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) {
536 return false;
537 }
538 Node *branch = b->get_node(branch_idx);
539 if (branch->is_Catch()) {
540 return true;
541 }
542 if (branch->is_Mach()) {
543 if (branch->is_MachNullCheck()) {
544 return true;
545 }
546 int iop = branch->as_Mach()->ideal_Opcode();
547 if (iop == Op_FastLock || iop == Op_FastUnlock) {
548 return true;
549 }
550 // Don't flip if branch has an implicit check.
551 if (branch->as_Mach()->is_TrapBasedCheckNode()) {
552 return true;
553 }
554 }
555 return false;
556 }
557
558 // Check for NeverBranch at block end. This needs to become a GOTO to the
559 // true target. NeverBranch are treated as a conditional branch that always
560 // goes the same direction for most of the optimizer and are used to give a
561 // fake exit path to infinite loops. At this late stage they need to turn
562 // into Goto's so that when you enter the infinite loop you indeed hang.
563 void PhaseCFG::convert_NeverBranch_to_Goto(Block *b) {
564 // Find true target
565 int end_idx = b->end_idx();
566 int idx = b->get_node(end_idx+1)->as_Proj()->_con;
567 Block *succ = b->_succs[idx];
568 Node* gto = _goto->clone(); // get a new goto node
569 gto->set_req(0, b->head());
570 Node *bp = b->get_node(end_idx);
571 b->map_node(gto, end_idx); // Slam over NeverBranch
572 map_node_to_block(gto, b);
573 C->regalloc()->set_bad(gto->_idx);
574 b->pop_node(); // Yank projections
575 b->pop_node(); // Yank projections
576 b->_succs.map(0,succ); // Map only successor
577 b->_num_succs = 1;
578 // remap successor's predecessors if necessary
579 uint j;
580 for( j = 1; j < succ->num_preds(); j++)
581 if( succ->pred(j)->in(0) == bp )
582 succ->head()->set_req(j, gto);
583 // Kill alternate exit path
584 Block *dead = b->_succs[1-idx];
585 for( j = 1; j < dead->num_preds(); j++)
586 if( dead->pred(j)->in(0) == bp )
587 break;
588 // Scan through block, yanking dead path from
589 // all regions and phis.
590 dead->head()->del_req(j);
591 for( int k = 1; dead->get_node(k)->is_Phi(); k++ )
592 dead->get_node(k)->del_req(j);
593 }
594
595 // Helper function to move block bx to the slot following b_index. Return
596 // true if the move is successful, otherwise false
597 bool PhaseCFG::move_to_next(Block* bx, uint b_index) {
598 if (bx == NULL) return false;
599
600 // Return false if bx is already scheduled.
601 uint bx_index = bx->_pre_order;
602 if ((bx_index <= b_index) && (get_block(bx_index) == bx)) {
603 return false;
604 }
605
606 // Find the current index of block bx on the block list
607 bx_index = b_index + 1;
608 while (bx_index < number_of_blocks() && get_block(bx_index) != bx) {
609 bx_index++;
610 }
611 assert(get_block(bx_index) == bx, "block not found");
612
613 // If the previous block conditionally falls into bx, return false,
614 // because moving bx will create an extra jump.
615 for(uint k = 1; k < bx->num_preds(); k++ ) {
616 Block* pred = get_block_for_node(bx->pred(k));
617 if (pred == get_block(bx_index - 1)) {
618 if (pred->_num_succs != 1) {
619 return false;
620 }
621 }
622 }
623
624 // Reinsert bx just past block 'b'
625 _blocks.remove(bx_index);
626 _blocks.insert(b_index + 1, bx);
627 return true;
628 }
629
630 // Move empty and uncommon blocks to the end.
631 void PhaseCFG::move_to_end(Block *b, uint i) {
632 int e = b->is_Empty();
633 if (e != Block::not_empty) {
634 if (e == Block::empty_with_goto) {
635 // Remove the goto, but leave the block.
636 b->pop_node();
637 }
638 // Mark this block as a connector block, which will cause it to be
639 // ignored in certain functions such as non_connector_successor().
640 b->set_connector();
641 }
642 // Move the empty block to the end, and don't recheck.
643 _blocks.remove(i);
644 _blocks.push(b);
645 }
646
647 // Set loop alignment for every block
648 void PhaseCFG::set_loop_alignment() {
649 uint last = number_of_blocks();
650 assert(get_block(0) == get_root_block(), "");
651
652 for (uint i = 1; i < last; i++) {
653 Block* block = get_block(i);
654 if (block->head()->is_Loop()) {
655 block->set_loop_alignment(block);
656 }
657 }
658 }
659
660 // Make empty basic blocks to be "connector" blocks, Move uncommon blocks
661 // to the end.
662 void PhaseCFG::remove_empty_blocks() {
663 // Move uncommon blocks to the end
664 uint last = number_of_blocks();
665 assert(get_block(0) == get_root_block(), "");
666
667 for (uint i = 1; i < last; i++) {
668 Block* block = get_block(i);
669 if (block->is_connector()) {
670 break;
671 }
672
673 // Check for NeverBranch at block end. This needs to become a GOTO to the
674 // true target. NeverBranch are treated as a conditional branch that
675 // always goes the same direction for most of the optimizer and are used
676 // to give a fake exit path to infinite loops. At this late stage they
677 // need to turn into Goto's so that when you enter the infinite loop you
678 // indeed hang.
679 if (block->get_node(block->end_idx())->Opcode() == Op_NeverBranch) {
680 convert_NeverBranch_to_Goto(block);
681 }
682
683 // Look for uncommon blocks and move to end.
684 if (!C->do_freq_based_layout()) {
685 if (is_uncommon(block)) {
686 move_to_end(block, i);
687 last--; // No longer check for being uncommon!
688 if (no_flip_branch(block)) { // Fall-thru case must follow?
689 // Find the fall-thru block
690 block = get_block(i);
691 move_to_end(block, i);
692 last--;
693 }
694 // backup block counter post-increment
695 i--;
696 }
697 }
698 }
699
700 // Move empty blocks to the end
701 last = number_of_blocks();
702 for (uint i = 1; i < last; i++) {
703 Block* block = get_block(i);
704 if (block->is_Empty() != Block::not_empty) {
705 move_to_end(block, i);
706 last--;
707 i--;
708 }
709 } // End of for all blocks
710 }
711
712 Block *PhaseCFG::fixup_trap_based_check(Node *branch, Block *block, int block_pos, Block *bnext) {
713 // Trap based checks must fall through to the successor with
714 // PROB_ALWAYS.
715 // They should be an If with 2 successors.
716 assert(branch->is_MachIf(), "must be If");
717 assert(block->_num_succs == 2, "must have 2 successors");
718
719 // Get the If node and the projection for the first successor.
720 MachIfNode *iff = block->get_node(block->number_of_nodes()-3)->as_MachIf();
721 ProjNode *proj0 = block->get_node(block->number_of_nodes()-2)->as_Proj();
722 ProjNode *proj1 = block->get_node(block->number_of_nodes()-1)->as_Proj();
723 ProjNode *projt = (proj0->Opcode() == Op_IfTrue) ? proj0 : proj1;
724 ProjNode *projf = (proj0->Opcode() == Op_IfFalse) ? proj0 : proj1;
725
726 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1].
727 assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0");
728 assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1");
729
730 ProjNode *proj_always;
731 ProjNode *proj_never;
732 // We must negate the branch if the implicit check doesn't follow
733 // the branch's TRUE path. Then, the new TRUE branch target will
734 // be the old FALSE branch target.
735 if (iff->_prob <= 2*PROB_NEVER) { // There are small rounding errors.
736 proj_never = projt;
737 proj_always = projf;
738 } else {
739 // We must negate the branch if the trap doesn't follow the
740 // branch's TRUE path. Then, the new TRUE branch target will
741 // be the old FALSE branch target.
742 proj_never = projf;
743 proj_always = projt;
744 iff->negate();
745 }
746 assert(iff->_prob <= 2*PROB_NEVER, "Trap based checks are expected to trap never!");
747 // Map the successors properly
748 block->_succs.map(0, get_block_for_node(proj_never ->raw_out(0))); // The target of the trap.
749 block->_succs.map(1, get_block_for_node(proj_always->raw_out(0))); // The fall through target.
750
751 // Place the fall through block after this block.
752 Block *bs1 = block->non_connector_successor(1);
753 if (bs1 != bnext && move_to_next(bs1, block_pos)) {
754 bnext = bs1;
755 }
756 // If the fall through block still is not the next block, insert a goto.
757 if (bs1 != bnext) {
758 insert_goto_at(block_pos, 1);
759 }
760 return bnext;
761 }
762
763 // Fix up the final control flow for basic blocks.
764 void PhaseCFG::fixup_flow() {
765 // Fixup final control flow for the blocks. Remove jump-to-next
766 // block. If neither arm of an IF follows the conditional branch, we
767 // have to add a second jump after the conditional. We place the
768 // TRUE branch target in succs[0] for both GOTOs and IFs.
769 for (uint i = 0; i < number_of_blocks(); i++) {
770 Block* block = get_block(i);
771 block->_pre_order = i; // turn pre-order into block-index
772
773 // Connector blocks need no further processing.
774 if (block->is_connector()) {
775 assert((i+1) == number_of_blocks() || get_block(i + 1)->is_connector(), "All connector blocks should sink to the end");
776 continue;
777 }
778 assert(block->is_Empty() != Block::completely_empty, "Empty blocks should be connectors");
779
780 Block* bnext = (i < number_of_blocks() - 1) ? get_block(i + 1) : NULL;
781 Block* bs0 = block->non_connector_successor(0);
782
783 // Check for multi-way branches where I cannot negate the test to
784 // exchange the true and false targets.
785 if (no_flip_branch(block)) {
786 // Find fall through case - if must fall into its target.
787 // Get the index of the branch's first successor.
788 int branch_idx = block->number_of_nodes() - block->_num_succs;
789
790 // The branch is 1 before the branch's first successor.
791 Node *branch = block->get_node(branch_idx-1);
792
793 // Handle no-flip branches which have implicit checks and which require
794 // special block ordering and individual semantics of the 'fall through
795 // case'.
796 if ((TrapBasedNullChecks || TrapBasedRangeChecks) &&
797 branch->is_Mach() && branch->as_Mach()->is_TrapBasedCheckNode()) {
798 bnext = fixup_trap_based_check(branch, block, i, bnext);
799 } else {
800 // Else, default handling for no-flip branches
801 for (uint j2 = 0; j2 < block->_num_succs; j2++) {
802 const ProjNode* p = block->get_node(branch_idx + j2)->as_Proj();
803 if (p->_con == 0) {
804 // successor j2 is fall through case
805 if (block->non_connector_successor(j2) != bnext) {
806 // but it is not the next block => insert a goto
807 insert_goto_at(i, j2);
808 }
809 // Put taken branch in slot 0
810 if (j2 == 0 && block->_num_succs == 2) {
811 // Flip targets in succs map
812 Block *tbs0 = block->_succs[0];
813 Block *tbs1 = block->_succs[1];
814 block->_succs.map(0, tbs1);
815 block->_succs.map(1, tbs0);
816 }
817 break;
818 }
819 }
820 }
821
822 // Remove all CatchProjs
823 for (uint j = 0; j < block->_num_succs; j++) {
824 block->pop_node();
825 }
826
827 } else if (block->_num_succs == 1) {
828 // Block ends in a Goto?
829 if (bnext == bs0) {
830 // We fall into next block; remove the Goto
831 block->pop_node();
832 }
833
834 } else if(block->_num_succs == 2) { // Block ends in a If?
835 // Get opcode of 1st projection (matches _succs[0])
836 // Note: Since this basic block has 2 exits, the last 2 nodes must
837 // be projections (in any order), the 3rd last node must be
838 // the IfNode (we have excluded other 2-way exits such as
839 // CatchNodes already).
840 MachNode* iff = block->get_node(block->number_of_nodes() - 3)->as_Mach();
841 ProjNode* proj0 = block->get_node(block->number_of_nodes() - 2)->as_Proj();
842 ProjNode* proj1 = block->get_node(block->number_of_nodes() - 1)->as_Proj();
843
844 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1].
845 assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0");
846 assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1");
847
848 Block* bs1 = block->non_connector_successor(1);
849
850 // Check for neither successor block following the current
851 // block ending in a conditional. If so, move one of the
852 // successors after the current one, provided that the
853 // successor was previously unscheduled, but moveable
854 // (i.e., all paths to it involve a branch).
855 if (!C->do_freq_based_layout() && bnext != bs0 && bnext != bs1) {
856 // Choose the more common successor based on the probability
857 // of the conditional branch.
858 Block* bx = bs0;
859 Block* by = bs1;
860
861 // _prob is the probability of taking the true path. Make
862 // p the probability of taking successor #1.
863 float p = iff->as_MachIf()->_prob;
864 if (proj0->Opcode() == Op_IfTrue) {
865 p = 1.0 - p;
866 }
867
868 // Prefer successor #1 if p > 0.5
869 if (p > PROB_FAIR) {
870 bx = bs1;
871 by = bs0;
872 }
873
874 // Attempt the more common successor first
875 if (move_to_next(bx, i)) {
876 bnext = bx;
877 } else if (move_to_next(by, i)) {
878 bnext = by;
879 }
880 }
881
882 // Check for conditional branching the wrong way. Negate
883 // conditional, if needed, so it falls into the following block
884 // and branches to the not-following block.
885
886 // Check for the next block being in succs[0]. We are going to branch
887 // to succs[0], so we want the fall-thru case as the next block in
888 // succs[1].
889 if (bnext == bs0) {
890 // Fall-thru case in succs[0], so flip targets in succs map
891 Block* tbs0 = block->_succs[0];
892 Block* tbs1 = block->_succs[1];
893 block->_succs.map(0, tbs1);
894 block->_succs.map(1, tbs0);
895 // Flip projection for each target
896 ProjNode* tmp = proj0;
897 proj0 = proj1;
898 proj1 = tmp;
899
900 } else if(bnext != bs1) {
901 // Need a double-branch
902 // The existing conditional branch need not change.
903 // Add a unconditional branch to the false target.
904 // Alas, it must appear in its own block and adding a
905 // block this late in the game is complicated. Sigh.
906 insert_goto_at(i, 1);
907 }
908
909 // Make sure we TRUE branch to the target
910 if (proj0->Opcode() == Op_IfFalse) {
911 iff->as_MachIf()->negate();
912 }
913
914 block->pop_node(); // Remove IfFalse & IfTrue projections
915 block->pop_node();
916
917 } else {
918 // Multi-exit block, e.g. a switch statement
919 // But we don't need to do anything here
920 }
921 } // End of for all blocks
922 }
923
924
925 // postalloc_expand: Expand nodes after register allocation.
926 //
927 // postalloc_expand has to be called after register allocation, just
928 // before output (i.e. scheduling). It only gets called if
929 // Matcher::require_postalloc_expand is true.
930 //
931 // Background:
932 //
933 // Nodes that are expandend (one compound node requiring several
934 // assembler instructions to be implemented split into two or more
935 // non-compound nodes) after register allocation are not as nice as
936 // the ones expanded before register allocation - they don't
937 // participate in optimizations as global code motion. But after
938 // register allocation we can expand nodes that use registers which
939 // are not spillable or registers that are not allocated, because the
940 // old compound node is simply replaced (in its location in the basic
941 // block) by a new subgraph which does not contain compound nodes any
942 // more. The scheduler called during output can later on process these
943 // non-compound nodes.
944 //
945 // Implementation:
946 //
947 // Nodes requiring postalloc expand are specified in the ad file by using
948 // a postalloc_expand statement instead of ins_encode. A postalloc_expand
949 // contains a single call to an encoding, as does an ins_encode
950 // statement. Instead of an emit() function a postalloc_expand() function
951 // is generated that doesn't emit assembler but creates a new
952 // subgraph. The code below calls this postalloc_expand function for each
953 // node with the appropriate attribute. This function returns the new
954 // nodes generated in an array passed in the call. The old node,
955 // potential MachTemps before and potential Projs after it then get
956 // disconnected and replaced by the new nodes. The instruction
957 // generating the result has to be the last one in the array. In
958 // general it is assumed that Projs after the node expanded are
959 // kills. These kills are not required any more after expanding as
960 // there are now explicitly visible def-use chains and the Projs are
961 // removed. This does not hold for calls: They do not only have
962 // kill-Projs but also Projs defining values. Therefore Projs after
963 // the node expanded are removed for all but for calls. If a node is
964 // to be reused, it must be added to the nodes list returned, and it
965 // will be added again.
966 //
967 // Implementing the postalloc_expand function for a node in an enc_class
968 // is rather tedious. It requires knowledge about many node details, as
969 // the nodes and the subgraph must be hand crafted. To simplify this,
970 // adlc generates some utility variables into the postalloc_expand function,
971 // e.g., holding the operands as specified by the postalloc_expand encoding
972 // specification, e.g.:
973 // * unsigned idx_<par_name> holding the index of the node in the ins
974 // * Node *n_<par_name> holding the node loaded from the ins
975 // * MachOpnd *op_<par_name> holding the corresponding operand
976 //
977 // The ordering of operands can not be determined by looking at a
978 // rule. Especially if a match rule matches several different trees,
979 // several nodes are generated from one instruct specification with
980 // different operand orderings. In this case the adlc generated
981 // variables are the only way to access the ins and operands
982 // deterministically.
983 //
984 // If assigning a register to a node that contains an oop, don't
985 // forget to call ra_->set_oop() for the node.
986 void PhaseCFG::postalloc_expand(PhaseRegAlloc* _ra) {
987 GrowableArray <Node *> new_nodes(32); // Array with new nodes filled by postalloc_expand function of node.
988 GrowableArray <Node *> remove(32);
989 GrowableArray <Node *> succs(32);
990 unsigned int max_idx = C->unique(); // Remember to distinguish new from old nodes.
991 DEBUG_ONLY(bool foundNode = false);
992
993 // for all blocks
994 for (uint i = 0; i < number_of_blocks(); i++) {
995 Block *b = _blocks[i];
996 // For all instructions in the current block.
997 for (uint j = 0; j < b->number_of_nodes(); j++) {
998 Node *n = b->get_node(j);
999 if (n->is_Mach() && n->as_Mach()->requires_postalloc_expand()) {
1000 #ifdef ASSERT
1001 if (TracePostallocExpand) {
1002 if (!foundNode) {
1003 foundNode = true;
1004 tty->print("POSTALLOC EXPANDING %d %s\n", C->compile_id(),
1005 C->method() ? C->method()->name()->as_utf8() : C->stub_name());
1006 }
1007 tty->print(" postalloc expanding "); n->dump();
1008 if (Verbose) {
1009 tty->print(" with ins:\n");
1010 for (uint k = 0; k < n->len(); ++k) {
1011 if (n->in(k)) { tty->print(" "); n->in(k)->dump(); }
1012 }
1013 }
1014 }
1015 #endif
1016 new_nodes.clear();
1017 // Collect nodes that have to be removed from the block later on.
1018 uint req = n->req();
1019 remove.clear();
1020 for (uint k = 0; k < req; ++k) {
1021 if (n->in(k) && n->in(k)->is_MachTemp()) {
1022 remove.push(n->in(k)); // MachTemps which are inputs to the old node have to be removed.
1023 n->in(k)->del_req(0);
1024 j--;
1025 }
1026 }
1027
1028 // Check whether we can allocate enough nodes. We set a fix limit for
1029 // the size of postalloc expands with this.
1030 uint unique_limit = C->unique() + 40;
1031 if (unique_limit >= _ra->node_regs_max_index()) {
1032 Compile::current()->record_failure("out of nodes in postalloc expand");
1033 return;
1034 }
1035
1036 // Emit (i.e. generate new nodes).
1037 n->as_Mach()->postalloc_expand(&new_nodes, _ra);
1038
1039 assert(C->unique() < unique_limit, "You allocated too many nodes in your postalloc expand.");
1040
1041 // Disconnect the inputs of the old node.
1042 //
1043 // We reuse MachSpillCopy nodes. If we need to expand them, there
1044 // are many, so reusing pays off. If reused, the node already
1045 // has the new ins. n must be the last node on new_nodes list.
1046 if (!n->is_MachSpillCopy()) {
1047 for (int k = req - 1; k >= 0; --k) {
1048 n->del_req(k);
1049 }
1050 }
1051
1052 #ifdef ASSERT
1053 // Check that all nodes have proper operands.
1054 for (int k = 0; k < new_nodes.length(); ++k) {
1055 if (new_nodes.at(k)->_idx < max_idx || !new_nodes.at(k)->is_Mach()) continue; // old node, Proj ...
1056 MachNode *m = new_nodes.at(k)->as_Mach();
1057 for (unsigned int l = 0; l < m->num_opnds(); ++l) {
1058 if (MachOper::notAnOper(m->_opnds[l])) {
1059 outputStream *os = tty;
1060 os->print("Node %s ", m->Name());
1061 os->print("has invalid opnd %d: %p\n", l, m->_opnds[l]);
1062 assert(0, "Invalid operands, see inline trace in hs_err_pid file.");
1063 }
1064 }
1065 }
1066 #endif
1067
1068 // Collect succs of old node in remove (for projections) and in succs (for
1069 // all other nodes) do _not_ collect projections in remove (but in succs)
1070 // in case the node is a call. We need the projections for calls as they are
1071 // associated with registes (i.e. they are defs).
1072 succs.clear();
1073 for (DUIterator k = n->outs(); n->has_out(k); k++) {
1074 if (n->out(k)->is_Proj() && !n->is_MachCall() && !n->is_MachBranch()) {
1075 remove.push(n->out(k));
1076 } else {
1077 succs.push(n->out(k));
1078 }
1079 }
1080 // Replace old node n as input of its succs by last of the new nodes.
1081 for (int k = 0; k < succs.length(); ++k) {
1082 Node *succ = succs.at(k);
1083 for (uint l = 0; l < succ->req(); ++l) {
1084 if (succ->in(l) == n) {
1085 succ->set_req(l, new_nodes.at(new_nodes.length() - 1));
1086 }
1087 }
1088 for (uint l = succ->req(); l < succ->len(); ++l) {
1089 if (succ->in(l) == n) {
1090 succ->set_prec(l, new_nodes.at(new_nodes.length() - 1));
1091 }
1092 }
1093 }
1094
1095 // Index of old node in block.
1096 uint index = b->find_node(n);
1097 // Insert new nodes into block and map them in nodes->blocks array
1098 // and remember last node in n2.
1099 Node *n2 = NULL;
1100 for (int k = 0; k < new_nodes.length(); ++k) {
1101 n2 = new_nodes.at(k);
1102 b->insert_node(n2, ++index);
1103 map_node_to_block(n2, b);
1104 }
1105
1106 // Add old node n to remove and remove them all from block.
1107 remove.push(n);
1108 j--;
1109 #ifdef ASSERT
1110 if (TracePostallocExpand && Verbose) {
1111 tty->print(" removing:\n");
1112 for (int k = 0; k < remove.length(); ++k) {
1113 tty->print(" "); remove.at(k)->dump();
1114 }
1115 tty->print(" inserting:\n");
1116 for (int k = 0; k < new_nodes.length(); ++k) {
1117 tty->print(" "); new_nodes.at(k)->dump();
1118 }
1119 }
1120 #endif
1121 for (int k = 0; k < remove.length(); ++k) {
1122 if (b->contains(remove.at(k))) {
1123 b->find_remove(remove.at(k));
1124 } else {
1125 assert(remove.at(k)->is_Proj() && (remove.at(k)->in(0)->is_MachBranch()), "");
1126 }
1127 }
1128 // If anything has been inserted (n2 != NULL), continue after last node inserted.
1129 // This does not always work. Some postalloc expands don't insert any nodes, if they
1130 // do optimizations (e.g., max(x,x)). In this case we decrement j accordingly.
1131 j = n2 ? b->find_node(n2) : j;
1132 }
1133 }
1134 }
1135
1136 #ifdef ASSERT
1137 if (foundNode) {
1138 tty->print("FINISHED %d %s\n", C->compile_id(),
1139 C->method() ? C->method()->name()->as_utf8() : C->stub_name());
1140 tty->flush();
1141 }
1142 #endif
1143 }
1144
1145
1146 //------------------------------dump-------------------------------------------
1147 #ifndef PRODUCT
1148 void PhaseCFG::_dump_cfg( const Node *end, VectorSet &visited ) const {
1149 const Node *x = end->is_block_proj();
1150 assert( x, "not a CFG" );
1151
1152 // Do not visit this block again
1153 if( visited.test_set(x->_idx) ) return;
1154
1155 // Skip through this block
1156 const Node *p = x;
1157 do {
1158 p = p->in(0); // Move control forward
1159 assert( !p->is_block_proj() || p->is_Root(), "not a CFG" );
1160 } while( !p->is_block_start() );
1161
1162 // Recursively visit
1163 for (uint i = 1; i < p->req(); i++) {
1164 _dump_cfg(p->in(i), visited);
1165 }
1166
1167 // Dump the block
1168 get_block_for_node(p)->dump(this);
1169 }
1170
1171 void PhaseCFG::dump( ) const {
1172 tty->print("\n--- CFG --- %d BBs\n", number_of_blocks());
1173 if (_blocks.size()) { // Did we do basic-block layout?
1174 for (uint i = 0; i < number_of_blocks(); i++) {
1175 const Block* block = get_block(i);
1176 block->dump(this);
1177 }
1178 } else { // Else do it with a DFS
1179 VectorSet visited(_block_arena);
1180 _dump_cfg(_root,visited);
1181 }
1182 }
1183
1184 void PhaseCFG::dump_headers() {
1185 for (uint i = 0; i < number_of_blocks(); i++) {
1186 Block* block = get_block(i);
1187 if (block != NULL) {
1188 block->dump_head(this);
1189 }
1190 }
1191 }
1192
1193 void PhaseCFG::verify() const {
1194 #ifdef ASSERT
1195 // Verify sane CFG
1196 for (uint i = 0; i < number_of_blocks(); i++) {
1197 Block* block = get_block(i);
1198 uint cnt = block->number_of_nodes();
1199 uint j;
1200 for (j = 0; j < cnt; j++) {
1201 Node *n = block->get_node(j);
1202 assert(get_block_for_node(n) == block, "");
1203 if (j >= 1 && n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CreateEx) {
1204 assert(j == 1 || block->get_node(j-1)->is_Phi(), "CreateEx must be first instruction in block");
1205 }
1206 for (uint k = 0; k < n->req(); k++) {
1207 Node *def = n->in(k);
1208 if (def && def != n) {
1209 assert(get_block_for_node(def) || def->is_Con(), "must have block; constants for debug info ok");
1210 // Verify that instructions in the block is in correct order.
1211 // Uses must follow their definition if they are at the same block.
1212 // Mostly done to check that MachSpillCopy nodes are placed correctly
1213 // when CreateEx node is moved in build_ifg_physical().
1214 if (get_block_for_node(def) == block && !(block->head()->is_Loop() && n->is_Phi()) &&
1215 // See (+++) comment in reg_split.cpp
1216 !(n->jvms() != NULL && n->jvms()->is_monitor_use(k))) {
1217 bool is_loop = false;
1218 if (n->is_Phi()) {
1219 for (uint l = 1; l < def->req(); l++) {
1220 if (n == def->in(l)) {
1221 is_loop = true;
1222 break; // Some kind of loop
1223 }
1224 }
1225 }
1226 assert(is_loop || block->find_node(def) < j, "uses must follow definitions");
1227 }
1228 }
1229 }
1230 }
1231
1232 j = block->end_idx();
1233 Node* bp = (Node*)block->get_node(block->number_of_nodes() - 1)->is_block_proj();
1234 assert(bp, "last instruction must be a block proj");
1235 assert(bp == block->get_node(j), "wrong number of successors for this block");
1236 if (bp->is_Catch()) {
1237 while (block->get_node(--j)->is_MachProj()) {
1238 ;
1239 }
1240 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1241 } else if (bp->is_Mach() && bp->as_Mach()->ideal_Opcode() == Op_If) {
1242 assert(block->_num_succs == 2, "Conditional branch must have two targets");
1243 }
1244 }
1245 #endif
1246 }
1247 #endif
1248
1249 UnionFind::UnionFind( uint max ) : _cnt(max), _max(max), _indices(NEW_RESOURCE_ARRAY(uint,max)) {
1250 Copy::zero_to_bytes( _indices, sizeof(uint)*max );
1251 }
1252
1253 void UnionFind::extend( uint from_idx, uint to_idx ) {
1254 _nesting.check();
1255 if( from_idx >= _max ) {
1256 uint size = 16;
1257 while( size <= from_idx ) size <<=1;
1258 _indices = REALLOC_RESOURCE_ARRAY( uint, _indices, _max, size );
1259 _max = size;
1260 }
1261 while( _cnt <= from_idx ) _indices[_cnt++] = 0;
1262 _indices[from_idx] = to_idx;
1263 }
1264
1265 void UnionFind::reset( uint max ) {
1266 assert( max <= max_uint, "Must fit within uint" );
1267 // Force the Union-Find mapping to be at least this large
1268 extend(max,0);
1269 // Initialize to be the ID mapping.
1270 for( uint i=0; i<max; i++ ) map(i,i);
1271 }
1272
1273 // Straight out of Tarjan's union-find algorithm
1274 uint UnionFind::Find_compress( uint idx ) {
1275 uint cur = idx;
1276 uint next = lookup(cur);
1277 while( next != cur ) { // Scan chain of equivalences
1278 assert( next < cur, "always union smaller" );
1279 cur = next; // until find a fixed-point
1280 next = lookup(cur);
1281 }
1282 // Core of union-find algorithm: update chain of
1283 // equivalences to be equal to the root.
1284 while( idx != next ) {
1285 uint tmp = lookup(idx);
1286 map(idx, next);
1287 idx = tmp;
1288 }
1289 return idx;
1290 }
1291
1292 // Like Find above, but no path compress, so bad asymptotic behavior
1293 uint UnionFind::Find_const( uint idx ) const {
1294 if( idx == 0 ) return idx; // Ignore the zero idx
1295 // Off the end? This can happen during debugging dumps
1296 // when data structures have not finished being updated.
1297 if( idx >= _max ) return idx;
1298 uint next = lookup(idx);
1299 while( next != idx ) { // Scan chain of equivalences
1300 idx = next; // until find a fixed-point
1301 next = lookup(idx);
1302 }
1303 return next;
1304 }
1305
1306 // union 2 sets together.
1307 void UnionFind::Union( uint idx1, uint idx2 ) {
1308 uint src = Find(idx1);
1309 uint dst = Find(idx2);
1310 assert( src, "" );
1311 assert( dst, "" );
1312 assert( src < _max, "oob" );
1313 assert( dst < _max, "oob" );
1314 assert( src < dst, "always union smaller" );
1315 map(dst,src);
1316 }
1317
1318 #ifndef PRODUCT
1319 void Trace::dump( ) const {
1320 tty->print_cr("Trace (freq %f)", first_block()->_freq);
1321 for (Block *b = first_block(); b != NULL; b = next(b)) {
1322 tty->print(" B%d", b->_pre_order);
1323 if (b->head()->is_Loop()) {
1324 tty->print(" (L%d)", b->compute_loop_alignment());
1325 }
1326 if (b->has_loop_alignment()) {
1327 tty->print(" (T%d)", b->code_alignment());
1328 }
1329 }
1330 tty->cr();
1331 }
1332
1333 void CFGEdge::dump( ) const {
1334 tty->print(" B%d --> B%d Freq: %f out:%3d%% in:%3d%% State: ",
1335 from()->_pre_order, to()->_pre_order, freq(), _from_pct, _to_pct);
1336 switch(state()) {
1337 case connected:
1338 tty->print("connected");
1339 break;
1340 case open:
1341 tty->print("open");
1342 break;
1343 case interior:
1344 tty->print("interior");
1345 break;
1346 }
1347 if (infrequent()) {
1348 tty->print(" infrequent");
1349 }
1350 tty->cr();
1351 }
1352 #endif
1353
1354 // Comparison function for edges
1355 static int edge_order(CFGEdge **e0, CFGEdge **e1) {
1356 float freq0 = (*e0)->freq();
1357 float freq1 = (*e1)->freq();
1358 if (freq0 != freq1) {
1359 return freq0 > freq1 ? -1 : 1;
1360 }
1361
1362 int dist0 = (*e0)->to()->_rpo - (*e0)->from()->_rpo;
1363 int dist1 = (*e1)->to()->_rpo - (*e1)->from()->_rpo;
1364
1365 return dist1 - dist0;
1366 }
1367
1368 // Comparison function for edges
1369 extern "C" int trace_frequency_order(const void *p0, const void *p1) {
1370 Trace *tr0 = *(Trace **) p0;
1371 Trace *tr1 = *(Trace **) p1;
1372 Block *b0 = tr0->first_block();
1373 Block *b1 = tr1->first_block();
1374
1375 // The trace of connector blocks goes at the end;
1376 // we only expect one such trace
1377 if (b0->is_connector() != b1->is_connector()) {
1378 return b1->is_connector() ? -1 : 1;
1379 }
1380
1381 // Pull more frequently executed blocks to the beginning
1382 float freq0 = b0->_freq;
1383 float freq1 = b1->_freq;
1384 if (freq0 != freq1) {
1385 return freq0 > freq1 ? -1 : 1;
1386 }
1387
1388 int diff = tr0->first_block()->_rpo - tr1->first_block()->_rpo;
1389
1390 return diff;
1391 }
1392
1393 // Find edges of interest, i.e, those which can fall through. Presumes that
1394 // edges which don't fall through are of low frequency and can be generally
1395 // ignored. Initialize the list of traces.
1396 void PhaseBlockLayout::find_edges() {
1397 // Walk the blocks, creating edges and Traces
1398 uint i;
1399 Trace *tr = NULL;
1400 for (i = 0; i < _cfg.number_of_blocks(); i++) {
1401 Block* b = _cfg.get_block(i);
1402 tr = new Trace(b, next, prev);
1403 traces[tr->id()] = tr;
1404
1405 // All connector blocks should be at the end of the list
1406 if (b->is_connector()) break;
1407
1408 // If this block and the next one have a one-to-one successor
1409 // predecessor relationship, simply append the next block
1410 int nfallthru = b->num_fall_throughs();
1411 while (nfallthru == 1 &&
1412 b->succ_fall_through(0)) {
1413 Block *n = b->_succs[0];
1414
1415 // Skip over single-entry connector blocks, we don't want to
1416 // add them to the trace.
1417 while (n->is_connector() && n->num_preds() == 1) {
1418 n = n->_succs[0];
1419 }
1420
1421 // We see a merge point, so stop search for the next block
1422 if (n->num_preds() != 1) break;
1423
1424 i++;
1425 assert(n = _cfg.get_block(i), "expecting next block");
1426 tr->append(n);
1427 uf->map(n->_pre_order, tr->id());
1428 traces[n->_pre_order] = NULL;
1429 nfallthru = b->num_fall_throughs();
1430 b = n;
1431 }
1432
1433 if (nfallthru > 0) {
1434 // Create a CFGEdge for each outgoing
1435 // edge that could be a fall-through.
1436 for (uint j = 0; j < b->_num_succs; j++ ) {
1437 if (b->succ_fall_through(j)) {
1438 Block *target = b->non_connector_successor(j);
1439 float freq = b->_freq * b->succ_prob(j);
1440 int from_pct = (int) ((100 * freq) / b->_freq);
1441 int to_pct = (int) ((100 * freq) / target->_freq);
1442 edges->append(new CFGEdge(b, target, freq, from_pct, to_pct));
1443 }
1444 }
1445 }
1446 }
1447
1448 // Group connector blocks into one trace
1449 for (i++; i < _cfg.number_of_blocks(); i++) {
1450 Block *b = _cfg.get_block(i);
1451 assert(b->is_connector(), "connector blocks at the end");
1452 tr->append(b);
1453 uf->map(b->_pre_order, tr->id());
1454 traces[b->_pre_order] = NULL;
1455 }
1456 }
1457
1458 // Union two traces together in uf, and null out the trace in the list
1459 void PhaseBlockLayout::union_traces(Trace* updated_trace, Trace* old_trace) {
1460 uint old_id = old_trace->id();
1461 uint updated_id = updated_trace->id();
1462
1463 uint lo_id = updated_id;
1464 uint hi_id = old_id;
1465
1466 // If from is greater than to, swap values to meet
1467 // UnionFind guarantee.
1468 if (updated_id > old_id) {
1469 lo_id = old_id;
1470 hi_id = updated_id;
1471
1472 // Fix up the trace ids
1473 traces[lo_id] = traces[updated_id];
1474 updated_trace->set_id(lo_id);
1475 }
1476
1477 // Union the lower with the higher and remove the pointer
1478 // to the higher.
1479 uf->Union(lo_id, hi_id);
1480 traces[hi_id] = NULL;
1481 }
1482
1483 // Append traces together via the most frequently executed edges
1484 void PhaseBlockLayout::grow_traces() {
1485 // Order the edges, and drive the growth of Traces via the most
1486 // frequently executed edges.
1487 edges->sort(edge_order);
1488 for (int i = 0; i < edges->length(); i++) {
1489 CFGEdge *e = edges->at(i);
1490
1491 if (e->state() != CFGEdge::open) continue;
1492
1493 Block *src_block = e->from();
1494 Block *targ_block = e->to();
1495
1496 // Don't grow traces along backedges?
1497 if (!BlockLayoutRotateLoops) {
1498 if (targ_block->_rpo <= src_block->_rpo) {
1499 targ_block->set_loop_alignment(targ_block);
1500 continue;
1501 }
1502 }
1503
1504 Trace *src_trace = trace(src_block);
1505 Trace *targ_trace = trace(targ_block);
1506
1507 // If the edge in question can join two traces at their ends,
1508 // append one trace to the other.
1509 if (src_trace->last_block() == src_block) {
1510 if (src_trace == targ_trace) {
1511 e->set_state(CFGEdge::interior);
1512 if (targ_trace->backedge(e)) {
1513 // Reset i to catch any newly eligible edge
1514 // (Or we could remember the first "open" edge, and reset there)
1515 i = 0;
1516 }
1517 } else if (targ_trace->first_block() == targ_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 // Embed one trace into another, if the fork or join points are sufficiently
1527 // balanced.
1528 void PhaseBlockLayout::merge_traces(bool fall_thru_only) {
1529 // Walk the edge list a another time, looking at unprocessed edges.
1530 // Fold in diamonds
1531 for (int i = 0; i < edges->length(); i++) {
1532 CFGEdge *e = edges->at(i);
1533
1534 if (e->state() != CFGEdge::open) continue;
1535 if (fall_thru_only) {
1536 if (e->infrequent()) continue;
1537 }
1538
1539 Block *src_block = e->from();
1540 Trace *src_trace = trace(src_block);
1541 bool src_at_tail = src_trace->last_block() == src_block;
1542
1543 Block *targ_block = e->to();
1544 Trace *targ_trace = trace(targ_block);
1545 bool targ_at_start = targ_trace->first_block() == targ_block;
1546
1547 if (src_trace == targ_trace) {
1548 // This may be a loop, but we can't do much about it.
1549 e->set_state(CFGEdge::interior);
1550 continue;
1551 }
1552
1553 if (fall_thru_only) {
1554 // If the edge links the middle of two traces, we can't do anything.
1555 // Mark the edge and continue.
1556 if (!src_at_tail & !targ_at_start) {
1557 continue;
1558 }
1559
1560 // Don't grow traces along backedges?
1561 if (!BlockLayoutRotateLoops && (targ_block->_rpo <= src_block->_rpo)) {
1562 continue;
1563 }
1564
1565 // If both ends of the edge are available, why didn't we handle it earlier?
1566 assert(src_at_tail ^ targ_at_start, "Should have caught this edge earlier.");
1567
1568 if (targ_at_start) {
1569 // Insert the "targ" trace in the "src" trace if the insertion point
1570 // is a two way branch.
1571 // Better profitability check possible, but may not be worth it.
1572 // Someday, see if the this "fork" has an associated "join";
1573 // then make a policy on merging this trace at the fork or join.
1574 // For example, other things being equal, it may be better to place this
1575 // trace at the join point if the "src" trace ends in a two-way, but
1576 // the insertion point is one-way.
1577 assert(src_block->num_fall_throughs() == 2, "unexpected diamond");
1578 e->set_state(CFGEdge::connected);
1579 src_trace->insert_after(src_block, targ_trace);
1580 union_traces(src_trace, targ_trace);
1581 } else if (src_at_tail) {
1582 if (src_trace != trace(_cfg.get_root_block())) {
1583 e->set_state(CFGEdge::connected);
1584 targ_trace->insert_before(targ_block, src_trace);
1585 union_traces(targ_trace, src_trace);
1586 }
1587 }
1588 } else if (e->state() == CFGEdge::open) {
1589 // Append traces, even without a fall-thru connection.
1590 // But leave root entry at the beginning of the block list.
1591 if (targ_trace != trace(_cfg.get_root_block())) {
1592 e->set_state(CFGEdge::connected);
1593 src_trace->append(targ_trace);
1594 union_traces(src_trace, targ_trace);
1595 }
1596 }
1597 }
1598 }
1599
1600 // Order the sequence of the traces in some desirable way, and fixup the
1601 // jumps at the end of each block.
1602 void PhaseBlockLayout::reorder_traces(int count) {
1603 ResourceArea *area = Thread::current()->resource_area();
1604 Trace ** new_traces = NEW_ARENA_ARRAY(area, Trace *, count);
1605 Block_List worklist;
1606 int new_count = 0;
1607
1608 // Compact the traces.
1609 for (int i = 0; i < count; i++) {
1610 Trace *tr = traces[i];
1611 if (tr != NULL) {
1612 new_traces[new_count++] = tr;
1613 }
1614 }
1615
1616 // The entry block should be first on the new trace list.
1617 Trace *tr = trace(_cfg.get_root_block());
1618 assert(tr == new_traces[0], "entry trace misplaced");
1619
1620 // Sort the new trace list by frequency
1621 qsort(new_traces + 1, new_count - 1, sizeof(new_traces[0]), trace_frequency_order);
1622
1623 // Patch up the successor blocks
1624 _cfg.clear_blocks();
1625 for (int i = 0; i < new_count; i++) {
1626 Trace *tr = new_traces[i];
1627 if (tr != NULL) {
1628 tr->fixup_blocks(_cfg);
1629 }
1630 }
1631 }
1632
1633 // Order basic blocks based on frequency
1634 PhaseBlockLayout::PhaseBlockLayout(PhaseCFG &cfg)
1635 : Phase(BlockLayout)
1636 , _cfg(cfg) {
1637 ResourceMark rm;
1638 ResourceArea *area = Thread::current()->resource_area();
1639
1640 // List of traces
1641 int size = _cfg.number_of_blocks() + 1;
1642 traces = NEW_ARENA_ARRAY(area, Trace *, size);
1643 memset(traces, 0, size*sizeof(Trace*));
1644 next = NEW_ARENA_ARRAY(area, Block *, size);
1645 memset(next, 0, size*sizeof(Block *));
1646 prev = NEW_ARENA_ARRAY(area, Block *, size);
1647 memset(prev , 0, size*sizeof(Block *));
1648
1649 // List of edges
1650 edges = new GrowableArray<CFGEdge*>;
1651
1652 // Mapping block index --> block_trace
1653 uf = new UnionFind(size);
1654 uf->reset(size);
1655
1656 // Find edges and create traces.
1657 find_edges();
1658
1659 // Grow traces at their ends via most frequent edges.
1660 grow_traces();
1661
1662 // Merge one trace into another, but only at fall-through points.
1663 // This may make diamonds and other related shapes in a trace.
1664 merge_traces(true);
1665
1666 // Run merge again, allowing two traces to be catenated, even if
1667 // one does not fall through into the other. This appends loosely
1668 // related traces to be near each other.
1669 merge_traces(false);
1670
1671 // Re-order all the remaining traces by frequency
1672 reorder_traces(size);
1673
1674 assert(_cfg.number_of_blocks() >= (uint) (size - 1), "number of blocks can not shrink");
1675 }
1676
1677
1678 // Edge e completes a loop in a trace. If the target block is head of the
1679 // loop, rotate the loop block so that the loop ends in a conditional branch.
1680 bool Trace::backedge(CFGEdge *e) {
1681 bool loop_rotated = false;
1682 Block *src_block = e->from();
1683 Block *targ_block = e->to();
1684
1685 assert(last_block() == src_block, "loop discovery at back branch");
1686 if (first_block() == targ_block) {
1687 if (BlockLayoutRotateLoops && last_block()->num_fall_throughs() < 2) {
1688 // Find the last block in the trace that has a conditional
1689 // branch.
1690 Block *b;
1691 for (b = last_block(); b != NULL; b = prev(b)) {
1692 if (b->num_fall_throughs() == 2) {
1693 break;
1694 }
1695 }
1696
1697 if (b != last_block() && b != NULL) {
1698 loop_rotated = true;
1699
1700 // Rotate the loop by doing two-part linked-list surgery.
1701 append(first_block());
1702 break_loop_after(b);
1703 }
1704 }
1705
1706 // Backbranch to the top of a trace
1707 // Scroll forward through the trace from the targ_block. If we find
1708 // a loop head before another loop top, use the the loop head alignment.
1709 for (Block *b = targ_block; b != NULL; b = next(b)) {
1710 if (b->has_loop_alignment()) {
1711 break;
1712 }
1713 if (b->head()->is_Loop()) {
1714 targ_block = b;
1715 break;
1716 }
1717 }
1718
1719 first_block()->set_loop_alignment(targ_block);
1720
1721 } else {
1722 // Backbranch into the middle of a trace
1723 targ_block->set_loop_alignment(targ_block);
1724 }
1725
1726 return loop_rotated;
1727 }
1728
1729 // push blocks onto the CFG list
1730 // ensure that blocks have the correct two-way branch sense
1731 void Trace::fixup_blocks(PhaseCFG &cfg) {
1732 Block *last = last_block();
1733 for (Block *b = first_block(); b != NULL; b = next(b)) {
1734 cfg.add_block(b);
1735 if (!b->is_connector()) {
1736 int nfallthru = b->num_fall_throughs();
1737 if (b != last) {
1738 if (nfallthru == 2) {
1739 // Ensure that the sense of the branch is correct
1740 Block *bnext = next(b);
1741 Block *bs0 = b->non_connector_successor(0);
1742
1743 MachNode *iff = b->get_node(b->number_of_nodes() - 3)->as_Mach();
1744 ProjNode *proj0 = b->get_node(b->number_of_nodes() - 2)->as_Proj();
1745 ProjNode *proj1 = b->get_node(b->number_of_nodes() - 1)->as_Proj();
1746
1747 if (bnext == bs0) {
1748 // Fall-thru case in succs[0], should be in succs[1]
1749
1750 // Flip targets in _succs map
1751 Block *tbs0 = b->_succs[0];
1752 Block *tbs1 = b->_succs[1];
1753 b->_succs.map( 0, tbs1 );
1754 b->_succs.map( 1, tbs0 );
1755
1756 // Flip projections to match targets
1757 b->map_node(proj1, b->number_of_nodes() - 2);
1758 b->map_node(proj0, b->number_of_nodes() - 1);
1759 }
1760 }
1761 }
1762 }
1763 }
1764 }