src/share/vm/opto/block.cpp

Tue, 03 Aug 2010 15:55:03 -0700

author
kvn
date
Tue, 03 Aug 2010 15:55:03 -0700
changeset 2040
0e35fa8ebccd
parent 1907
c18cbe5936b8
child 2314
f95d63e2154a
permissions
-rw-r--r--

6973963: SEGV in ciBlock::start_bci() with EA
Summary: Added more checks into ResourceObj and growableArray to verify correctness of allocation type.
Reviewed-by: never, coleenp, dholmes

duke@435 1 /*
trims@1907 2 * Copyright (c) 1997, 2009, Oracle and/or its affiliates. All rights reserved.
duke@435 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@435 4 *
duke@435 5 * This code is free software; you can redistribute it and/or modify it
duke@435 6 * under the terms of the GNU General Public License version 2 only, as
duke@435 7 * published by the Free Software Foundation.
duke@435 8 *
duke@435 9 * This code is distributed in the hope that it will be useful, but WITHOUT
duke@435 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@435 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@435 12 * version 2 for more details (a copy is included in the LICENSE file that
duke@435 13 * accompanied this code).
duke@435 14 *
duke@435 15 * You should have received a copy of the GNU General Public License version
duke@435 16 * 2 along with this work; if not, write to the Free Software Foundation,
duke@435 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@435 18 *
trims@1907 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1907 20 * or visit www.oracle.com if you need additional information or have any
trims@1907 21 * questions.
duke@435 22 *
duke@435 23 */
duke@435 24
duke@435 25 // Optimization - Graph Style
duke@435 26
duke@435 27 #include "incls/_precompiled.incl"
duke@435 28 #include "incls/_block.cpp.incl"
duke@435 29
duke@435 30
duke@435 31 //-----------------------------------------------------------------------------
duke@435 32 void Block_Array::grow( uint i ) {
duke@435 33 assert(i >= Max(), "must be an overflow");
duke@435 34 debug_only(_limit = i+1);
duke@435 35 if( i < _size ) return;
duke@435 36 if( !_size ) {
duke@435 37 _size = 1;
duke@435 38 _blocks = (Block**)_arena->Amalloc( _size * sizeof(Block*) );
duke@435 39 _blocks[0] = NULL;
duke@435 40 }
duke@435 41 uint old = _size;
duke@435 42 while( i >= _size ) _size <<= 1; // Double to fit
duke@435 43 _blocks = (Block**)_arena->Arealloc( _blocks, old*sizeof(Block*),_size*sizeof(Block*));
duke@435 44 Copy::zero_to_bytes( &_blocks[old], (_size-old)*sizeof(Block*) );
duke@435 45 }
duke@435 46
duke@435 47 //=============================================================================
duke@435 48 void Block_List::remove(uint i) {
duke@435 49 assert(i < _cnt, "index out of bounds");
duke@435 50 Copy::conjoint_words_to_lower((HeapWord*)&_blocks[i+1], (HeapWord*)&_blocks[i], ((_cnt-i-1)*sizeof(Block*)));
duke@435 51 pop(); // shrink list by one block
duke@435 52 }
duke@435 53
duke@435 54 void Block_List::insert(uint i, Block *b) {
duke@435 55 push(b); // grow list by one block
duke@435 56 Copy::conjoint_words_to_higher((HeapWord*)&_blocks[i], (HeapWord*)&_blocks[i+1], ((_cnt-i-1)*sizeof(Block*)));
duke@435 57 _blocks[i] = b;
duke@435 58 }
duke@435 59
rasbold@853 60 #ifndef PRODUCT
rasbold@853 61 void Block_List::print() {
rasbold@853 62 for (uint i=0; i < size(); i++) {
rasbold@853 63 tty->print("B%d ", _blocks[i]->_pre_order);
rasbold@853 64 }
rasbold@853 65 tty->print("size = %d\n", size());
rasbold@853 66 }
rasbold@853 67 #endif
duke@435 68
duke@435 69 //=============================================================================
duke@435 70
duke@435 71 uint Block::code_alignment() {
duke@435 72 // Check for Root block
duke@435 73 if( _pre_order == 0 ) return CodeEntryAlignment;
duke@435 74 // Check for Start block
duke@435 75 if( _pre_order == 1 ) return InteriorEntryAlignment;
duke@435 76 // Check for loop alignment
rasbold@853 77 if (has_loop_alignment()) return loop_alignment();
rasbold@853 78
rasbold@853 79 return 1; // no particular alignment
rasbold@853 80 }
rasbold@853 81
rasbold@853 82 uint Block::compute_loop_alignment() {
duke@435 83 Node *h = head();
duke@435 84 if( h->is_Loop() && h->as_Loop()->is_inner_loop() ) {
duke@435 85 // Pre- and post-loops have low trip count so do not bother with
duke@435 86 // NOPs for align loop head. The constants are hidden from tuning
duke@435 87 // but only because my "divide by 4" heuristic surely gets nearly
duke@435 88 // all possible gain (a "do not align at all" heuristic has a
duke@435 89 // chance of getting a really tiny gain).
duke@435 90 if( h->is_CountedLoop() && (h->as_CountedLoop()->is_pre_loop() ||
duke@435 91 h->as_CountedLoop()->is_post_loop()) )
duke@435 92 return (OptoLoopAlignment > 4) ? (OptoLoopAlignment>>2) : 1;
duke@435 93 // Loops with low backedge frequency should not be aligned.
duke@435 94 Node *n = h->in(LoopNode::LoopBackControl)->in(0);
duke@435 95 if( n->is_MachIf() && n->as_MachIf()->_prob < 0.01 ) {
duke@435 96 return 1; // Loop does not loop, more often than not!
duke@435 97 }
duke@435 98 return OptoLoopAlignment; // Otherwise align loop head
duke@435 99 }
rasbold@853 100
duke@435 101 return 1; // no particular alignment
duke@435 102 }
duke@435 103
duke@435 104 //-----------------------------------------------------------------------------
duke@435 105 // Compute the size of first 'inst_cnt' instructions in this block.
duke@435 106 // Return the number of instructions left to compute if the block has
rasbold@853 107 // less then 'inst_cnt' instructions. Stop, and return 0 if sum_size
rasbold@853 108 // exceeds OptoLoopAlignment.
duke@435 109 uint Block::compute_first_inst_size(uint& sum_size, uint inst_cnt,
duke@435 110 PhaseRegAlloc* ra) {
duke@435 111 uint last_inst = _nodes.size();
duke@435 112 for( uint j = 0; j < last_inst && inst_cnt > 0; j++ ) {
duke@435 113 uint inst_size = _nodes[j]->size(ra);
duke@435 114 if( inst_size > 0 ) {
duke@435 115 inst_cnt--;
duke@435 116 uint sz = sum_size + inst_size;
duke@435 117 if( sz <= (uint)OptoLoopAlignment ) {
duke@435 118 // Compute size of instructions which fit into fetch buffer only
duke@435 119 // since all inst_cnt instructions will not fit even if we align them.
duke@435 120 sum_size = sz;
duke@435 121 } else {
duke@435 122 return 0;
duke@435 123 }
duke@435 124 }
duke@435 125 }
duke@435 126 return inst_cnt;
duke@435 127 }
duke@435 128
duke@435 129 //-----------------------------------------------------------------------------
duke@435 130 uint Block::find_node( const Node *n ) const {
duke@435 131 for( uint i = 0; i < _nodes.size(); i++ ) {
duke@435 132 if( _nodes[i] == n )
duke@435 133 return i;
duke@435 134 }
duke@435 135 ShouldNotReachHere();
duke@435 136 return 0;
duke@435 137 }
duke@435 138
duke@435 139 // Find and remove n from block list
duke@435 140 void Block::find_remove( const Node *n ) {
duke@435 141 _nodes.remove(find_node(n));
duke@435 142 }
duke@435 143
duke@435 144 //------------------------------is_Empty---------------------------------------
duke@435 145 // Return empty status of a block. Empty blocks contain only the head, other
duke@435 146 // ideal nodes, and an optional trailing goto.
duke@435 147 int Block::is_Empty() const {
duke@435 148
duke@435 149 // Root or start block is not considered empty
duke@435 150 if (head()->is_Root() || head()->is_Start()) {
duke@435 151 return not_empty;
duke@435 152 }
duke@435 153
duke@435 154 int success_result = completely_empty;
duke@435 155 int end_idx = _nodes.size()-1;
duke@435 156
duke@435 157 // Check for ending goto
duke@435 158 if ((end_idx > 0) && (_nodes[end_idx]->is_Goto())) {
duke@435 159 success_result = empty_with_goto;
duke@435 160 end_idx--;
duke@435 161 }
duke@435 162
duke@435 163 // Unreachable blocks are considered empty
duke@435 164 if (num_preds() <= 1) {
duke@435 165 return success_result;
duke@435 166 }
duke@435 167
duke@435 168 // Ideal nodes are allowable in empty blocks: skip them Only MachNodes
duke@435 169 // turn directly into code, because only MachNodes have non-trivial
duke@435 170 // emit() functions.
duke@435 171 while ((end_idx > 0) && !_nodes[end_idx]->is_Mach()) {
duke@435 172 end_idx--;
duke@435 173 }
duke@435 174
duke@435 175 // No room for any interesting instructions?
duke@435 176 if (end_idx == 0) {
duke@435 177 return success_result;
duke@435 178 }
duke@435 179
duke@435 180 return not_empty;
duke@435 181 }
duke@435 182
duke@435 183 //------------------------------has_uncommon_code------------------------------
twisti@1040 184 // Return true if the block's code implies that it is likely to be
duke@435 185 // executed infrequently. Check to see if the block ends in a Halt or
duke@435 186 // a low probability call.
duke@435 187 bool Block::has_uncommon_code() const {
duke@435 188 Node* en = end();
duke@435 189
duke@435 190 if (en->is_Goto())
duke@435 191 en = en->in(0);
duke@435 192 if (en->is_Catch())
duke@435 193 en = en->in(0);
duke@435 194 if (en->is_Proj() && en->in(0)->is_MachCall()) {
duke@435 195 MachCallNode* call = en->in(0)->as_MachCall();
duke@435 196 if (call->cnt() != COUNT_UNKNOWN && call->cnt() <= PROB_UNLIKELY_MAG(4)) {
duke@435 197 // This is true for slow-path stubs like new_{instance,array},
duke@435 198 // slow_arraycopy, complete_monitor_locking, uncommon_trap.
duke@435 199 // The magic number corresponds to the probability of an uncommon_trap,
duke@435 200 // even though it is a count not a probability.
duke@435 201 return true;
duke@435 202 }
duke@435 203 }
duke@435 204
duke@435 205 int op = en->is_Mach() ? en->as_Mach()->ideal_Opcode() : en->Opcode();
duke@435 206 return op == Op_Halt;
duke@435 207 }
duke@435 208
duke@435 209 //------------------------------is_uncommon------------------------------------
duke@435 210 // True if block is low enough frequency or guarded by a test which
duke@435 211 // mostly does not go here.
duke@435 212 bool Block::is_uncommon( Block_Array &bbs ) const {
duke@435 213 // Initial blocks must never be moved, so are never uncommon.
duke@435 214 if (head()->is_Root() || head()->is_Start()) return false;
duke@435 215
duke@435 216 // Check for way-low freq
duke@435 217 if( _freq < BLOCK_FREQUENCY(0.00001f) ) return true;
duke@435 218
duke@435 219 // Look for code shape indicating uncommon_trap or slow path
duke@435 220 if (has_uncommon_code()) return true;
duke@435 221
duke@435 222 const float epsilon = 0.05f;
duke@435 223 const float guard_factor = PROB_UNLIKELY_MAG(4) / (1.f - epsilon);
duke@435 224 uint uncommon_preds = 0;
duke@435 225 uint freq_preds = 0;
duke@435 226 uint uncommon_for_freq_preds = 0;
duke@435 227
duke@435 228 for( uint i=1; i<num_preds(); i++ ) {
duke@435 229 Block* guard = bbs[pred(i)->_idx];
duke@435 230 // Check to see if this block follows its guard 1 time out of 10000
duke@435 231 // or less.
duke@435 232 //
duke@435 233 // See list of magnitude-4 unlikely probabilities in cfgnode.hpp which
duke@435 234 // we intend to be "uncommon", such as slow-path TLE allocation,
duke@435 235 // predicted call failure, and uncommon trap triggers.
duke@435 236 //
duke@435 237 // Use an epsilon value of 5% to allow for variability in frequency
duke@435 238 // predictions and floating point calculations. The net effect is
duke@435 239 // that guard_factor is set to 9500.
duke@435 240 //
duke@435 241 // Ignore low-frequency blocks.
duke@435 242 // The next check is (guard->_freq < 1.e-5 * 9500.).
duke@435 243 if(guard->_freq*BLOCK_FREQUENCY(guard_factor) < BLOCK_FREQUENCY(0.00001f)) {
duke@435 244 uncommon_preds++;
duke@435 245 } else {
duke@435 246 freq_preds++;
duke@435 247 if( _freq < guard->_freq * guard_factor ) {
duke@435 248 uncommon_for_freq_preds++;
duke@435 249 }
duke@435 250 }
duke@435 251 }
duke@435 252 if( num_preds() > 1 &&
duke@435 253 // The block is uncommon if all preds are uncommon or
duke@435 254 (uncommon_preds == (num_preds()-1) ||
duke@435 255 // it is uncommon for all frequent preds.
duke@435 256 uncommon_for_freq_preds == freq_preds) ) {
duke@435 257 return true;
duke@435 258 }
duke@435 259 return false;
duke@435 260 }
duke@435 261
duke@435 262 //------------------------------dump-------------------------------------------
duke@435 263 #ifndef PRODUCT
duke@435 264 void Block::dump_bidx(const Block* orig) const {
duke@435 265 if (_pre_order) tty->print("B%d",_pre_order);
duke@435 266 else tty->print("N%d", head()->_idx);
duke@435 267
duke@435 268 if (Verbose && orig != this) {
duke@435 269 // Dump the original block's idx
duke@435 270 tty->print(" (");
duke@435 271 orig->dump_bidx(orig);
duke@435 272 tty->print(")");
duke@435 273 }
duke@435 274 }
duke@435 275
duke@435 276 void Block::dump_pred(const Block_Array *bbs, Block* orig) const {
duke@435 277 if (is_connector()) {
duke@435 278 for (uint i=1; i<num_preds(); i++) {
duke@435 279 Block *p = ((*bbs)[pred(i)->_idx]);
duke@435 280 p->dump_pred(bbs, orig);
duke@435 281 }
duke@435 282 } else {
duke@435 283 dump_bidx(orig);
duke@435 284 tty->print(" ");
duke@435 285 }
duke@435 286 }
duke@435 287
duke@435 288 void Block::dump_head( const Block_Array *bbs ) const {
duke@435 289 // Print the basic block
duke@435 290 dump_bidx(this);
duke@435 291 tty->print(": #\t");
duke@435 292
duke@435 293 // Print the incoming CFG edges and the outgoing CFG edges
duke@435 294 for( uint i=0; i<_num_succs; i++ ) {
duke@435 295 non_connector_successor(i)->dump_bidx(_succs[i]);
duke@435 296 tty->print(" ");
duke@435 297 }
duke@435 298 tty->print("<- ");
duke@435 299 if( head()->is_block_start() ) {
duke@435 300 for (uint i=1; i<num_preds(); i++) {
duke@435 301 Node *s = pred(i);
duke@435 302 if (bbs) {
duke@435 303 Block *p = (*bbs)[s->_idx];
duke@435 304 p->dump_pred(bbs, p);
duke@435 305 } else {
duke@435 306 while (!s->is_block_start())
duke@435 307 s = s->in(0);
duke@435 308 tty->print("N%d ", s->_idx );
duke@435 309 }
duke@435 310 }
duke@435 311 } else
duke@435 312 tty->print("BLOCK HEAD IS JUNK ");
duke@435 313
duke@435 314 // Print loop, if any
duke@435 315 const Block *bhead = this; // Head of self-loop
duke@435 316 Node *bh = bhead->head();
duke@435 317 if( bbs && bh->is_Loop() && !head()->is_Root() ) {
duke@435 318 LoopNode *loop = bh->as_Loop();
duke@435 319 const Block *bx = (*bbs)[loop->in(LoopNode::LoopBackControl)->_idx];
duke@435 320 while (bx->is_connector()) {
duke@435 321 bx = (*bbs)[bx->pred(1)->_idx];
duke@435 322 }
duke@435 323 tty->print("\tLoop: B%d-B%d ", bhead->_pre_order, bx->_pre_order);
duke@435 324 // Dump any loop-specific bits, especially for CountedLoops.
duke@435 325 loop->dump_spec(tty);
rasbold@853 326 } else if (has_loop_alignment()) {
rasbold@853 327 tty->print(" top-of-loop");
duke@435 328 }
duke@435 329 tty->print(" Freq: %g",_freq);
duke@435 330 if( Verbose || WizardMode ) {
duke@435 331 tty->print(" IDom: %d/#%d", _idom ? _idom->_pre_order : 0, _dom_depth);
duke@435 332 tty->print(" RegPressure: %d",_reg_pressure);
duke@435 333 tty->print(" IHRP Index: %d",_ihrp_index);
duke@435 334 tty->print(" FRegPressure: %d",_freg_pressure);
duke@435 335 tty->print(" FHRP Index: %d",_fhrp_index);
duke@435 336 }
duke@435 337 tty->print_cr("");
duke@435 338 }
duke@435 339
duke@435 340 void Block::dump() const { dump(0); }
duke@435 341
duke@435 342 void Block::dump( const Block_Array *bbs ) const {
duke@435 343 dump_head(bbs);
duke@435 344 uint cnt = _nodes.size();
duke@435 345 for( uint i=0; i<cnt; i++ )
duke@435 346 _nodes[i]->dump();
duke@435 347 tty->print("\n");
duke@435 348 }
duke@435 349 #endif
duke@435 350
duke@435 351 //=============================================================================
duke@435 352 //------------------------------PhaseCFG---------------------------------------
duke@435 353 PhaseCFG::PhaseCFG( Arena *a, RootNode *r, Matcher &m ) :
duke@435 354 Phase(CFG),
duke@435 355 _bbs(a),
kvn@2040 356 _root(r),
kvn@2040 357 _node_latency(NULL)
duke@435 358 #ifndef PRODUCT
duke@435 359 , _trace_opto_pipelining(TraceOptoPipelining || C->method_has_option("TraceOptoPipelining"))
duke@435 360 #endif
kvn@1268 361 #ifdef ASSERT
kvn@1268 362 , _raw_oops(a)
kvn@1268 363 #endif
duke@435 364 {
duke@435 365 ResourceMark rm;
duke@435 366 // I'll need a few machine-specific GotoNodes. Make an Ideal GotoNode,
duke@435 367 // then Match it into a machine-specific Node. Then clone the machine
duke@435 368 // Node on demand.
duke@435 369 Node *x = new (C, 1) GotoNode(NULL);
duke@435 370 x->init_req(0, x);
duke@435 371 _goto = m.match_tree(x);
duke@435 372 assert(_goto != NULL, "");
duke@435 373 _goto->set_req(0,_goto);
duke@435 374
duke@435 375 // Build the CFG in Reverse Post Order
duke@435 376 _num_blocks = build_cfg();
duke@435 377 _broot = _bbs[_root->_idx];
duke@435 378 }
duke@435 379
duke@435 380 //------------------------------build_cfg--------------------------------------
duke@435 381 // Build a proper looking CFG. Make every block begin with either a StartNode
duke@435 382 // or a RegionNode. Make every block end with either a Goto, If or Return.
duke@435 383 // The RootNode both starts and ends it's own block. Do this with a recursive
duke@435 384 // backwards walk over the control edges.
duke@435 385 uint PhaseCFG::build_cfg() {
duke@435 386 Arena *a = Thread::current()->resource_area();
duke@435 387 VectorSet visited(a);
duke@435 388
duke@435 389 // Allocate stack with enough space to avoid frequent realloc
duke@435 390 Node_Stack nstack(a, C->unique() >> 1);
duke@435 391 nstack.push(_root, 0);
duke@435 392 uint sum = 0; // Counter for blocks
duke@435 393
duke@435 394 while (nstack.is_nonempty()) {
duke@435 395 // node and in's index from stack's top
duke@435 396 // 'np' is _root (see above) or RegionNode, StartNode: we push on stack
duke@435 397 // only nodes which point to the start of basic block (see below).
duke@435 398 Node *np = nstack.node();
duke@435 399 // idx > 0, except for the first node (_root) pushed on stack
duke@435 400 // at the beginning when idx == 0.
duke@435 401 // We will use the condition (idx == 0) later to end the build.
duke@435 402 uint idx = nstack.index();
duke@435 403 Node *proj = np->in(idx);
duke@435 404 const Node *x = proj->is_block_proj();
duke@435 405 // Does the block end with a proper block-ending Node? One of Return,
duke@435 406 // If or Goto? (This check should be done for visited nodes also).
duke@435 407 if (x == NULL) { // Does not end right...
duke@435 408 Node *g = _goto->clone(); // Force it to end in a Goto
duke@435 409 g->set_req(0, proj);
duke@435 410 np->set_req(idx, g);
duke@435 411 x = proj = g;
duke@435 412 }
duke@435 413 if (!visited.test_set(x->_idx)) { // Visit this block once
duke@435 414 // Skip any control-pinned middle'in stuff
duke@435 415 Node *p = proj;
duke@435 416 do {
duke@435 417 proj = p; // Update pointer to last Control
duke@435 418 p = p->in(0); // Move control forward
duke@435 419 } while( !p->is_block_proj() &&
duke@435 420 !p->is_block_start() );
duke@435 421 // Make the block begin with one of Region or StartNode.
duke@435 422 if( !p->is_block_start() ) {
duke@435 423 RegionNode *r = new (C, 2) RegionNode( 2 );
duke@435 424 r->init_req(1, p); // Insert RegionNode in the way
duke@435 425 proj->set_req(0, r); // Insert RegionNode in the way
duke@435 426 p = r;
duke@435 427 }
duke@435 428 // 'p' now points to the start of this basic block
duke@435 429
duke@435 430 // Put self in array of basic blocks
duke@435 431 Block *bb = new (_bbs._arena) Block(_bbs._arena,p);
duke@435 432 _bbs.map(p->_idx,bb);
duke@435 433 _bbs.map(x->_idx,bb);
duke@435 434 if( x != p ) // Only for root is x == p
duke@435 435 bb->_nodes.push((Node*)x);
duke@435 436
duke@435 437 // Now handle predecessors
duke@435 438 ++sum; // Count 1 for self block
duke@435 439 uint cnt = bb->num_preds();
duke@435 440 for (int i = (cnt - 1); i > 0; i-- ) { // For all predecessors
duke@435 441 Node *prevproj = p->in(i); // Get prior input
duke@435 442 assert( !prevproj->is_Con(), "dead input not removed" );
duke@435 443 // Check to see if p->in(i) is a "control-dependent" CFG edge -
duke@435 444 // i.e., it splits at the source (via an IF or SWITCH) and merges
duke@435 445 // at the destination (via a many-input Region).
duke@435 446 // This breaks critical edges. The RegionNode to start the block
duke@435 447 // will be added when <p,i> is pulled off the node stack
duke@435 448 if ( cnt > 2 ) { // Merging many things?
duke@435 449 assert( prevproj== bb->pred(i),"");
duke@435 450 if(prevproj->is_block_proj() != prevproj) { // Control-dependent edge?
duke@435 451 // Force a block on the control-dependent edge
duke@435 452 Node *g = _goto->clone(); // Force it to end in a Goto
duke@435 453 g->set_req(0,prevproj);
duke@435 454 p->set_req(i,g);
duke@435 455 }
duke@435 456 }
duke@435 457 nstack.push(p, i); // 'p' is RegionNode or StartNode
duke@435 458 }
duke@435 459 } else { // Post-processing visited nodes
duke@435 460 nstack.pop(); // remove node from stack
duke@435 461 // Check if it the fist node pushed on stack at the beginning.
duke@435 462 if (idx == 0) break; // end of the build
duke@435 463 // Find predecessor basic block
duke@435 464 Block *pb = _bbs[x->_idx];
duke@435 465 // Insert into nodes array, if not already there
duke@435 466 if( !_bbs.lookup(proj->_idx) ) {
duke@435 467 assert( x != proj, "" );
duke@435 468 // Map basic block of projection
duke@435 469 _bbs.map(proj->_idx,pb);
duke@435 470 pb->_nodes.push(proj);
duke@435 471 }
duke@435 472 // Insert self as a child of my predecessor block
duke@435 473 pb->_succs.map(pb->_num_succs++, _bbs[np->_idx]);
duke@435 474 assert( pb->_nodes[ pb->_nodes.size() - pb->_num_succs ]->is_block_proj(),
duke@435 475 "too many control users, not a CFG?" );
duke@435 476 }
duke@435 477 }
duke@435 478 // Return number of basic blocks for all children and self
duke@435 479 return sum;
duke@435 480 }
duke@435 481
duke@435 482 //------------------------------insert_goto_at---------------------------------
duke@435 483 // Inserts a goto & corresponding basic block between
duke@435 484 // block[block_no] and its succ_no'th successor block
duke@435 485 void PhaseCFG::insert_goto_at(uint block_no, uint succ_no) {
duke@435 486 // get block with block_no
duke@435 487 assert(block_no < _num_blocks, "illegal block number");
duke@435 488 Block* in = _blocks[block_no];
duke@435 489 // get successor block succ_no
duke@435 490 assert(succ_no < in->_num_succs, "illegal successor number");
duke@435 491 Block* out = in->_succs[succ_no];
rasbold@743 492 // Compute frequency of the new block. Do this before inserting
rasbold@743 493 // new block in case succ_prob() needs to infer the probability from
rasbold@743 494 // surrounding blocks.
rasbold@743 495 float freq = in->_freq * in->succ_prob(succ_no);
duke@435 496 // get ProjNode corresponding to the succ_no'th successor of the in block
duke@435 497 ProjNode* proj = in->_nodes[in->_nodes.size() - in->_num_succs + succ_no]->as_Proj();
duke@435 498 // create region for basic block
duke@435 499 RegionNode* region = new (C, 2) RegionNode(2);
duke@435 500 region->init_req(1, proj);
duke@435 501 // setup corresponding basic block
duke@435 502 Block* block = new (_bbs._arena) Block(_bbs._arena, region);
duke@435 503 _bbs.map(region->_idx, block);
duke@435 504 C->regalloc()->set_bad(region->_idx);
duke@435 505 // add a goto node
duke@435 506 Node* gto = _goto->clone(); // get a new goto node
duke@435 507 gto->set_req(0, region);
duke@435 508 // add it to the basic block
duke@435 509 block->_nodes.push(gto);
duke@435 510 _bbs.map(gto->_idx, block);
duke@435 511 C->regalloc()->set_bad(gto->_idx);
duke@435 512 // hook up successor block
duke@435 513 block->_succs.map(block->_num_succs++, out);
duke@435 514 // remap successor's predecessors if necessary
duke@435 515 for (uint i = 1; i < out->num_preds(); i++) {
duke@435 516 if (out->pred(i) == proj) out->head()->set_req(i, gto);
duke@435 517 }
duke@435 518 // remap predecessor's successor to new block
duke@435 519 in->_succs.map(succ_no, block);
rasbold@743 520 // Set the frequency of the new block
rasbold@743 521 block->_freq = freq;
duke@435 522 // add new basic block to basic block list
duke@435 523 _blocks.insert(block_no + 1, block);
duke@435 524 _num_blocks++;
duke@435 525 }
duke@435 526
duke@435 527 //------------------------------no_flip_branch---------------------------------
duke@435 528 // Does this block end in a multiway branch that cannot have the default case
duke@435 529 // flipped for another case?
duke@435 530 static bool no_flip_branch( Block *b ) {
duke@435 531 int branch_idx = b->_nodes.size() - b->_num_succs-1;
duke@435 532 if( branch_idx < 1 ) return false;
duke@435 533 Node *bra = b->_nodes[branch_idx];
rasbold@853 534 if( bra->is_Catch() )
rasbold@853 535 return true;
duke@435 536 if( bra->is_Mach() ) {
rasbold@853 537 if( bra->is_MachNullCheck() )
rasbold@853 538 return true;
duke@435 539 int iop = bra->as_Mach()->ideal_Opcode();
duke@435 540 if( iop == Op_FastLock || iop == Op_FastUnlock )
duke@435 541 return true;
duke@435 542 }
duke@435 543 return false;
duke@435 544 }
duke@435 545
duke@435 546 //------------------------------convert_NeverBranch_to_Goto--------------------
duke@435 547 // Check for NeverBranch at block end. This needs to become a GOTO to the
duke@435 548 // true target. NeverBranch are treated as a conditional branch that always
duke@435 549 // goes the same direction for most of the optimizer and are used to give a
duke@435 550 // fake exit path to infinite loops. At this late stage they need to turn
duke@435 551 // into Goto's so that when you enter the infinite loop you indeed hang.
duke@435 552 void PhaseCFG::convert_NeverBranch_to_Goto(Block *b) {
duke@435 553 // Find true target
duke@435 554 int end_idx = b->end_idx();
duke@435 555 int idx = b->_nodes[end_idx+1]->as_Proj()->_con;
duke@435 556 Block *succ = b->_succs[idx];
duke@435 557 Node* gto = _goto->clone(); // get a new goto node
duke@435 558 gto->set_req(0, b->head());
duke@435 559 Node *bp = b->_nodes[end_idx];
duke@435 560 b->_nodes.map(end_idx,gto); // Slam over NeverBranch
duke@435 561 _bbs.map(gto->_idx, b);
duke@435 562 C->regalloc()->set_bad(gto->_idx);
duke@435 563 b->_nodes.pop(); // Yank projections
duke@435 564 b->_nodes.pop(); // Yank projections
duke@435 565 b->_succs.map(0,succ); // Map only successor
duke@435 566 b->_num_succs = 1;
duke@435 567 // remap successor's predecessors if necessary
duke@435 568 uint j;
duke@435 569 for( j = 1; j < succ->num_preds(); j++)
duke@435 570 if( succ->pred(j)->in(0) == bp )
duke@435 571 succ->head()->set_req(j, gto);
duke@435 572 // Kill alternate exit path
duke@435 573 Block *dead = b->_succs[1-idx];
duke@435 574 for( j = 1; j < dead->num_preds(); j++)
duke@435 575 if( dead->pred(j)->in(0) == bp )
duke@435 576 break;
duke@435 577 // Scan through block, yanking dead path from
duke@435 578 // all regions and phis.
duke@435 579 dead->head()->del_req(j);
duke@435 580 for( int k = 1; dead->_nodes[k]->is_Phi(); k++ )
duke@435 581 dead->_nodes[k]->del_req(j);
duke@435 582 }
duke@435 583
rasbold@853 584 //------------------------------move_to_next-----------------------------------
duke@435 585 // Helper function to move block bx to the slot following b_index. Return
duke@435 586 // true if the move is successful, otherwise false
rasbold@853 587 bool PhaseCFG::move_to_next(Block* bx, uint b_index) {
duke@435 588 if (bx == NULL) return false;
duke@435 589
duke@435 590 // Return false if bx is already scheduled.
duke@435 591 uint bx_index = bx->_pre_order;
duke@435 592 if ((bx_index <= b_index) && (_blocks[bx_index] == bx)) {
duke@435 593 return false;
duke@435 594 }
duke@435 595
duke@435 596 // Find the current index of block bx on the block list
duke@435 597 bx_index = b_index + 1;
duke@435 598 while( bx_index < _num_blocks && _blocks[bx_index] != bx ) bx_index++;
duke@435 599 assert(_blocks[bx_index] == bx, "block not found");
duke@435 600
duke@435 601 // If the previous block conditionally falls into bx, return false,
duke@435 602 // because moving bx will create an extra jump.
duke@435 603 for(uint k = 1; k < bx->num_preds(); k++ ) {
duke@435 604 Block* pred = _bbs[bx->pred(k)->_idx];
duke@435 605 if (pred == _blocks[bx_index-1]) {
duke@435 606 if (pred->_num_succs != 1) {
duke@435 607 return false;
duke@435 608 }
duke@435 609 }
duke@435 610 }
duke@435 611
duke@435 612 // Reinsert bx just past block 'b'
duke@435 613 _blocks.remove(bx_index);
duke@435 614 _blocks.insert(b_index + 1, bx);
duke@435 615 return true;
duke@435 616 }
duke@435 617
rasbold@853 618 //------------------------------move_to_end------------------------------------
duke@435 619 // Move empty and uncommon blocks to the end.
rasbold@853 620 void PhaseCFG::move_to_end(Block *b, uint i) {
duke@435 621 int e = b->is_Empty();
duke@435 622 if (e != Block::not_empty) {
duke@435 623 if (e == Block::empty_with_goto) {
duke@435 624 // Remove the goto, but leave the block.
duke@435 625 b->_nodes.pop();
duke@435 626 }
duke@435 627 // Mark this block as a connector block, which will cause it to be
duke@435 628 // ignored in certain functions such as non_connector_successor().
duke@435 629 b->set_connector();
duke@435 630 }
duke@435 631 // Move the empty block to the end, and don't recheck.
duke@435 632 _blocks.remove(i);
duke@435 633 _blocks.push(b);
duke@435 634 }
duke@435 635
rasbold@853 636 //---------------------------set_loop_alignment--------------------------------
rasbold@853 637 // Set loop alignment for every block
rasbold@853 638 void PhaseCFG::set_loop_alignment() {
rasbold@853 639 uint last = _num_blocks;
rasbold@853 640 assert( _blocks[0] == _broot, "" );
rasbold@853 641
rasbold@853 642 for (uint i = 1; i < last; i++ ) {
rasbold@853 643 Block *b = _blocks[i];
rasbold@853 644 if (b->head()->is_Loop()) {
rasbold@853 645 b->set_loop_alignment(b);
rasbold@853 646 }
rasbold@853 647 }
rasbold@853 648 }
rasbold@853 649
rasbold@853 650 //-----------------------------remove_empty------------------------------------
rasbold@853 651 // Make empty basic blocks to be "connector" blocks, Move uncommon blocks
rasbold@853 652 // to the end.
rasbold@853 653 void PhaseCFG::remove_empty() {
duke@435 654 // Move uncommon blocks to the end
duke@435 655 uint last = _num_blocks;
duke@435 656 assert( _blocks[0] == _broot, "" );
rasbold@853 657
rasbold@853 658 for (uint i = 1; i < last; i++) {
duke@435 659 Block *b = _blocks[i];
rasbold@853 660 if (b->is_connector()) break;
duke@435 661
duke@435 662 // Check for NeverBranch at block end. This needs to become a GOTO to the
duke@435 663 // true target. NeverBranch are treated as a conditional branch that
duke@435 664 // always goes the same direction for most of the optimizer and are used
duke@435 665 // to give a fake exit path to infinite loops. At this late stage they
duke@435 666 // need to turn into Goto's so that when you enter the infinite loop you
duke@435 667 // indeed hang.
duke@435 668 if( b->_nodes[b->end_idx()]->Opcode() == Op_NeverBranch )
duke@435 669 convert_NeverBranch_to_Goto(b);
duke@435 670
duke@435 671 // Look for uncommon blocks and move to end.
rasbold@853 672 if (!C->do_freq_based_layout()) {
rasbold@853 673 if( b->is_uncommon(_bbs) ) {
rasbold@853 674 move_to_end(b, i);
rasbold@853 675 last--; // No longer check for being uncommon!
rasbold@853 676 if( no_flip_branch(b) ) { // Fall-thru case must follow?
rasbold@853 677 b = _blocks[i]; // Find the fall-thru block
rasbold@853 678 move_to_end(b, i);
rasbold@853 679 last--;
rasbold@853 680 }
rasbold@853 681 i--; // backup block counter post-increment
duke@435 682 }
duke@435 683 }
duke@435 684 }
duke@435 685
rasbold@853 686 // Move empty blocks to the end
duke@435 687 last = _num_blocks;
rasbold@853 688 for (uint i = 1; i < last; i++) {
duke@435 689 Block *b = _blocks[i];
rasbold@853 690 if (b->is_Empty() != Block::not_empty) {
rasbold@853 691 move_to_end(b, i);
rasbold@853 692 last--;
rasbold@853 693 i--;
duke@435 694 }
duke@435 695 } // End of for all blocks
rasbold@853 696 }
duke@435 697
rasbold@853 698 //-----------------------------fixup_flow--------------------------------------
rasbold@853 699 // Fix up the final control flow for basic blocks.
rasbold@853 700 void PhaseCFG::fixup_flow() {
duke@435 701 // Fixup final control flow for the blocks. Remove jump-to-next
duke@435 702 // block. If neither arm of a IF follows the conditional branch, we
duke@435 703 // have to add a second jump after the conditional. We place the
duke@435 704 // TRUE branch target in succs[0] for both GOTOs and IFs.
rasbold@853 705 for (uint i=0; i < _num_blocks; i++) {
duke@435 706 Block *b = _blocks[i];
duke@435 707 b->_pre_order = i; // turn pre-order into block-index
duke@435 708
duke@435 709 // Connector blocks need no further processing.
duke@435 710 if (b->is_connector()) {
duke@435 711 assert((i+1) == _num_blocks || _blocks[i+1]->is_connector(),
duke@435 712 "All connector blocks should sink to the end");
duke@435 713 continue;
duke@435 714 }
duke@435 715 assert(b->is_Empty() != Block::completely_empty,
duke@435 716 "Empty blocks should be connectors");
duke@435 717
duke@435 718 Block *bnext = (i < _num_blocks-1) ? _blocks[i+1] : NULL;
duke@435 719 Block *bs0 = b->non_connector_successor(0);
duke@435 720
duke@435 721 // Check for multi-way branches where I cannot negate the test to
duke@435 722 // exchange the true and false targets.
duke@435 723 if( no_flip_branch( b ) ) {
duke@435 724 // Find fall through case - if must fall into its target
duke@435 725 int branch_idx = b->_nodes.size() - b->_num_succs;
duke@435 726 for (uint j2 = 0; j2 < b->_num_succs; j2++) {
duke@435 727 const ProjNode* p = b->_nodes[branch_idx + j2]->as_Proj();
duke@435 728 if (p->_con == 0) {
duke@435 729 // successor j2 is fall through case
duke@435 730 if (b->non_connector_successor(j2) != bnext) {
duke@435 731 // but it is not the next block => insert a goto
duke@435 732 insert_goto_at(i, j2);
duke@435 733 }
duke@435 734 // Put taken branch in slot 0
duke@435 735 if( j2 == 0 && b->_num_succs == 2) {
duke@435 736 // Flip targets in succs map
duke@435 737 Block *tbs0 = b->_succs[0];
duke@435 738 Block *tbs1 = b->_succs[1];
duke@435 739 b->_succs.map( 0, tbs1 );
duke@435 740 b->_succs.map( 1, tbs0 );
duke@435 741 }
duke@435 742 break;
duke@435 743 }
duke@435 744 }
duke@435 745 // Remove all CatchProjs
rasbold@853 746 for (uint j1 = 0; j1 < b->_num_succs; j1++) b->_nodes.pop();
duke@435 747
duke@435 748 } else if (b->_num_succs == 1) {
duke@435 749 // Block ends in a Goto?
duke@435 750 if (bnext == bs0) {
duke@435 751 // We fall into next block; remove the Goto
duke@435 752 b->_nodes.pop();
duke@435 753 }
duke@435 754
duke@435 755 } else if( b->_num_succs == 2 ) { // Block ends in a If?
duke@435 756 // Get opcode of 1st projection (matches _succs[0])
duke@435 757 // Note: Since this basic block has 2 exits, the last 2 nodes must
duke@435 758 // be projections (in any order), the 3rd last node must be
duke@435 759 // the IfNode (we have excluded other 2-way exits such as
duke@435 760 // CatchNodes already).
duke@435 761 MachNode *iff = b->_nodes[b->_nodes.size()-3]->as_Mach();
duke@435 762 ProjNode *proj0 = b->_nodes[b->_nodes.size()-2]->as_Proj();
duke@435 763 ProjNode *proj1 = b->_nodes[b->_nodes.size()-1]->as_Proj();
duke@435 764
duke@435 765 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1].
duke@435 766 assert(proj0->raw_out(0) == b->_succs[0]->head(), "Mismatch successor 0");
duke@435 767 assert(proj1->raw_out(0) == b->_succs[1]->head(), "Mismatch successor 1");
duke@435 768
duke@435 769 Block *bs1 = b->non_connector_successor(1);
duke@435 770
duke@435 771 // Check for neither successor block following the current
duke@435 772 // block ending in a conditional. If so, move one of the
duke@435 773 // successors after the current one, provided that the
duke@435 774 // successor was previously unscheduled, but moveable
duke@435 775 // (i.e., all paths to it involve a branch).
rasbold@853 776 if( !C->do_freq_based_layout() && bnext != bs0 && bnext != bs1 ) {
duke@435 777 // Choose the more common successor based on the probability
duke@435 778 // of the conditional branch.
duke@435 779 Block *bx = bs0;
duke@435 780 Block *by = bs1;
duke@435 781
duke@435 782 // _prob is the probability of taking the true path. Make
duke@435 783 // p the probability of taking successor #1.
duke@435 784 float p = iff->as_MachIf()->_prob;
duke@435 785 if( proj0->Opcode() == Op_IfTrue ) {
duke@435 786 p = 1.0 - p;
duke@435 787 }
duke@435 788
duke@435 789 // Prefer successor #1 if p > 0.5
duke@435 790 if (p > PROB_FAIR) {
duke@435 791 bx = bs1;
duke@435 792 by = bs0;
duke@435 793 }
duke@435 794
duke@435 795 // Attempt the more common successor first
rasbold@853 796 if (move_to_next(bx, i)) {
duke@435 797 bnext = bx;
rasbold@853 798 } else if (move_to_next(by, i)) {
duke@435 799 bnext = by;
duke@435 800 }
duke@435 801 }
duke@435 802
duke@435 803 // Check for conditional branching the wrong way. Negate
duke@435 804 // conditional, if needed, so it falls into the following block
duke@435 805 // and branches to the not-following block.
duke@435 806
duke@435 807 // Check for the next block being in succs[0]. We are going to branch
duke@435 808 // to succs[0], so we want the fall-thru case as the next block in
duke@435 809 // succs[1].
duke@435 810 if (bnext == bs0) {
duke@435 811 // Fall-thru case in succs[0], so flip targets in succs map
duke@435 812 Block *tbs0 = b->_succs[0];
duke@435 813 Block *tbs1 = b->_succs[1];
duke@435 814 b->_succs.map( 0, tbs1 );
duke@435 815 b->_succs.map( 1, tbs0 );
duke@435 816 // Flip projection for each target
duke@435 817 { ProjNode *tmp = proj0; proj0 = proj1; proj1 = tmp; }
duke@435 818
rasbold@853 819 } else if( bnext != bs1 ) {
rasbold@853 820 // Need a double-branch
duke@435 821 // The existing conditional branch need not change.
duke@435 822 // Add a unconditional branch to the false target.
duke@435 823 // Alas, it must appear in its own block and adding a
duke@435 824 // block this late in the game is complicated. Sigh.
duke@435 825 insert_goto_at(i, 1);
duke@435 826 }
duke@435 827
duke@435 828 // Make sure we TRUE branch to the target
rasbold@853 829 if( proj0->Opcode() == Op_IfFalse ) {
duke@435 830 iff->negate();
rasbold@853 831 }
duke@435 832
duke@435 833 b->_nodes.pop(); // Remove IfFalse & IfTrue projections
duke@435 834 b->_nodes.pop();
duke@435 835
duke@435 836 } else {
duke@435 837 // Multi-exit block, e.g. a switch statement
duke@435 838 // But we don't need to do anything here
duke@435 839 }
duke@435 840 } // End of for all blocks
duke@435 841 }
duke@435 842
duke@435 843
duke@435 844 //------------------------------dump-------------------------------------------
duke@435 845 #ifndef PRODUCT
duke@435 846 void PhaseCFG::_dump_cfg( const Node *end, VectorSet &visited ) const {
duke@435 847 const Node *x = end->is_block_proj();
duke@435 848 assert( x, "not a CFG" );
duke@435 849
duke@435 850 // Do not visit this block again
duke@435 851 if( visited.test_set(x->_idx) ) return;
duke@435 852
duke@435 853 // Skip through this block
duke@435 854 const Node *p = x;
duke@435 855 do {
duke@435 856 p = p->in(0); // Move control forward
duke@435 857 assert( !p->is_block_proj() || p->is_Root(), "not a CFG" );
duke@435 858 } while( !p->is_block_start() );
duke@435 859
duke@435 860 // Recursively visit
duke@435 861 for( uint i=1; i<p->req(); i++ )
duke@435 862 _dump_cfg(p->in(i),visited);
duke@435 863
duke@435 864 // Dump the block
duke@435 865 _bbs[p->_idx]->dump(&_bbs);
duke@435 866 }
duke@435 867
duke@435 868 void PhaseCFG::dump( ) const {
duke@435 869 tty->print("\n--- CFG --- %d BBs\n",_num_blocks);
duke@435 870 if( _blocks.size() ) { // Did we do basic-block layout?
duke@435 871 for( uint i=0; i<_num_blocks; i++ )
duke@435 872 _blocks[i]->dump(&_bbs);
duke@435 873 } else { // Else do it with a DFS
duke@435 874 VectorSet visited(_bbs._arena);
duke@435 875 _dump_cfg(_root,visited);
duke@435 876 }
duke@435 877 }
duke@435 878
duke@435 879 void PhaseCFG::dump_headers() {
duke@435 880 for( uint i = 0; i < _num_blocks; i++ ) {
duke@435 881 if( _blocks[i] == NULL ) continue;
duke@435 882 _blocks[i]->dump_head(&_bbs);
duke@435 883 }
duke@435 884 }
duke@435 885
duke@435 886 void PhaseCFG::verify( ) const {
kvn@1001 887 #ifdef ASSERT
duke@435 888 // Verify sane CFG
duke@435 889 for( uint i = 0; i < _num_blocks; i++ ) {
duke@435 890 Block *b = _blocks[i];
duke@435 891 uint cnt = b->_nodes.size();
duke@435 892 uint j;
duke@435 893 for( j = 0; j < cnt; j++ ) {
duke@435 894 Node *n = b->_nodes[j];
duke@435 895 assert( _bbs[n->_idx] == b, "" );
duke@435 896 if( j >= 1 && n->is_Mach() &&
duke@435 897 n->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
duke@435 898 assert( j == 1 || b->_nodes[j-1]->is_Phi(),
duke@435 899 "CreateEx must be first instruction in block" );
duke@435 900 }
duke@435 901 for( uint k = 0; k < n->req(); k++ ) {
kvn@1001 902 Node *def = n->in(k);
kvn@1001 903 if( def && def != n ) {
kvn@1001 904 assert( _bbs[def->_idx] || def->is_Con(),
duke@435 905 "must have block; constants for debug info ok" );
kvn@1001 906 // Verify that instructions in the block is in correct order.
kvn@1001 907 // Uses must follow their definition if they are at the same block.
kvn@1001 908 // Mostly done to check that MachSpillCopy nodes are placed correctly
kvn@1001 909 // when CreateEx node is moved in build_ifg_physical().
kvn@1001 910 if( _bbs[def->_idx] == b &&
kvn@1001 911 !(b->head()->is_Loop() && n->is_Phi()) &&
kvn@1001 912 // See (+++) comment in reg_split.cpp
kvn@1001 913 !(n->jvms() != NULL && n->jvms()->is_monitor_use(k)) ) {
kvn@1328 914 bool is_loop = false;
kvn@1328 915 if (n->is_Phi()) {
kvn@1328 916 for( uint l = 1; l < def->req(); l++ ) {
kvn@1328 917 if (n == def->in(l)) {
kvn@1328 918 is_loop = true;
kvn@1328 919 break; // Some kind of loop
kvn@1328 920 }
kvn@1328 921 }
kvn@1328 922 }
kvn@1328 923 assert( is_loop || b->find_node(def) < j, "uses must follow definitions" );
kvn@1001 924 }
kvn@1036 925 if( def->is_SafePointScalarObject() ) {
kvn@1036 926 assert(_bbs[def->_idx] == b, "SafePointScalarObject Node should be at the same block as its SafePoint node");
kvn@1036 927 assert(_bbs[def->_idx] == _bbs[def->in(0)->_idx], "SafePointScalarObject Node should be at the same block as its control edge");
kvn@1036 928 }
duke@435 929 }
duke@435 930 }
duke@435 931 }
duke@435 932
duke@435 933 j = b->end_idx();
duke@435 934 Node *bp = (Node*)b->_nodes[b->_nodes.size()-1]->is_block_proj();
duke@435 935 assert( bp, "last instruction must be a block proj" );
duke@435 936 assert( bp == b->_nodes[j], "wrong number of successors for this block" );
duke@435 937 if( bp->is_Catch() ) {
duke@435 938 while( b->_nodes[--j]->Opcode() == Op_MachProj ) ;
duke@435 939 assert( b->_nodes[j]->is_Call(), "CatchProj must follow call" );
duke@435 940 }
duke@435 941 else if( bp->is_Mach() && bp->as_Mach()->ideal_Opcode() == Op_If ) {
duke@435 942 assert( b->_num_succs == 2, "Conditional branch must have two targets");
duke@435 943 }
duke@435 944 }
kvn@1001 945 #endif
duke@435 946 }
duke@435 947 #endif
duke@435 948
duke@435 949 //=============================================================================
duke@435 950 //------------------------------UnionFind--------------------------------------
duke@435 951 UnionFind::UnionFind( uint max ) : _cnt(max), _max(max), _indices(NEW_RESOURCE_ARRAY(uint,max)) {
duke@435 952 Copy::zero_to_bytes( _indices, sizeof(uint)*max );
duke@435 953 }
duke@435 954
duke@435 955 void UnionFind::extend( uint from_idx, uint to_idx ) {
duke@435 956 _nesting.check();
duke@435 957 if( from_idx >= _max ) {
duke@435 958 uint size = 16;
duke@435 959 while( size <= from_idx ) size <<=1;
duke@435 960 _indices = REALLOC_RESOURCE_ARRAY( uint, _indices, _max, size );
duke@435 961 _max = size;
duke@435 962 }
duke@435 963 while( _cnt <= from_idx ) _indices[_cnt++] = 0;
duke@435 964 _indices[from_idx] = to_idx;
duke@435 965 }
duke@435 966
duke@435 967 void UnionFind::reset( uint max ) {
duke@435 968 assert( max <= max_uint, "Must fit within uint" );
duke@435 969 // Force the Union-Find mapping to be at least this large
duke@435 970 extend(max,0);
duke@435 971 // Initialize to be the ID mapping.
rasbold@853 972 for( uint i=0; i<max; i++ ) map(i,i);
duke@435 973 }
duke@435 974
duke@435 975 //------------------------------Find_compress----------------------------------
duke@435 976 // Straight out of Tarjan's union-find algorithm
duke@435 977 uint UnionFind::Find_compress( uint idx ) {
duke@435 978 uint cur = idx;
duke@435 979 uint next = lookup(cur);
duke@435 980 while( next != cur ) { // Scan chain of equivalences
duke@435 981 assert( next < cur, "always union smaller" );
duke@435 982 cur = next; // until find a fixed-point
duke@435 983 next = lookup(cur);
duke@435 984 }
duke@435 985 // Core of union-find algorithm: update chain of
duke@435 986 // equivalences to be equal to the root.
duke@435 987 while( idx != next ) {
duke@435 988 uint tmp = lookup(idx);
duke@435 989 map(idx, next);
duke@435 990 idx = tmp;
duke@435 991 }
duke@435 992 return idx;
duke@435 993 }
duke@435 994
duke@435 995 //------------------------------Find_const-------------------------------------
duke@435 996 // Like Find above, but no path compress, so bad asymptotic behavior
duke@435 997 uint UnionFind::Find_const( uint idx ) const {
duke@435 998 if( idx == 0 ) return idx; // Ignore the zero idx
duke@435 999 // Off the end? This can happen during debugging dumps
duke@435 1000 // when data structures have not finished being updated.
duke@435 1001 if( idx >= _max ) return idx;
duke@435 1002 uint next = lookup(idx);
duke@435 1003 while( next != idx ) { // Scan chain of equivalences
duke@435 1004 idx = next; // until find a fixed-point
duke@435 1005 next = lookup(idx);
duke@435 1006 }
duke@435 1007 return next;
duke@435 1008 }
duke@435 1009
duke@435 1010 //------------------------------Union------------------------------------------
duke@435 1011 // union 2 sets together.
duke@435 1012 void UnionFind::Union( uint idx1, uint idx2 ) {
duke@435 1013 uint src = Find(idx1);
duke@435 1014 uint dst = Find(idx2);
duke@435 1015 assert( src, "" );
duke@435 1016 assert( dst, "" );
duke@435 1017 assert( src < _max, "oob" );
duke@435 1018 assert( dst < _max, "oob" );
duke@435 1019 assert( src < dst, "always union smaller" );
duke@435 1020 map(dst,src);
duke@435 1021 }
rasbold@853 1022
rasbold@853 1023 #ifndef PRODUCT
rasbold@853 1024 static void edge_dump(GrowableArray<CFGEdge *> *edges) {
rasbold@853 1025 tty->print_cr("---- Edges ----");
rasbold@853 1026 for (int i = 0; i < edges->length(); i++) {
rasbold@853 1027 CFGEdge *e = edges->at(i);
rasbold@853 1028 if (e != NULL) {
rasbold@853 1029 edges->at(i)->dump();
rasbold@853 1030 }
rasbold@853 1031 }
rasbold@853 1032 }
rasbold@853 1033
rasbold@853 1034 static void trace_dump(Trace *traces[], int count) {
rasbold@853 1035 tty->print_cr("---- Traces ----");
rasbold@853 1036 for (int i = 0; i < count; i++) {
rasbold@853 1037 Trace *tr = traces[i];
rasbold@853 1038 if (tr != NULL) {
rasbold@853 1039 tr->dump();
rasbold@853 1040 }
rasbold@853 1041 }
rasbold@853 1042 }
rasbold@853 1043
rasbold@853 1044 void Trace::dump( ) const {
rasbold@853 1045 tty->print_cr("Trace (freq %f)", first_block()->_freq);
rasbold@853 1046 for (Block *b = first_block(); b != NULL; b = next(b)) {
rasbold@853 1047 tty->print(" B%d", b->_pre_order);
rasbold@853 1048 if (b->head()->is_Loop()) {
rasbold@853 1049 tty->print(" (L%d)", b->compute_loop_alignment());
rasbold@853 1050 }
rasbold@853 1051 if (b->has_loop_alignment()) {
rasbold@853 1052 tty->print(" (T%d)", b->code_alignment());
rasbold@853 1053 }
rasbold@853 1054 }
rasbold@853 1055 tty->cr();
rasbold@853 1056 }
rasbold@853 1057
rasbold@853 1058 void CFGEdge::dump( ) const {
rasbold@853 1059 tty->print(" B%d --> B%d Freq: %f out:%3d%% in:%3d%% State: ",
rasbold@853 1060 from()->_pre_order, to()->_pre_order, freq(), _from_pct, _to_pct);
rasbold@853 1061 switch(state()) {
rasbold@853 1062 case connected:
rasbold@853 1063 tty->print("connected");
rasbold@853 1064 break;
rasbold@853 1065 case open:
rasbold@853 1066 tty->print("open");
rasbold@853 1067 break;
rasbold@853 1068 case interior:
rasbold@853 1069 tty->print("interior");
rasbold@853 1070 break;
rasbold@853 1071 }
rasbold@853 1072 if (infrequent()) {
rasbold@853 1073 tty->print(" infrequent");
rasbold@853 1074 }
rasbold@853 1075 tty->cr();
rasbold@853 1076 }
rasbold@853 1077 #endif
rasbold@853 1078
rasbold@853 1079 //=============================================================================
rasbold@853 1080
rasbold@853 1081 //------------------------------edge_order-------------------------------------
rasbold@853 1082 // Comparison function for edges
rasbold@853 1083 static int edge_order(CFGEdge **e0, CFGEdge **e1) {
rasbold@853 1084 float freq0 = (*e0)->freq();
rasbold@853 1085 float freq1 = (*e1)->freq();
rasbold@853 1086 if (freq0 != freq1) {
rasbold@853 1087 return freq0 > freq1 ? -1 : 1;
rasbold@853 1088 }
rasbold@853 1089
rasbold@853 1090 int dist0 = (*e0)->to()->_rpo - (*e0)->from()->_rpo;
rasbold@853 1091 int dist1 = (*e1)->to()->_rpo - (*e1)->from()->_rpo;
rasbold@853 1092
rasbold@853 1093 return dist1 - dist0;
rasbold@853 1094 }
rasbold@853 1095
rasbold@853 1096 //------------------------------trace_frequency_order--------------------------
rasbold@853 1097 // Comparison function for edges
rasbold@853 1098 static int trace_frequency_order(const void *p0, const void *p1) {
rasbold@853 1099 Trace *tr0 = *(Trace **) p0;
rasbold@853 1100 Trace *tr1 = *(Trace **) p1;
rasbold@853 1101 Block *b0 = tr0->first_block();
rasbold@853 1102 Block *b1 = tr1->first_block();
rasbold@853 1103
rasbold@853 1104 // The trace of connector blocks goes at the end;
rasbold@853 1105 // we only expect one such trace
rasbold@853 1106 if (b0->is_connector() != b1->is_connector()) {
rasbold@853 1107 return b1->is_connector() ? -1 : 1;
rasbold@853 1108 }
rasbold@853 1109
rasbold@853 1110 // Pull more frequently executed blocks to the beginning
rasbold@853 1111 float freq0 = b0->_freq;
rasbold@853 1112 float freq1 = b1->_freq;
rasbold@853 1113 if (freq0 != freq1) {
rasbold@853 1114 return freq0 > freq1 ? -1 : 1;
rasbold@853 1115 }
rasbold@853 1116
rasbold@853 1117 int diff = tr0->first_block()->_rpo - tr1->first_block()->_rpo;
rasbold@853 1118
rasbold@853 1119 return diff;
rasbold@853 1120 }
rasbold@853 1121
rasbold@853 1122 //------------------------------find_edges-------------------------------------
rasbold@853 1123 // Find edges of interest, i.e, those which can fall through. Presumes that
rasbold@853 1124 // edges which don't fall through are of low frequency and can be generally
rasbold@853 1125 // ignored. Initialize the list of traces.
rasbold@853 1126 void PhaseBlockLayout::find_edges()
rasbold@853 1127 {
rasbold@853 1128 // Walk the blocks, creating edges and Traces
rasbold@853 1129 uint i;
rasbold@853 1130 Trace *tr = NULL;
rasbold@853 1131 for (i = 0; i < _cfg._num_blocks; i++) {
rasbold@853 1132 Block *b = _cfg._blocks[i];
rasbold@853 1133 tr = new Trace(b, next, prev);
rasbold@853 1134 traces[tr->id()] = tr;
rasbold@853 1135
rasbold@853 1136 // All connector blocks should be at the end of the list
rasbold@853 1137 if (b->is_connector()) break;
rasbold@853 1138
rasbold@853 1139 // If this block and the next one have a one-to-one successor
rasbold@853 1140 // predecessor relationship, simply append the next block
rasbold@853 1141 int nfallthru = b->num_fall_throughs();
rasbold@853 1142 while (nfallthru == 1 &&
rasbold@853 1143 b->succ_fall_through(0)) {
rasbold@853 1144 Block *n = b->_succs[0];
rasbold@853 1145
rasbold@853 1146 // Skip over single-entry connector blocks, we don't want to
rasbold@853 1147 // add them to the trace.
rasbold@853 1148 while (n->is_connector() && n->num_preds() == 1) {
rasbold@853 1149 n = n->_succs[0];
rasbold@853 1150 }
rasbold@853 1151
rasbold@853 1152 // We see a merge point, so stop search for the next block
rasbold@853 1153 if (n->num_preds() != 1) break;
rasbold@853 1154
rasbold@853 1155 i++;
rasbold@853 1156 assert(n = _cfg._blocks[i], "expecting next block");
rasbold@853 1157 tr->append(n);
rasbold@853 1158 uf->map(n->_pre_order, tr->id());
rasbold@853 1159 traces[n->_pre_order] = NULL;
rasbold@853 1160 nfallthru = b->num_fall_throughs();
rasbold@853 1161 b = n;
rasbold@853 1162 }
rasbold@853 1163
rasbold@853 1164 if (nfallthru > 0) {
rasbold@853 1165 // Create a CFGEdge for each outgoing
rasbold@853 1166 // edge that could be a fall-through.
rasbold@853 1167 for (uint j = 0; j < b->_num_succs; j++ ) {
rasbold@853 1168 if (b->succ_fall_through(j)) {
rasbold@853 1169 Block *target = b->non_connector_successor(j);
rasbold@853 1170 float freq = b->_freq * b->succ_prob(j);
rasbold@853 1171 int from_pct = (int) ((100 * freq) / b->_freq);
rasbold@853 1172 int to_pct = (int) ((100 * freq) / target->_freq);
rasbold@853 1173 edges->append(new CFGEdge(b, target, freq, from_pct, to_pct));
rasbold@853 1174 }
rasbold@853 1175 }
rasbold@853 1176 }
rasbold@853 1177 }
rasbold@853 1178
rasbold@853 1179 // Group connector blocks into one trace
rasbold@853 1180 for (i++; i < _cfg._num_blocks; i++) {
rasbold@853 1181 Block *b = _cfg._blocks[i];
rasbold@853 1182 assert(b->is_connector(), "connector blocks at the end");
rasbold@853 1183 tr->append(b);
rasbold@853 1184 uf->map(b->_pre_order, tr->id());
rasbold@853 1185 traces[b->_pre_order] = NULL;
rasbold@853 1186 }
rasbold@853 1187 }
rasbold@853 1188
rasbold@853 1189 //------------------------------union_traces----------------------------------
rasbold@853 1190 // Union two traces together in uf, and null out the trace in the list
rasbold@853 1191 void PhaseBlockLayout::union_traces(Trace* updated_trace, Trace* old_trace)
rasbold@853 1192 {
rasbold@853 1193 uint old_id = old_trace->id();
rasbold@853 1194 uint updated_id = updated_trace->id();
rasbold@853 1195
rasbold@853 1196 uint lo_id = updated_id;
rasbold@853 1197 uint hi_id = old_id;
rasbold@853 1198
rasbold@853 1199 // If from is greater than to, swap values to meet
rasbold@853 1200 // UnionFind guarantee.
rasbold@853 1201 if (updated_id > old_id) {
rasbold@853 1202 lo_id = old_id;
rasbold@853 1203 hi_id = updated_id;
rasbold@853 1204
rasbold@853 1205 // Fix up the trace ids
rasbold@853 1206 traces[lo_id] = traces[updated_id];
rasbold@853 1207 updated_trace->set_id(lo_id);
rasbold@853 1208 }
rasbold@853 1209
rasbold@853 1210 // Union the lower with the higher and remove the pointer
rasbold@853 1211 // to the higher.
rasbold@853 1212 uf->Union(lo_id, hi_id);
rasbold@853 1213 traces[hi_id] = NULL;
rasbold@853 1214 }
rasbold@853 1215
rasbold@853 1216 //------------------------------grow_traces-------------------------------------
rasbold@853 1217 // Append traces together via the most frequently executed edges
rasbold@853 1218 void PhaseBlockLayout::grow_traces()
rasbold@853 1219 {
rasbold@853 1220 // Order the edges, and drive the growth of Traces via the most
rasbold@853 1221 // frequently executed edges.
rasbold@853 1222 edges->sort(edge_order);
rasbold@853 1223 for (int i = 0; i < edges->length(); i++) {
rasbold@853 1224 CFGEdge *e = edges->at(i);
rasbold@853 1225
rasbold@853 1226 if (e->state() != CFGEdge::open) continue;
rasbold@853 1227
rasbold@853 1228 Block *src_block = e->from();
rasbold@853 1229 Block *targ_block = e->to();
rasbold@853 1230
rasbold@853 1231 // Don't grow traces along backedges?
rasbold@853 1232 if (!BlockLayoutRotateLoops) {
rasbold@853 1233 if (targ_block->_rpo <= src_block->_rpo) {
rasbold@853 1234 targ_block->set_loop_alignment(targ_block);
rasbold@853 1235 continue;
rasbold@853 1236 }
rasbold@853 1237 }
rasbold@853 1238
rasbold@853 1239 Trace *src_trace = trace(src_block);
rasbold@853 1240 Trace *targ_trace = trace(targ_block);
rasbold@853 1241
rasbold@853 1242 // If the edge in question can join two traces at their ends,
rasbold@853 1243 // append one trace to the other.
rasbold@853 1244 if (src_trace->last_block() == src_block) {
rasbold@853 1245 if (src_trace == targ_trace) {
rasbold@853 1246 e->set_state(CFGEdge::interior);
rasbold@853 1247 if (targ_trace->backedge(e)) {
rasbold@853 1248 // Reset i to catch any newly eligible edge
rasbold@853 1249 // (Or we could remember the first "open" edge, and reset there)
rasbold@853 1250 i = 0;
rasbold@853 1251 }
rasbold@853 1252 } else if (targ_trace->first_block() == targ_block) {
rasbold@853 1253 e->set_state(CFGEdge::connected);
rasbold@853 1254 src_trace->append(targ_trace);
rasbold@853 1255 union_traces(src_trace, targ_trace);
rasbold@853 1256 }
rasbold@853 1257 }
rasbold@853 1258 }
rasbold@853 1259 }
rasbold@853 1260
rasbold@853 1261 //------------------------------merge_traces-----------------------------------
rasbold@853 1262 // Embed one trace into another, if the fork or join points are sufficiently
rasbold@853 1263 // balanced.
rasbold@853 1264 void PhaseBlockLayout::merge_traces(bool fall_thru_only)
rasbold@853 1265 {
rasbold@853 1266 // Walk the edge list a another time, looking at unprocessed edges.
rasbold@853 1267 // Fold in diamonds
rasbold@853 1268 for (int i = 0; i < edges->length(); i++) {
rasbold@853 1269 CFGEdge *e = edges->at(i);
rasbold@853 1270
rasbold@853 1271 if (e->state() != CFGEdge::open) continue;
rasbold@853 1272 if (fall_thru_only) {
rasbold@853 1273 if (e->infrequent()) continue;
rasbold@853 1274 }
rasbold@853 1275
rasbold@853 1276 Block *src_block = e->from();
rasbold@853 1277 Trace *src_trace = trace(src_block);
rasbold@853 1278 bool src_at_tail = src_trace->last_block() == src_block;
rasbold@853 1279
rasbold@853 1280 Block *targ_block = e->to();
rasbold@853 1281 Trace *targ_trace = trace(targ_block);
rasbold@853 1282 bool targ_at_start = targ_trace->first_block() == targ_block;
rasbold@853 1283
rasbold@853 1284 if (src_trace == targ_trace) {
rasbold@853 1285 // This may be a loop, but we can't do much about it.
rasbold@853 1286 e->set_state(CFGEdge::interior);
rasbold@853 1287 continue;
rasbold@853 1288 }
rasbold@853 1289
rasbold@853 1290 if (fall_thru_only) {
rasbold@853 1291 // If the edge links the middle of two traces, we can't do anything.
rasbold@853 1292 // Mark the edge and continue.
rasbold@853 1293 if (!src_at_tail & !targ_at_start) {
rasbold@853 1294 continue;
rasbold@853 1295 }
rasbold@853 1296
rasbold@853 1297 // Don't grow traces along backedges?
rasbold@853 1298 if (!BlockLayoutRotateLoops && (targ_block->_rpo <= src_block->_rpo)) {
rasbold@853 1299 continue;
rasbold@853 1300 }
rasbold@853 1301
rasbold@853 1302 // If both ends of the edge are available, why didn't we handle it earlier?
rasbold@853 1303 assert(src_at_tail ^ targ_at_start, "Should have caught this edge earlier.");
rasbold@853 1304
rasbold@853 1305 if (targ_at_start) {
rasbold@853 1306 // Insert the "targ" trace in the "src" trace if the insertion point
rasbold@853 1307 // is a two way branch.
rasbold@853 1308 // Better profitability check possible, but may not be worth it.
rasbold@853 1309 // Someday, see if the this "fork" has an associated "join";
rasbold@853 1310 // then make a policy on merging this trace at the fork or join.
rasbold@853 1311 // For example, other things being equal, it may be better to place this
rasbold@853 1312 // trace at the join point if the "src" trace ends in a two-way, but
rasbold@853 1313 // the insertion point is one-way.
rasbold@853 1314 assert(src_block->num_fall_throughs() == 2, "unexpected diamond");
rasbold@853 1315 e->set_state(CFGEdge::connected);
rasbold@853 1316 src_trace->insert_after(src_block, targ_trace);
rasbold@853 1317 union_traces(src_trace, targ_trace);
rasbold@853 1318 } else if (src_at_tail) {
rasbold@853 1319 if (src_trace != trace(_cfg._broot)) {
rasbold@853 1320 e->set_state(CFGEdge::connected);
rasbold@853 1321 targ_trace->insert_before(targ_block, src_trace);
rasbold@853 1322 union_traces(targ_trace, src_trace);
rasbold@853 1323 }
rasbold@853 1324 }
rasbold@853 1325 } else if (e->state() == CFGEdge::open) {
rasbold@853 1326 // Append traces, even without a fall-thru connection.
twisti@1040 1327 // But leave root entry at the beginning of the block list.
rasbold@853 1328 if (targ_trace != trace(_cfg._broot)) {
rasbold@853 1329 e->set_state(CFGEdge::connected);
rasbold@853 1330 src_trace->append(targ_trace);
rasbold@853 1331 union_traces(src_trace, targ_trace);
rasbold@853 1332 }
rasbold@853 1333 }
rasbold@853 1334 }
rasbold@853 1335 }
rasbold@853 1336
rasbold@853 1337 //----------------------------reorder_traces-----------------------------------
rasbold@853 1338 // Order the sequence of the traces in some desirable way, and fixup the
rasbold@853 1339 // jumps at the end of each block.
rasbold@853 1340 void PhaseBlockLayout::reorder_traces(int count)
rasbold@853 1341 {
rasbold@853 1342 ResourceArea *area = Thread::current()->resource_area();
rasbold@853 1343 Trace ** new_traces = NEW_ARENA_ARRAY(area, Trace *, count);
rasbold@853 1344 Block_List worklist;
rasbold@853 1345 int new_count = 0;
rasbold@853 1346
rasbold@853 1347 // Compact the traces.
rasbold@853 1348 for (int i = 0; i < count; i++) {
rasbold@853 1349 Trace *tr = traces[i];
rasbold@853 1350 if (tr != NULL) {
rasbold@853 1351 new_traces[new_count++] = tr;
rasbold@853 1352 }
rasbold@853 1353 }
rasbold@853 1354
rasbold@853 1355 // The entry block should be first on the new trace list.
rasbold@853 1356 Trace *tr = trace(_cfg._broot);
rasbold@853 1357 assert(tr == new_traces[0], "entry trace misplaced");
rasbold@853 1358
rasbold@853 1359 // Sort the new trace list by frequency
rasbold@853 1360 qsort(new_traces + 1, new_count - 1, sizeof(new_traces[0]), trace_frequency_order);
rasbold@853 1361
rasbold@853 1362 // Patch up the successor blocks
rasbold@853 1363 _cfg._blocks.reset();
rasbold@853 1364 _cfg._num_blocks = 0;
rasbold@853 1365 for (int i = 0; i < new_count; i++) {
rasbold@853 1366 Trace *tr = new_traces[i];
rasbold@853 1367 if (tr != NULL) {
rasbold@853 1368 tr->fixup_blocks(_cfg);
rasbold@853 1369 }
rasbold@853 1370 }
rasbold@853 1371 }
rasbold@853 1372
rasbold@853 1373 //------------------------------PhaseBlockLayout-------------------------------
rasbold@853 1374 // Order basic blocks based on frequency
rasbold@853 1375 PhaseBlockLayout::PhaseBlockLayout(PhaseCFG &cfg) :
rasbold@853 1376 Phase(BlockLayout),
rasbold@853 1377 _cfg(cfg)
rasbold@853 1378 {
rasbold@853 1379 ResourceMark rm;
rasbold@853 1380 ResourceArea *area = Thread::current()->resource_area();
rasbold@853 1381
rasbold@853 1382 // List of traces
rasbold@853 1383 int size = _cfg._num_blocks + 1;
rasbold@853 1384 traces = NEW_ARENA_ARRAY(area, Trace *, size);
rasbold@853 1385 memset(traces, 0, size*sizeof(Trace*));
rasbold@853 1386 next = NEW_ARENA_ARRAY(area, Block *, size);
rasbold@853 1387 memset(next, 0, size*sizeof(Block *));
rasbold@853 1388 prev = NEW_ARENA_ARRAY(area, Block *, size);
rasbold@853 1389 memset(prev , 0, size*sizeof(Block *));
rasbold@853 1390
rasbold@853 1391 // List of edges
rasbold@853 1392 edges = new GrowableArray<CFGEdge*>;
rasbold@853 1393
rasbold@853 1394 // Mapping block index --> block_trace
rasbold@853 1395 uf = new UnionFind(size);
rasbold@853 1396 uf->reset(size);
rasbold@853 1397
rasbold@853 1398 // Find edges and create traces.
rasbold@853 1399 find_edges();
rasbold@853 1400
rasbold@853 1401 // Grow traces at their ends via most frequent edges.
rasbold@853 1402 grow_traces();
rasbold@853 1403
rasbold@853 1404 // Merge one trace into another, but only at fall-through points.
rasbold@853 1405 // This may make diamonds and other related shapes in a trace.
rasbold@853 1406 merge_traces(true);
rasbold@853 1407
rasbold@853 1408 // Run merge again, allowing two traces to be catenated, even if
rasbold@853 1409 // one does not fall through into the other. This appends loosely
rasbold@853 1410 // related traces to be near each other.
rasbold@853 1411 merge_traces(false);
rasbold@853 1412
rasbold@853 1413 // Re-order all the remaining traces by frequency
rasbold@853 1414 reorder_traces(size);
rasbold@853 1415
rasbold@853 1416 assert(_cfg._num_blocks >= (uint) (size - 1), "number of blocks can not shrink");
rasbold@853 1417 }
rasbold@853 1418
rasbold@853 1419
rasbold@853 1420 //------------------------------backedge---------------------------------------
rasbold@853 1421 // Edge e completes a loop in a trace. If the target block is head of the
rasbold@853 1422 // loop, rotate the loop block so that the loop ends in a conditional branch.
rasbold@853 1423 bool Trace::backedge(CFGEdge *e) {
rasbold@853 1424 bool loop_rotated = false;
rasbold@853 1425 Block *src_block = e->from();
rasbold@853 1426 Block *targ_block = e->to();
rasbold@853 1427
rasbold@853 1428 assert(last_block() == src_block, "loop discovery at back branch");
rasbold@853 1429 if (first_block() == targ_block) {
rasbold@853 1430 if (BlockLayoutRotateLoops && last_block()->num_fall_throughs() < 2) {
rasbold@853 1431 // Find the last block in the trace that has a conditional
rasbold@853 1432 // branch.
rasbold@853 1433 Block *b;
rasbold@853 1434 for (b = last_block(); b != NULL; b = prev(b)) {
rasbold@853 1435 if (b->num_fall_throughs() == 2) {
rasbold@853 1436 break;
rasbold@853 1437 }
rasbold@853 1438 }
rasbold@853 1439
rasbold@853 1440 if (b != last_block() && b != NULL) {
rasbold@853 1441 loop_rotated = true;
rasbold@853 1442
rasbold@853 1443 // Rotate the loop by doing two-part linked-list surgery.
rasbold@853 1444 append(first_block());
rasbold@853 1445 break_loop_after(b);
rasbold@853 1446 }
rasbold@853 1447 }
rasbold@853 1448
rasbold@853 1449 // Backbranch to the top of a trace
twisti@1040 1450 // Scroll forward through the trace from the targ_block. If we find
rasbold@853 1451 // a loop head before another loop top, use the the loop head alignment.
rasbold@853 1452 for (Block *b = targ_block; b != NULL; b = next(b)) {
rasbold@853 1453 if (b->has_loop_alignment()) {
rasbold@853 1454 break;
rasbold@853 1455 }
rasbold@853 1456 if (b->head()->is_Loop()) {
rasbold@853 1457 targ_block = b;
rasbold@853 1458 break;
rasbold@853 1459 }
rasbold@853 1460 }
rasbold@853 1461
rasbold@853 1462 first_block()->set_loop_alignment(targ_block);
rasbold@853 1463
rasbold@853 1464 } else {
rasbold@853 1465 // Backbranch into the middle of a trace
rasbold@853 1466 targ_block->set_loop_alignment(targ_block);
rasbold@853 1467 }
rasbold@853 1468
rasbold@853 1469 return loop_rotated;
rasbold@853 1470 }
rasbold@853 1471
rasbold@853 1472 //------------------------------fixup_blocks-----------------------------------
rasbold@853 1473 // push blocks onto the CFG list
rasbold@853 1474 // ensure that blocks have the correct two-way branch sense
rasbold@853 1475 void Trace::fixup_blocks(PhaseCFG &cfg) {
rasbold@853 1476 Block *last = last_block();
rasbold@853 1477 for (Block *b = first_block(); b != NULL; b = next(b)) {
rasbold@853 1478 cfg._blocks.push(b);
rasbold@853 1479 cfg._num_blocks++;
rasbold@853 1480 if (!b->is_connector()) {
rasbold@853 1481 int nfallthru = b->num_fall_throughs();
rasbold@853 1482 if (b != last) {
rasbold@853 1483 if (nfallthru == 2) {
rasbold@853 1484 // Ensure that the sense of the branch is correct
rasbold@853 1485 Block *bnext = next(b);
rasbold@853 1486 Block *bs0 = b->non_connector_successor(0);
rasbold@853 1487
rasbold@853 1488 MachNode *iff = b->_nodes[b->_nodes.size()-3]->as_Mach();
rasbold@853 1489 ProjNode *proj0 = b->_nodes[b->_nodes.size()-2]->as_Proj();
rasbold@853 1490 ProjNode *proj1 = b->_nodes[b->_nodes.size()-1]->as_Proj();
rasbold@853 1491
rasbold@853 1492 if (bnext == bs0) {
rasbold@853 1493 // Fall-thru case in succs[0], should be in succs[1]
rasbold@853 1494
rasbold@853 1495 // Flip targets in _succs map
rasbold@853 1496 Block *tbs0 = b->_succs[0];
rasbold@853 1497 Block *tbs1 = b->_succs[1];
rasbold@853 1498 b->_succs.map( 0, tbs1 );
rasbold@853 1499 b->_succs.map( 1, tbs0 );
rasbold@853 1500
rasbold@853 1501 // Flip projections to match targets
rasbold@853 1502 b->_nodes.map(b->_nodes.size()-2, proj1);
rasbold@853 1503 b->_nodes.map(b->_nodes.size()-1, proj0);
rasbold@853 1504 }
rasbold@853 1505 }
rasbold@853 1506 }
rasbold@853 1507 }
rasbold@853 1508 }
rasbold@853 1509 }

mercurial