1 /* 2 * Copyright (c) 1997, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "libadt/vectset.hpp" 27 #include "memory/allocation.inline.hpp" 28 #include "compiler/compilerDirectives.hpp" 29 #include "opto/block.hpp" 30 #include "opto/cfgnode.hpp" 31 #include "opto/chaitin.hpp" 32 #include "opto/loopnode.hpp" 33 #include "opto/machnode.hpp" 34 #include "opto/matcher.hpp" 35 #include "opto/opcodes.hpp" 36 #include "opto/rootnode.hpp" 37 #include "utilities/copy.hpp" 38 39 void Block_Array::grow( uint i ) { 40 assert(i >= Max(), "must be an overflow"); 41 debug_only(_limit = i+1); 42 if( i < _size ) return; 43 if( !_size ) { 44 _size = 1; 45 _blocks = (Block**)_arena->Amalloc( _size * sizeof(Block*) ); 46 _blocks[0] = NULL; 47 } 48 uint old = _size; 49 while( i >= _size ) _size <<= 1; // Double to fit 50 _blocks = (Block**)_arena->Arealloc( _blocks, old*sizeof(Block*),_size*sizeof(Block*)); 51 Copy::zero_to_bytes( &_blocks[old], (_size-old)*sizeof(Block*) ); 52 } 53 54 void Block_List::remove(uint i) { 55 assert(i < _cnt, "index out of bounds"); 56 Copy::conjoint_words_to_lower((HeapWord*)&_blocks[i+1], (HeapWord*)&_blocks[i], ((_cnt-i-1)*sizeof(Block*))); 57 pop(); // shrink list by one block 58 } 59 60 void Block_List::insert(uint i, Block *b) { 61 push(b); // grow list by one block 62 Copy::conjoint_words_to_higher((HeapWord*)&_blocks[i], (HeapWord*)&_blocks[i+1], ((_cnt-i-1)*sizeof(Block*))); 63 _blocks[i] = b; 64 } 65 66 #ifndef PRODUCT 67 void Block_List::print() { 68 for (uint i=0; i < size(); i++) { 69 tty->print("B%d ", _blocks[i]->_pre_order); 70 } 71 tty->print("size = %d\n", size()); 72 } 73 #endif 74 75 uint Block::code_alignment() { 76 // Check for Root block 77 if (_pre_order == 0) return CodeEntryAlignment; 78 // Check for Start block 79 if (_pre_order == 1) return InteriorEntryAlignment; 80 // Check for loop alignment 81 if (has_loop_alignment()) return loop_alignment(); 82 83 return relocInfo::addr_unit(); // no particular alignment 84 } 85 86 uint Block::compute_loop_alignment() { 87 Node *h = head(); 88 int unit_sz = relocInfo::addr_unit(); 89 if (h->is_Loop() && h->as_Loop()->is_inner_loop()) { 90 // Pre- and post-loops have low trip count so do not bother with 91 // NOPs for align loop head. The constants are hidden from tuning 92 // but only because my "divide by 4" heuristic surely gets nearly 93 // all possible gain (a "do not align at all" heuristic has a 94 // chance of getting a really tiny gain). 95 if (h->is_CountedLoop() && (h->as_CountedLoop()->is_pre_loop() || 96 h->as_CountedLoop()->is_post_loop())) { 97 return (OptoLoopAlignment > 4*unit_sz) ? (OptoLoopAlignment>>2) : unit_sz; 98 } 99 // Loops with low backedge frequency should not be aligned. 100 Node *n = h->in(LoopNode::LoopBackControl)->in(0); 101 if (n->is_MachIf() && n->as_MachIf()->_prob < 0.01) { 102 return unit_sz; // Loop does not loop, more often than not! 103 } 104 return OptoLoopAlignment; // Otherwise align loop head 105 } 106 107 return unit_sz; // no particular alignment 108 } 109 110 // Compute the size of first 'inst_cnt' instructions in this block. 111 // Return the number of instructions left to compute if the block has 112 // less then 'inst_cnt' instructions. Stop, and return 0 if sum_size 113 // exceeds OptoLoopAlignment. 114 uint Block::compute_first_inst_size(uint& sum_size, uint inst_cnt, 115 PhaseRegAlloc* ra) { 116 uint last_inst = number_of_nodes(); 117 for( uint j = 0; j < last_inst && inst_cnt > 0; j++ ) { 118 uint inst_size = get_node(j)->size(ra); 119 if( inst_size > 0 ) { 120 inst_cnt--; 121 uint sz = sum_size + inst_size; 122 if( sz <= (uint)OptoLoopAlignment ) { 123 // Compute size of instructions which fit into fetch buffer only 124 // since all inst_cnt instructions will not fit even if we align them. 125 sum_size = sz; 126 } else { 127 return 0; 128 } 129 } 130 } 131 return inst_cnt; 132 } 133 134 uint Block::find_node( const Node *n ) const { 135 for( uint i = 0; i < number_of_nodes(); i++ ) { 136 if( get_node(i) == n ) 137 return i; 138 } 139 ShouldNotReachHere(); 140 return 0; 141 } 142 143 // Find and remove n from block list 144 void Block::find_remove( const Node *n ) { 145 remove_node(find_node(n)); 146 } 147 148 bool Block::contains(const Node *n) const { 149 return _nodes.contains(n); 150 } 151 152 // Return empty status of a block. Empty blocks contain only the head, other 153 // ideal nodes, and an optional trailing goto. 154 int Block::is_Empty() const { 155 156 // Root or start block is not considered empty 157 if (head()->is_Root() || head()->is_Start()) { 158 return not_empty; 159 } 160 161 int success_result = completely_empty; 162 int end_idx = number_of_nodes() - 1; 163 164 // Check for ending goto 165 if ((end_idx > 0) && (get_node(end_idx)->is_MachGoto())) { 166 success_result = empty_with_goto; 167 end_idx--; 168 } 169 170 // Unreachable blocks are considered empty 171 if (num_preds() <= 1) { 172 return success_result; 173 } 174 175 // Ideal nodes are allowable in empty blocks: skip them Only MachNodes 176 // turn directly into code, because only MachNodes have non-trivial 177 // emit() functions. 178 while ((end_idx > 0) && !get_node(end_idx)->is_Mach()) { 179 end_idx--; 180 } 181 182 // No room for any interesting instructions? 183 if (end_idx == 0) { 184 return success_result; 185 } 186 187 return not_empty; 188 } 189 190 // Return true if the block's code implies that it is likely to be 191 // executed infrequently. Check to see if the block ends in a Halt or 192 // a low probability call. 193 bool Block::has_uncommon_code() const { 194 Node* en = end(); 195 196 if (en->is_MachGoto()) 197 en = en->in(0); 198 if (en->is_Catch()) 199 en = en->in(0); 200 if (en->is_MachProj() && en->in(0)->is_MachCall()) { 201 MachCallNode* call = en->in(0)->as_MachCall(); 202 if (call->cnt() != COUNT_UNKNOWN && call->cnt() <= PROB_UNLIKELY_MAG(4)) { 203 // This is true for slow-path stubs like new_{instance,array}, 204 // slow_arraycopy, complete_monitor_locking, uncommon_trap. 205 // The magic number corresponds to the probability of an uncommon_trap, 206 // even though it is a count not a probability. 207 return true; 208 } 209 } 210 211 int op = en->is_Mach() ? en->as_Mach()->ideal_Opcode() : en->Opcode(); 212 return op == Op_Halt; 213 } 214 215 // True if block is low enough frequency or guarded by a test which 216 // mostly does not go here. 217 bool PhaseCFG::is_uncommon(const Block* block) { 218 // Initial blocks must never be moved, so are never uncommon. 219 if (block->head()->is_Root() || block->head()->is_Start()) return false; 220 221 // Check for way-low freq 222 if(block->_freq < BLOCK_FREQUENCY(0.00001f) ) return true; 223 224 // Look for code shape indicating uncommon_trap or slow path 225 if (block->has_uncommon_code()) return true; 226 227 const float epsilon = 0.05f; 228 const float guard_factor = PROB_UNLIKELY_MAG(4) / (1.f - epsilon); 229 uint uncommon_preds = 0; 230 uint freq_preds = 0; 231 uint uncommon_for_freq_preds = 0; 232 233 for( uint i=1; i< block->num_preds(); i++ ) { 234 Block* guard = get_block_for_node(block->pred(i)); 235 // Check to see if this block follows its guard 1 time out of 10000 236 // or less. 237 // 238 // See list of magnitude-4 unlikely probabilities in cfgnode.hpp which 239 // we intend to be "uncommon", such as slow-path TLE allocation, 240 // predicted call failure, and uncommon trap triggers. 241 // 242 // Use an epsilon value of 5% to allow for variability in frequency 243 // predictions and floating point calculations. The net effect is 244 // that guard_factor is set to 9500. 245 // 246 // Ignore low-frequency blocks. 247 // The next check is (guard->_freq < 1.e-5 * 9500.). 248 if(guard->_freq*BLOCK_FREQUENCY(guard_factor) < BLOCK_FREQUENCY(0.00001f)) { 249 uncommon_preds++; 250 } else { 251 freq_preds++; 252 if(block->_freq < guard->_freq * guard_factor ) { 253 uncommon_for_freq_preds++; 254 } 255 } 256 } 257 if( block->num_preds() > 1 && 258 // The block is uncommon if all preds are uncommon or 259 (uncommon_preds == (block->num_preds()-1) || 260 // it is uncommon for all frequent preds. 261 uncommon_for_freq_preds == freq_preds) ) { 262 return true; 263 } 264 return false; 265 } 266 267 #ifndef PRODUCT 268 void Block::dump_bidx(const Block* orig, outputStream* st) const { 269 if (_pre_order) st->print("B%d",_pre_order); 270 else st->print("N%d", head()->_idx); 271 272 if (Verbose && orig != this) { 273 // Dump the original block's idx 274 st->print(" ("); 275 orig->dump_bidx(orig, st); 276 st->print(")"); 277 } 278 } 279 280 void Block::dump_pred(const PhaseCFG* cfg, Block* orig, outputStream* st) const { 281 if (is_connector()) { 282 for (uint i=1; i<num_preds(); i++) { 283 Block *p = cfg->get_block_for_node(pred(i)); 284 p->dump_pred(cfg, orig, st); 285 } 286 } else { 287 dump_bidx(orig, st); 288 st->print(" "); 289 } 290 } 291 292 void Block::dump_head(const PhaseCFG* cfg, outputStream* st) const { 293 // Print the basic block 294 dump_bidx(this, st); 295 st->print(": #\t"); 296 297 // Print the incoming CFG edges and the outgoing CFG edges 298 for( uint i=0; i<_num_succs; i++ ) { 299 non_connector_successor(i)->dump_bidx(_succs[i], st); 300 st->print(" "); 301 } 302 st->print("<- "); 303 if( head()->is_block_start() ) { 304 for (uint i=1; i<num_preds(); i++) { 305 Node *s = pred(i); 306 if (cfg != NULL) { 307 Block *p = cfg->get_block_for_node(s); 308 p->dump_pred(cfg, p, st); 309 } else { 310 while (!s->is_block_start()) 311 s = s->in(0); 312 st->print("N%d ", s->_idx ); 313 } 314 } 315 } else { 316 st->print("BLOCK HEAD IS JUNK "); 317 } 318 319 // Print loop, if any 320 const Block *bhead = this; // Head of self-loop 321 Node *bh = bhead->head(); 322 323 if ((cfg != NULL) && bh->is_Loop() && !head()->is_Root()) { 324 LoopNode *loop = bh->as_Loop(); 325 const Block *bx = cfg->get_block_for_node(loop->in(LoopNode::LoopBackControl)); 326 while (bx->is_connector()) { 327 bx = cfg->get_block_for_node(bx->pred(1)); 328 } 329 st->print("\tLoop: B%d-B%d ", bhead->_pre_order, bx->_pre_order); 330 // Dump any loop-specific bits, especially for CountedLoops. 331 loop->dump_spec(st); 332 } else if (has_loop_alignment()) { 333 st->print(" top-of-loop"); 334 } 335 st->print(" Freq: %g",_freq); 336 if( Verbose || WizardMode ) { 337 st->print(" IDom: %d/#%d", _idom ? _idom->_pre_order : 0, _dom_depth); 338 st->print(" RegPressure: %d",_reg_pressure); 339 st->print(" IHRP Index: %d",_ihrp_index); 340 st->print(" FRegPressure: %d",_freg_pressure); 341 st->print(" FHRP Index: %d",_fhrp_index); 342 } 343 st->cr(); 344 } 345 346 void Block::dump() const { 347 dump(NULL); 348 } 349 350 void Block::dump(const PhaseCFG* cfg) const { 351 dump_head(cfg); 352 for (uint i=0; i< number_of_nodes(); i++) { 353 get_node(i)->dump(); 354 } 355 tty->print("\n"); 356 } 357 #endif 358 359 PhaseCFG::PhaseCFG(Arena* arena, RootNode* root, Matcher& matcher) 360 : Phase(CFG) 361 , _block_arena(arena) 362 , _regalloc(NULL) 363 , _scheduling_for_pressure(false) 364 , _root(root) 365 , _matcher(matcher) 366 , _node_to_block_mapping(arena) 367 , _node_latency(NULL) 368 #ifndef PRODUCT 369 , _trace_opto_pipelining(C->directive()->TraceOptoPipeliningOption) 370 #endif 371 #ifdef ASSERT 372 , _raw_oops(arena) 373 #endif 374 { 375 ResourceMark rm; 376 // I'll need a few machine-specific GotoNodes. Make an Ideal GotoNode, 377 // then Match it into a machine-specific Node. Then clone the machine 378 // Node on demand. 379 Node *x = new GotoNode(NULL); 380 x->init_req(0, x); 381 _goto = matcher.match_tree(x); 382 assert(_goto != NULL, ""); 383 _goto->set_req(0,_goto); 384 385 // Build the CFG in Reverse Post Order 386 _number_of_blocks = build_cfg(); 387 _root_block = get_block_for_node(_root); 388 } 389 390 // Build a proper looking CFG. Make every block begin with either a StartNode 391 // or a RegionNode. Make every block end with either a Goto, If or Return. 392 // The RootNode both starts and ends it's own block. Do this with a recursive 393 // backwards walk over the control edges. 394 uint PhaseCFG::build_cfg() { 395 Arena *a = Thread::current()->resource_area(); 396 VectorSet visited(a); 397 398 // Allocate stack with enough space to avoid frequent realloc 399 Node_Stack nstack(a, C->live_nodes() >> 1); 400 nstack.push(_root, 0); 401 uint sum = 0; // Counter for blocks 402 403 while (nstack.is_nonempty()) { 404 // node and in's index from stack's top 405 // 'np' is _root (see above) or RegionNode, StartNode: we push on stack 406 // only nodes which point to the start of basic block (see below). 407 Node *np = nstack.node(); 408 // idx > 0, except for the first node (_root) pushed on stack 409 // at the beginning when idx == 0. 410 // We will use the condition (idx == 0) later to end the build. 411 uint idx = nstack.index(); 412 Node *proj = np->in(idx); 413 const Node *x = proj->is_block_proj(); 414 // Does the block end with a proper block-ending Node? One of Return, 415 // If or Goto? (This check should be done for visited nodes also). 416 if (x == NULL) { // Does not end right... 417 Node *g = _goto->clone(); // Force it to end in a Goto 418 g->set_req(0, proj); 419 np->set_req(idx, g); 420 x = proj = g; 421 } 422 if (!visited.test_set(x->_idx)) { // Visit this block once 423 // Skip any control-pinned middle'in stuff 424 Node *p = proj; 425 do { 426 proj = p; // Update pointer to last Control 427 p = p->in(0); // Move control forward 428 } while( !p->is_block_proj() && 429 !p->is_block_start() ); 430 // Make the block begin with one of Region or StartNode. 431 if( !p->is_block_start() ) { 432 RegionNode *r = new RegionNode( 2 ); 433 r->init_req(1, p); // Insert RegionNode in the way 434 proj->set_req(0, r); // Insert RegionNode in the way 435 p = r; 436 } 437 // 'p' now points to the start of this basic block 438 439 // Put self in array of basic blocks 440 Block *bb = new (_block_arena) Block(_block_arena, p); 441 map_node_to_block(p, bb); 442 map_node_to_block(x, bb); 443 if( x != p ) { // Only for root is x == p 444 bb->push_node((Node*)x); 445 } 446 // Now handle predecessors 447 ++sum; // Count 1 for self block 448 uint cnt = bb->num_preds(); 449 for (int i = (cnt - 1); i > 0; i-- ) { // For all predecessors 450 Node *prevproj = p->in(i); // Get prior input 451 assert( !prevproj->is_Con(), "dead input not removed" ); 452 // Check to see if p->in(i) is a "control-dependent" CFG edge - 453 // i.e., it splits at the source (via an IF or SWITCH) and merges 454 // at the destination (via a many-input Region). 455 // This breaks critical edges. The RegionNode to start the block 456 // will be added when <p,i> is pulled off the node stack 457 if ( cnt > 2 ) { // Merging many things? 458 assert( prevproj== bb->pred(i),""); 459 if(prevproj->is_block_proj() != prevproj) { // Control-dependent edge? 460 // Force a block on the control-dependent edge 461 Node *g = _goto->clone(); // Force it to end in a Goto 462 g->set_req(0,prevproj); 463 p->set_req(i,g); 464 } 465 } 466 nstack.push(p, i); // 'p' is RegionNode or StartNode 467 } 468 } else { // Post-processing visited nodes 469 nstack.pop(); // remove node from stack 470 // Check if it the fist node pushed on stack at the beginning. 471 if (idx == 0) break; // end of the build 472 // Find predecessor basic block 473 Block *pb = get_block_for_node(x); 474 // Insert into nodes array, if not already there 475 if (!has_block(proj)) { 476 assert( x != proj, "" ); 477 // Map basic block of projection 478 map_node_to_block(proj, pb); 479 pb->push_node(proj); 480 } 481 // Insert self as a child of my predecessor block 482 pb->_succs.map(pb->_num_succs++, get_block_for_node(np)); 483 assert( pb->get_node(pb->number_of_nodes() - pb->_num_succs)->is_block_proj(), 484 "too many control users, not a CFG?" ); 485 } 486 } 487 // Return number of basic blocks for all children and self 488 return sum; 489 } 490 491 // Inserts a goto & corresponding basic block between 492 // block[block_no] and its succ_no'th successor block 493 void PhaseCFG::insert_goto_at(uint block_no, uint succ_no) { 494 // get block with block_no 495 assert(block_no < number_of_blocks(), "illegal block number"); 496 Block* in = get_block(block_no); 497 // get successor block succ_no 498 assert(succ_no < in->_num_succs, "illegal successor number"); 499 Block* out = in->_succs[succ_no]; 500 // Compute frequency of the new block. Do this before inserting 501 // new block in case succ_prob() needs to infer the probability from 502 // surrounding blocks. 503 float freq = in->_freq * in->succ_prob(succ_no); 504 // get ProjNode corresponding to the succ_no'th successor of the in block 505 ProjNode* proj = in->get_node(in->number_of_nodes() - in->_num_succs + succ_no)->as_Proj(); 506 // create region for basic block 507 RegionNode* region = new RegionNode(2); 508 region->init_req(1, proj); 509 // setup corresponding basic block 510 Block* block = new (_block_arena) Block(_block_arena, region); 511 map_node_to_block(region, block); 512 C->regalloc()->set_bad(region->_idx); 513 // add a goto node 514 Node* gto = _goto->clone(); // get a new goto node 515 gto->set_req(0, region); 516 // add it to the basic block 517 block->push_node(gto); 518 map_node_to_block(gto, block); 519 C->regalloc()->set_bad(gto->_idx); 520 // hook up successor block 521 block->_succs.map(block->_num_succs++, out); 522 // remap successor's predecessors if necessary 523 for (uint i = 1; i < out->num_preds(); i++) { 524 if (out->pred(i) == proj) out->head()->set_req(i, gto); 525 } 526 // remap predecessor's successor to new block 527 in->_succs.map(succ_no, block); 528 // Set the frequency of the new block 529 block->_freq = freq; 530 // add new basic block to basic block list 531 add_block_at(block_no + 1, block); 532 } 533 534 // Does this block end in a multiway branch that cannot have the default case 535 // flipped for another case? 536 static bool no_flip_branch(Block *b) { 537 int branch_idx = b->number_of_nodes() - b->_num_succs-1; 538 if (branch_idx < 1) { 539 return false; 540 } 541 Node *branch = b->get_node(branch_idx); 542 if (branch->is_Catch()) { 543 return true; 544 } 545 if (branch->is_Mach()) { 546 if (branch->is_MachNullCheck()) { 547 return true; 548 } 549 int iop = branch->as_Mach()->ideal_Opcode(); 550 if (iop == Op_FastLock || iop == Op_FastUnlock) { 551 return true; 552 } 553 // Don't flip if branch has an implicit check. 554 if (branch->as_Mach()->is_TrapBasedCheckNode()) { 555 return true; 556 } 557 } 558 return false; 559 } 560 561 // Check for NeverBranch at block end. This needs to become a GOTO to the 562 // true target. NeverBranch are treated as a conditional branch that always 563 // goes the same direction for most of the optimizer and are used to give a 564 // fake exit path to infinite loops. At this late stage they need to turn 565 // into Goto's so that when you enter the infinite loop you indeed hang. 566 void PhaseCFG::convert_NeverBranch_to_Goto(Block *b) { 567 // Find true target 568 int end_idx = b->end_idx(); 569 int idx = b->get_node(end_idx+1)->as_Proj()->_con; 570 Block *succ = b->_succs[idx]; 571 Node* gto = _goto->clone(); // get a new goto node 572 gto->set_req(0, b->head()); 573 Node *bp = b->get_node(end_idx); 574 b->map_node(gto, end_idx); // Slam over NeverBranch 575 map_node_to_block(gto, b); 576 C->regalloc()->set_bad(gto->_idx); 577 b->pop_node(); // Yank projections 578 b->pop_node(); // Yank projections 579 b->_succs.map(0,succ); // Map only successor 580 b->_num_succs = 1; 581 // remap successor's predecessors if necessary 582 uint j; 583 for( j = 1; j < succ->num_preds(); j++) 584 if( succ->pred(j)->in(0) == bp ) 585 succ->head()->set_req(j, gto); 586 // Kill alternate exit path 587 Block *dead = b->_succs[1-idx]; 588 for( j = 1; j < dead->num_preds(); j++) 589 if( dead->pred(j)->in(0) == bp ) 590 break; 591 // Scan through block, yanking dead path from 592 // all regions and phis. 593 dead->head()->del_req(j); 594 for( int k = 1; dead->get_node(k)->is_Phi(); k++ ) 595 dead->get_node(k)->del_req(j); 596 } 597 598 // Helper function to move block bx to the slot following b_index. Return 599 // true if the move is successful, otherwise false 600 bool PhaseCFG::move_to_next(Block* bx, uint b_index) { 601 if (bx == NULL) return false; 602 603 // Return false if bx is already scheduled. 604 uint bx_index = bx->_pre_order; 605 if ((bx_index <= b_index) && (get_block(bx_index) == bx)) { 606 return false; 607 } 608 609 // Find the current index of block bx on the block list 610 bx_index = b_index + 1; 611 while (bx_index < number_of_blocks() && get_block(bx_index) != bx) { 612 bx_index++; 613 } 614 assert(get_block(bx_index) == bx, "block not found"); 615 616 // If the previous block conditionally falls into bx, return false, 617 // because moving bx will create an extra jump. 618 for(uint k = 1; k < bx->num_preds(); k++ ) { 619 Block* pred = get_block_for_node(bx->pred(k)); 620 if (pred == get_block(bx_index - 1)) { 621 if (pred->_num_succs != 1) { 622 return false; 623 } 624 } 625 } 626 627 // Reinsert bx just past block 'b' 628 _blocks.remove(bx_index); 629 _blocks.insert(b_index + 1, bx); 630 return true; 631 } 632 633 // Move empty and uncommon blocks to the end. 634 void PhaseCFG::move_to_end(Block *b, uint i) { 635 int e = b->is_Empty(); 636 if (e != Block::not_empty) { 637 if (e == Block::empty_with_goto) { 638 // Remove the goto, but leave the block. 639 b->pop_node(); 640 } 641 // Mark this block as a connector block, which will cause it to be 642 // ignored in certain functions such as non_connector_successor(). 643 b->set_connector(); 644 } 645 // Move the empty block to the end, and don't recheck. 646 _blocks.remove(i); 647 _blocks.push(b); 648 } 649 650 // Set loop alignment for every block 651 void PhaseCFG::set_loop_alignment() { 652 uint last = number_of_blocks(); 653 assert(get_block(0) == get_root_block(), ""); 654 655 for (uint i = 1; i < last; i++) { 656 Block* block = get_block(i); 657 if (block->head()->is_Loop()) { 658 block->set_loop_alignment(block); 659 } 660 } 661 } 662 663 // Make empty basic blocks to be "connector" blocks, Move uncommon blocks 664 // to the end. 665 void PhaseCFG::remove_empty_blocks() { 666 // Move uncommon blocks to the end 667 uint last = number_of_blocks(); 668 assert(get_block(0) == get_root_block(), ""); 669 670 for (uint i = 1; i < last; i++) { 671 Block* block = get_block(i); 672 if (block->is_connector()) { 673 break; 674 } 675 676 // Check for NeverBranch at block end. This needs to become a GOTO to the 677 // true target. NeverBranch are treated as a conditional branch that 678 // always goes the same direction for most of the optimizer and are used 679 // to give a fake exit path to infinite loops. At this late stage they 680 // need to turn into Goto's so that when you enter the infinite loop you 681 // indeed hang. 682 if (block->get_node(block->end_idx())->Opcode() == Op_NeverBranch) { 683 convert_NeverBranch_to_Goto(block); 684 } 685 686 // Look for uncommon blocks and move to end. 687 if (!C->do_freq_based_layout()) { 688 if (is_uncommon(block)) { 689 move_to_end(block, i); 690 last--; // No longer check for being uncommon! 691 if (no_flip_branch(block)) { // Fall-thru case must follow? 692 // Find the fall-thru block 693 block = get_block(i); 694 move_to_end(block, i); 695 last--; 696 } 697 // backup block counter post-increment 698 i--; 699 } 700 } 701 } 702 703 // Move empty blocks to the end 704 last = number_of_blocks(); 705 for (uint i = 1; i < last; i++) { 706 Block* block = get_block(i); 707 if (block->is_Empty() != Block::not_empty) { 708 move_to_end(block, i); 709 last--; 710 i--; 711 } 712 } // End of for all blocks 713 } 714 715 Block *PhaseCFG::fixup_trap_based_check(Node *branch, Block *block, int block_pos, Block *bnext) { 716 // Trap based checks must fall through to the successor with 717 // PROB_ALWAYS. 718 // They should be an If with 2 successors. 719 assert(branch->is_MachIf(), "must be If"); 720 assert(block->_num_succs == 2, "must have 2 successors"); 721 722 // Get the If node and the projection for the first successor. 723 MachIfNode *iff = block->get_node(block->number_of_nodes()-3)->as_MachIf(); 724 ProjNode *proj0 = block->get_node(block->number_of_nodes()-2)->as_Proj(); 725 ProjNode *proj1 = block->get_node(block->number_of_nodes()-1)->as_Proj(); 726 ProjNode *projt = (proj0->Opcode() == Op_IfTrue) ? proj0 : proj1; 727 ProjNode *projf = (proj0->Opcode() == Op_IfFalse) ? proj0 : proj1; 728 729 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1]. 730 assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0"); 731 assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1"); 732 733 ProjNode *proj_always; 734 ProjNode *proj_never; 735 // We must negate the branch if the implicit check doesn't follow 736 // the branch's TRUE path. Then, the new TRUE branch target will 737 // be the old FALSE branch target. 738 if (iff->_prob <= 2*PROB_NEVER) { // There are small rounding errors. 739 proj_never = projt; 740 proj_always = projf; 741 } else { 742 // We must negate the branch if the trap doesn't follow the 743 // branch's TRUE path. Then, the new TRUE branch target will 744 // be the old FALSE branch target. 745 proj_never = projf; 746 proj_always = projt; 747 iff->negate(); 748 } 749 assert(iff->_prob <= 2*PROB_NEVER, "Trap based checks are expected to trap never!"); 750 // Map the successors properly 751 block->_succs.map(0, get_block_for_node(proj_never ->raw_out(0))); // The target of the trap. 752 block->_succs.map(1, get_block_for_node(proj_always->raw_out(0))); // The fall through target. 753 754 if (block->get_node(block->number_of_nodes() - block->_num_succs + 1) != proj_always) { 755 block->map_node(proj_never, block->number_of_nodes() - block->_num_succs + 0); 756 block->map_node(proj_always, block->number_of_nodes() - block->_num_succs + 1); 757 } 758 759 // Place the fall through block after this block. 760 Block *bs1 = block->non_connector_successor(1); 761 if (bs1 != bnext && move_to_next(bs1, block_pos)) { 762 bnext = bs1; 763 } 764 // If the fall through block still is not the next block, insert a goto. 765 if (bs1 != bnext) { 766 insert_goto_at(block_pos, 1); 767 } 768 return bnext; 769 } 770 771 // Fix up the final control flow for basic blocks. 772 void PhaseCFG::fixup_flow() { 773 // Fixup final control flow for the blocks. Remove jump-to-next 774 // block. If neither arm of an IF follows the conditional branch, we 775 // have to add a second jump after the conditional. We place the 776 // TRUE branch target in succs[0] for both GOTOs and IFs. 777 for (uint i = 0; i < number_of_blocks(); i++) { 778 Block* block = get_block(i); 779 block->_pre_order = i; // turn pre-order into block-index 780 781 // Connector blocks need no further processing. 782 if (block->is_connector()) { 783 assert((i+1) == number_of_blocks() || get_block(i + 1)->is_connector(), "All connector blocks should sink to the end"); 784 continue; 785 } 786 assert(block->is_Empty() != Block::completely_empty, "Empty blocks should be connectors"); 787 788 Block* bnext = (i < number_of_blocks() - 1) ? get_block(i + 1) : NULL; 789 Block* bs0 = block->non_connector_successor(0); 790 791 // Check for multi-way branches where I cannot negate the test to 792 // exchange the true and false targets. 793 if (no_flip_branch(block)) { 794 // Find fall through case - if must fall into its target. 795 // Get the index of the branch's first successor. 796 int branch_idx = block->number_of_nodes() - block->_num_succs; 797 798 // The branch is 1 before the branch's first successor. 799 Node *branch = block->get_node(branch_idx-1); 800 801 // Handle no-flip branches which have implicit checks and which require 802 // special block ordering and individual semantics of the 'fall through 803 // case'. 804 if ((TrapBasedNullChecks || TrapBasedRangeChecks) && 805 branch->is_Mach() && branch->as_Mach()->is_TrapBasedCheckNode()) { 806 bnext = fixup_trap_based_check(branch, block, i, bnext); 807 } else { 808 // Else, default handling for no-flip branches 809 for (uint j2 = 0; j2 < block->_num_succs; j2++) { 810 const ProjNode* p = block->get_node(branch_idx + j2)->as_Proj(); 811 if (p->_con == 0) { 812 // successor j2 is fall through case 813 if (block->non_connector_successor(j2) != bnext) { 814 // but it is not the next block => insert a goto 815 insert_goto_at(i, j2); 816 } 817 // Put taken branch in slot 0 818 if (j2 == 0 && block->_num_succs == 2) { 819 // Flip targets in succs map 820 Block *tbs0 = block->_succs[0]; 821 Block *tbs1 = block->_succs[1]; 822 block->_succs.map(0, tbs1); 823 block->_succs.map(1, tbs0); 824 } 825 break; 826 } 827 } 828 } 829 830 // Remove all CatchProjs 831 for (uint j = 0; j < block->_num_succs; j++) { 832 block->pop_node(); 833 } 834 835 } else if (block->_num_succs == 1) { 836 // Block ends in a Goto? 837 if (bnext == bs0) { 838 // We fall into next block; remove the Goto 839 block->pop_node(); 840 } 841 842 } else if(block->_num_succs == 2) { // Block ends in a If? 843 // Get opcode of 1st projection (matches _succs[0]) 844 // Note: Since this basic block has 2 exits, the last 2 nodes must 845 // be projections (in any order), the 3rd last node must be 846 // the IfNode (we have excluded other 2-way exits such as 847 // CatchNodes already). 848 MachNode* iff = block->get_node(block->number_of_nodes() - 3)->as_Mach(); 849 ProjNode* proj0 = block->get_node(block->number_of_nodes() - 2)->as_Proj(); 850 ProjNode* proj1 = block->get_node(block->number_of_nodes() - 1)->as_Proj(); 851 852 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1]. 853 assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0"); 854 assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1"); 855 856 Block* bs1 = block->non_connector_successor(1); 857 858 // Check for neither successor block following the current 859 // block ending in a conditional. If so, move one of the 860 // successors after the current one, provided that the 861 // successor was previously unscheduled, but moveable 862 // (i.e., all paths to it involve a branch). 863 if (!C->do_freq_based_layout() && bnext != bs0 && bnext != bs1) { 864 // Choose the more common successor based on the probability 865 // of the conditional branch. 866 Block* bx = bs0; 867 Block* by = bs1; 868 869 // _prob is the probability of taking the true path. Make 870 // p the probability of taking successor #1. 871 float p = iff->as_MachIf()->_prob; 872 if (proj0->Opcode() == Op_IfTrue) { 873 p = 1.0 - p; 874 } 875 876 // Prefer successor #1 if p > 0.5 877 if (p > PROB_FAIR) { 878 bx = bs1; 879 by = bs0; 880 } 881 882 // Attempt the more common successor first 883 if (move_to_next(bx, i)) { 884 bnext = bx; 885 } else if (move_to_next(by, i)) { 886 bnext = by; 887 } 888 } 889 890 // Check for conditional branching the wrong way. Negate 891 // conditional, if needed, so it falls into the following block 892 // and branches to the not-following block. 893 894 // Check for the next block being in succs[0]. We are going to branch 895 // to succs[0], so we want the fall-thru case as the next block in 896 // succs[1]. 897 if (bnext == bs0) { 898 // Fall-thru case in succs[0], so flip targets in succs map 899 Block* tbs0 = block->_succs[0]; 900 Block* tbs1 = block->_succs[1]; 901 block->_succs.map(0, tbs1); 902 block->_succs.map(1, tbs0); 903 // Flip projection for each target 904 ProjNode* tmp = proj0; 905 proj0 = proj1; 906 proj1 = tmp; 907 908 } else if(bnext != bs1) { 909 // Need a double-branch 910 // The existing conditional branch need not change. 911 // Add a unconditional branch to the false target. 912 // Alas, it must appear in its own block and adding a 913 // block this late in the game is complicated. Sigh. 914 insert_goto_at(i, 1); 915 } 916 917 // Make sure we TRUE branch to the target 918 if (proj0->Opcode() == Op_IfFalse) { 919 iff->as_MachIf()->negate(); 920 } 921 922 block->pop_node(); // Remove IfFalse & IfTrue projections 923 block->pop_node(); 924 925 } else { 926 // Multi-exit block, e.g. a switch statement 927 // But we don't need to do anything here 928 } 929 } // End of for all blocks 930 } 931 932 933 // postalloc_expand: Expand nodes after register allocation. 934 // 935 // postalloc_expand has to be called after register allocation, just 936 // before output (i.e. scheduling). It only gets called if 937 // Matcher::require_postalloc_expand is true. 938 // 939 // Background: 940 // 941 // Nodes that are expandend (one compound node requiring several 942 // assembler instructions to be implemented split into two or more 943 // non-compound nodes) after register allocation are not as nice as 944 // the ones expanded before register allocation - they don't 945 // participate in optimizations as global code motion. But after 946 // register allocation we can expand nodes that use registers which 947 // are not spillable or registers that are not allocated, because the 948 // old compound node is simply replaced (in its location in the basic 949 // block) by a new subgraph which does not contain compound nodes any 950 // more. The scheduler called during output can later on process these 951 // non-compound nodes. 952 // 953 // Implementation: 954 // 955 // Nodes requiring postalloc expand are specified in the ad file by using 956 // a postalloc_expand statement instead of ins_encode. A postalloc_expand 957 // contains a single call to an encoding, as does an ins_encode 958 // statement. Instead of an emit() function a postalloc_expand() function 959 // is generated that doesn't emit assembler but creates a new 960 // subgraph. The code below calls this postalloc_expand function for each 961 // node with the appropriate attribute. This function returns the new 962 // nodes generated in an array passed in the call. The old node, 963 // potential MachTemps before and potential Projs after it then get 964 // disconnected and replaced by the new nodes. The instruction 965 // generating the result has to be the last one in the array. In 966 // general it is assumed that Projs after the node expanded are 967 // kills. These kills are not required any more after expanding as 968 // there are now explicitly visible def-use chains and the Projs are 969 // removed. This does not hold for calls: They do not only have 970 // kill-Projs but also Projs defining values. Therefore Projs after 971 // the node expanded are removed for all but for calls. If a node is 972 // to be reused, it must be added to the nodes list returned, and it 973 // will be added again. 974 // 975 // Implementing the postalloc_expand function for a node in an enc_class 976 // is rather tedious. It requires knowledge about many node details, as 977 // the nodes and the subgraph must be hand crafted. To simplify this, 978 // adlc generates some utility variables into the postalloc_expand function, 979 // e.g., holding the operands as specified by the postalloc_expand encoding 980 // specification, e.g.: 981 // * unsigned idx_<par_name> holding the index of the node in the ins 982 // * Node *n_<par_name> holding the node loaded from the ins 983 // * MachOpnd *op_<par_name> holding the corresponding operand 984 // 985 // The ordering of operands can not be determined by looking at a 986 // rule. Especially if a match rule matches several different trees, 987 // several nodes are generated from one instruct specification with 988 // different operand orderings. In this case the adlc generated 989 // variables are the only way to access the ins and operands 990 // deterministically. 991 // 992 // If assigning a register to a node that contains an oop, don't 993 // forget to call ra_->set_oop() for the node. 994 void PhaseCFG::postalloc_expand(PhaseRegAlloc* _ra) { 995 GrowableArray <Node *> new_nodes(32); // Array with new nodes filled by postalloc_expand function of node. 996 GrowableArray <Node *> remove(32); 997 GrowableArray <Node *> succs(32); 998 unsigned int max_idx = C->unique(); // Remember to distinguish new from old nodes. 999 DEBUG_ONLY(bool foundNode = false); 1000 1001 // for all blocks 1002 for (uint i = 0; i < number_of_blocks(); i++) { 1003 Block *b = _blocks[i]; 1004 // For all instructions in the current block. 1005 for (uint j = 0; j < b->number_of_nodes(); j++) { 1006 Node *n = b->get_node(j); 1007 if (n->is_Mach() && n->as_Mach()->requires_postalloc_expand()) { 1008 #ifdef ASSERT 1009 if (TracePostallocExpand) { 1010 if (!foundNode) { 1011 foundNode = true; 1012 tty->print("POSTALLOC EXPANDING %d %s\n", C->compile_id(), 1013 C->method() ? C->method()->name()->as_utf8() : C->stub_name()); 1014 } 1015 tty->print(" postalloc expanding "); n->dump(); 1016 if (Verbose) { 1017 tty->print(" with ins:\n"); 1018 for (uint k = 0; k < n->len(); ++k) { 1019 if (n->in(k)) { tty->print(" "); n->in(k)->dump(); } 1020 } 1021 } 1022 } 1023 #endif 1024 new_nodes.clear(); 1025 // Collect nodes that have to be removed from the block later on. 1026 uint req = n->req(); 1027 remove.clear(); 1028 for (uint k = 0; k < req; ++k) { 1029 if (n->in(k) && n->in(k)->is_MachTemp()) { 1030 remove.push(n->in(k)); // MachTemps which are inputs to the old node have to be removed. 1031 n->in(k)->del_req(0); 1032 j--; 1033 } 1034 } 1035 1036 // Check whether we can allocate enough nodes. We set a fix limit for 1037 // the size of postalloc expands with this. 1038 uint unique_limit = C->unique() + 40; 1039 if (unique_limit >= _ra->node_regs_max_index()) { 1040 Compile::current()->record_failure("out of nodes in postalloc expand"); 1041 return; 1042 } 1043 1044 // Emit (i.e. generate new nodes). 1045 n->as_Mach()->postalloc_expand(&new_nodes, _ra); 1046 1047 assert(C->unique() < unique_limit, "You allocated too many nodes in your postalloc expand."); 1048 1049 // Disconnect the inputs of the old node. 1050 // 1051 // We reuse MachSpillCopy nodes. If we need to expand them, there 1052 // are many, so reusing pays off. If reused, the node already 1053 // has the new ins. n must be the last node on new_nodes list. 1054 if (!n->is_MachSpillCopy()) { 1055 for (int k = req - 1; k >= 0; --k) { 1056 n->del_req(k); 1057 } 1058 } 1059 1060 #ifdef ASSERT 1061 // Check that all nodes have proper operands. 1062 for (int k = 0; k < new_nodes.length(); ++k) { 1063 if (new_nodes.at(k)->_idx < max_idx || !new_nodes.at(k)->is_Mach()) continue; // old node, Proj ... 1064 MachNode *m = new_nodes.at(k)->as_Mach(); 1065 for (unsigned int l = 0; l < m->num_opnds(); ++l) { 1066 if (MachOper::notAnOper(m->_opnds[l])) { 1067 outputStream *os = tty; 1068 os->print("Node %s ", m->Name()); 1069 os->print("has invalid opnd %d: %p\n", l, m->_opnds[l]); 1070 assert(0, "Invalid operands, see inline trace in hs_err_pid file."); 1071 } 1072 } 1073 } 1074 #endif 1075 1076 // Collect succs of old node in remove (for projections) and in succs (for 1077 // all other nodes) do _not_ collect projections in remove (but in succs) 1078 // in case the node is a call. We need the projections for calls as they are 1079 // associated with registes (i.e. they are defs). 1080 succs.clear(); 1081 for (DUIterator k = n->outs(); n->has_out(k); k++) { 1082 if (n->out(k)->is_Proj() && !n->is_MachCall() && !n->is_MachBranch()) { 1083 remove.push(n->out(k)); 1084 } else { 1085 succs.push(n->out(k)); 1086 } 1087 } 1088 // Replace old node n as input of its succs by last of the new nodes. 1089 for (int k = 0; k < succs.length(); ++k) { 1090 Node *succ = succs.at(k); 1091 for (uint l = 0; l < succ->req(); ++l) { 1092 if (succ->in(l) == n) { 1093 succ->set_req(l, new_nodes.at(new_nodes.length() - 1)); 1094 } 1095 } 1096 for (uint l = succ->req(); l < succ->len(); ++l) { 1097 if (succ->in(l) == n) { 1098 succ->set_prec(l, new_nodes.at(new_nodes.length() - 1)); 1099 } 1100 } 1101 } 1102 1103 // Index of old node in block. 1104 uint index = b->find_node(n); 1105 // Insert new nodes into block and map them in nodes->blocks array 1106 // and remember last node in n2. 1107 Node *n2 = NULL; 1108 for (int k = 0; k < new_nodes.length(); ++k) { 1109 n2 = new_nodes.at(k); 1110 b->insert_node(n2, ++index); 1111 map_node_to_block(n2, b); 1112 } 1113 1114 // Add old node n to remove and remove them all from block. 1115 remove.push(n); 1116 j--; 1117 #ifdef ASSERT 1118 if (TracePostallocExpand && Verbose) { 1119 tty->print(" removing:\n"); 1120 for (int k = 0; k < remove.length(); ++k) { 1121 tty->print(" "); remove.at(k)->dump(); 1122 } 1123 tty->print(" inserting:\n"); 1124 for (int k = 0; k < new_nodes.length(); ++k) { 1125 tty->print(" "); new_nodes.at(k)->dump(); 1126 } 1127 } 1128 #endif 1129 for (int k = 0; k < remove.length(); ++k) { 1130 if (b->contains(remove.at(k))) { 1131 b->find_remove(remove.at(k)); 1132 } else { 1133 assert(remove.at(k)->is_Proj() && (remove.at(k)->in(0)->is_MachBranch()), ""); 1134 } 1135 } 1136 // If anything has been inserted (n2 != NULL), continue after last node inserted. 1137 // This does not always work. Some postalloc expands don't insert any nodes, if they 1138 // do optimizations (e.g., max(x,x)). In this case we decrement j accordingly. 1139 j = n2 ? b->find_node(n2) : j; 1140 } 1141 } 1142 } 1143 1144 #ifdef ASSERT 1145 if (foundNode) { 1146 tty->print("FINISHED %d %s\n", C->compile_id(), 1147 C->method() ? C->method()->name()->as_utf8() : C->stub_name()); 1148 tty->flush(); 1149 } 1150 #endif 1151 } 1152 1153 1154 //------------------------------dump------------------------------------------- 1155 #ifndef PRODUCT 1156 void PhaseCFG::_dump_cfg( const Node *end, VectorSet &visited ) const { 1157 const Node *x = end->is_block_proj(); 1158 assert( x, "not a CFG" ); 1159 1160 // Do not visit this block again 1161 if( visited.test_set(x->_idx) ) return; 1162 1163 // Skip through this block 1164 const Node *p = x; 1165 do { 1166 p = p->in(0); // Move control forward 1167 assert( !p->is_block_proj() || p->is_Root(), "not a CFG" ); 1168 } while( !p->is_block_start() ); 1169 1170 // Recursively visit 1171 for (uint i = 1; i < p->req(); i++) { 1172 _dump_cfg(p->in(i), visited); 1173 } 1174 1175 // Dump the block 1176 get_block_for_node(p)->dump(this); 1177 } 1178 1179 void PhaseCFG::dump( ) const { 1180 tty->print("\n--- CFG --- %d BBs\n", number_of_blocks()); 1181 if (_blocks.size()) { // Did we do basic-block layout? 1182 for (uint i = 0; i < number_of_blocks(); i++) { 1183 const Block* block = get_block(i); 1184 block->dump(this); 1185 } 1186 } else { // Else do it with a DFS 1187 VectorSet visited(_block_arena); 1188 _dump_cfg(_root,visited); 1189 } 1190 } 1191 1192 void PhaseCFG::dump_headers() { 1193 for (uint i = 0; i < number_of_blocks(); i++) { 1194 Block* block = get_block(i); 1195 if (block != NULL) { 1196 block->dump_head(this); 1197 } 1198 } 1199 } 1200 1201 void PhaseCFG::verify() const { 1202 #ifdef ASSERT 1203 // Verify sane CFG 1204 for (uint i = 0; i < number_of_blocks(); i++) { 1205 Block* block = get_block(i); 1206 uint cnt = block->number_of_nodes(); 1207 uint j; 1208 for (j = 0; j < cnt; j++) { 1209 Node *n = block->get_node(j); 1210 assert(get_block_for_node(n) == block, ""); 1211 if (j >= 1 && n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CreateEx) { 1212 assert(j == 1 || block->get_node(j-1)->is_Phi(), "CreateEx must be first instruction in block"); 1213 } 1214 for (uint k = 0; k < n->req(); k++) { 1215 Node *def = n->in(k); 1216 if (def && def != n) { 1217 assert(get_block_for_node(def) || def->is_Con(), "must have block; constants for debug info ok"); 1218 // Verify that instructions in the block is in correct order. 1219 // Uses must follow their definition if they are at the same block. 1220 // Mostly done to check that MachSpillCopy nodes are placed correctly 1221 // when CreateEx node is moved in build_ifg_physical(). 1222 if (get_block_for_node(def) == block && !(block->head()->is_Loop() && n->is_Phi()) && 1223 // See (+++) comment in reg_split.cpp 1224 !(n->jvms() != NULL && n->jvms()->is_monitor_use(k))) { 1225 bool is_loop = false; 1226 if (n->is_Phi()) { 1227 for (uint l = 1; l < def->req(); l++) { 1228 if (n == def->in(l)) { 1229 is_loop = true; 1230 break; // Some kind of loop 1231 } 1232 } 1233 } 1234 assert(is_loop || block->find_node(def) < j, "uses must follow definitions"); 1235 } 1236 } 1237 } 1238 } 1239 1240 j = block->end_idx(); 1241 Node* bp = (Node*)block->get_node(block->number_of_nodes() - 1)->is_block_proj(); 1242 assert(bp, "last instruction must be a block proj"); 1243 assert(bp == block->get_node(j), "wrong number of successors for this block"); 1244 if (bp->is_Catch()) { 1245 while (block->get_node(--j)->is_MachProj()) { 1246 ; 1247 } 1248 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call"); 1249 } else if (bp->is_Mach() && bp->as_Mach()->ideal_Opcode() == Op_If) { 1250 assert(block->_num_succs == 2, "Conditional branch must have two targets"); 1251 } 1252 } 1253 #endif 1254 } 1255 #endif 1256 1257 UnionFind::UnionFind( uint max ) : _cnt(max), _max(max), _indices(NEW_RESOURCE_ARRAY(uint,max)) { 1258 Copy::zero_to_bytes( _indices, sizeof(uint)*max ); 1259 } 1260 1261 void UnionFind::extend( uint from_idx, uint to_idx ) { 1262 _nesting.check(); 1263 if( from_idx >= _max ) { 1264 uint size = 16; 1265 while( size <= from_idx ) size <<=1; 1266 _indices = REALLOC_RESOURCE_ARRAY( uint, _indices, _max, size ); 1267 _max = size; 1268 } 1269 while( _cnt <= from_idx ) _indices[_cnt++] = 0; 1270 _indices[from_idx] = to_idx; 1271 } 1272 1273 void UnionFind::reset( uint max ) { 1274 // Force the Union-Find mapping to be at least this large 1275 extend(max,0); 1276 // Initialize to be the ID mapping. 1277 for( uint i=0; i<max; i++ ) map(i,i); 1278 } 1279 1280 // Straight out of Tarjan's union-find algorithm 1281 uint UnionFind::Find_compress( uint idx ) { 1282 uint cur = idx; 1283 uint next = lookup(cur); 1284 while( next != cur ) { // Scan chain of equivalences 1285 assert( next < cur, "always union smaller" ); 1286 cur = next; // until find a fixed-point 1287 next = lookup(cur); 1288 } 1289 // Core of union-find algorithm: update chain of 1290 // equivalences to be equal to the root. 1291 while( idx != next ) { 1292 uint tmp = lookup(idx); 1293 map(idx, next); 1294 idx = tmp; 1295 } 1296 return idx; 1297 } 1298 1299 // Like Find above, but no path compress, so bad asymptotic behavior 1300 uint UnionFind::Find_const( uint idx ) const { 1301 if( idx == 0 ) return idx; // Ignore the zero idx 1302 // Off the end? This can happen during debugging dumps 1303 // when data structures have not finished being updated. 1304 if( idx >= _max ) return idx; 1305 uint next = lookup(idx); 1306 while( next != idx ) { // Scan chain of equivalences 1307 idx = next; // until find a fixed-point 1308 next = lookup(idx); 1309 } 1310 return next; 1311 } 1312 1313 // union 2 sets together. 1314 void UnionFind::Union( uint idx1, uint idx2 ) { 1315 uint src = Find(idx1); 1316 uint dst = Find(idx2); 1317 assert( src, "" ); 1318 assert( dst, "" ); 1319 assert( src < _max, "oob" ); 1320 assert( dst < _max, "oob" ); 1321 assert( src < dst, "always union smaller" ); 1322 map(dst,src); 1323 } 1324 1325 #ifndef PRODUCT 1326 void Trace::dump( ) const { 1327 tty->print_cr("Trace (freq %f)", first_block()->_freq); 1328 for (Block *b = first_block(); b != NULL; b = next(b)) { 1329 tty->print(" B%d", b->_pre_order); 1330 if (b->head()->is_Loop()) { 1331 tty->print(" (L%d)", b->compute_loop_alignment()); 1332 } 1333 if (b->has_loop_alignment()) { 1334 tty->print(" (T%d)", b->code_alignment()); 1335 } 1336 } 1337 tty->cr(); 1338 } 1339 1340 void CFGEdge::dump( ) const { 1341 tty->print(" B%d --> B%d Freq: %f out:%3d%% in:%3d%% State: ", 1342 from()->_pre_order, to()->_pre_order, freq(), _from_pct, _to_pct); 1343 switch(state()) { 1344 case connected: 1345 tty->print("connected"); 1346 break; 1347 case open: 1348 tty->print("open"); 1349 break; 1350 case interior: 1351 tty->print("interior"); 1352 break; 1353 } 1354 if (infrequent()) { 1355 tty->print(" infrequent"); 1356 } 1357 tty->cr(); 1358 } 1359 #endif 1360 1361 // Comparison function for edges 1362 static int edge_order(CFGEdge **e0, CFGEdge **e1) { 1363 float freq0 = (*e0)->freq(); 1364 float freq1 = (*e1)->freq(); 1365 if (freq0 != freq1) { 1366 return freq0 > freq1 ? -1 : 1; 1367 } 1368 1369 int dist0 = (*e0)->to()->_rpo - (*e0)->from()->_rpo; 1370 int dist1 = (*e1)->to()->_rpo - (*e1)->from()->_rpo; 1371 1372 return dist1 - dist0; 1373 } 1374 1375 // Comparison function for edges 1376 extern "C" int trace_frequency_order(const void *p0, const void *p1) { 1377 Trace *tr0 = *(Trace **) p0; 1378 Trace *tr1 = *(Trace **) p1; 1379 Block *b0 = tr0->first_block(); 1380 Block *b1 = tr1->first_block(); 1381 1382 // The trace of connector blocks goes at the end; 1383 // we only expect one such trace 1384 if (b0->is_connector() != b1->is_connector()) { 1385 return b1->is_connector() ? -1 : 1; 1386 } 1387 1388 // Pull more frequently executed blocks to the beginning 1389 float freq0 = b0->_freq; 1390 float freq1 = b1->_freq; 1391 if (freq0 != freq1) { 1392 return freq0 > freq1 ? -1 : 1; 1393 } 1394 1395 int diff = tr0->first_block()->_rpo - tr1->first_block()->_rpo; 1396 1397 return diff; 1398 } 1399 1400 // Find edges of interest, i.e, those which can fall through. Presumes that 1401 // edges which don't fall through are of low frequency and can be generally 1402 // ignored. Initialize the list of traces. 1403 void PhaseBlockLayout::find_edges() { 1404 // Walk the blocks, creating edges and Traces 1405 uint i; 1406 Trace *tr = NULL; 1407 for (i = 0; i < _cfg.number_of_blocks(); i++) { 1408 Block* b = _cfg.get_block(i); 1409 tr = new Trace(b, next, prev); 1410 traces[tr->id()] = tr; 1411 1412 // All connector blocks should be at the end of the list 1413 if (b->is_connector()) break; 1414 1415 // If this block and the next one have a one-to-one successor 1416 // predecessor relationship, simply append the next block 1417 int nfallthru = b->num_fall_throughs(); 1418 while (nfallthru == 1 && 1419 b->succ_fall_through(0)) { 1420 Block *n = b->_succs[0]; 1421 1422 // Skip over single-entry connector blocks, we don't want to 1423 // add them to the trace. 1424 while (n->is_connector() && n->num_preds() == 1) { 1425 n = n->_succs[0]; 1426 } 1427 1428 // We see a merge point, so stop search for the next block 1429 if (n->num_preds() != 1) break; 1430 1431 i++; 1432 assert(n = _cfg.get_block(i), "expecting next block"); 1433 tr->append(n); 1434 uf->map(n->_pre_order, tr->id()); 1435 traces[n->_pre_order] = NULL; 1436 nfallthru = b->num_fall_throughs(); 1437 b = n; 1438 } 1439 1440 if (nfallthru > 0) { 1441 // Create a CFGEdge for each outgoing 1442 // edge that could be a fall-through. 1443 for (uint j = 0; j < b->_num_succs; j++ ) { 1444 if (b->succ_fall_through(j)) { 1445 Block *target = b->non_connector_successor(j); 1446 float freq = b->_freq * b->succ_prob(j); 1447 int from_pct = (int) ((100 * freq) / b->_freq); 1448 int to_pct = (int) ((100 * freq) / target->_freq); 1449 edges->append(new CFGEdge(b, target, freq, from_pct, to_pct)); 1450 } 1451 } 1452 } 1453 } 1454 1455 // Group connector blocks into one trace 1456 for (i++; i < _cfg.number_of_blocks(); i++) { 1457 Block *b = _cfg.get_block(i); 1458 assert(b->is_connector(), "connector blocks at the end"); 1459 tr->append(b); 1460 uf->map(b->_pre_order, tr->id()); 1461 traces[b->_pre_order] = NULL; 1462 } 1463 } 1464 1465 // Union two traces together in uf, and null out the trace in the list 1466 void PhaseBlockLayout::union_traces(Trace* updated_trace, Trace* old_trace) { 1467 uint old_id = old_trace->id(); 1468 uint updated_id = updated_trace->id(); 1469 1470 uint lo_id = updated_id; 1471 uint hi_id = old_id; 1472 1473 // If from is greater than to, swap values to meet 1474 // UnionFind guarantee. 1475 if (updated_id > old_id) { 1476 lo_id = old_id; 1477 hi_id = updated_id; 1478 1479 // Fix up the trace ids 1480 traces[lo_id] = traces[updated_id]; 1481 updated_trace->set_id(lo_id); 1482 } 1483 1484 // Union the lower with the higher and remove the pointer 1485 // to the higher. 1486 uf->Union(lo_id, hi_id); 1487 traces[hi_id] = NULL; 1488 } 1489 1490 // Append traces together via the most frequently executed edges 1491 void PhaseBlockLayout::grow_traces() { 1492 // Order the edges, and drive the growth of Traces via the most 1493 // frequently executed edges. 1494 edges->sort(edge_order); 1495 for (int i = 0; i < edges->length(); i++) { 1496 CFGEdge *e = edges->at(i); 1497 1498 if (e->state() != CFGEdge::open) continue; 1499 1500 Block *src_block = e->from(); 1501 Block *targ_block = e->to(); 1502 1503 // Don't grow traces along backedges? 1504 if (!BlockLayoutRotateLoops) { 1505 if (targ_block->_rpo <= src_block->_rpo) { 1506 targ_block->set_loop_alignment(targ_block); 1507 continue; 1508 } 1509 } 1510 1511 Trace *src_trace = trace(src_block); 1512 Trace *targ_trace = trace(targ_block); 1513 1514 // If the edge in question can join two traces at their ends, 1515 // append one trace to the other. 1516 if (src_trace->last_block() == src_block) { 1517 if (src_trace == targ_trace) { 1518 e->set_state(CFGEdge::interior); 1519 if (targ_trace->backedge(e)) { 1520 // Reset i to catch any newly eligible edge 1521 // (Or we could remember the first "open" edge, and reset there) 1522 i = 0; 1523 } 1524 } else if (targ_trace->first_block() == targ_block) { 1525 e->set_state(CFGEdge::connected); 1526 src_trace->append(targ_trace); 1527 union_traces(src_trace, targ_trace); 1528 } 1529 } 1530 } 1531 } 1532 1533 // Embed one trace into another, if the fork or join points are sufficiently 1534 // balanced. 1535 void PhaseBlockLayout::merge_traces(bool fall_thru_only) { 1536 // Walk the edge list a another time, looking at unprocessed edges. 1537 // Fold in diamonds 1538 for (int i = 0; i < edges->length(); i++) { 1539 CFGEdge *e = edges->at(i); 1540 1541 if (e->state() != CFGEdge::open) continue; 1542 if (fall_thru_only) { 1543 if (e->infrequent()) continue; 1544 } 1545 1546 Block *src_block = e->from(); 1547 Trace *src_trace = trace(src_block); 1548 bool src_at_tail = src_trace->last_block() == src_block; 1549 1550 Block *targ_block = e->to(); 1551 Trace *targ_trace = trace(targ_block); 1552 bool targ_at_start = targ_trace->first_block() == targ_block; 1553 1554 if (src_trace == targ_trace) { 1555 // This may be a loop, but we can't do much about it. 1556 e->set_state(CFGEdge::interior); 1557 continue; 1558 } 1559 1560 if (fall_thru_only) { 1561 // If the edge links the middle of two traces, we can't do anything. 1562 // Mark the edge and continue. 1563 if (!src_at_tail & !targ_at_start) { 1564 continue; 1565 } 1566 1567 // Don't grow traces along backedges? 1568 if (!BlockLayoutRotateLoops && (targ_block->_rpo <= src_block->_rpo)) { 1569 continue; 1570 } 1571 1572 // If both ends of the edge are available, why didn't we handle it earlier? 1573 assert(src_at_tail ^ targ_at_start, "Should have caught this edge earlier."); 1574 1575 if (targ_at_start) { 1576 // Insert the "targ" trace in the "src" trace if the insertion point 1577 // is a two way branch. 1578 // Better profitability check possible, but may not be worth it. 1579 // Someday, see if the this "fork" has an associated "join"; 1580 // then make a policy on merging this trace at the fork or join. 1581 // For example, other things being equal, it may be better to place this 1582 // trace at the join point if the "src" trace ends in a two-way, but 1583 // the insertion point is one-way. 1584 assert(src_block->num_fall_throughs() == 2, "unexpected diamond"); 1585 e->set_state(CFGEdge::connected); 1586 src_trace->insert_after(src_block, targ_trace); 1587 union_traces(src_trace, targ_trace); 1588 } else if (src_at_tail) { 1589 if (src_trace != trace(_cfg.get_root_block())) { 1590 e->set_state(CFGEdge::connected); 1591 targ_trace->insert_before(targ_block, src_trace); 1592 union_traces(targ_trace, src_trace); 1593 } 1594 } 1595 } else if (e->state() == CFGEdge::open) { 1596 // Append traces, even without a fall-thru connection. 1597 // But leave root entry at the beginning of the block list. 1598 if (targ_trace != trace(_cfg.get_root_block())) { 1599 e->set_state(CFGEdge::connected); 1600 src_trace->append(targ_trace); 1601 union_traces(src_trace, targ_trace); 1602 } 1603 } 1604 } 1605 } 1606 1607 // Order the sequence of the traces in some desirable way, and fixup the 1608 // jumps at the end of each block. 1609 void PhaseBlockLayout::reorder_traces(int count) { 1610 ResourceArea *area = Thread::current()->resource_area(); 1611 Trace ** new_traces = NEW_ARENA_ARRAY(area, Trace *, count); 1612 Block_List worklist; 1613 int new_count = 0; 1614 1615 // Compact the traces. 1616 for (int i = 0; i < count; i++) { 1617 Trace *tr = traces[i]; 1618 if (tr != NULL) { 1619 new_traces[new_count++] = tr; 1620 } 1621 } 1622 1623 // The entry block should be first on the new trace list. 1624 Trace *tr = trace(_cfg.get_root_block()); 1625 assert(tr == new_traces[0], "entry trace misplaced"); 1626 1627 // Sort the new trace list by frequency 1628 qsort(new_traces + 1, new_count - 1, sizeof(new_traces[0]), trace_frequency_order); 1629 1630 // Patch up the successor blocks 1631 _cfg.clear_blocks(); 1632 for (int i = 0; i < new_count; i++) { 1633 Trace *tr = new_traces[i]; 1634 if (tr != NULL) { 1635 tr->fixup_blocks(_cfg); 1636 } 1637 } 1638 } 1639 1640 // Order basic blocks based on frequency 1641 PhaseBlockLayout::PhaseBlockLayout(PhaseCFG &cfg) 1642 : Phase(BlockLayout) 1643 , _cfg(cfg) { 1644 ResourceMark rm; 1645 ResourceArea *area = Thread::current()->resource_area(); 1646 1647 // List of traces 1648 int size = _cfg.number_of_blocks() + 1; 1649 traces = NEW_ARENA_ARRAY(area, Trace *, size); 1650 memset(traces, 0, size*sizeof(Trace*)); 1651 next = NEW_ARENA_ARRAY(area, Block *, size); 1652 memset(next, 0, size*sizeof(Block *)); 1653 prev = NEW_ARENA_ARRAY(area, Block *, size); 1654 memset(prev , 0, size*sizeof(Block *)); 1655 1656 // List of edges 1657 edges = new GrowableArray<CFGEdge*>; 1658 1659 // Mapping block index --> block_trace 1660 uf = new UnionFind(size); 1661 uf->reset(size); 1662 1663 // Find edges and create traces. 1664 find_edges(); 1665 1666 // Grow traces at their ends via most frequent edges. 1667 grow_traces(); 1668 1669 // Merge one trace into another, but only at fall-through points. 1670 // This may make diamonds and other related shapes in a trace. 1671 merge_traces(true); 1672 1673 // Run merge again, allowing two traces to be catenated, even if 1674 // one does not fall through into the other. This appends loosely 1675 // related traces to be near each other. 1676 merge_traces(false); 1677 1678 // Re-order all the remaining traces by frequency 1679 reorder_traces(size); 1680 1681 assert(_cfg.number_of_blocks() >= (uint) (size - 1), "number of blocks can not shrink"); 1682 } 1683 1684 1685 // Edge e completes a loop in a trace. If the target block is head of the 1686 // loop, rotate the loop block so that the loop ends in a conditional branch. 1687 bool Trace::backedge(CFGEdge *e) { 1688 bool loop_rotated = false; 1689 Block *src_block = e->from(); 1690 Block *targ_block = e->to(); 1691 1692 assert(last_block() == src_block, "loop discovery at back branch"); 1693 if (first_block() == targ_block) { 1694 if (BlockLayoutRotateLoops && last_block()->num_fall_throughs() < 2) { 1695 // Find the last block in the trace that has a conditional 1696 // branch. 1697 Block *b; 1698 for (b = last_block(); b != NULL; b = prev(b)) { 1699 if (b->num_fall_throughs() == 2) { 1700 break; 1701 } 1702 } 1703 1704 if (b != last_block() && b != NULL) { 1705 loop_rotated = true; 1706 1707 // Rotate the loop by doing two-part linked-list surgery. 1708 append(first_block()); 1709 break_loop_after(b); 1710 } 1711 } 1712 1713 // Backbranch to the top of a trace 1714 // Scroll forward through the trace from the targ_block. If we find 1715 // a loop head before another loop top, use the the loop head alignment. 1716 for (Block *b = targ_block; b != NULL; b = next(b)) { 1717 if (b->has_loop_alignment()) { 1718 break; 1719 } 1720 if (b->head()->is_Loop()) { 1721 targ_block = b; 1722 break; 1723 } 1724 } 1725 1726 first_block()->set_loop_alignment(targ_block); 1727 1728 } else { 1729 // Backbranch into the middle of a trace 1730 targ_block->set_loop_alignment(targ_block); 1731 } 1732 1733 return loop_rotated; 1734 } 1735 1736 // push blocks onto the CFG list 1737 // ensure that blocks have the correct two-way branch sense 1738 void Trace::fixup_blocks(PhaseCFG &cfg) { 1739 Block *last = last_block(); 1740 for (Block *b = first_block(); b != NULL; b = next(b)) { 1741 cfg.add_block(b); 1742 if (!b->is_connector()) { 1743 int nfallthru = b->num_fall_throughs(); 1744 if (b != last) { 1745 if (nfallthru == 2) { 1746 // Ensure that the sense of the branch is correct 1747 Block *bnext = next(b); 1748 Block *bs0 = b->non_connector_successor(0); 1749 1750 MachNode *iff = b->get_node(b->number_of_nodes() - 3)->as_Mach(); 1751 ProjNode *proj0 = b->get_node(b->number_of_nodes() - 2)->as_Proj(); 1752 ProjNode *proj1 = b->get_node(b->number_of_nodes() - 1)->as_Proj(); 1753 1754 if (bnext == bs0) { 1755 // Fall-thru case in succs[0], should be in succs[1] 1756 1757 // Flip targets in _succs map 1758 Block *tbs0 = b->_succs[0]; 1759 Block *tbs1 = b->_succs[1]; 1760 b->_succs.map( 0, tbs1 ); 1761 b->_succs.map( 1, tbs0 ); 1762 1763 // Flip projections to match targets 1764 b->map_node(proj1, b->number_of_nodes() - 2); 1765 b->map_node(proj0, b->number_of_nodes() - 1); 1766 } 1767 } 1768 } 1769 } 1770 } 1771 }