1 /*
2 * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "libadt/vectset.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "opto/block.hpp"
29 #include "opto/cfgnode.hpp"
30 #include "opto/chaitin.hpp"
31 #include "opto/loopnode.hpp"
32 #include "opto/machnode.hpp"
33 #include "opto/matcher.hpp"
34 #include "opto/opcodes.hpp"
35 #include "opto/rootnode.hpp"
36 #include "utilities/copy.hpp"
37
38 void Block_Array::grow( uint i ) {
39 assert(i >= Max(), "must be an overflow");
40 debug_only(_limit = i+1);
41 if( i < _size ) return;
42 if( !_size ) {
43 _size = 1;
44 _blocks = (Block**)_arena->Amalloc( _size * sizeof(Block*) );
45 _blocks[0] = NULL;
46 }
47 uint old = _size;
48 while( i >= _size ) _size <<= 1; // Double to fit
49 _blocks = (Block**)_arena->Arealloc( _blocks, old*sizeof(Block*),_size*sizeof(Block*));
50 Copy::zero_to_bytes( &_blocks[old], (_size-old)*sizeof(Block*) );
51 }
52
53 void Block_List::remove(uint i) {
54 assert(i < _cnt, "index out of bounds");
55 Copy::conjoint_words_to_lower((HeapWord*)&_blocks[i+1], (HeapWord*)&_blocks[i], ((_cnt-i-1)*sizeof(Block*)));
56 pop(); // shrink list by one block
57 }
58
59 void Block_List::insert(uint i, Block *b) {
60 push(b); // grow list by one block
61 Copy::conjoint_words_to_higher((HeapWord*)&_blocks[i], (HeapWord*)&_blocks[i+1], ((_cnt-i-1)*sizeof(Block*)));
62 _blocks[i] = b;
63 }
64
65 #ifndef PRODUCT
66 void Block_List::print() {
67 for (uint i=0; i < size(); i++) {
68 tty->print("B%d ", _blocks[i]->_pre_order);
69 }
70 tty->print("size = %d\n", size());
71 }
72 #endif
73
74 uint Block::code_alignment() {
75 // Check for Root block
76 if (_pre_order == 0) return CodeEntryAlignment;
77 // Check for Start block
78 if (_pre_order == 1) return InteriorEntryAlignment;
79 // Check for loop alignment
80 if (has_loop_alignment()) return loop_alignment();
81
82 return relocInfo::addr_unit(); // no particular alignment
83 }
84
85 uint Block::compute_loop_alignment() {
86 Node *h = head();
87 int unit_sz = relocInfo::addr_unit();
88 if (h->is_Loop() && h->as_Loop()->is_inner_loop()) {
89 // Pre- and post-loops have low trip count so do not bother with
90 // NOPs for align loop head. The constants are hidden from tuning
91 // but only because my "divide by 4" heuristic surely gets nearly
92 // all possible gain (a "do not align at all" heuristic has a
93 // chance of getting a really tiny gain).
94 if (h->is_CountedLoop() && (h->as_CountedLoop()->is_pre_loop() ||
95 h->as_CountedLoop()->is_post_loop())) {
96 return (OptoLoopAlignment > 4*unit_sz) ? (OptoLoopAlignment>>2) : unit_sz;
97 }
98 // Loops with low backedge frequency should not be aligned.
99 Node *n = h->in(LoopNode::LoopBackControl)->in(0);
100 if (n->is_MachIf() && n->as_MachIf()->_prob < 0.01) {
101 return unit_sz; // Loop does not loop, more often than not!
102 }
103 return OptoLoopAlignment; // Otherwise align loop head
104 }
105
106 return unit_sz; // no particular alignment
107 }
108
109 // Compute the size of first 'inst_cnt' instructions in this block.
110 // Return the number of instructions left to compute if the block has
111 // less then 'inst_cnt' instructions. Stop, and return 0 if sum_size
112 // exceeds OptoLoopAlignment.
113 uint Block::compute_first_inst_size(uint& sum_size, uint inst_cnt,
114 PhaseRegAlloc* ra) {
115 uint last_inst = number_of_nodes();
116 for( uint j = 0; j < last_inst && inst_cnt > 0; j++ ) {
117 uint inst_size = get_node(j)->size(ra);
118 if( inst_size > 0 ) {
119 inst_cnt--;
120 uint sz = sum_size + inst_size;
121 if( sz <= (uint)OptoLoopAlignment ) {
122 // Compute size of instructions which fit into fetch buffer only
123 // since all inst_cnt instructions will not fit even if we align them.
124 sum_size = sz;
125 } else {
126 return 0;
127 }
128 }
129 }
130 return inst_cnt;
131 }
132
133 uint Block::find_node( const Node *n ) const {
134 for( uint i = 0; i < number_of_nodes(); i++ ) {
135 if( get_node(i) == n )
136 return i;
137 }
138 ShouldNotReachHere();
139 return 0;
140 }
141
142 // Find and remove n from block list
143 void Block::find_remove( const Node *n ) {
144 remove_node(find_node(n));
145 }
146
147 bool Block::contains(const Node *n) const {
148 return _nodes.contains(n);
149 }
150
151 // Return empty status of a block. Empty blocks contain only the head, other
152 // ideal nodes, and an optional trailing goto.
153 int Block::is_Empty() const {
154
155 // Root or start block is not considered empty
156 if (head()->is_Root() || head()->is_Start()) {
157 return not_empty;
158 }
159
160 int success_result = completely_empty;
161 int end_idx = number_of_nodes() - 1;
162
163 // Check for ending goto
164 if ((end_idx > 0) && (get_node(end_idx)->is_MachGoto())) {
165 success_result = empty_with_goto;
166 end_idx--;
167 }
168
169 // Unreachable blocks are considered empty
170 if (num_preds() <= 1) {
171 return success_result;
172 }
173
174 // Ideal nodes are allowable in empty blocks: skip them Only MachNodes
175 // turn directly into code, because only MachNodes have non-trivial
176 // emit() functions.
177 while ((end_idx > 0) && !get_node(end_idx)->is_Mach()) {
178 end_idx--;
179 }
180
181 // No room for any interesting instructions?
182 if (end_idx == 0) {
183 return success_result;
184 }
185
186 return not_empty;
187 }
188
189 // Return true if the block's code implies that it is likely to be
190 // executed infrequently. Check to see if the block ends in a Halt or
191 // a low probability call.
192 bool Block::has_uncommon_code() const {
193 Node* en = end();
194
195 if (en->is_MachGoto())
196 en = en->in(0);
197 if (en->is_Catch())
198 en = en->in(0);
199 if (en->is_MachProj() && en->in(0)->is_MachCall()) {
200 MachCallNode* call = en->in(0)->as_MachCall();
201 if (call->cnt() != COUNT_UNKNOWN && call->cnt() <= PROB_UNLIKELY_MAG(4)) {
202 // This is true for slow-path stubs like new_{instance,array},
203 // slow_arraycopy, complete_monitor_locking, uncommon_trap.
204 // The magic number corresponds to the probability of an uncommon_trap,
205 // even though it is a count not a probability.
206 return true;
207 }
208 }
209
210 int op = en->is_Mach() ? en->as_Mach()->ideal_Opcode() : en->Opcode();
211 return op == Op_Halt;
212 }
213
214 // True if block is low enough frequency or guarded by a test which
215 // mostly does not go here.
216 bool PhaseCFG::is_uncommon(const Block* block) {
217 // Initial blocks must never be moved, so are never uncommon.
218 if (block->head()->is_Root() || block->head()->is_Start()) return false;
219
220 // Check for way-low freq
221 if(block->_freq < BLOCK_FREQUENCY(0.00001f) ) return true;
222
223 // Look for code shape indicating uncommon_trap or slow path
224 if (block->has_uncommon_code()) return true;
225
226 const float epsilon = 0.05f;
227 const float guard_factor = PROB_UNLIKELY_MAG(4) / (1.f - epsilon);
228 uint uncommon_preds = 0;
229 uint freq_preds = 0;
230 uint uncommon_for_freq_preds = 0;
231
232 for( uint i=1; i< block->num_preds(); i++ ) {
233 Block* guard = get_block_for_node(block->pred(i));
234 // Check to see if this block follows its guard 1 time out of 10000
235 // or less.
236 //
237 // See list of magnitude-4 unlikely probabilities in cfgnode.hpp which
238 // we intend to be "uncommon", such as slow-path TLE allocation,
239 // predicted call failure, and uncommon trap triggers.
240 //
241 // Use an epsilon value of 5% to allow for variability in frequency
242 // predictions and floating point calculations. The net effect is
243 // that guard_factor is set to 9500.
244 //
245 // Ignore low-frequency blocks.
246 // The next check is (guard->_freq < 1.e-5 * 9500.).
247 if(guard->_freq*BLOCK_FREQUENCY(guard_factor) < BLOCK_FREQUENCY(0.00001f)) {
248 uncommon_preds++;
249 } else {
250 freq_preds++;
251 if(block->_freq < guard->_freq * guard_factor ) {
252 uncommon_for_freq_preds++;
253 }
254 }
255 }
256 if( block->num_preds() > 1 &&
257 // The block is uncommon if all preds are uncommon or
258 (uncommon_preds == (block->num_preds()-1) ||
259 // it is uncommon for all frequent preds.
260 uncommon_for_freq_preds == freq_preds) ) {
261 return true;
262 }
263 return false;
264 }
265
266 #ifndef PRODUCT
267 void Block::dump_bidx(const Block* orig, outputStream* st) const {
268 if (_pre_order) st->print("B%d",_pre_order);
269 else st->print("N%d", head()->_idx);
270
271 if (Verbose && orig != this) {
272 // Dump the original block's idx
273 st->print(" (");
274 orig->dump_bidx(orig, st);
275 st->print(")");
276 }
277 }
278
279 void Block::dump_pred(const PhaseCFG* cfg, Block* orig, outputStream* st) const {
280 if (is_connector()) {
281 for (uint i=1; i<num_preds(); i++) {
282 Block *p = cfg->get_block_for_node(pred(i));
283 p->dump_pred(cfg, orig, st);
284 }
285 } else {
286 dump_bidx(orig, st);
287 st->print(" ");
288 }
289 }
290
291 void Block::dump_head(const PhaseCFG* cfg, outputStream* st) const {
292 // Print the basic block
293 dump_bidx(this, st);
294 st->print(": #\t");
295
296 // Print the incoming CFG edges and the outgoing CFG edges
297 for( uint i=0; i<_num_succs; i++ ) {
298 non_connector_successor(i)->dump_bidx(_succs[i], st);
299 st->print(" ");
300 }
301 st->print("<- ");
302 if( head()->is_block_start() ) {
303 for (uint i=1; i<num_preds(); i++) {
304 Node *s = pred(i);
305 if (cfg != NULL) {
306 Block *p = cfg->get_block_for_node(s);
307 p->dump_pred(cfg, p, st);
308 } else {
309 while (!s->is_block_start())
310 s = s->in(0);
311 st->print("N%d ", s->_idx );
312 }
313 }
314 } else {
315 st->print("BLOCK HEAD IS JUNK ");
316 }
317
318 // Print loop, if any
319 const Block *bhead = this; // Head of self-loop
320 Node *bh = bhead->head();
321
322 if ((cfg != NULL) && bh->is_Loop() && !head()->is_Root()) {
323 LoopNode *loop = bh->as_Loop();
324 const Block *bx = cfg->get_block_for_node(loop->in(LoopNode::LoopBackControl));
325 while (bx->is_connector()) {
326 bx = cfg->get_block_for_node(bx->pred(1));
327 }
328 st->print("\tLoop: B%d-B%d ", bhead->_pre_order, bx->_pre_order);
329 // Dump any loop-specific bits, especially for CountedLoops.
330 loop->dump_spec(st);
331 } else if (has_loop_alignment()) {
332 st->print(" top-of-loop");
333 }
334 st->print(" Freq: %g",_freq);
335 if( Verbose || WizardMode ) {
336 st->print(" IDom: %d/#%d", _idom ? _idom->_pre_order : 0, _dom_depth);
337 st->print(" RegPressure: %d",_reg_pressure);
338 st->print(" IHRP Index: %d",_ihrp_index);
339 st->print(" FRegPressure: %d",_freg_pressure);
340 st->print(" FHRP Index: %d",_fhrp_index);
341 }
342 st->print_cr("");
343 }
344
345 void Block::dump() const {
346 dump(NULL);
347 }
348
349 void Block::dump(const PhaseCFG* cfg) const {
350 dump_head(cfg);
351 for (uint i=0; i< number_of_nodes(); i++) {
352 get_node(i)->dump();
353 }
354 tty->print("\n");
355 }
356 #endif
357
358 PhaseCFG::PhaseCFG(Arena* arena, RootNode* root, Matcher& matcher)
359 : Phase(CFG)
360 , _block_arena(arena)
361 , _root(root)
362 , _matcher(matcher)
363 , _node_to_block_mapping(arena)
364 , _node_latency(NULL)
365 #ifndef PRODUCT
366 , _trace_opto_pipelining(TraceOptoPipelining || C->method_has_option("TraceOptoPipelining"))
367 #endif
368 #ifdef ASSERT
369 , _raw_oops(arena)
370 #endif
371 {
372 ResourceMark rm;
373 // I'll need a few machine-specific GotoNodes. Make an Ideal GotoNode,
374 // then Match it into a machine-specific Node. Then clone the machine
375 // Node on demand.
376 Node *x = new (C) GotoNode(NULL);
377 x->init_req(0, x);
378 _goto = matcher.match_tree(x);
379 assert(_goto != NULL, "");
380 _goto->set_req(0,_goto);
381
382 // Build the CFG in Reverse Post Order
383 _number_of_blocks = build_cfg();
384 _root_block = get_block_for_node(_root);
385 }
386
387 // Build a proper looking CFG. Make every block begin with either a StartNode
388 // or a RegionNode. Make every block end with either a Goto, If or Return.
389 // The RootNode both starts and ends it's own block. Do this with a recursive
390 // backwards walk over the control edges.
391 uint PhaseCFG::build_cfg() {
392 Arena *a = Thread::current()->resource_area();
393 VectorSet visited(a);
394
395 // Allocate stack with enough space to avoid frequent realloc
396 Node_Stack nstack(a, C->unique() >> 1);
397 nstack.push(_root, 0);
398 uint sum = 0; // Counter for blocks
399
400 while (nstack.is_nonempty()) {
401 // node and in's index from stack's top
402 // 'np' is _root (see above) or RegionNode, StartNode: we push on stack
403 // only nodes which point to the start of basic block (see below).
404 Node *np = nstack.node();
405 // idx > 0, except for the first node (_root) pushed on stack
406 // at the beginning when idx == 0.
407 // We will use the condition (idx == 0) later to end the build.
408 uint idx = nstack.index();
409 Node *proj = np->in(idx);
410 const Node *x = proj->is_block_proj();
411 // Does the block end with a proper block-ending Node? One of Return,
412 // If or Goto? (This check should be done for visited nodes also).
413 if (x == NULL) { // Does not end right...
414 Node *g = _goto->clone(); // Force it to end in a Goto
415 g->set_req(0, proj);
416 np->set_req(idx, g);
417 x = proj = g;
418 }
419 if (!visited.test_set(x->_idx)) { // Visit this block once
420 // Skip any control-pinned middle'in stuff
421 Node *p = proj;
422 do {
423 proj = p; // Update pointer to last Control
424 p = p->in(0); // Move control forward
425 } while( !p->is_block_proj() &&
426 !p->is_block_start() );
427 // Make the block begin with one of Region or StartNode.
428 if( !p->is_block_start() ) {
429 RegionNode *r = new (C) RegionNode( 2 );
430 r->init_req(1, p); // Insert RegionNode in the way
431 proj->set_req(0, r); // Insert RegionNode in the way
432 p = r;
433 }
434 // 'p' now points to the start of this basic block
435
436 // Put self in array of basic blocks
437 Block *bb = new (_block_arena) Block(_block_arena, p);
438 map_node_to_block(p, bb);
439 map_node_to_block(x, bb);
440 if( x != p ) { // Only for root is x == p
441 bb->push_node((Node*)x);
442 }
443 // Now handle predecessors
444 ++sum; // Count 1 for self block
445 uint cnt = bb->num_preds();
446 for (int i = (cnt - 1); i > 0; i-- ) { // For all predecessors
447 Node *prevproj = p->in(i); // Get prior input
448 assert( !prevproj->is_Con(), "dead input not removed" );
449 // Check to see if p->in(i) is a "control-dependent" CFG edge -
450 // i.e., it splits at the source (via an IF or SWITCH) and merges
451 // at the destination (via a many-input Region).
452 // This breaks critical edges. The RegionNode to start the block
453 // will be added when <p,i> is pulled off the node stack
454 if ( cnt > 2 ) { // Merging many things?
455 assert( prevproj== bb->pred(i),"");
456 if(prevproj->is_block_proj() != prevproj) { // Control-dependent edge?
457 // Force a block on the control-dependent edge
458 Node *g = _goto->clone(); // Force it to end in a Goto
459 g->set_req(0,prevproj);
460 p->set_req(i,g);
461 }
462 }
463 nstack.push(p, i); // 'p' is RegionNode or StartNode
464 }
465 } else { // Post-processing visited nodes
466 nstack.pop(); // remove node from stack
467 // Check if it the fist node pushed on stack at the beginning.
468 if (idx == 0) break; // end of the build
469 // Find predecessor basic block
470 Block *pb = get_block_for_node(x);
471 // Insert into nodes array, if not already there
472 if (!has_block(proj)) {
473 assert( x != proj, "" );
474 // Map basic block of projection
475 map_node_to_block(proj, pb);
476 pb->push_node(proj);
477 }
478 // Insert self as a child of my predecessor block
479 pb->_succs.map(pb->_num_succs++, get_block_for_node(np));
480 assert( pb->get_node(pb->number_of_nodes() - pb->_num_succs)->is_block_proj(),
481 "too many control users, not a CFG?" );
482 }
483 }
484 // Return number of basic blocks for all children and self
485 return sum;
486 }
487
488 // Inserts a goto & corresponding basic block between
489 // block[block_no] and its succ_no'th successor block
490 void PhaseCFG::insert_goto_at(uint block_no, uint succ_no) {
491 // get block with block_no
492 assert(block_no < number_of_blocks(), "illegal block number");
493 Block* in = get_block(block_no);
494 // get successor block succ_no
495 assert(succ_no < in->_num_succs, "illegal successor number");
496 Block* out = in->_succs[succ_no];
497 // Compute frequency of the new block. Do this before inserting
498 // new block in case succ_prob() needs to infer the probability from
499 // surrounding blocks.
500 float freq = in->_freq * in->succ_prob(succ_no);
501 // get ProjNode corresponding to the succ_no'th successor of the in block
502 ProjNode* proj = in->get_node(in->number_of_nodes() - in->_num_succs + succ_no)->as_Proj();
503 // create region for basic block
504 RegionNode* region = new (C) RegionNode(2);
505 region->init_req(1, proj);
506 // setup corresponding basic block
507 Block* block = new (_block_arena) Block(_block_arena, region);
508 map_node_to_block(region, block);
509 C->regalloc()->set_bad(region->_idx);
510 // add a goto node
511 Node* gto = _goto->clone(); // get a new goto node
512 gto->set_req(0, region);
513 // add it to the basic block
514 block->push_node(gto);
515 map_node_to_block(gto, block);
516 C->regalloc()->set_bad(gto->_idx);
517 // hook up successor block
518 block->_succs.map(block->_num_succs++, out);
519 // remap successor's predecessors if necessary
520 for (uint i = 1; i < out->num_preds(); i++) {
521 if (out->pred(i) == proj) out->head()->set_req(i, gto);
522 }
523 // remap predecessor's successor to new block
524 in->_succs.map(succ_no, block);
525 // Set the frequency of the new block
526 block->_freq = freq;
527 // add new basic block to basic block list
528 add_block_at(block_no + 1, block);
529 }
530
531 // Does this block end in a multiway branch that cannot have the default case
532 // flipped for another case?
533 static bool no_flip_branch( Block *b ) {
534 int branch_idx = b->number_of_nodes() - b->_num_succs-1;
535 if( branch_idx < 1 ) return false;
536 Node *bra = b->get_node(branch_idx);
537 if( bra->is_Catch() )
538 return true;
539 if( bra->is_Mach() ) {
540 if( bra->is_MachNullCheck() )
541 return true;
542 int iop = bra->as_Mach()->ideal_Opcode();
543 if( iop == Op_FastLock || iop == Op_FastUnlock )
544 return true;
545 }
546 return false;
547 }
548
549 // Check for NeverBranch at block end. This needs to become a GOTO to the
550 // true target. NeverBranch are treated as a conditional branch that always
551 // goes the same direction for most of the optimizer and are used to give a
552 // fake exit path to infinite loops. At this late stage they need to turn
553 // into Goto's so that when you enter the infinite loop you indeed hang.
554 void PhaseCFG::convert_NeverBranch_to_Goto(Block *b) {
555 // Find true target
556 int end_idx = b->end_idx();
557 int idx = b->get_node(end_idx+1)->as_Proj()->_con;
558 Block *succ = b->_succs[idx];
559 Node* gto = _goto->clone(); // get a new goto node
560 gto->set_req(0, b->head());
561 Node *bp = b->get_node(end_idx);
562 b->map_node(gto, end_idx); // Slam over NeverBranch
563 map_node_to_block(gto, b);
564 C->regalloc()->set_bad(gto->_idx);
565 b->pop_node(); // Yank projections
566 b->pop_node(); // Yank projections
567 b->_succs.map(0,succ); // Map only successor
568 b->_num_succs = 1;
569 // remap successor's predecessors if necessary
570 uint j;
571 for( j = 1; j < succ->num_preds(); j++)
572 if( succ->pred(j)->in(0) == bp )
573 succ->head()->set_req(j, gto);
574 // Kill alternate exit path
575 Block *dead = b->_succs[1-idx];
576 for( j = 1; j < dead->num_preds(); j++)
577 if( dead->pred(j)->in(0) == bp )
578 break;
579 // Scan through block, yanking dead path from
580 // all regions and phis.
581 dead->head()->del_req(j);
582 for( int k = 1; dead->get_node(k)->is_Phi(); k++ )
583 dead->get_node(k)->del_req(j);
584 }
585
586 // Helper function to move block bx to the slot following b_index. Return
587 // true if the move is successful, otherwise false
588 bool PhaseCFG::move_to_next(Block* bx, uint b_index) {
589 if (bx == NULL) return false;
590
591 // Return false if bx is already scheduled.
592 uint bx_index = bx->_pre_order;
593 if ((bx_index <= b_index) && (get_block(bx_index) == bx)) {
594 return false;
595 }
596
597 // Find the current index of block bx on the block list
598 bx_index = b_index + 1;
599 while (bx_index < number_of_blocks() && get_block(bx_index) != bx) {
600 bx_index++;
601 }
602 assert(get_block(bx_index) == bx, "block not found");
603
604 // If the previous block conditionally falls into bx, return false,
605 // because moving bx will create an extra jump.
606 for(uint k = 1; k < bx->num_preds(); k++ ) {
607 Block* pred = get_block_for_node(bx->pred(k));
608 if (pred == get_block(bx_index - 1)) {
609 if (pred->_num_succs != 1) {
610 return false;
611 }
612 }
613 }
614
615 // Reinsert bx just past block 'b'
616 _blocks.remove(bx_index);
617 _blocks.insert(b_index + 1, bx);
618 return true;
619 }
620
621 // Move empty and uncommon blocks to the end.
622 void PhaseCFG::move_to_end(Block *b, uint i) {
623 int e = b->is_Empty();
624 if (e != Block::not_empty) {
625 if (e == Block::empty_with_goto) {
626 // Remove the goto, but leave the block.
627 b->pop_node();
628 }
629 // Mark this block as a connector block, which will cause it to be
630 // ignored in certain functions such as non_connector_successor().
631 b->set_connector();
632 }
633 // Move the empty block to the end, and don't recheck.
634 _blocks.remove(i);
635 _blocks.push(b);
636 }
637
638 // Set loop alignment for every block
639 void PhaseCFG::set_loop_alignment() {
640 uint last = number_of_blocks();
641 assert(get_block(0) == get_root_block(), "");
642
643 for (uint i = 1; i < last; i++) {
644 Block* block = get_block(i);
645 if (block->head()->is_Loop()) {
646 block->set_loop_alignment(block);
647 }
648 }
649 }
650
651 // Make empty basic blocks to be "connector" blocks, Move uncommon blocks
652 // to the end.
653 void PhaseCFG::remove_empty_blocks() {
654 // Move uncommon blocks to the end
655 uint last = number_of_blocks();
656 assert(get_block(0) == get_root_block(), "");
657
658 for (uint i = 1; i < last; i++) {
659 Block* block = get_block(i);
660 if (block->is_connector()) {
661 break;
662 }
663
664 // Check for NeverBranch at block end. This needs to become a GOTO to the
665 // true target. NeverBranch are treated as a conditional branch that
666 // always goes the same direction for most of the optimizer and are used
667 // to give a fake exit path to infinite loops. At this late stage they
668 // need to turn into Goto's so that when you enter the infinite loop you
669 // indeed hang.
670 if (block->get_node(block->end_idx())->Opcode() == Op_NeverBranch) {
671 convert_NeverBranch_to_Goto(block);
672 }
673
674 // Look for uncommon blocks and move to end.
675 if (!C->do_freq_based_layout()) {
676 if (is_uncommon(block)) {
677 move_to_end(block, i);
678 last--; // No longer check for being uncommon!
679 if (no_flip_branch(block)) { // Fall-thru case must follow?
680 // Find the fall-thru block
681 block = get_block(i);
682 move_to_end(block, i);
683 last--;
684 }
685 // backup block counter post-increment
686 i--;
687 }
688 }
689 }
690
691 // Move empty blocks to the end
692 last = number_of_blocks();
693 for (uint i = 1; i < last; i++) {
694 Block* block = get_block(i);
695 if (block->is_Empty() != Block::not_empty) {
696 move_to_end(block, i);
697 last--;
698 i--;
699 }
700 } // End of for all blocks
701 }
702
703 // Fix up the final control flow for basic blocks.
704 void PhaseCFG::fixup_flow() {
705 // Fixup final control flow for the blocks. Remove jump-to-next
706 // block. If neither arm of an IF follows the conditional branch, we
707 // have to add a second jump after the conditional. We place the
708 // TRUE branch target in succs[0] for both GOTOs and IFs.
709 for (uint i = 0; i < number_of_blocks(); i++) {
710 Block* block = get_block(i);
711 block->_pre_order = i; // turn pre-order into block-index
712
713 // Connector blocks need no further processing.
714 if (block->is_connector()) {
715 assert((i+1) == number_of_blocks() || get_block(i + 1)->is_connector(), "All connector blocks should sink to the end");
716 continue;
717 }
718 assert(block->is_Empty() != Block::completely_empty, "Empty blocks should be connectors");
719
720 Block* bnext = (i < number_of_blocks() - 1) ? get_block(i + 1) : NULL;
721 Block* bs0 = block->non_connector_successor(0);
722
723 // Check for multi-way branches where I cannot negate the test to
724 // exchange the true and false targets.
725 if (no_flip_branch(block)) {
726 // Find fall through case - if must fall into its target
727 int branch_idx = block->number_of_nodes() - block->_num_succs;
728 for (uint j2 = 0; j2 < block->_num_succs; j2++) {
729 const ProjNode* p = block->get_node(branch_idx + j2)->as_Proj();
730 if (p->_con == 0) {
731 // successor j2 is fall through case
732 if (block->non_connector_successor(j2) != bnext) {
733 // but it is not the next block => insert a goto
734 insert_goto_at(i, j2);
735 }
736 // Put taken branch in slot 0
737 if (j2 == 0 && block->_num_succs == 2) {
738 // Flip targets in succs map
739 Block *tbs0 = block->_succs[0];
740 Block *tbs1 = block->_succs[1];
741 block->_succs.map(0, tbs1);
742 block->_succs.map(1, tbs0);
743 }
744 break;
745 }
746 }
747
748 // Remove all CatchProjs
749 for (uint j = 0; j < block->_num_succs; j++) {
750 block->pop_node();
751 }
752
753 } else if (block->_num_succs == 1) {
754 // Block ends in a Goto?
755 if (bnext == bs0) {
756 // We fall into next block; remove the Goto
757 block->pop_node();
758 }
759
760 } else if(block->_num_succs == 2) { // Block ends in a If?
761 // Get opcode of 1st projection (matches _succs[0])
762 // Note: Since this basic block has 2 exits, the last 2 nodes must
763 // be projections (in any order), the 3rd last node must be
764 // the IfNode (we have excluded other 2-way exits such as
765 // CatchNodes already).
766 MachNode* iff = block->get_node(block->number_of_nodes() - 3)->as_Mach();
767 ProjNode* proj0 = block->get_node(block->number_of_nodes() - 2)->as_Proj();
768 ProjNode* proj1 = block->get_node(block->number_of_nodes() - 1)->as_Proj();
769
770 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1].
771 assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0");
772 assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1");
773
774 Block* bs1 = block->non_connector_successor(1);
775
776 // Check for neither successor block following the current
777 // block ending in a conditional. If so, move one of the
778 // successors after the current one, provided that the
779 // successor was previously unscheduled, but moveable
780 // (i.e., all paths to it involve a branch).
781 if (!C->do_freq_based_layout() && bnext != bs0 && bnext != bs1) {
782 // Choose the more common successor based on the probability
783 // of the conditional branch.
784 Block* bx = bs0;
785 Block* by = bs1;
786
787 // _prob is the probability of taking the true path. Make
788 // p the probability of taking successor #1.
789 float p = iff->as_MachIf()->_prob;
790 if (proj0->Opcode() == Op_IfTrue) {
791 p = 1.0 - p;
792 }
793
794 // Prefer successor #1 if p > 0.5
795 if (p > PROB_FAIR) {
796 bx = bs1;
797 by = bs0;
798 }
799
800 // Attempt the more common successor first
801 if (move_to_next(bx, i)) {
802 bnext = bx;
803 } else if (move_to_next(by, i)) {
804 bnext = by;
805 }
806 }
807
808 // Check for conditional branching the wrong way. Negate
809 // conditional, if needed, so it falls into the following block
810 // and branches to the not-following block.
811
812 // Check for the next block being in succs[0]. We are going to branch
813 // to succs[0], so we want the fall-thru case as the next block in
814 // succs[1].
815 if (bnext == bs0) {
816 // Fall-thru case in succs[0], so flip targets in succs map
817 Block* tbs0 = block->_succs[0];
818 Block* tbs1 = block->_succs[1];
819 block->_succs.map(0, tbs1);
820 block->_succs.map(1, tbs0);
821 // Flip projection for each target
822 ProjNode* tmp = proj0;
823 proj0 = proj1;
824 proj1 = tmp;
825
826 } else if(bnext != bs1) {
827 // Need a double-branch
828 // The existing conditional branch need not change.
829 // Add a unconditional branch to the false target.
830 // Alas, it must appear in its own block and adding a
831 // block this late in the game is complicated. Sigh.
832 insert_goto_at(i, 1);
833 }
834
835 // Make sure we TRUE branch to the target
836 if (proj0->Opcode() == Op_IfFalse) {
837 iff->as_MachIf()->negate();
838 }
839
840 block->pop_node(); // Remove IfFalse & IfTrue projections
841 block->pop_node();
842
843 } else {
844 // Multi-exit block, e.g. a switch statement
845 // But we don't need to do anything here
846 }
847 } // End of for all blocks
848 }
849
850
851 // postalloc_expand: Expand nodes after register allocation.
852 //
853 // postalloc_expand has to be called after register allocation, just
854 // before output (i.e. scheduling). It only gets called if
855 // Matcher::require_postalloc_expand is true.
856 //
857 // Background:
858 //
859 // Nodes that are expandend (one compound node requiring several
860 // assembler instructions to be implemented split into two or more
861 // non-compound nodes) after register allocation are not as nice as
862 // the ones expanded before register allocation - they don't
863 // participate in optimizations as global code motion. But after
864 // register allocation we can expand nodes that use registers which
865 // are not spillable or registers that are not allocated, because the
866 // old compound node is simply replaced (in its location in the basic
867 // block) by a new subgraph which does not contain compound nodes any
868 // more. The scheduler called during output can later on process these
869 // non-compound nodes.
870 //
871 // Implementation:
872 //
873 // Nodes requiring postalloc expand are specified in the ad file by using
874 // a postalloc_expand statement instead of ins_encode. A postalloc_expand
875 // contains a single call to an encoding, as does an ins_encode
876 // statement. Instead of an emit() function a postalloc_expand() function
877 // is generated that doesn't emit assembler but creates a new
878 // subgraph. The code below calls this postalloc_expand function for each
879 // node with the appropriate attribute. This function returns the new
880 // nodes generated in an array passed in the call. The old node,
881 // potential MachTemps before and potential Projs after it then get
882 // disconnected and replaced by the new nodes. The instruction
883 // generating the result has to be the last one in the array. In
884 // general it is assumed that Projs after the node expanded are
885 // kills. These kills are not required any more after expanding as
886 // there are now explicitly visible def-use chains and the Projs are
887 // removed. This does not hold for calls: They do not only have
888 // kill-Projs but also Projs defining values. Therefore Projs after
889 // the node expanded are removed for all but for calls. If a node is
890 // to be reused, it must be added to the nodes list returned, and it
891 // will be added again.
892 //
893 // Implementing the postalloc_expand function for a node in an enc_class
894 // is rather tedious. It requires knowledge about many node details, as
895 // the nodes and the subgraph must be hand crafted. To simplify this,
896 // adlc generates some utility variables into the postalloc_expand function,
897 // e.g., holding the operands as specified by the postalloc_expand encoding
898 // specification, e.g.:
899 // * unsigned idx_<par_name> holding the index of the node in the ins
900 // * Node *n_<par_name> holding the node loaded from the ins
901 // * MachOpnd *op_<par_name> holding the corresponding operand
902 //
903 // The ordering of operands can not be determined by looking at a
904 // rule. Especially if a match rule matches several different trees,
905 // several nodes are generated from one instruct specification with
906 // different operand orderings. In this case the adlc generated
907 // variables are the only way to access the ins and operands
908 // deterministically.
909 //
910 // If assigning a register to a node that contains an oop, don't
911 // forget to call ra_->set_oop() for the node.
912 void PhaseCFG::postalloc_expand(PhaseRegAlloc* _ra) {
913 GrowableArray <Node *> new_nodes(32); // Array with new nodes filled by postalloc_expand function of node.
914 GrowableArray <Node *> remove(32);
915 GrowableArray <Node *> succs(32);
916 unsigned int max_idx = C->unique(); // Remember to distinguish new from old nodes.
917 DEBUG_ONLY(bool foundNode = false);
918
919 // for all blocks
920 for (uint i = 0; i < number_of_blocks(); i++) {
921 Block *b = _blocks[i];
922 // For all instructions in the current block.
923 for (uint j = 0; j < b->number_of_nodes(); j++) {
924 Node *n = b->get_node(j);
925 if (n->is_Mach() && n->as_Mach()->requires_postalloc_expand()) {
926 #ifdef ASSERT
927 if (TracePostallocExpand) {
928 if (!foundNode) {
929 foundNode = true;
930 tty->print("POSTALLOC EXPANDING %d %s\n", C->compile_id(),
931 C->method() ? C->method()->name()->as_utf8() : C->stub_name());
932 }
933 tty->print(" postalloc expanding "); n->dump();
934 if (Verbose) {
935 tty->print(" with ins:\n");
936 for (uint k = 0; k < n->len(); ++k) {
937 if (n->in(k)) { tty->print(" "); n->in(k)->dump(); }
938 }
939 }
940 }
941 #endif
942 new_nodes.clear();
943 // Collect nodes that have to be removed from the block later on.
944 uint req = n->req();
945 remove.clear();
946 for (uint k = 0; k < req; ++k) {
947 if (n->in(k) && n->in(k)->is_MachTemp()) {
948 remove.push(n->in(k)); // MachTemps which are inputs to the old node have to be removed.
949 n->in(k)->del_req(0);
950 j--;
951 }
952 }
953
954 // Check whether we can allocate enough nodes. We set a fix limit for
955 // the size of postalloc expands with this.
956 uint unique_limit = C->unique() + 40;
957 if (unique_limit >= _ra->node_regs_max_index()) {
958 Compile::current()->record_failure("out of nodes in postalloc expand");
959 return;
960 }
961
962 // Emit (i.e. generate new nodes).
963 n->as_Mach()->postalloc_expand(&new_nodes, _ra);
964
965 assert(C->unique() < unique_limit, "You allocated too many nodes in your postalloc expand.");
966
967 // Disconnect the inputs of the old node.
968 //
969 // We reuse MachSpillCopy nodes. If we need to expand them, there
970 // are many, so reusing pays off. If reused, the node already
971 // has the new ins. n must be the last node on new_nodes list.
972 if (!n->is_MachSpillCopy()) {
973 for (int k = req - 1; k >= 0; --k) {
974 n->del_req(k);
975 }
976 }
977
978 #ifdef ASSERT
979 // Check that all nodes have proper operands.
980 for (int k = 0; k < new_nodes.length(); ++k) {
981 if (new_nodes.at(k)->_idx < max_idx || !new_nodes.at(k)->is_Mach()) continue; // old node, Proj ...
982 MachNode *m = new_nodes.at(k)->as_Mach();
983 for (unsigned int l = 0; l < m->num_opnds(); ++l) {
984 if (MachOper::notAnOper(m->_opnds[l])) {
985 outputStream *os = tty;
986 os->print("Node %s ", m->Name());
987 os->print("has invalid opnd %d: %p\n", l, m->_opnds[l]);
988 assert(0, "Invalid operands, see inline trace in hs_err_pid file.");
989 }
990 }
991 }
992 #endif
993
994 // Collect succs of old node in remove (for projections) and in succs (for
995 // all other nodes) do _not_ collect projections in remove (but in succs)
996 // in case the node is a call. We need the projections for calls as they are
997 // associated with registes (i.e. they are defs).
998 succs.clear();
999 for (DUIterator k = n->outs(); n->has_out(k); k++) {
1000 if (n->out(k)->is_Proj() && !n->is_MachCall() && !n->is_MachBranch()) {
1001 remove.push(n->out(k));
1002 } else {
1003 succs.push(n->out(k));
1004 }
1005 }
1006 // Replace old node n as input of its succs by last of the new nodes.
1007 for (int k = 0; k < succs.length(); ++k) {
1008 Node *succ = succs.at(k);
1009 for (uint l = 0; l < succ->req(); ++l) {
1010 if (succ->in(l) == n) {
1011 succ->set_req(l, new_nodes.at(new_nodes.length() - 1));
1012 }
1013 }
1014 for (uint l = succ->req(); l < succ->len(); ++l) {
1015 if (succ->in(l) == n) {
1016 succ->set_prec(l, new_nodes.at(new_nodes.length() - 1));
1017 }
1018 }
1019 }
1020
1021 // Index of old node in block.
1022 uint index = b->find_node(n);
1023 // Insert new nodes into block and map them in nodes->blocks array
1024 // and remember last node in n2.
1025 Node *n2 = NULL;
1026 for (int k = 0; k < new_nodes.length(); ++k) {
1027 n2 = new_nodes.at(k);
1028 b->insert_node(n2, ++index);
1029 map_node_to_block(n2, b);
1030 }
1031
1032 // Add old node n to remove and remove them all from block.
1033 remove.push(n);
1034 j--;
1035 #ifdef ASSERT
1036 if (TracePostallocExpand && Verbose) {
1037 tty->print(" removing:\n");
1038 for (int k = 0; k < remove.length(); ++k) {
1039 tty->print(" "); remove.at(k)->dump();
1040 }
1041 tty->print(" inserting:\n");
1042 for (int k = 0; k < new_nodes.length(); ++k) {
1043 tty->print(" "); new_nodes.at(k)->dump();
1044 }
1045 }
1046 #endif
1047 for (int k = 0; k < remove.length(); ++k) {
1048 if (b->contains(remove.at(k))) {
1049 b->find_remove(remove.at(k));
1050 } else {
1051 assert(remove.at(k)->is_Proj() && (remove.at(k)->in(0)->is_MachBranch()), "");
1052 }
1053 }
1054 // If anything has been inserted (n2 != NULL), continue after last node inserted.
1055 // This does not always work. Some postalloc expands don't insert any nodes, if they
1056 // do optimizations (e.g., max(x,x)). In this case we decrement j accordingly.
1057 j = n2 ? b->find_node(n2) : j;
1058 }
1059 }
1060 }
1061
1062 #ifdef ASSERT
1063 if (foundNode) {
1064 tty->print("FINISHED %d %s\n", C->compile_id(),
1065 C->method() ? C->method()->name()->as_utf8() : C->stub_name());
1066 tty->flush();
1067 }
1068 #endif
1069 }
1070
1071
1072 //------------------------------dump-------------------------------------------
1073 #ifndef PRODUCT
1074 void PhaseCFG::_dump_cfg( const Node *end, VectorSet &visited ) const {
1075 const Node *x = end->is_block_proj();
1076 assert( x, "not a CFG" );
1077
1078 // Do not visit this block again
1079 if( visited.test_set(x->_idx) ) return;
1080
1081 // Skip through this block
1082 const Node *p = x;
1083 do {
1084 p = p->in(0); // Move control forward
1085 assert( !p->is_block_proj() || p->is_Root(), "not a CFG" );
1086 } while( !p->is_block_start() );
1087
1088 // Recursively visit
1089 for (uint i = 1; i < p->req(); i++) {
1090 _dump_cfg(p->in(i), visited);
1091 }
1092
1093 // Dump the block
1094 get_block_for_node(p)->dump(this);
1095 }
1096
1097 void PhaseCFG::dump( ) const {
1098 tty->print("\n--- CFG --- %d BBs\n", number_of_blocks());
1099 if (_blocks.size()) { // Did we do basic-block layout?
1100 for (uint i = 0; i < number_of_blocks(); i++) {
1101 const Block* block = get_block(i);
1102 block->dump(this);
1103 }
1104 } else { // Else do it with a DFS
1105 VectorSet visited(_block_arena);
1106 _dump_cfg(_root,visited);
1107 }
1108 }
1109
1110 void PhaseCFG::dump_headers() {
1111 for (uint i = 0; i < number_of_blocks(); i++) {
1112 Block* block = get_block(i);
1113 if (block != NULL) {
1114 block->dump_head(this);
1115 }
1116 }
1117 }
1118
1119 void PhaseCFG::verify() const {
1120 #ifdef ASSERT
1121 // Verify sane CFG
1122 for (uint i = 0; i < number_of_blocks(); i++) {
1123 Block* block = get_block(i);
1124 uint cnt = block->number_of_nodes();
1125 uint j;
1126 for (j = 0; j < cnt; j++) {
1127 Node *n = block->get_node(j);
1128 assert(get_block_for_node(n) == block, "");
1129 if (j >= 1 && n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CreateEx) {
1130 assert(j == 1 || block->get_node(j-1)->is_Phi(), "CreateEx must be first instruction in block");
1131 }
1132 for (uint k = 0; k < n->req(); k++) {
1133 Node *def = n->in(k);
1134 if (def && def != n) {
1135 assert(get_block_for_node(def) || def->is_Con(), "must have block; constants for debug info ok");
1136 // Verify that instructions in the block is in correct order.
1137 // Uses must follow their definition if they are at the same block.
1138 // Mostly done to check that MachSpillCopy nodes are placed correctly
1139 // when CreateEx node is moved in build_ifg_physical().
1140 if (get_block_for_node(def) == block && !(block->head()->is_Loop() && n->is_Phi()) &&
1141 // See (+++) comment in reg_split.cpp
1142 !(n->jvms() != NULL && n->jvms()->is_monitor_use(k))) {
1143 bool is_loop = false;
1144 if (n->is_Phi()) {
1145 for (uint l = 1; l < def->req(); l++) {
1146 if (n == def->in(l)) {
1147 is_loop = true;
1148 break; // Some kind of loop
1149 }
1150 }
1151 }
1152 assert(is_loop || block->find_node(def) < j, "uses must follow definitions");
1153 }
1154 }
1155 }
1156 }
1157
1158 j = block->end_idx();
1159 Node* bp = (Node*)block->get_node(block->number_of_nodes() - 1)->is_block_proj();
1160 assert(bp, "last instruction must be a block proj");
1161 assert(bp == block->get_node(j), "wrong number of successors for this block");
1162 if (bp->is_Catch()) {
1163 while (block->get_node(--j)->is_MachProj()) {
1164 ;
1165 }
1166 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1167 } else if (bp->is_Mach() && bp->as_Mach()->ideal_Opcode() == Op_If) {
1168 assert(block->_num_succs == 2, "Conditional branch must have two targets");
1169 }
1170 }
1171 #endif
1172 }
1173 #endif
1174
1175 UnionFind::UnionFind( uint max ) : _cnt(max), _max(max), _indices(NEW_RESOURCE_ARRAY(uint,max)) {
1176 Copy::zero_to_bytes( _indices, sizeof(uint)*max );
1177 }
1178
1179 void UnionFind::extend( uint from_idx, uint to_idx ) {
1180 _nesting.check();
1181 if( from_idx >= _max ) {
1182 uint size = 16;
1183 while( size <= from_idx ) size <<=1;
1184 _indices = REALLOC_RESOURCE_ARRAY( uint, _indices, _max, size );
1185 _max = size;
1186 }
1187 while( _cnt <= from_idx ) _indices[_cnt++] = 0;
1188 _indices[from_idx] = to_idx;
1189 }
1190
1191 void UnionFind::reset( uint max ) {
1192 assert( max <= max_uint, "Must fit within uint" );
1193 // Force the Union-Find mapping to be at least this large
1194 extend(max,0);
1195 // Initialize to be the ID mapping.
1196 for( uint i=0; i<max; i++ ) map(i,i);
1197 }
1198
1199 // Straight out of Tarjan's union-find algorithm
1200 uint UnionFind::Find_compress( uint idx ) {
1201 uint cur = idx;
1202 uint next = lookup(cur);
1203 while( next != cur ) { // Scan chain of equivalences
1204 assert( next < cur, "always union smaller" );
1205 cur = next; // until find a fixed-point
1206 next = lookup(cur);
1207 }
1208 // Core of union-find algorithm: update chain of
1209 // equivalences to be equal to the root.
1210 while( idx != next ) {
1211 uint tmp = lookup(idx);
1212 map(idx, next);
1213 idx = tmp;
1214 }
1215 return idx;
1216 }
1217
1218 // Like Find above, but no path compress, so bad asymptotic behavior
1219 uint UnionFind::Find_const( uint idx ) const {
1220 if( idx == 0 ) return idx; // Ignore the zero idx
1221 // Off the end? This can happen during debugging dumps
1222 // when data structures have not finished being updated.
1223 if( idx >= _max ) return idx;
1224 uint next = lookup(idx);
1225 while( next != idx ) { // Scan chain of equivalences
1226 idx = next; // until find a fixed-point
1227 next = lookup(idx);
1228 }
1229 return next;
1230 }
1231
1232 // union 2 sets together.
1233 void UnionFind::Union( uint idx1, uint idx2 ) {
1234 uint src = Find(idx1);
1235 uint dst = Find(idx2);
1236 assert( src, "" );
1237 assert( dst, "" );
1238 assert( src < _max, "oob" );
1239 assert( dst < _max, "oob" );
1240 assert( src < dst, "always union smaller" );
1241 map(dst,src);
1242 }
1243
1244 #ifndef PRODUCT
1245 void Trace::dump( ) const {
1246 tty->print_cr("Trace (freq %f)", first_block()->_freq);
1247 for (Block *b = first_block(); b != NULL; b = next(b)) {
1248 tty->print(" B%d", b->_pre_order);
1249 if (b->head()->is_Loop()) {
1250 tty->print(" (L%d)", b->compute_loop_alignment());
1251 }
1252 if (b->has_loop_alignment()) {
1253 tty->print(" (T%d)", b->code_alignment());
1254 }
1255 }
1256 tty->cr();
1257 }
1258
1259 void CFGEdge::dump( ) const {
1260 tty->print(" B%d --> B%d Freq: %f out:%3d%% in:%3d%% State: ",
1261 from()->_pre_order, to()->_pre_order, freq(), _from_pct, _to_pct);
1262 switch(state()) {
1263 case connected:
1264 tty->print("connected");
1265 break;
1266 case open:
1267 tty->print("open");
1268 break;
1269 case interior:
1270 tty->print("interior");
1271 break;
1272 }
1273 if (infrequent()) {
1274 tty->print(" infrequent");
1275 }
1276 tty->cr();
1277 }
1278 #endif
1279
1280 // Comparison function for edges
1281 static int edge_order(CFGEdge **e0, CFGEdge **e1) {
1282 float freq0 = (*e0)->freq();
1283 float freq1 = (*e1)->freq();
1284 if (freq0 != freq1) {
1285 return freq0 > freq1 ? -1 : 1;
1286 }
1287
1288 int dist0 = (*e0)->to()->_rpo - (*e0)->from()->_rpo;
1289 int dist1 = (*e1)->to()->_rpo - (*e1)->from()->_rpo;
1290
1291 return dist1 - dist0;
1292 }
1293
1294 // Comparison function for edges
1295 extern "C" int trace_frequency_order(const void *p0, const void *p1) {
1296 Trace *tr0 = *(Trace **) p0;
1297 Trace *tr1 = *(Trace **) p1;
1298 Block *b0 = tr0->first_block();
1299 Block *b1 = tr1->first_block();
1300
1301 // The trace of connector blocks goes at the end;
1302 // we only expect one such trace
1303 if (b0->is_connector() != b1->is_connector()) {
1304 return b1->is_connector() ? -1 : 1;
1305 }
1306
1307 // Pull more frequently executed blocks to the beginning
1308 float freq0 = b0->_freq;
1309 float freq1 = b1->_freq;
1310 if (freq0 != freq1) {
1311 return freq0 > freq1 ? -1 : 1;
1312 }
1313
1314 int diff = tr0->first_block()->_rpo - tr1->first_block()->_rpo;
1315
1316 return diff;
1317 }
1318
1319 // Find edges of interest, i.e, those which can fall through. Presumes that
1320 // edges which don't fall through are of low frequency and can be generally
1321 // ignored. Initialize the list of traces.
1322 void PhaseBlockLayout::find_edges() {
1323 // Walk the blocks, creating edges and Traces
1324 uint i;
1325 Trace *tr = NULL;
1326 for (i = 0; i < _cfg.number_of_blocks(); i++) {
1327 Block* b = _cfg.get_block(i);
1328 tr = new Trace(b, next, prev);
1329 traces[tr->id()] = tr;
1330
1331 // All connector blocks should be at the end of the list
1332 if (b->is_connector()) break;
1333
1334 // If this block and the next one have a one-to-one successor
1335 // predecessor relationship, simply append the next block
1336 int nfallthru = b->num_fall_throughs();
1337 while (nfallthru == 1 &&
1338 b->succ_fall_through(0)) {
1339 Block *n = b->_succs[0];
1340
1341 // Skip over single-entry connector blocks, we don't want to
1342 // add them to the trace.
1343 while (n->is_connector() && n->num_preds() == 1) {
1344 n = n->_succs[0];
1345 }
1346
1347 // We see a merge point, so stop search for the next block
1348 if (n->num_preds() != 1) break;
1349
1350 i++;
1351 assert(n = _cfg.get_block(i), "expecting next block");
1352 tr->append(n);
1353 uf->map(n->_pre_order, tr->id());
1354 traces[n->_pre_order] = NULL;
1355 nfallthru = b->num_fall_throughs();
1356 b = n;
1357 }
1358
1359 if (nfallthru > 0) {
1360 // Create a CFGEdge for each outgoing
1361 // edge that could be a fall-through.
1362 for (uint j = 0; j < b->_num_succs; j++ ) {
1363 if (b->succ_fall_through(j)) {
1364 Block *target = b->non_connector_successor(j);
1365 float freq = b->_freq * b->succ_prob(j);
1366 int from_pct = (int) ((100 * freq) / b->_freq);
1367 int to_pct = (int) ((100 * freq) / target->_freq);
1368 edges->append(new CFGEdge(b, target, freq, from_pct, to_pct));
1369 }
1370 }
1371 }
1372 }
1373
1374 // Group connector blocks into one trace
1375 for (i++; i < _cfg.number_of_blocks(); i++) {
1376 Block *b = _cfg.get_block(i);
1377 assert(b->is_connector(), "connector blocks at the end");
1378 tr->append(b);
1379 uf->map(b->_pre_order, tr->id());
1380 traces[b->_pre_order] = NULL;
1381 }
1382 }
1383
1384 // Union two traces together in uf, and null out the trace in the list
1385 void PhaseBlockLayout::union_traces(Trace* updated_trace, Trace* old_trace) {
1386 uint old_id = old_trace->id();
1387 uint updated_id = updated_trace->id();
1388
1389 uint lo_id = updated_id;
1390 uint hi_id = old_id;
1391
1392 // If from is greater than to, swap values to meet
1393 // UnionFind guarantee.
1394 if (updated_id > old_id) {
1395 lo_id = old_id;
1396 hi_id = updated_id;
1397
1398 // Fix up the trace ids
1399 traces[lo_id] = traces[updated_id];
1400 updated_trace->set_id(lo_id);
1401 }
1402
1403 // Union the lower with the higher and remove the pointer
1404 // to the higher.
1405 uf->Union(lo_id, hi_id);
1406 traces[hi_id] = NULL;
1407 }
1408
1409 // Append traces together via the most frequently executed edges
1410 void PhaseBlockLayout::grow_traces() {
1411 // Order the edges, and drive the growth of Traces via the most
1412 // frequently executed edges.
1413 edges->sort(edge_order);
1414 for (int i = 0; i < edges->length(); i++) {
1415 CFGEdge *e = edges->at(i);
1416
1417 if (e->state() != CFGEdge::open) continue;
1418
1419 Block *src_block = e->from();
1420 Block *targ_block = e->to();
1421
1422 // Don't grow traces along backedges?
1423 if (!BlockLayoutRotateLoops) {
1424 if (targ_block->_rpo <= src_block->_rpo) {
1425 targ_block->set_loop_alignment(targ_block);
1426 continue;
1427 }
1428 }
1429
1430 Trace *src_trace = trace(src_block);
1431 Trace *targ_trace = trace(targ_block);
1432
1433 // If the edge in question can join two traces at their ends,
1434 // append one trace to the other.
1435 if (src_trace->last_block() == src_block) {
1436 if (src_trace == targ_trace) {
1437 e->set_state(CFGEdge::interior);
1438 if (targ_trace->backedge(e)) {
1439 // Reset i to catch any newly eligible edge
1440 // (Or we could remember the first "open" edge, and reset there)
1441 i = 0;
1442 }
1443 } else if (targ_trace->first_block() == targ_block) {
1444 e->set_state(CFGEdge::connected);
1445 src_trace->append(targ_trace);
1446 union_traces(src_trace, targ_trace);
1447 }
1448 }
1449 }
1450 }
1451
1452 // Embed one trace into another, if the fork or join points are sufficiently
1453 // balanced.
1454 void PhaseBlockLayout::merge_traces(bool fall_thru_only) {
1455 // Walk the edge list a another time, looking at unprocessed edges.
1456 // Fold in diamonds
1457 for (int i = 0; i < edges->length(); i++) {
1458 CFGEdge *e = edges->at(i);
1459
1460 if (e->state() != CFGEdge::open) continue;
1461 if (fall_thru_only) {
1462 if (e->infrequent()) continue;
1463 }
1464
1465 Block *src_block = e->from();
1466 Trace *src_trace = trace(src_block);
1467 bool src_at_tail = src_trace->last_block() == src_block;
1468
1469 Block *targ_block = e->to();
1470 Trace *targ_trace = trace(targ_block);
1471 bool targ_at_start = targ_trace->first_block() == targ_block;
1472
1473 if (src_trace == targ_trace) {
1474 // This may be a loop, but we can't do much about it.
1475 e->set_state(CFGEdge::interior);
1476 continue;
1477 }
1478
1479 if (fall_thru_only) {
1480 // If the edge links the middle of two traces, we can't do anything.
1481 // Mark the edge and continue.
1482 if (!src_at_tail & !targ_at_start) {
1483 continue;
1484 }
1485
1486 // Don't grow traces along backedges?
1487 if (!BlockLayoutRotateLoops && (targ_block->_rpo <= src_block->_rpo)) {
1488 continue;
1489 }
1490
1491 // If both ends of the edge are available, why didn't we handle it earlier?
1492 assert(src_at_tail ^ targ_at_start, "Should have caught this edge earlier.");
1493
1494 if (targ_at_start) {
1495 // Insert the "targ" trace in the "src" trace if the insertion point
1496 // is a two way branch.
1497 // Better profitability check possible, but may not be worth it.
1498 // Someday, see if the this "fork" has an associated "join";
1499 // then make a policy on merging this trace at the fork or join.
1500 // For example, other things being equal, it may be better to place this
1501 // trace at the join point if the "src" trace ends in a two-way, but
1502 // the insertion point is one-way.
1503 assert(src_block->num_fall_throughs() == 2, "unexpected diamond");
1504 e->set_state(CFGEdge::connected);
1505 src_trace->insert_after(src_block, targ_trace);
1506 union_traces(src_trace, targ_trace);
1507 } else if (src_at_tail) {
1508 if (src_trace != trace(_cfg.get_root_block())) {
1509 e->set_state(CFGEdge::connected);
1510 targ_trace->insert_before(targ_block, src_trace);
1511 union_traces(targ_trace, src_trace);
1512 }
1513 }
1514 } else if (e->state() == CFGEdge::open) {
1515 // Append traces, even without a fall-thru connection.
1516 // But leave root entry at the beginning of the block list.
1517 if (targ_trace != trace(_cfg.get_root_block())) {
1518 e->set_state(CFGEdge::connected);
1519 src_trace->append(targ_trace);
1520 union_traces(src_trace, targ_trace);
1521 }
1522 }
1523 }
1524 }
1525
1526 // Order the sequence of the traces in some desirable way, and fixup the
1527 // jumps at the end of each block.
1528 void PhaseBlockLayout::reorder_traces(int count) {
1529 ResourceArea *area = Thread::current()->resource_area();
1530 Trace ** new_traces = NEW_ARENA_ARRAY(area, Trace *, count);
1531 Block_List worklist;
1532 int new_count = 0;
1533
1534 // Compact the traces.
1535 for (int i = 0; i < count; i++) {
1536 Trace *tr = traces[i];
1537 if (tr != NULL) {
1538 new_traces[new_count++] = tr;
1539 }
1540 }
1541
1542 // The entry block should be first on the new trace list.
1543 Trace *tr = trace(_cfg.get_root_block());
1544 assert(tr == new_traces[0], "entry trace misplaced");
1545
1546 // Sort the new trace list by frequency
1547 qsort(new_traces + 1, new_count - 1, sizeof(new_traces[0]), trace_frequency_order);
1548
1549 // Patch up the successor blocks
1550 _cfg.clear_blocks();
1551 for (int i = 0; i < new_count; i++) {
1552 Trace *tr = new_traces[i];
1553 if (tr != NULL) {
1554 tr->fixup_blocks(_cfg);
1555 }
1556 }
1557 }
1558
1559 // Order basic blocks based on frequency
1560 PhaseBlockLayout::PhaseBlockLayout(PhaseCFG &cfg)
1561 : Phase(BlockLayout)
1562 , _cfg(cfg) {
1563 ResourceMark rm;
1564 ResourceArea *area = Thread::current()->resource_area();
1565
1566 // List of traces
1567 int size = _cfg.number_of_blocks() + 1;
1568 traces = NEW_ARENA_ARRAY(area, Trace *, size);
1569 memset(traces, 0, size*sizeof(Trace*));
1570 next = NEW_ARENA_ARRAY(area, Block *, size);
1571 memset(next, 0, size*sizeof(Block *));
1572 prev = NEW_ARENA_ARRAY(area, Block *, size);
1573 memset(prev , 0, size*sizeof(Block *));
1574
1575 // List of edges
1576 edges = new GrowableArray<CFGEdge*>;
1577
1578 // Mapping block index --> block_trace
1579 uf = new UnionFind(size);
1580 uf->reset(size);
1581
1582 // Find edges and create traces.
1583 find_edges();
1584
1585 // Grow traces at their ends via most frequent edges.
1586 grow_traces();
1587
1588 // Merge one trace into another, but only at fall-through points.
1589 // This may make diamonds and other related shapes in a trace.
1590 merge_traces(true);
1591
1592 // Run merge again, allowing two traces to be catenated, even if
1593 // one does not fall through into the other. This appends loosely
1594 // related traces to be near each other.
1595 merge_traces(false);
1596
1597 // Re-order all the remaining traces by frequency
1598 reorder_traces(size);
1599
1600 assert(_cfg.number_of_blocks() >= (uint) (size - 1), "number of blocks can not shrink");
1601 }
1602
1603
1604 // Edge e completes a loop in a trace. If the target block is head of the
1605 // loop, rotate the loop block so that the loop ends in a conditional branch.
1606 bool Trace::backedge(CFGEdge *e) {
1607 bool loop_rotated = false;
1608 Block *src_block = e->from();
1609 Block *targ_block = e->to();
1610
1611 assert(last_block() == src_block, "loop discovery at back branch");
1612 if (first_block() == targ_block) {
1613 if (BlockLayoutRotateLoops && last_block()->num_fall_throughs() < 2) {
1614 // Find the last block in the trace that has a conditional
1615 // branch.
1616 Block *b;
1617 for (b = last_block(); b != NULL; b = prev(b)) {
1618 if (b->num_fall_throughs() == 2) {
1619 break;
1620 }
1621 }
1622
1623 if (b != last_block() && b != NULL) {
1624 loop_rotated = true;
1625
1626 // Rotate the loop by doing two-part linked-list surgery.
1627 append(first_block());
1628 break_loop_after(b);
1629 }
1630 }
1631
1632 // Backbranch to the top of a trace
1633 // Scroll forward through the trace from the targ_block. If we find
1634 // a loop head before another loop top, use the the loop head alignment.
1635 for (Block *b = targ_block; b != NULL; b = next(b)) {
1636 if (b->has_loop_alignment()) {
1637 break;
1638 }
1639 if (b->head()->is_Loop()) {
1640 targ_block = b;
1641 break;
1642 }
1643 }
1644
1645 first_block()->set_loop_alignment(targ_block);
1646
1647 } else {
1648 // Backbranch into the middle of a trace
1649 targ_block->set_loop_alignment(targ_block);
1650 }
1651
1652 return loop_rotated;
1653 }
1654
1655 // push blocks onto the CFG list
1656 // ensure that blocks have the correct two-way branch sense
1657 void Trace::fixup_blocks(PhaseCFG &cfg) {
1658 Block *last = last_block();
1659 for (Block *b = first_block(); b != NULL; b = next(b)) {
1660 cfg.add_block(b);
1661 if (!b->is_connector()) {
1662 int nfallthru = b->num_fall_throughs();
1663 if (b != last) {
1664 if (nfallthru == 2) {
1665 // Ensure that the sense of the branch is correct
1666 Block *bnext = next(b);
1667 Block *bs0 = b->non_connector_successor(0);
1668
1669 MachNode *iff = b->get_node(b->number_of_nodes() - 3)->as_Mach();
1670 ProjNode *proj0 = b->get_node(b->number_of_nodes() - 2)->as_Proj();
1671 ProjNode *proj1 = b->get_node(b->number_of_nodes() - 1)->as_Proj();
1672
1673 if (bnext == bs0) {
1674 // Fall-thru case in succs[0], should be in succs[1]
1675
1676 // Flip targets in _succs map
1677 Block *tbs0 = b->_succs[0];
1678 Block *tbs1 = b->_succs[1];
1679 b->_succs.map( 0, tbs1 );
1680 b->_succs.map( 1, tbs0 );
1681
1682 // Flip projections to match targets
1683 b->map_node(proj1, b->number_of_nodes() - 2);
1684 b->map_node(proj0, b->number_of_nodes() - 1);
1685 }
1686 }
1687 }
1688 }
1689 }
1690 }