> *edges) { 1035 tty->print_cr("---- Edges ----"); 1036 for (int i = 0; i < edges->length(); i++) { 1037 CFGEdge *e = edges->at(i); 1038 if (e != NULL) { 1039 edges->at(i)->dump(); 1040 } 1041 } 1042 } 1043 1044 static void trace_dump(Trace *traces[], int count) { 1045 tty->print_cr("---- Traces ----"); 1046 for (int i = 0; i < count; i++) { 1047 Trace *tr = traces[i]; 1048 if (tr != NULL) { 1049 tr->dump(); 1050 } 1051 } 1052 } 1053 1054 void Trace::dump( ) const { 1055 tty->print_cr("Trace (freq %f)", first_block()->_freq); 1056 for (Block *b = first_block(); b != NULL; b = next(b)) { 1057 tty->print(" B%d", b->_pre_order); 1058 if (b->head()->is_Loop()) { 1059 tty->print(" (L%d)", b->compute_loop_alignment()); 1060 } 1061 if (b->has_loop_alignment()) { 1062 tty->print(" (T%d)", b->code_alignment()); 1063 } 1064 } 1065 tty->cr(); 1066 } 1067 1068 void CFGEdge::dump( ) const { 1069 tty->print(" B%d --> B%d Freq: %f out:%3d%% in:%3d%% State: ", 1070 from()->_pre_order, to()->_pre_order, freq(), _from_pct, _to_pct); 1071 switch(state()) { 1072 case connected: 1073 tty->print("connected"); 1074 break; 1075 case open: 1076 tty->print("open"); 1077 break; 1078 case interior: 1079 tty->print("interior"); 1080 break; 1081 } 1082 if (infrequent()) { 1083 tty->print(" infrequent"); 1084 } 1085 tty->cr(); 1086 } 1087 #endif 1088 1089 //============================================================================= 1090 1091 //------------------------------edge_order------------------------------------- 1092 // Comparison function for edges 1093 static int edge_order(CFGEdge **e0, CFGEdge **e1) { 1094 float freq0 = (*e0)->freq(); 1095 float freq1 = (*e1)->freq(); 1096 if (freq0 != freq1) { 1097 return freq0 > freq1 ? -1 : 1; 1098 } 1099 1100 int dist0 = (*e0)->to()->_rpo - (*e0)->from()->_rpo; 1101 int dist1 = (*e1)->to()->_rpo - (*e1)->from()->_rpo; 1102 1103 return dist1 - dist0; 1104 } 1105 1106 //------------------------------trace_frequency_order-------------------------- 1107 // Comparison function for edges 1108 static int trace_frequency_order(const void *p0, const void *p1) { 1109 Trace *tr0 = *(Trace **) p0; 1110 Trace *tr1 = *(Trace **) p1; 1111 Block *b0 = tr0->first_block(); 1112 Block *b1 = tr1->first_block(); 1113 1114 // The trace of connector blocks goes at the end; 1115 // we only expect one such trace 1116 if (b0->is_connector() != b1->is_connector()) { 1117 return b1->is_connector() ? -1 : 1; 1118 } 1119 1120 // Pull more frequently executed blocks to the beginning 1121 float freq0 = b0->_freq; 1122 float freq1 = b1->_freq; 1123 if (freq0 != freq1) { 1124 return freq0 > freq1 ? -1 : 1; 1125 } 1126 1127 int diff = tr0->first_block()->_rpo - tr1->first_block()->_rpo; 1128 1129 return diff; 1130 } 1131 1132 //------------------------------find_edges------------------------------------- 1133 // Find edges of interest, i.e, those which can fall through. Presumes that 1134 // edges which don't fall through are of low frequency and can be generally 1135 // ignored. Initialize the list of traces. 1136 void PhaseBlockLayout::find_edges() 1137 { 1138 // Walk the blocks, creating edges and Traces 1139 uint i; 1140 Trace *tr = NULL; 1141 for (i = 0; i < _cfg._num_blocks; i++) { 1142 Block *b = _cfg._blocks[i]; 1143 tr = new Trace(b, next, prev); 1144 traces[tr->id()] = tr; 1145 1146 // All connector blocks should be at the end of the list 1147 if (b->is_connector()) break; 1148 1149 // If this block and the next one have a one-to-one successor 1150 // predecessor relationship, simply append the next block 1151 int nfallthru = b->num_fall_throughs(); 1152 while (nfallthru == 1 && 1153 b->succ_fall_through(0)) { 1154 Block *n = b->_succs[0]; 1155 1156 // Skip over single-entry connector blocks, we don't want to 1157 // add them to the trace. 1158 while (n->is_connector() && n->num_preds() == 1) { 1159 n = n->_succs[0]; 1160 } 1161 1162 // We see a merge point, so stop search for the next block 1163 if (n->num_preds() != 1) break; 1164 1165 i++; 1166 assert(n = _cfg._blocks[i], "expecting next block"); 1167 tr->append(n); 1168 uf->map(n->_pre_order, tr->id()); 1169 traces[n->_pre_order] = NULL; 1170 nfallthru = b->num_fall_throughs(); 1171 b = n; 1172 } 1173 1174 if (nfallthru > 0) { 1175 // Create a CFGEdge for each outgoing 1176 // edge that could be a fall-through. 1177 for (uint j = 0; j < b->_num_succs; j++ ) { 1178 if (b->succ_fall_through(j)) { 1179 Block *target = b->non_connector_successor(j); 1180 float freq = b->_freq * b->succ_prob(j); 1181 int from_pct = (int) ((100 * freq) / b->_freq); 1182 int to_pct = (int) ((100 * freq) / target->_freq); 1183 edges->append(new CFGEdge(b, target, freq, from_pct, to_pct)); 1184 } 1185 } 1186 } 1187 } 1188 1189 // Group connector blocks into one trace 1190 for (i++; i < _cfg._num_blocks; i++) { 1191 Block *b = _cfg._blocks[i]; 1192 assert(b->is_connector(), "connector blocks at the end"); 1193 tr->append(b); 1194 uf->map(b->_pre_order, tr->id()); 1195 traces[b->_pre_order] = NULL; 1196 } 1197 } 1198 1199 //------------------------------union_traces---------------------------------- 1200 // Union two traces together in uf, and null out the trace in the list 1201 void PhaseBlockLayout::union_traces(Trace* updated_trace, Trace* old_trace) 1202 { 1203 uint old_id = old_trace->id(); 1204 uint updated_id = updated_trace->id(); 1205 1206 uint lo_id = updated_id; 1207 uint hi_id = old_id; 1208 1209 // If from is greater than to, swap values to meet 1210 // UnionFind guarantee. 1211 if (updated_id > old_id) { 1212 lo_id = old_id; 1213 hi_id = updated_id; 1214 1215 // Fix up the trace ids 1216 traces[lo_id] = traces[updated_id]; 1217 updated_trace->set_id(lo_id); 1218 } 1219 1220 // Union the lower with the higher and remove the pointer 1221 // to the higher. 1222 uf->Union(lo_id, hi_id); 1223 traces[hi_id] = NULL; 1224 } 1225 1226 //------------------------------grow_traces------------------------------------- 1227 // Append traces together via the most frequently executed edges 1228 void PhaseBlockLayout::grow_traces() 1229 { 1230 // Order the edges, and drive the growth of Traces via the most 1231 // frequently executed edges. 1232 edges->sort(edge_order); 1233 for (int i = 0; i < edges->length(); i++) { 1234 CFGEdge *e = edges->at(i); 1235 1236 if (e->state() != CFGEdge::open) continue; 1237 1238 Block *src_block = e->from(); 1239 Block *targ_block = e->to(); 1240 1241 // Don't grow traces along backedges? 1242 if (!BlockLayoutRotateLoops) { 1243 if (targ_block->_rpo <= src_block->_rpo) { 1244 targ_block->set_loop_alignment(targ_block); 1245 continue; 1246 } 1247 } 1248 1249 Trace *src_trace = trace(src_block); 1250 Trace *targ_trace = trace(targ_block); 1251 1252 // If the edge in question can join two traces at their ends, 1253 // append one trace to the other. 1254 if (src_trace->last_block() == src_block) { 1255 if (src_trace == targ_trace) { 1256 e->set_state(CFGEdge::interior); 1257 if (targ_trace->backedge(e)) { 1258 // Reset i to catch any newly eligible edge 1259 // (Or we could remember the first "open" edge, and reset there) 1260 i = 0; 1261 } 1262 } else if (targ_trace->first_block() == targ_block) { 1263 e->set_state(CFGEdge::connected); 1264 src_trace->append(targ_trace); 1265 union_traces(src_trace, targ_trace); 1266 } 1267 } 1268 } 1269 } 1270 1271 //------------------------------merge_traces----------------------------------- 1272 // Embed one trace into another, if the fork or join points are sufficiently 1273 // balanced. 1274 void PhaseBlockLayout::merge_traces(bool fall_thru_only) 1275 { 1276 // Walk the edge list a another time, looking at unprocessed edges. 1277 // Fold in diamonds 1278 for (int i = 0; i < edges->length(); i++) { 1279 CFGEdge *e = edges->at(i); 1280 1281 if (e->state() != CFGEdge::open) continue; 1282 if (fall_thru_only) { 1283 if (e->infrequent()) continue; 1284 } 1285 1286 Block *src_block = e->from(); 1287 Trace *src_trace = trace(src_block); 1288 bool src_at_tail = src_trace->last_block() == src_block; 1289 1290 Block *targ_block = e->to(); 1291 Trace *targ_trace = trace(targ_block); 1292 bool targ_at_start = targ_trace->first_block() == targ_block; 1293 1294 if (src_trace == targ_trace) { 1295 // This may be a loop, but we can't do much about it. 1296 e->set_state(CFGEdge::interior); 1297 continue; 1298 } 1299 1300 if (fall_thru_only) { 1301 // If the edge links the middle of two traces, we can't do anything. 1302 // Mark the edge and continue. 1303 if (!src_at_tail & !targ_at_start) { 1304 continue; 1305 } 1306 1307 // Don't grow traces along backedges? 1308 if (!BlockLayoutRotateLoops && (targ_block->_rpo <= src_block->_rpo)) { 1309 continue; 1310 } 1311 1312 // If both ends of the edge are available, why didn't we handle it earlier? 1313 assert(src_at_tail ^ targ_at_start, "Should have caught this edge earlier."); 1314 1315 if (targ_at_start) { 1316 // Insert the "targ" trace in the "src" trace if the insertion point 1317 // is a two way branch. 1318 // Better profitability check possible, but may not be worth it. 1319 // Someday, see if the this "fork" has an associated "join"; 1320 // then make a policy on merging this trace at the fork or join. 1321 // For example, other things being equal, it may be better to place this 1322 // trace at the join point if the "src" trace ends in a two-way, but 1323 // the insertion point is one-way. 1324 assert(src_block->num_fall_throughs() == 2, "unexpected diamond"); 1325 e->set_state(CFGEdge::connected); 1326 src_trace->insert_after(src_block, targ_trace); 1327 union_traces(src_trace, targ_trace); 1328 } else if (src_at_tail) { 1329 if (src_trace != trace(_cfg._broot)) { 1330 e->set_state(CFGEdge::connected); 1331 targ_trace->insert_before(targ_block, src_trace); 1332 union_traces(targ_trace, src_trace); 1333 } 1334 } 1335 } else if (e->state() == CFGEdge::open) { 1336 // Append traces, even without a fall-thru connection. 1337 // But leave root entry at the beginning of the block list. 1338 if (targ_trace != trace(_cfg._broot)) { 1339 e->set_state(CFGEdge::connected); 1340 src_trace->append(targ_trace); 1341 union_traces(src_trace, targ_trace); 1342 } 1343 } 1344 } 1345 } 1346 1347 //----------------------------reorder_traces----------------------------------- 1348 // Order the sequence of the traces in some desirable way, and fixup the 1349 // jumps at the end of each block. 1350 void PhaseBlockLayout::reorder_traces(int count) 1351 { 1352 ResourceArea *area = Thread::current()->resource_area(); 1353 Trace ** new_traces = NEW_ARENA_ARRAY(area, Trace *, count); 1354 Block_List worklist; 1355 int new_count = 0; 1356 1357 // Compact the traces. 1358 for (int i = 0; i < count; i++) { 1359 Trace *tr = traces[i]; 1360 if (tr != NULL) { 1361 new_traces[new_count++] = tr; 1362 } 1363 } 1364 1365 // The entry block should be first on the new trace list. 1366 Trace *tr = trace(_cfg._broot); 1367 assert(tr == new_traces[0], "entry trace misplaced"); 1368 1369 // Sort the new trace list by frequency 1370 qsort(new_traces + 1, new_count - 1, sizeof(new_traces[0]), trace_frequency_order); 1371 1372 // Patch up the successor blocks 1373 _cfg._blocks.reset(); 1374 _cfg._num_blocks = 0; 1375 for (int i = 0; i < new_count; i++) { 1376 Trace *tr = new_traces[i]; 1377 if (tr != NULL) { 1378 tr->fixup_blocks(_cfg); 1379 } 1380 } 1381 } 1382 1383 //------------------------------PhaseBlockLayout------------------------------- 1384 // Order basic blocks based on frequency 1385 PhaseBlockLayout::PhaseBlockLayout(PhaseCFG &cfg) : 1386 Phase(BlockLayout), 1387 _cfg(cfg) 1388 { 1389 ResourceMark rm; 1390 ResourceArea *area = Thread::current()->resource_area(); 1391 1392 // List of traces 1393 int size = _cfg._num_blocks + 1; 1394 traces = NEW_ARENA_ARRAY(area, Trace *, size); 1395 memset(traces, 0, size*sizeof(Trace*)); 1396 next = NEW_ARENA_ARRAY(area, Block *, size); 1397 memset(next, 0, size*sizeof(Block *)); 1398 prev = NEW_ARENA_ARRAY(area, Block *, size); 1399 memset(prev , 0, size*sizeof(Block *)); 1400 1401 // List of edges 1402 edges = new GrowableArray<CFGEdge*>; 1403 1404 // Mapping block index --> block_trace 1405 uf = new UnionFind(size); 1406 uf->reset(size); 1407 1408 // Find edges and create traces. 1409 find_edges(); 1410 1411 // Grow traces at their ends via most frequent edges. 1412 grow_traces(); 1413 1414 // Merge one trace into another, but only at fall-through points. 1415 // This may make diamonds and other related shapes in a trace. 1416 merge_traces(true); 1417 1418 // Run merge again, allowing two traces to be catenated, even if 1419 // one does not fall through into the other. This appends loosely 1420 // related traces to be near each other. 1421 merge_traces(false); 1422 1423 // Re-order all the remaining traces by frequency 1424 reorder_traces(size); 1425 1426 assert(_cfg._num_blocks >= (uint) (size - 1), "number of blocks can not shrink"); 1427 } 1428 1429 1430 //------------------------------backedge--------------------------------------- 1431 // Edge e completes a loop in a trace. If the target block is head of the 1432 // loop, rotate the loop block so that the loop ends in a conditional branch. 1433 bool Trace::backedge(CFGEdge *e) { 1434 bool loop_rotated = false; 1435 Block *src_block = e->from(); 1436 Block *targ_block = e->to(); 1437 1438 assert(last_block() == src_block, "loop discovery at back branch"); 1439 if (first_block() == targ_block) { 1440 if (BlockLayoutRotateLoops && last_block()->num_fall_throughs() < 2) { 1441 // Find the last block in the trace that has a conditional 1442 // branch. 1443 Block *b; 1444 for (b = last_block(); b != NULL; b = prev(b)) { 1445 if (b->num_fall_throughs() == 2) { 1446 break; 1447 } 1448 } 1449 1450 if (b != last_block() && b != NULL) { 1451 loop_rotated = true; 1452 1453 // Rotate the loop by doing two-part linked-list surgery. 1454 append(first_block()); 1455 break_loop_after(b); 1456 } 1457 } 1458 1459 // Backbranch to the top of a trace 1460 // Scroll forward through the trace from the targ_block. If we find 1461 // a loop head before another loop top, use the the loop head alignment. 1462 for (Block *b = targ_block; b != NULL; b = next(b)) { 1463 if (b->has_loop_alignment()) { 1464 break; 1465 } 1466 if (b->head()->is_Loop()) { 1467 targ_block = b; 1468 break; 1469 } 1470 } 1471 1472 first_block()->set_loop_alignment(targ_block); 1473 1474 } else { 1475 // Backbranch into the middle of a trace 1476 targ_block->set_loop_alignment(targ_block); 1477 } 1478 1479 return loop_rotated; 1480 } 1481 1482 //------------------------------fixup_blocks----------------------------------- 1483 // push blocks onto the CFG list 1484 // ensure that blocks have the correct two-way branch sense 1485 void Trace::fixup_blocks(PhaseCFG &cfg) { 1486 Block *last = last_block(); 1487 for (Block *b = first_block(); b != NULL; b = next(b)) { 1488 cfg._blocks.push(b); 1489 cfg._num_blocks++; 1490 if (!b->is_connector()) { 1491 int nfallthru = b->num_fall_throughs(); 1492 if (b != last) { 1493 if (nfallthru == 2) { 1494 // Ensure that the sense of the branch is correct 1495 Block *bnext = next(b); 1496 Block *bs0 = b->non_connector_successor(0); 1497 1498 MachNode *iff = b->_nodes[b->_nodes.size()-3]->as_Mach(); 1499 ProjNode *proj0 = b->_nodes[b->_nodes.size()-2]->as_Proj(); 1500 ProjNode *proj1 = b->_nodes[b->_nodes.size()-1]->as_Proj(); 1501 1502 if (bnext == bs0) { 1503 // Fall-thru case in succs[0], should be in succs[1] 1504 1505 // Flip targets in _succs map 1506 Block *tbs0 = b->_succs[0]; 1507 Block *tbs1 = b->_succs[1]; 1508 b->_succs.map( 0, tbs1 ); 1509 b->_succs.map( 1, tbs0 ); 1510 1511 // Flip projections to match targets 1512 b->_nodes.map(b->_nodes.size()-2, proj1); 1513 b->_nodes.map(b->_nodes.size()-1, proj0); 1514 } 1515 } 1516 } 1517 } 1518 } 1519 }