1/*
2 * Copyright (c) 1997, 2019, 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 "memory/resourceArea.hpp"
29#include "compiler/compilerDirectives.hpp"
30#include "opto/block.hpp"
31#include "opto/cfgnode.hpp"
32#include "opto/chaitin.hpp"
33#include "opto/loopnode.hpp"
34#include "opto/machnode.hpp"
35#include "opto/matcher.hpp"
36#include "opto/opcodes.hpp"
37#include "opto/rootnode.hpp"
38#include "utilities/copy.hpp"
39
40void Block_Array::grow( uint i ) {
41 assert(i >= Max(), "must be an overflow");
42 debug_only(_limit = i+1);
43 if( i < _size ) return;
44 if( !_size ) {
45 _size = 1;
46 _blocks = (Block**)_arena->Amalloc( _size * sizeof(Block*) );
47 _blocks[0] = NULL;
48 }
49 uint old = _size;
50 while( i >= _size ) _size <<= 1; // Double to fit
51 _blocks = (Block**)_arena->Arealloc( _blocks, old*sizeof(Block*),_size*sizeof(Block*));
52 Copy::zero_to_bytes( &_blocks[old], (_size-old)*sizeof(Block*) );
53}
54
55void Block_List::remove(uint i) {
56 assert(i < _cnt, "index out of bounds");
57 Copy::conjoint_words_to_lower((HeapWord*)&_blocks[i+1], (HeapWord*)&_blocks[i], ((_cnt-i-1)*sizeof(Block*)));
58 pop(); // shrink list by one block
59}
60
61void Block_List::insert(uint i, Block *b) {
62 push(b); // grow list by one block
63 Copy::conjoint_words_to_higher((HeapWord*)&_blocks[i], (HeapWord*)&_blocks[i+1], ((_cnt-i-1)*sizeof(Block*)));
64 _blocks[i] = b;
65}
66
67#ifndef PRODUCT
68void Block_List::print() {
69 for (uint i=0; i < size(); i++) {
70 tty->print("B%d ", _blocks[i]->_pre_order);
71 }
72 tty->print("size = %d\n", size());
73}
74#endif
75
76uint Block::code_alignment() const {
77 // Check for Root block
78 if (_pre_order == 0) return CodeEntryAlignment;
79 // Check for Start block
80 if (_pre_order == 1) return InteriorEntryAlignment;
81 // Check for loop alignment
82 if (has_loop_alignment()) return loop_alignment();
83
84 return relocInfo::addr_unit(); // no particular alignment
85}
86
87uint Block::compute_loop_alignment() {
88 Node *h = head();
89 int unit_sz = relocInfo::addr_unit();
90 if (h->is_Loop() && h->as_Loop()->is_inner_loop()) {
91 // Pre- and post-loops have low trip count so do not bother with
92 // NOPs for align loop head. The constants are hidden from tuning
93 // but only because my "divide by 4" heuristic surely gets nearly
94 // all possible gain (a "do not align at all" heuristic has a
95 // chance of getting a really tiny gain).
96 if (h->is_CountedLoop() && (h->as_CountedLoop()->is_pre_loop() ||
97 h->as_CountedLoop()->is_post_loop())) {
98 return (OptoLoopAlignment > 4*unit_sz) ? (OptoLoopAlignment>>2) : unit_sz;
99 }
100 // Loops with low backedge frequency should not be aligned.
101 Node *n = h->in(LoopNode::LoopBackControl)->in(0);
102 if (n->is_MachIf() && n->as_MachIf()->_prob < 0.01) {
103 return unit_sz; // Loop does not loop, more often than not!
104 }
105 return OptoLoopAlignment; // Otherwise align loop head
106 }
107
108 return unit_sz; // no particular alignment
109}
110
111// Compute the size of first 'inst_cnt' instructions in this block.
112// Return the number of instructions left to compute if the block has
113// less then 'inst_cnt' instructions. Stop, and return 0 if sum_size
114// exceeds OptoLoopAlignment.
115uint Block::compute_first_inst_size(uint& sum_size, uint inst_cnt,
116 PhaseRegAlloc* ra) {
117 uint last_inst = number_of_nodes();
118 for( uint j = 0; j < last_inst && inst_cnt > 0; j++ ) {
119 uint inst_size = get_node(j)->size(ra);
120 if( inst_size > 0 ) {
121 inst_cnt--;
122 uint sz = sum_size + inst_size;
123 if( sz <= (uint)OptoLoopAlignment ) {
124 // Compute size of instructions which fit into fetch buffer only
125 // since all inst_cnt instructions will not fit even if we align them.
126 sum_size = sz;
127 } else {
128 return 0;
129 }
130 }
131 }
132 return inst_cnt;
133}
134
135uint Block::find_node( const Node *n ) const {
136 for( uint i = 0; i < number_of_nodes(); i++ ) {
137 if( get_node(i) == n )
138 return i;
139 }
140 ShouldNotReachHere();
141 return 0;
142}
143
144// Find and remove n from block list
145void Block::find_remove( const Node *n ) {
146 remove_node(find_node(n));
147}
148
149bool Block::contains(const Node *n) const {
150 return _nodes.contains(n);
151}
152
153// Return empty status of a block. Empty blocks contain only the head, other
154// ideal nodes, and an optional trailing goto.
155int Block::is_Empty() const {
156
157 // Root or start block is not considered empty
158 if (head()->is_Root() || head()->is_Start()) {
159 return not_empty;
160 }
161
162 int success_result = completely_empty;
163 int end_idx = number_of_nodes() - 1;
164
165 // Check for ending goto
166 if ((end_idx > 0) && (get_node(end_idx)->is_MachGoto())) {
167 success_result = empty_with_goto;
168 end_idx--;
169 }
170
171 // Unreachable blocks are considered empty
172 if (num_preds() <= 1) {
173 return success_result;
174 }
175
176 // Ideal nodes are allowable in empty blocks: skip them Only MachNodes
177 // turn directly into code, because only MachNodes have non-trivial
178 // emit() functions.
179 while ((end_idx > 0) && !get_node(end_idx)->is_Mach()) {
180 end_idx--;
181 }
182
183 // No room for any interesting instructions?
184 if (end_idx == 0) {
185 return success_result;
186 }
187
188 return not_empty;
189}
190
191// Return true if the block's code implies that it is likely to be
192// executed infrequently. Check to see if the block ends in a Halt or
193// a low probability call.
194bool Block::has_uncommon_code() const {
195 Node* en = end();
196
197 if (en->is_MachGoto())
198 en = en->in(0);
199 if (en->is_Catch())
200 en = en->in(0);
201 if (en->is_MachProj() && en->in(0)->is_MachCall()) {
202 MachCallNode* call = en->in(0)->as_MachCall();
203 if (call->cnt() != COUNT_UNKNOWN && call->cnt() <= PROB_UNLIKELY_MAG(4)) {
204 // This is true for slow-path stubs like new_{instance,array},
205 // slow_arraycopy, complete_monitor_locking, uncommon_trap.
206 // The magic number corresponds to the probability of an uncommon_trap,
207 // even though it is a count not a probability.
208 return true;
209 }
210 }
211
212 int op = en->is_Mach() ? en->as_Mach()->ideal_Opcode() : en->Opcode();
213 return op == Op_Halt;
214}
215
216// True if block is low enough frequency or guarded by a test which
217// mostly does not go here.
218bool PhaseCFG::is_uncommon(const Block* block) {
219 // Initial blocks must never be moved, so are never uncommon.
220 if (block->head()->is_Root() || block->head()->is_Start()) return false;
221
222 // Check for way-low freq
223 if(block->_freq < BLOCK_FREQUENCY(0.00001f) ) return true;
224
225 // Look for code shape indicating uncommon_trap or slow path
226 if (block->has_uncommon_code()) return true;
227
228 const float epsilon = 0.05f;
229 const float guard_factor = PROB_UNLIKELY_MAG(4) / (1.f - epsilon);
230 uint uncommon_preds = 0;
231 uint freq_preds = 0;
232 uint uncommon_for_freq_preds = 0;
233
234 for( uint i=1; i< block->num_preds(); i++ ) {
235 Block* guard = get_block_for_node(block->pred(i));
236 // Check to see if this block follows its guard 1 time out of 10000
237 // or less.
238 //
239 // See list of magnitude-4 unlikely probabilities in cfgnode.hpp which
240 // we intend to be "uncommon", such as slow-path TLE allocation,
241 // predicted call failure, and uncommon trap triggers.
242 //
243 // Use an epsilon value of 5% to allow for variability in frequency
244 // predictions and floating point calculations. The net effect is
245 // that guard_factor is set to 9500.
246 //
247 // Ignore low-frequency blocks.
248 // The next check is (guard->_freq < 1.e-5 * 9500.).
249 if(guard->_freq*BLOCK_FREQUENCY(guard_factor) < BLOCK_FREQUENCY(0.00001f)) {
250 uncommon_preds++;
251 } else {
252 freq_preds++;
253 if(block->_freq < guard->_freq * guard_factor ) {
254 uncommon_for_freq_preds++;
255 }
256 }
257 }
258 if( block->num_preds() > 1 &&
259 // The block is uncommon if all preds are uncommon or
260 (uncommon_preds == (block->num_preds()-1) ||
261 // it is uncommon for all frequent preds.
262 uncommon_for_freq_preds == freq_preds) ) {
263 return true;
264 }
265 return false;
266}
267
268#ifndef PRODUCT
269void Block::dump_bidx(const Block* orig, outputStream* st) const {
270 if (_pre_order) st->print("B%d", _pre_order);
271 else st->print("N%d", head()->_idx);
272
273 if (Verbose && orig != this) {
274 // Dump the original block's idx
275 st->print(" (");
276 orig->dump_bidx(orig, st);
277 st->print(")");
278 }
279}
280
281void Block::dump_pred(const PhaseCFG* cfg, Block* orig, outputStream* st) const {
282 if (is_connector()) {
283 for (uint i=1; i<num_preds(); i++) {
284 Block *p = cfg->get_block_for_node(pred(i));
285 p->dump_pred(cfg, orig, st);
286 }
287 } else {
288 dump_bidx(orig, st);
289 st->print(" ");
290 }
291}
292
293void Block::dump_head(const PhaseCFG* cfg, outputStream* st) const {
294 // Print the basic block.
295 dump_bidx(this, st);
296 st->print(": ");
297
298 // Print the outgoing CFG edges.
299 st->print("#\tout( ");
300 for( uint i=0; i<_num_succs; i++ ) {
301 non_connector_successor(i)->dump_bidx(_succs[i], st);
302 st->print(" ");
303 }
304
305 // Print the incoming CFG edges.
306 st->print(") <- ");
307 if( head()->is_block_start() ) {
308 st->print("in( ");
309 for (uint i=1; i<num_preds(); i++) {
310 Node *s = pred(i);
311 if (cfg != NULL) {
312 Block *p = cfg->get_block_for_node(s);
313 p->dump_pred(cfg, p, st);
314 } else {
315 while (!s->is_block_start()) {
316 s = s->in(0);
317 }
318 st->print("N%d ", s->_idx );
319 }
320 }
321 st->print(") ");
322 } else {
323 st->print("BLOCK HEAD IS JUNK ");
324 }
325
326 // Print loop, if any
327 const Block *bhead = this; // Head of self-loop
328 Node *bh = bhead->head();
329
330 if ((cfg != NULL) && bh->is_Loop() && !head()->is_Root()) {
331 LoopNode *loop = bh->as_Loop();
332 const Block *bx = cfg->get_block_for_node(loop->in(LoopNode::LoopBackControl));
333 while (bx->is_connector()) {
334 bx = cfg->get_block_for_node(bx->pred(1));
335 }
336 st->print("Loop( B%d-B%d ", bhead->_pre_order, bx->_pre_order);
337 // Dump any loop-specific bits, especially for CountedLoops.
338 loop->dump_spec(st);
339 st->print(")");
340 } else if (has_loop_alignment()) {
341 st->print("top-of-loop");
342 }
343
344 // Print frequency and other optimization-relevant information
345 st->print(" Freq: %g",_freq);
346 if( Verbose || WizardMode ) {
347 st->print(" IDom: %d/#%d", _idom ? _idom->_pre_order : 0, _dom_depth);
348 st->print(" RegPressure: %d",_reg_pressure);
349 st->print(" IHRP Index: %d",_ihrp_index);
350 st->print(" FRegPressure: %d",_freg_pressure);
351 st->print(" FHRP Index: %d",_fhrp_index);
352 }
353 st->cr();
354}
355
356void Block::dump() const {
357 dump(NULL);
358}
359
360void Block::dump(const PhaseCFG* cfg) const {
361 dump_head(cfg);
362 for (uint i=0; i< number_of_nodes(); i++) {
363 get_node(i)->dump();
364 }
365 tty->print("\n");
366}
367#endif
368
369PhaseCFG::PhaseCFG(Arena* arena, RootNode* root, Matcher& matcher)
370: Phase(CFG)
371, _root(root)
372, _block_arena(arena)
373, _regalloc(NULL)
374, _scheduling_for_pressure(false)
375, _matcher(matcher)
376, _node_to_block_mapping(arena)
377, _node_latency(NULL)
378#ifndef PRODUCT
379, _trace_opto_pipelining(C->directive()->TraceOptoPipeliningOption)
380#endif
381#ifdef ASSERT
382, _raw_oops(arena)
383#endif
384{
385 ResourceMark rm;
386 // I'll need a few machine-specific GotoNodes. Make an Ideal GotoNode,
387 // then Match it into a machine-specific Node. Then clone the machine
388 // Node on demand.
389 Node *x = new GotoNode(NULL);
390 x->init_req(0, x);
391 _goto = matcher.match_tree(x);
392 assert(_goto != NULL, "");
393 _goto->set_req(0,_goto);
394
395 // Build the CFG in Reverse Post Order
396 _number_of_blocks = build_cfg();
397 _root_block = get_block_for_node(_root);
398}
399
400// Build a proper looking CFG. Make every block begin with either a StartNode
401// or a RegionNode. Make every block end with either a Goto, If or Return.
402// The RootNode both starts and ends it's own block. Do this with a recursive
403// backwards walk over the control edges.
404uint PhaseCFG::build_cfg() {
405 Arena *a = Thread::current()->resource_area();
406 VectorSet visited(a);
407
408 // Allocate stack with enough space to avoid frequent realloc
409 Node_Stack nstack(a, C->live_nodes() >> 1);
410 nstack.push(_root, 0);
411 uint sum = 0; // Counter for blocks
412
413 while (nstack.is_nonempty()) {
414 // node and in's index from stack's top
415 // 'np' is _root (see above) or RegionNode, StartNode: we push on stack
416 // only nodes which point to the start of basic block (see below).
417 Node *np = nstack.node();
418 // idx > 0, except for the first node (_root) pushed on stack
419 // at the beginning when idx == 0.
420 // We will use the condition (idx == 0) later to end the build.
421 uint idx = nstack.index();
422 Node *proj = np->in(idx);
423 const Node *x = proj->is_block_proj();
424 // Does the block end with a proper block-ending Node? One of Return,
425 // If or Goto? (This check should be done for visited nodes also).
426 if (x == NULL) { // Does not end right...
427 Node *g = _goto->clone(); // Force it to end in a Goto
428 g->set_req(0, proj);
429 np->set_req(idx, g);
430 x = proj = g;
431 }
432 if (!visited.test_set(x->_idx)) { // Visit this block once
433 // Skip any control-pinned middle'in stuff
434 Node *p = proj;
435 do {
436 proj = p; // Update pointer to last Control
437 p = p->in(0); // Move control forward
438 } while( !p->is_block_proj() &&
439 !p->is_block_start() );
440 // Make the block begin with one of Region or StartNode.
441 if( !p->is_block_start() ) {
442 RegionNode *r = new RegionNode( 2 );
443 r->init_req(1, p); // Insert RegionNode in the way
444 proj->set_req(0, r); // Insert RegionNode in the way
445 p = r;
446 }
447 // 'p' now points to the start of this basic block
448
449 // Put self in array of basic blocks
450 Block *bb = new (_block_arena) Block(_block_arena, p);
451 map_node_to_block(p, bb);
452 map_node_to_block(x, bb);
453 if( x != p ) { // Only for root is x == p
454 bb->push_node((Node*)x);
455 }
456 // Now handle predecessors
457 ++sum; // Count 1 for self block
458 uint cnt = bb->num_preds();
459 for (int i = (cnt - 1); i > 0; i-- ) { // For all predecessors
460 Node *prevproj = p->in(i); // Get prior input
461 assert( !prevproj->is_Con(), "dead input not removed" );
462 // Check to see if p->in(i) is a "control-dependent" CFG edge -
463 // i.e., it splits at the source (via an IF or SWITCH) and merges
464 // at the destination (via a many-input Region).
465 // This breaks critical edges. The RegionNode to start the block
466 // will be added when <p,i> is pulled off the node stack
467 if ( cnt > 2 ) { // Merging many things?
468 assert( prevproj== bb->pred(i),"");
469 if(prevproj->is_block_proj() != prevproj) { // Control-dependent edge?
470 // Force a block on the control-dependent edge
471 Node *g = _goto->clone(); // Force it to end in a Goto
472 g->set_req(0,prevproj);
473 p->set_req(i,g);
474 }
475 }
476 nstack.push(p, i); // 'p' is RegionNode or StartNode
477 }
478 } else { // Post-processing visited nodes
479 nstack.pop(); // remove node from stack
480 // Check if it the fist node pushed on stack at the beginning.
481 if (idx == 0) break; // end of the build
482 // Find predecessor basic block
483 Block *pb = get_block_for_node(x);
484 // Insert into nodes array, if not already there
485 if (!has_block(proj)) {
486 assert( x != proj, "" );
487 // Map basic block of projection
488 map_node_to_block(proj, pb);
489 pb->push_node(proj);
490 }
491 // Insert self as a child of my predecessor block
492 pb->_succs.map(pb->_num_succs++, get_block_for_node(np));
493 assert( pb->get_node(pb->number_of_nodes() - pb->_num_succs)->is_block_proj(),
494 "too many control users, not a CFG?" );
495 }
496 }
497 // Return number of basic blocks for all children and self
498 return sum;
499}
500
501// Inserts a goto & corresponding basic block between
502// block[block_no] and its succ_no'th successor block
503void PhaseCFG::insert_goto_at(uint block_no, uint succ_no) {
504 // get block with block_no
505 assert(block_no < number_of_blocks(), "illegal block number");
506 Block* in = get_block(block_no);
507 // get successor block succ_no
508 assert(succ_no < in->_num_succs, "illegal successor number");
509 Block* out = in->_succs[succ_no];
510 // Compute frequency of the new block. Do this before inserting
511 // new block in case succ_prob() needs to infer the probability from
512 // surrounding blocks.
513 float freq = in->_freq * in->succ_prob(succ_no);
514 // get ProjNode corresponding to the succ_no'th successor of the in block
515 ProjNode* proj = in->get_node(in->number_of_nodes() - in->_num_succs + succ_no)->as_Proj();
516 // create region for basic block
517 RegionNode* region = new RegionNode(2);
518 region->init_req(1, proj);
519 // setup corresponding basic block
520 Block* block = new (_block_arena) Block(_block_arena, region);
521 map_node_to_block(region, block);
522 C->regalloc()->set_bad(region->_idx);
523 // add a goto node
524 Node* gto = _goto->clone(); // get a new goto node
525 gto->set_req(0, region);
526 // add it to the basic block
527 block->push_node(gto);
528 map_node_to_block(gto, block);
529 C->regalloc()->set_bad(gto->_idx);
530 // hook up successor block
531 block->_succs.map(block->_num_succs++, out);
532 // remap successor's predecessors if necessary
533 for (uint i = 1; i < out->num_preds(); i++) {
534 if (out->pred(i) == proj) out->head()->set_req(i, gto);
535 }
536 // remap predecessor's successor to new block
537 in->_succs.map(succ_no, block);
538 // Set the frequency of the new block
539 block->_freq = freq;
540 // add new basic block to basic block list
541 add_block_at(block_no + 1, block);
542}
543
544// Does this block end in a multiway branch that cannot have the default case
545// flipped for another case?
546static bool no_flip_branch(Block *b) {
547 int branch_idx = b->number_of_nodes() - b->_num_succs-1;
548 if (branch_idx < 1) {
549 return false;
550 }
551 Node *branch = b->get_node(branch_idx);
552 if (branch->is_Catch()) {
553 return true;
554 }
555 if (branch->is_Mach()) {
556 if (branch->is_MachNullCheck()) {
557 return true;
558 }
559 int iop = branch->as_Mach()->ideal_Opcode();
560 if (iop == Op_FastLock || iop == Op_FastUnlock) {
561 return true;
562 }
563 // Don't flip if branch has an implicit check.
564 if (branch->as_Mach()->is_TrapBasedCheckNode()) {
565 return true;
566 }
567 }
568 return false;
569}
570
571// Check for NeverBranch at block end. This needs to become a GOTO to the
572// true target. NeverBranch are treated as a conditional branch that always
573// goes the same direction for most of the optimizer and are used to give a
574// fake exit path to infinite loops. At this late stage they need to turn
575// into Goto's so that when you enter the infinite loop you indeed hang.
576void PhaseCFG::convert_NeverBranch_to_Goto(Block *b) {
577 // Find true target
578 int end_idx = b->end_idx();
579 int idx = b->get_node(end_idx+1)->as_Proj()->_con;
580 Block *succ = b->_succs[idx];
581 Node* gto = _goto->clone(); // get a new goto node
582 gto->set_req(0, b->head());
583 Node *bp = b->get_node(end_idx);
584 b->map_node(gto, end_idx); // Slam over NeverBranch
585 map_node_to_block(gto, b);
586 C->regalloc()->set_bad(gto->_idx);
587 b->pop_node(); // Yank projections
588 b->pop_node(); // Yank projections
589 b->_succs.map(0,succ); // Map only successor
590 b->_num_succs = 1;
591 // remap successor's predecessors if necessary
592 uint j;
593 for( j = 1; j < succ->num_preds(); j++)
594 if( succ->pred(j)->in(0) == bp )
595 succ->head()->set_req(j, gto);
596 // Kill alternate exit path
597 Block *dead = b->_succs[1-idx];
598 for( j = 1; j < dead->num_preds(); j++)
599 if( dead->pred(j)->in(0) == bp )
600 break;
601 // Scan through block, yanking dead path from
602 // all regions and phis.
603 dead->head()->del_req(j);
604 for( int k = 1; dead->get_node(k)->is_Phi(); k++ )
605 dead->get_node(k)->del_req(j);
606}
607
608// Helper function to move block bx to the slot following b_index. Return
609// true if the move is successful, otherwise false
610bool PhaseCFG::move_to_next(Block* bx, uint b_index) {
611 if (bx == NULL) return false;
612
613 // Return false if bx is already scheduled.
614 uint bx_index = bx->_pre_order;
615 if ((bx_index <= b_index) && (get_block(bx_index) == bx)) {
616 return false;
617 }
618
619 // Find the current index of block bx on the block list
620 bx_index = b_index + 1;
621 while (bx_index < number_of_blocks() && get_block(bx_index) != bx) {
622 bx_index++;
623 }
624 assert(get_block(bx_index) == bx, "block not found");
625
626 // If the previous block conditionally falls into bx, return false,
627 // because moving bx will create an extra jump.
628 for(uint k = 1; k < bx->num_preds(); k++ ) {
629 Block* pred = get_block_for_node(bx->pred(k));
630 if (pred == get_block(bx_index - 1)) {
631 if (pred->_num_succs != 1) {
632 return false;
633 }
634 }
635 }
636
637 // Reinsert bx just past block 'b'
638 _blocks.remove(bx_index);
639 _blocks.insert(b_index + 1, bx);
640 return true;
641}
642
643// Move empty and uncommon blocks to the end.
644void PhaseCFG::move_to_end(Block *b, uint i) {
645 int e = b->is_Empty();
646 if (e != Block::not_empty) {
647 if (e == Block::empty_with_goto) {
648 // Remove the goto, but leave the block.
649 b->pop_node();
650 }
651 // Mark this block as a connector block, which will cause it to be
652 // ignored in certain functions such as non_connector_successor().
653 b->set_connector();
654 }
655 // Move the empty block to the end, and don't recheck.
656 _blocks.remove(i);
657 _blocks.push(b);
658}
659
660// Set loop alignment for every block
661void PhaseCFG::set_loop_alignment() {
662 uint last = number_of_blocks();
663 assert(get_block(0) == get_root_block(), "");
664
665 for (uint i = 1; i < last; i++) {
666 Block* block = get_block(i);
667 if (block->head()->is_Loop()) {
668 block->set_loop_alignment(block);
669 }
670 }
671}
672
673// Make empty basic blocks to be "connector" blocks, Move uncommon blocks
674// to the end.
675void PhaseCFG::remove_empty_blocks() {
676 // Move uncommon blocks to the end
677 uint last = number_of_blocks();
678 assert(get_block(0) == get_root_block(), "");
679
680 for (uint i = 1; i < last; i++) {
681 Block* block = get_block(i);
682 if (block->is_connector()) {
683 break;
684 }
685
686 // Check for NeverBranch at block end. This needs to become a GOTO to the
687 // true target. NeverBranch are treated as a conditional branch that
688 // always goes the same direction for most of the optimizer and are used
689 // to give a fake exit path to infinite loops. At this late stage they
690 // need to turn into Goto's so that when you enter the infinite loop you
691 // indeed hang.
692 if (block->get_node(block->end_idx())->Opcode() == Op_NeverBranch) {
693 convert_NeverBranch_to_Goto(block);
694 }
695
696 // Look for uncommon blocks and move to end.
697 if (!C->do_freq_based_layout()) {
698 if (is_uncommon(block)) {
699 move_to_end(block, i);
700 last--; // No longer check for being uncommon!
701 if (no_flip_branch(block)) { // Fall-thru case must follow?
702 // Find the fall-thru block
703 block = get_block(i);
704 move_to_end(block, i);
705 last--;
706 }
707 // backup block counter post-increment
708 i--;
709 }
710 }
711 }
712
713 // Move empty blocks to the end
714 last = number_of_blocks();
715 for (uint i = 1; i < last; i++) {
716 Block* block = get_block(i);
717 if (block->is_Empty() != Block::not_empty) {
718 move_to_end(block, i);
719 last--;
720 i--;
721 }
722 } // End of for all blocks
723}
724
725Block *PhaseCFG::fixup_trap_based_check(Node *branch, Block *block, int block_pos, Block *bnext) {
726 // Trap based checks must fall through to the successor with
727 // PROB_ALWAYS.
728 // They should be an If with 2 successors.
729 assert(branch->is_MachIf(), "must be If");
730 assert(block->_num_succs == 2, "must have 2 successors");
731
732 // Get the If node and the projection for the first successor.
733 MachIfNode *iff = block->get_node(block->number_of_nodes()-3)->as_MachIf();
734 ProjNode *proj0 = block->get_node(block->number_of_nodes()-2)->as_Proj();
735 ProjNode *proj1 = block->get_node(block->number_of_nodes()-1)->as_Proj();
736 ProjNode *projt = (proj0->Opcode() == Op_IfTrue) ? proj0 : proj1;
737 ProjNode *projf = (proj0->Opcode() == Op_IfFalse) ? proj0 : proj1;
738
739 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1].
740 assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0");
741 assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1");
742
743 ProjNode *proj_always;
744 ProjNode *proj_never;
745 // We must negate the branch if the implicit check doesn't follow
746 // the branch's TRUE path. Then, the new TRUE branch target will
747 // be the old FALSE branch target.
748 if (iff->_prob <= 2*PROB_NEVER) { // There are small rounding errors.
749 proj_never = projt;
750 proj_always = projf;
751 } else {
752 // We must negate the branch if the trap doesn't follow the
753 // branch's TRUE path. Then, the new TRUE branch target will
754 // be the old FALSE branch target.
755 proj_never = projf;
756 proj_always = projt;
757 iff->negate();
758 }
759 assert(iff->_prob <= 2*PROB_NEVER, "Trap based checks are expected to trap never!");
760 // Map the successors properly
761 block->_succs.map(0, get_block_for_node(proj_never ->raw_out(0))); // The target of the trap.
762 block->_succs.map(1, get_block_for_node(proj_always->raw_out(0))); // The fall through target.
763
764 if (block->get_node(block->number_of_nodes() - block->_num_succs + 1) != proj_always) {
765 block->map_node(proj_never, block->number_of_nodes() - block->_num_succs + 0);
766 block->map_node(proj_always, block->number_of_nodes() - block->_num_succs + 1);
767 }
768
769 // Place the fall through block after this block.
770 Block *bs1 = block->non_connector_successor(1);
771 if (bs1 != bnext && move_to_next(bs1, block_pos)) {
772 bnext = bs1;
773 }
774 // If the fall through block still is not the next block, insert a goto.
775 if (bs1 != bnext) {
776 insert_goto_at(block_pos, 1);
777 }
778 return bnext;
779}
780
781// Fix up the final control flow for basic blocks.
782void PhaseCFG::fixup_flow() {
783 // Fixup final control flow for the blocks. Remove jump-to-next
784 // block. If neither arm of an IF follows the conditional branch, we
785 // have to add a second jump after the conditional. We place the
786 // TRUE branch target in succs[0] for both GOTOs and IFs.
787 for (uint i = 0; i < number_of_blocks(); i++) {
788 Block* block = get_block(i);
789 block->_pre_order = i; // turn pre-order into block-index
790
791 // Connector blocks need no further processing.
792 if (block->is_connector()) {
793 assert((i+1) == number_of_blocks() || get_block(i + 1)->is_connector(), "All connector blocks should sink to the end");
794 continue;
795 }
796 assert(block->is_Empty() != Block::completely_empty, "Empty blocks should be connectors");
797
798 Block* bnext = (i < number_of_blocks() - 1) ? get_block(i + 1) : NULL;
799 Block* bs0 = block->non_connector_successor(0);
800
801 // Check for multi-way branches where I cannot negate the test to
802 // exchange the true and false targets.
803 if (no_flip_branch(block)) {
804 // Find fall through case - if must fall into its target.
805 // Get the index of the branch's first successor.
806 int branch_idx = block->number_of_nodes() - block->_num_succs;
807
808 // The branch is 1 before the branch's first successor.
809 Node *branch = block->get_node(branch_idx-1);
810
811 // Handle no-flip branches which have implicit checks and which require
812 // special block ordering and individual semantics of the 'fall through
813 // case'.
814 if ((TrapBasedNullChecks || TrapBasedRangeChecks) &&
815 branch->is_Mach() && branch->as_Mach()->is_TrapBasedCheckNode()) {
816 bnext = fixup_trap_based_check(branch, block, i, bnext);
817 } else {
818 // Else, default handling for no-flip branches
819 for (uint j2 = 0; j2 < block->_num_succs; j2++) {
820 const ProjNode* p = block->get_node(branch_idx + j2)->as_Proj();
821 if (p->_con == 0) {
822 // successor j2 is fall through case
823 if (block->non_connector_successor(j2) != bnext) {
824 // but it is not the next block => insert a goto
825 insert_goto_at(i, j2);
826 }
827 // Put taken branch in slot 0
828 if (j2 == 0 && block->_num_succs == 2) {
829 // Flip targets in succs map
830 Block *tbs0 = block->_succs[0];
831 Block *tbs1 = block->_succs[1];
832 block->_succs.map(0, tbs1);
833 block->_succs.map(1, tbs0);
834 }
835 break;
836 }
837 }
838 }
839
840 // Remove all CatchProjs
841 for (uint j = 0; j < block->_num_succs; j++) {
842 block->pop_node();
843 }
844
845 } else if (block->_num_succs == 1) {
846 // Block ends in a Goto?
847 if (bnext == bs0) {
848 // We fall into next block; remove the Goto
849 block->pop_node();
850 }
851
852 } else if(block->_num_succs == 2) { // Block ends in a If?
853 // Get opcode of 1st projection (matches _succs[0])
854 // Note: Since this basic block has 2 exits, the last 2 nodes must
855 // be projections (in any order), the 3rd last node must be
856 // the IfNode (we have excluded other 2-way exits such as
857 // CatchNodes already).
858 MachNode* iff = block->get_node(block->number_of_nodes() - 3)->as_Mach();
859 ProjNode* proj0 = block->get_node(block->number_of_nodes() - 2)->as_Proj();
860 ProjNode* proj1 = block->get_node(block->number_of_nodes() - 1)->as_Proj();
861
862 // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1].
863 assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0");
864 assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1");
865
866 Block* bs1 = block->non_connector_successor(1);
867
868 // Check for neither successor block following the current
869 // block ending in a conditional. If so, move one of the
870 // successors after the current one, provided that the
871 // successor was previously unscheduled, but moveable
872 // (i.e., all paths to it involve a branch).
873 if (!C->do_freq_based_layout() && bnext != bs0 && bnext != bs1) {
874 // Choose the more common successor based on the probability
875 // of the conditional branch.
876 Block* bx = bs0;
877 Block* by = bs1;
878
879 // _prob is the probability of taking the true path. Make
880 // p the probability of taking successor #1.
881 float p = iff->as_MachIf()->_prob;
882 if (proj0->Opcode() == Op_IfTrue) {
883 p = 1.0 - p;
884 }
885
886 // Prefer successor #1 if p > 0.5
887 if (p > PROB_FAIR) {
888 bx = bs1;
889 by = bs0;
890 }
891
892 // Attempt the more common successor first
893 if (move_to_next(bx, i)) {
894 bnext = bx;
895 } else if (move_to_next(by, i)) {
896 bnext = by;
897 }
898 }
899
900 // Check for conditional branching the wrong way. Negate
901 // conditional, if needed, so it falls into the following block
902 // and branches to the not-following block.
903
904 // Check for the next block being in succs[0]. We are going to branch
905 // to succs[0], so we want the fall-thru case as the next block in
906 // succs[1].
907 if (bnext == bs0) {
908 // Fall-thru case in succs[0], so flip targets in succs map
909 Block* tbs0 = block->_succs[0];
910 Block* tbs1 = block->_succs[1];
911 block->_succs.map(0, tbs1);
912 block->_succs.map(1, tbs0);
913 // Flip projection for each target
914 ProjNode* tmp = proj0;
915 proj0 = proj1;
916 proj1 = tmp;
917
918 } else if(bnext != bs1) {
919 // Need a double-branch
920 // The existing conditional branch need not change.
921 // Add a unconditional branch to the false target.
922 // Alas, it must appear in its own block and adding a
923 // block this late in the game is complicated. Sigh.
924 insert_goto_at(i, 1);
925 }
926
927 // Make sure we TRUE branch to the target
928 if (proj0->Opcode() == Op_IfFalse) {
929 iff->as_MachIf()->negate();
930 }
931
932 block->pop_node(); // Remove IfFalse & IfTrue projections
933 block->pop_node();
934
935 } else {
936 // Multi-exit block, e.g. a switch statement
937 // But we don't need to do anything here
938 }
939 } // End of for all blocks
940}
941
942
943// postalloc_expand: Expand nodes after register allocation.
944//
945// postalloc_expand has to be called after register allocation, just
946// before output (i.e. scheduling). It only gets called if
947// Matcher::require_postalloc_expand is true.
948//
949// Background:
950//
951// Nodes that are expandend (one compound node requiring several
952// assembler instructions to be implemented split into two or more
953// non-compound nodes) after register allocation are not as nice as
954// the ones expanded before register allocation - they don't
955// participate in optimizations as global code motion. But after
956// register allocation we can expand nodes that use registers which
957// are not spillable or registers that are not allocated, because the
958// old compound node is simply replaced (in its location in the basic
959// block) by a new subgraph which does not contain compound nodes any
960// more. The scheduler called during output can later on process these
961// non-compound nodes.
962//
963// Implementation:
964//
965// Nodes requiring postalloc expand are specified in the ad file by using
966// a postalloc_expand statement instead of ins_encode. A postalloc_expand
967// contains a single call to an encoding, as does an ins_encode
968// statement. Instead of an emit() function a postalloc_expand() function
969// is generated that doesn't emit assembler but creates a new
970// subgraph. The code below calls this postalloc_expand function for each
971// node with the appropriate attribute. This function returns the new
972// nodes generated in an array passed in the call. The old node,
973// potential MachTemps before and potential Projs after it then get
974// disconnected and replaced by the new nodes. The instruction
975// generating the result has to be the last one in the array. In
976// general it is assumed that Projs after the node expanded are
977// kills. These kills are not required any more after expanding as
978// there are now explicitly visible def-use chains and the Projs are
979// removed. This does not hold for calls: They do not only have
980// kill-Projs but also Projs defining values. Therefore Projs after
981// the node expanded are removed for all but for calls. If a node is
982// to be reused, it must be added to the nodes list returned, and it
983// will be added again.
984//
985// Implementing the postalloc_expand function for a node in an enc_class
986// is rather tedious. It requires knowledge about many node details, as
987// the nodes and the subgraph must be hand crafted. To simplify this,
988// adlc generates some utility variables into the postalloc_expand function,
989// e.g., holding the operands as specified by the postalloc_expand encoding
990// specification, e.g.:
991// * unsigned idx_<par_name> holding the index of the node in the ins
992// * Node *n_<par_name> holding the node loaded from the ins
993// * MachOpnd *op_<par_name> holding the corresponding operand
994//
995// The ordering of operands can not be determined by looking at a
996// rule. Especially if a match rule matches several different trees,
997// several nodes are generated from one instruct specification with
998// different operand orderings. In this case the adlc generated
999// variables are the only way to access the ins and operands
1000// deterministically.
1001//
1002// If assigning a register to a node that contains an oop, don't
1003// forget to call ra_->set_oop() for the node.
1004void PhaseCFG::postalloc_expand(PhaseRegAlloc* _ra) {
1005 GrowableArray <Node *> new_nodes(32); // Array with new nodes filled by postalloc_expand function of node.
1006 GrowableArray <Node *> remove(32);
1007 GrowableArray <Node *> succs(32);
1008 unsigned int max_idx = C->unique(); // Remember to distinguish new from old nodes.
1009 DEBUG_ONLY(bool foundNode = false);
1010
1011 // for all blocks
1012 for (uint i = 0; i < number_of_blocks(); i++) {
1013 Block *b = _blocks[i];
1014 // For all instructions in the current block.
1015 for (uint j = 0; j < b->number_of_nodes(); j++) {
1016 Node *n = b->get_node(j);
1017 if (n->is_Mach() && n->as_Mach()->requires_postalloc_expand()) {
1018#ifdef ASSERT
1019 if (TracePostallocExpand) {
1020 if (!foundNode) {
1021 foundNode = true;
1022 tty->print("POSTALLOC EXPANDING %d %s\n", C->compile_id(),
1023 C->method() ? C->method()->name()->as_utf8() : C->stub_name());
1024 }
1025 tty->print(" postalloc expanding "); n->dump();
1026 if (Verbose) {
1027 tty->print(" with ins:\n");
1028 for (uint k = 0; k < n->len(); ++k) {
1029 if (n->in(k)) { tty->print(" "); n->in(k)->dump(); }
1030 }
1031 }
1032 }
1033#endif
1034 new_nodes.clear();
1035 // Collect nodes that have to be removed from the block later on.
1036 uint req = n->req();
1037 remove.clear();
1038 for (uint k = 0; k < req; ++k) {
1039 if (n->in(k) && n->in(k)->is_MachTemp()) {
1040 remove.push(n->in(k)); // MachTemps which are inputs to the old node have to be removed.
1041 n->in(k)->del_req(0);
1042 j--;
1043 }
1044 }
1045
1046 // Check whether we can allocate enough nodes. We set a fix limit for
1047 // the size of postalloc expands with this.
1048 uint unique_limit = C->unique() + 40;
1049 if (unique_limit >= _ra->node_regs_max_index()) {
1050 Compile::current()->record_failure("out of nodes in postalloc expand");
1051 return;
1052 }
1053
1054 // Emit (i.e. generate new nodes).
1055 n->as_Mach()->postalloc_expand(&new_nodes, _ra);
1056
1057 assert(C->unique() < unique_limit, "You allocated too many nodes in your postalloc expand.");
1058
1059 // Disconnect the inputs of the old node.
1060 //
1061 // We reuse MachSpillCopy nodes. If we need to expand them, there
1062 // are many, so reusing pays off. If reused, the node already
1063 // has the new ins. n must be the last node on new_nodes list.
1064 if (!n->is_MachSpillCopy()) {
1065 for (int k = req - 1; k >= 0; --k) {
1066 n->del_req(k);
1067 }
1068 }
1069
1070#ifdef ASSERT
1071 // Check that all nodes have proper operands.
1072 for (int k = 0; k < new_nodes.length(); ++k) {
1073 if (new_nodes.at(k)->_idx < max_idx || !new_nodes.at(k)->is_Mach()) continue; // old node, Proj ...
1074 MachNode *m = new_nodes.at(k)->as_Mach();
1075 for (unsigned int l = 0; l < m->num_opnds(); ++l) {
1076 if (MachOper::notAnOper(m->_opnds[l])) {
1077 outputStream *os = tty;
1078 os->print("Node %s ", m->Name());
1079 os->print("has invalid opnd %d: %p\n", l, m->_opnds[l]);
1080 assert(0, "Invalid operands, see inline trace in hs_err_pid file.");
1081 }
1082 }
1083 }
1084#endif
1085
1086 // Collect succs of old node in remove (for projections) and in succs (for
1087 // all other nodes) do _not_ collect projections in remove (but in succs)
1088 // in case the node is a call. We need the projections for calls as they are
1089 // associated with registes (i.e. they are defs).
1090 succs.clear();
1091 for (DUIterator k = n->outs(); n->has_out(k); k++) {
1092 if (n->out(k)->is_Proj() && !n->is_MachCall() && !n->is_MachBranch()) {
1093 remove.push(n->out(k));
1094 } else {
1095 succs.push(n->out(k));
1096 }
1097 }
1098 // Replace old node n as input of its succs by last of the new nodes.
1099 for (int k = 0; k < succs.length(); ++k) {
1100 Node *succ = succs.at(k);
1101 for (uint l = 0; l < succ->req(); ++l) {
1102 if (succ->in(l) == n) {
1103 succ->set_req(l, new_nodes.at(new_nodes.length() - 1));
1104 }
1105 }
1106 for (uint l = succ->req(); l < succ->len(); ++l) {
1107 if (succ->in(l) == n) {
1108 succ->set_prec(l, new_nodes.at(new_nodes.length() - 1));
1109 }
1110 }
1111 }
1112
1113 // Index of old node in block.
1114 uint index = b->find_node(n);
1115 // Insert new nodes into block and map them in nodes->blocks array
1116 // and remember last node in n2.
1117 Node *n2 = NULL;
1118 for (int k = 0; k < new_nodes.length(); ++k) {
1119 n2 = new_nodes.at(k);
1120 b->insert_node(n2, ++index);
1121 map_node_to_block(n2, b);
1122 }
1123
1124 // Add old node n to remove and remove them all from block.
1125 remove.push(n);
1126 j--;
1127#ifdef ASSERT
1128 if (TracePostallocExpand && Verbose) {
1129 tty->print(" removing:\n");
1130 for (int k = 0; k < remove.length(); ++k) {
1131 tty->print(" "); remove.at(k)->dump();
1132 }
1133 tty->print(" inserting:\n");
1134 for (int k = 0; k < new_nodes.length(); ++k) {
1135 tty->print(" "); new_nodes.at(k)->dump();
1136 }
1137 }
1138#endif
1139 for (int k = 0; k < remove.length(); ++k) {
1140 if (b->contains(remove.at(k))) {
1141 b->find_remove(remove.at(k));
1142 } else {
1143 assert(remove.at(k)->is_Proj() && (remove.at(k)->in(0)->is_MachBranch()), "");
1144 }
1145 }
1146 // If anything has been inserted (n2 != NULL), continue after last node inserted.
1147 // This does not always work. Some postalloc expands don't insert any nodes, if they
1148 // do optimizations (e.g., max(x,x)). In this case we decrement j accordingly.
1149 j = n2 ? b->find_node(n2) : j;
1150 }
1151 }
1152 }
1153
1154#ifdef ASSERT
1155 if (foundNode) {
1156 tty->print("FINISHED %d %s\n", C->compile_id(),
1157 C->method() ? C->method()->name()->as_utf8() : C->stub_name());
1158 tty->flush();
1159 }
1160#endif
1161}
1162
1163
1164//------------------------------dump-------------------------------------------
1165#ifndef PRODUCT
1166void PhaseCFG::_dump_cfg( const Node *end, VectorSet &visited ) const {
1167 const Node *x = end->is_block_proj();
1168 assert( x, "not a CFG" );
1169
1170 // Do not visit this block again
1171 if( visited.test_set(x->_idx) ) return;
1172
1173 // Skip through this block
1174 const Node *p = x;
1175 do {
1176 p = p->in(0); // Move control forward
1177 assert( !p->is_block_proj() || p->is_Root(), "not a CFG" );
1178 } while( !p->is_block_start() );
1179
1180 // Recursively visit
1181 for (uint i = 1; i < p->req(); i++) {
1182 _dump_cfg(p->in(i), visited);
1183 }
1184
1185 // Dump the block
1186 get_block_for_node(p)->dump(this);
1187}
1188
1189void PhaseCFG::dump( ) const {
1190 tty->print("\n--- CFG --- %d BBs\n", number_of_blocks());
1191 if (_blocks.size()) { // Did we do basic-block layout?
1192 for (uint i = 0; i < number_of_blocks(); i++) {
1193 const Block* block = get_block(i);
1194 block->dump(this);
1195 }
1196 } else { // Else do it with a DFS
1197 VectorSet visited(_block_arena);
1198 _dump_cfg(_root,visited);
1199 }
1200}
1201
1202void PhaseCFG::dump_headers() {
1203 for (uint i = 0; i < number_of_blocks(); i++) {
1204 Block* block = get_block(i);
1205 if (block != NULL) {
1206 block->dump_head(this);
1207 }
1208 }
1209}
1210
1211void PhaseCFG::verify() const {
1212#ifdef ASSERT
1213 // Verify sane CFG
1214 for (uint i = 0; i < number_of_blocks(); i++) {
1215 Block* block = get_block(i);
1216 uint cnt = block->number_of_nodes();
1217 uint j;
1218 for (j = 0; j < cnt; j++) {
1219 Node *n = block->get_node(j);
1220 assert(get_block_for_node(n) == block, "");
1221 if (j >= 1 && n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CreateEx) {
1222 assert(j == 1 || block->get_node(j-1)->is_Phi(), "CreateEx must be first instruction in block");
1223 }
1224 if (n->needs_anti_dependence_check()) {
1225 verify_anti_dependences(block, n);
1226 }
1227 for (uint k = 0; k < n->req(); k++) {
1228 Node *def = n->in(k);
1229 if (def && def != n) {
1230 assert(get_block_for_node(def) || def->is_Con(), "must have block; constants for debug info ok");
1231 // Verify that instructions in the block is in correct order.
1232 // Uses must follow their definition if they are at the same block.
1233 // Mostly done to check that MachSpillCopy nodes are placed correctly
1234 // when CreateEx node is moved in build_ifg_physical().
1235 if (get_block_for_node(def) == block && !(block->head()->is_Loop() && n->is_Phi()) &&
1236 // See (+++) comment in reg_split.cpp
1237 !(n->jvms() != NULL && n->jvms()->is_monitor_use(k))) {
1238 bool is_loop = false;
1239 if (n->is_Phi()) {
1240 for (uint l = 1; l < def->req(); l++) {
1241 if (n == def->in(l)) {
1242 is_loop = true;
1243 break; // Some kind of loop
1244 }
1245 }
1246 }
1247 assert(is_loop || block->find_node(def) < j, "uses must follow definitions");
1248 }
1249 }
1250 }
1251 }
1252
1253 j = block->end_idx();
1254 Node* bp = (Node*)block->get_node(block->number_of_nodes() - 1)->is_block_proj();
1255 assert(bp, "last instruction must be a block proj");
1256 assert(bp == block->get_node(j), "wrong number of successors for this block");
1257 if (bp->is_Catch()) {
1258 while (block->get_node(--j)->is_MachProj()) {
1259 ;
1260 }
1261 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1262 } else if (bp->is_Mach() && bp->as_Mach()->ideal_Opcode() == Op_If) {
1263 assert(block->_num_succs == 2, "Conditional branch must have two targets");
1264 }
1265 }
1266#endif
1267}
1268#endif
1269
1270UnionFind::UnionFind( uint max ) : _cnt(max), _max(max), _indices(NEW_RESOURCE_ARRAY(uint,max)) {
1271 Copy::zero_to_bytes( _indices, sizeof(uint)*max );
1272}
1273
1274void UnionFind::extend( uint from_idx, uint to_idx ) {
1275 _nesting.check();
1276 if( from_idx >= _max ) {
1277 uint size = 16;
1278 while( size <= from_idx ) size <<=1;
1279 _indices = REALLOC_RESOURCE_ARRAY( uint, _indices, _max, size );
1280 _max = size;
1281 }
1282 while( _cnt <= from_idx ) _indices[_cnt++] = 0;
1283 _indices[from_idx] = to_idx;
1284}
1285
1286void UnionFind::reset( uint max ) {
1287 // Force the Union-Find mapping to be at least this large
1288 extend(max,0);
1289 // Initialize to be the ID mapping.
1290 for( uint i=0; i<max; i++ ) map(i,i);
1291}
1292
1293// Straight out of Tarjan's union-find algorithm
1294uint UnionFind::Find_compress( uint idx ) {
1295 uint cur = idx;
1296 uint next = lookup(cur);
1297 while( next != cur ) { // Scan chain of equivalences
1298 assert( next < cur, "always union smaller" );
1299 cur = next; // until find a fixed-point
1300 next = lookup(cur);
1301 }
1302 // Core of union-find algorithm: update chain of
1303 // equivalences to be equal to the root.
1304 while( idx != next ) {
1305 uint tmp = lookup(idx);
1306 map(idx, next);
1307 idx = tmp;
1308 }
1309 return idx;
1310}
1311
1312// Like Find above, but no path compress, so bad asymptotic behavior
1313uint UnionFind::Find_const( uint idx ) const {
1314 if( idx == 0 ) return idx; // Ignore the zero idx
1315 // Off the end? This can happen during debugging dumps
1316 // when data structures have not finished being updated.
1317 if( idx >= _max ) return idx;
1318 uint next = lookup(idx);
1319 while( next != idx ) { // Scan chain of equivalences
1320 idx = next; // until find a fixed-point
1321 next = lookup(idx);
1322 }
1323 return next;
1324}
1325
1326// union 2 sets together.
1327void UnionFind::Union( uint idx1, uint idx2 ) {
1328 uint src = Find(idx1);
1329 uint dst = Find(idx2);
1330 assert( src, "" );
1331 assert( dst, "" );
1332 assert( src < _max, "oob" );
1333 assert( dst < _max, "oob" );
1334 assert( src < dst, "always union smaller" );
1335 map(dst,src);
1336}
1337
1338#ifndef PRODUCT
1339void Trace::dump( ) const {
1340 tty->print_cr("Trace (freq %f)", first_block()->_freq);
1341 for (Block *b = first_block(); b != NULL; b = next(b)) {
1342 tty->print(" B%d", b->_pre_order);
1343 if (b->head()->is_Loop()) {
1344 tty->print(" (L%d)", b->compute_loop_alignment());
1345 }
1346 if (b->has_loop_alignment()) {
1347 tty->print(" (T%d)", b->code_alignment());
1348 }
1349 }
1350 tty->cr();
1351}
1352
1353void CFGEdge::dump( ) const {
1354 tty->print(" B%d --> B%d Freq: %f out:%3d%% in:%3d%% State: ",
1355 from()->_pre_order, to()->_pre_order, freq(), _from_pct, _to_pct);
1356 switch(state()) {
1357 case connected:
1358 tty->print("connected");
1359 break;
1360 case open:
1361 tty->print("open");
1362 break;
1363 case interior:
1364 tty->print("interior");
1365 break;
1366 }
1367 if (infrequent()) {
1368 tty->print(" infrequent");
1369 }
1370 tty->cr();
1371}
1372#endif
1373
1374// Comparison function for edges
1375static int edge_order(CFGEdge **e0, CFGEdge **e1) {
1376 float freq0 = (*e0)->freq();
1377 float freq1 = (*e1)->freq();
1378 if (freq0 != freq1) {
1379 return freq0 > freq1 ? -1 : 1;
1380 }
1381
1382 int dist0 = (*e0)->to()->_rpo - (*e0)->from()->_rpo;
1383 int dist1 = (*e1)->to()->_rpo - (*e1)->from()->_rpo;
1384
1385 return dist1 - dist0;
1386}
1387
1388// Comparison function for edges
1389extern "C" int trace_frequency_order(const void *p0, const void *p1) {
1390 Trace *tr0 = *(Trace **) p0;
1391 Trace *tr1 = *(Trace **) p1;
1392 Block *b0 = tr0->first_block();
1393 Block *b1 = tr1->first_block();
1394
1395 // The trace of connector blocks goes at the end;
1396 // we only expect one such trace
1397 if (b0->is_connector() != b1->is_connector()) {
1398 return b1->is_connector() ? -1 : 1;
1399 }
1400
1401 // Pull more frequently executed blocks to the beginning
1402 float freq0 = b0->_freq;
1403 float freq1 = b1->_freq;
1404 if (freq0 != freq1) {
1405 return freq0 > freq1 ? -1 : 1;
1406 }
1407
1408 int diff = tr0->first_block()->_rpo - tr1->first_block()->_rpo;
1409
1410 return diff;
1411}
1412
1413// Find edges of interest, i.e, those which can fall through. Presumes that
1414// edges which don't fall through are of low frequency and can be generally
1415// ignored. Initialize the list of traces.
1416void PhaseBlockLayout::find_edges() {
1417 // Walk the blocks, creating edges and Traces
1418 uint i;
1419 Trace *tr = NULL;
1420 for (i = 0; i < _cfg.number_of_blocks(); i++) {
1421 Block* b = _cfg.get_block(i);
1422 tr = new Trace(b, next, prev);
1423 traces[tr->id()] = tr;
1424
1425 // All connector blocks should be at the end of the list
1426 if (b->is_connector()) break;
1427
1428 // If this block and the next one have a one-to-one successor
1429 // predecessor relationship, simply append the next block
1430 int nfallthru = b->num_fall_throughs();
1431 while (nfallthru == 1 &&
1432 b->succ_fall_through(0)) {
1433 Block *n = b->_succs[0];
1434
1435 // Skip over single-entry connector blocks, we don't want to
1436 // add them to the trace.
1437 while (n->is_connector() && n->num_preds() == 1) {
1438 n = n->_succs[0];
1439 }
1440
1441 // We see a merge point, so stop search for the next block
1442 if (n->num_preds() != 1) break;
1443
1444 i++;
1445 assert(n == _cfg.get_block(i), "expecting next block");
1446 tr->append(n);
1447 uf->map(n->_pre_order, tr->id());
1448 traces[n->_pre_order] = NULL;
1449 nfallthru = b->num_fall_throughs();
1450 b = n;
1451 }
1452
1453 if (nfallthru > 0) {
1454 // Create a CFGEdge for each outgoing
1455 // edge that could be a fall-through.
1456 for (uint j = 0; j < b->_num_succs; j++ ) {
1457 if (b->succ_fall_through(j)) {
1458 Block *target = b->non_connector_successor(j);
1459 float freq = b->_freq * b->succ_prob(j);
1460 int from_pct = (int) ((100 * freq) / b->_freq);
1461 int to_pct = (int) ((100 * freq) / target->_freq);
1462 edges->append(new CFGEdge(b, target, freq, from_pct, to_pct));
1463 }
1464 }
1465 }
1466 }
1467
1468 // Group connector blocks into one trace
1469 for (i++; i < _cfg.number_of_blocks(); i++) {
1470 Block *b = _cfg.get_block(i);
1471 assert(b->is_connector(), "connector blocks at the end");
1472 tr->append(b);
1473 uf->map(b->_pre_order, tr->id());
1474 traces[b->_pre_order] = NULL;
1475 }
1476}
1477
1478// Union two traces together in uf, and null out the trace in the list
1479void PhaseBlockLayout::union_traces(Trace* updated_trace, Trace* old_trace) {
1480 uint old_id = old_trace->id();
1481 uint updated_id = updated_trace->id();
1482
1483 uint lo_id = updated_id;
1484 uint hi_id = old_id;
1485
1486 // If from is greater than to, swap values to meet
1487 // UnionFind guarantee.
1488 if (updated_id > old_id) {
1489 lo_id = old_id;
1490 hi_id = updated_id;
1491
1492 // Fix up the trace ids
1493 traces[lo_id] = traces[updated_id];
1494 updated_trace->set_id(lo_id);
1495 }
1496
1497 // Union the lower with the higher and remove the pointer
1498 // to the higher.
1499 uf->Union(lo_id, hi_id);
1500 traces[hi_id] = NULL;
1501}
1502
1503// Append traces together via the most frequently executed edges
1504void PhaseBlockLayout::grow_traces() {
1505 // Order the edges, and drive the growth of Traces via the most
1506 // frequently executed edges.
1507 edges->sort(edge_order);
1508 for (int i = 0; i < edges->length(); i++) {
1509 CFGEdge *e = edges->at(i);
1510
1511 if (e->state() != CFGEdge::open) continue;
1512
1513 Block *src_block = e->from();
1514 Block *targ_block = e->to();
1515
1516 // Don't grow traces along backedges?
1517 if (!BlockLayoutRotateLoops) {
1518 if (targ_block->_rpo <= src_block->_rpo) {
1519 targ_block->set_loop_alignment(targ_block);
1520 continue;
1521 }
1522 }
1523
1524 Trace *src_trace = trace(src_block);
1525 Trace *targ_trace = trace(targ_block);
1526
1527 // If the edge in question can join two traces at their ends,
1528 // append one trace to the other.
1529 if (src_trace->last_block() == src_block) {
1530 if (src_trace == targ_trace) {
1531 e->set_state(CFGEdge::interior);
1532 if (targ_trace->backedge(e)) {
1533 // Reset i to catch any newly eligible edge
1534 // (Or we could remember the first "open" edge, and reset there)
1535 i = 0;
1536 }
1537 } else if (targ_trace->first_block() == targ_block) {
1538 e->set_state(CFGEdge::connected);
1539 src_trace->append(targ_trace);
1540 union_traces(src_trace, targ_trace);
1541 }
1542 }
1543 }
1544}
1545
1546// Embed one trace into another, if the fork or join points are sufficiently
1547// balanced.
1548void PhaseBlockLayout::merge_traces(bool fall_thru_only) {
1549 // Walk the edge list a another time, looking at unprocessed edges.
1550 // Fold in diamonds
1551 for (int i = 0; i < edges->length(); i++) {
1552 CFGEdge *e = edges->at(i);
1553
1554 if (e->state() != CFGEdge::open) continue;
1555 if (fall_thru_only) {
1556 if (e->infrequent()) continue;
1557 }
1558
1559 Block *src_block = e->from();
1560 Trace *src_trace = trace(src_block);
1561 bool src_at_tail = src_trace->last_block() == src_block;
1562
1563 Block *targ_block = e->to();
1564 Trace *targ_trace = trace(targ_block);
1565 bool targ_at_start = targ_trace->first_block() == targ_block;
1566
1567 if (src_trace == targ_trace) {
1568 // This may be a loop, but we can't do much about it.
1569 e->set_state(CFGEdge::interior);
1570 continue;
1571 }
1572
1573 if (fall_thru_only) {
1574 // If the edge links the middle of two traces, we can't do anything.
1575 // Mark the edge and continue.
1576 if (!src_at_tail & !targ_at_start) {
1577 continue;
1578 }
1579
1580 // Don't grow traces along backedges?
1581 if (!BlockLayoutRotateLoops && (targ_block->_rpo <= src_block->_rpo)) {
1582 continue;
1583 }
1584
1585 // If both ends of the edge are available, why didn't we handle it earlier?
1586 assert(src_at_tail ^ targ_at_start, "Should have caught this edge earlier.");
1587
1588 if (targ_at_start) {
1589 // Insert the "targ" trace in the "src" trace if the insertion point
1590 // is a two way branch.
1591 // Better profitability check possible, but may not be worth it.
1592 // Someday, see if the this "fork" has an associated "join";
1593 // then make a policy on merging this trace at the fork or join.
1594 // For example, other things being equal, it may be better to place this
1595 // trace at the join point if the "src" trace ends in a two-way, but
1596 // the insertion point is one-way.
1597 assert(src_block->num_fall_throughs() == 2, "unexpected diamond");
1598 e->set_state(CFGEdge::connected);
1599 src_trace->insert_after(src_block, targ_trace);
1600 union_traces(src_trace, targ_trace);
1601 } else if (src_at_tail) {
1602 if (src_trace != trace(_cfg.get_root_block())) {
1603 e->set_state(CFGEdge::connected);
1604 targ_trace->insert_before(targ_block, src_trace);
1605 union_traces(targ_trace, src_trace);
1606 }
1607 }
1608 } else if (e->state() == CFGEdge::open) {
1609 // Append traces, even without a fall-thru connection.
1610 // But leave root entry at the beginning of the block list.
1611 if (targ_trace != trace(_cfg.get_root_block())) {
1612 e->set_state(CFGEdge::connected);
1613 src_trace->append(targ_trace);
1614 union_traces(src_trace, targ_trace);
1615 }
1616 }
1617 }
1618}
1619
1620// Order the sequence of the traces in some desirable way, and fixup the
1621// jumps at the end of each block.
1622void PhaseBlockLayout::reorder_traces(int count) {
1623 ResourceArea *area = Thread::current()->resource_area();
1624 Trace ** new_traces = NEW_ARENA_ARRAY(area, Trace *, count);
1625 Block_List worklist;
1626 int new_count = 0;
1627
1628 // Compact the traces.
1629 for (int i = 0; i < count; i++) {
1630 Trace *tr = traces[i];
1631 if (tr != NULL) {
1632 new_traces[new_count++] = tr;
1633 }
1634 }
1635
1636 // The entry block should be first on the new trace list.
1637 Trace *tr = trace(_cfg.get_root_block());
1638 assert(tr == new_traces[0], "entry trace misplaced");
1639
1640 // Sort the new trace list by frequency
1641 qsort(new_traces + 1, new_count - 1, sizeof(new_traces[0]), trace_frequency_order);
1642
1643 // Patch up the successor blocks
1644 _cfg.clear_blocks();
1645 for (int i = 0; i < new_count; i++) {
1646 Trace *tr = new_traces[i];
1647 if (tr != NULL) {
1648 tr->fixup_blocks(_cfg);
1649 }
1650 }
1651}
1652
1653// Order basic blocks based on frequency
1654PhaseBlockLayout::PhaseBlockLayout(PhaseCFG &cfg)
1655: Phase(BlockLayout)
1656, _cfg(cfg) {
1657 ResourceMark rm;
1658 ResourceArea *area = Thread::current()->resource_area();
1659
1660 // List of traces
1661 int size = _cfg.number_of_blocks() + 1;
1662 traces = NEW_ARENA_ARRAY(area, Trace *, size);
1663 memset(traces, 0, size*sizeof(Trace*));
1664 next = NEW_ARENA_ARRAY(area, Block *, size);
1665 memset(next, 0, size*sizeof(Block *));
1666 prev = NEW_ARENA_ARRAY(area, Block *, size);
1667 memset(prev , 0, size*sizeof(Block *));
1668
1669 // List of edges
1670 edges = new GrowableArray<CFGEdge*>;
1671
1672 // Mapping block index --> block_trace
1673 uf = new UnionFind(size);
1674 uf->reset(size);
1675
1676 // Find edges and create traces.
1677 find_edges();
1678
1679 // Grow traces at their ends via most frequent edges.
1680 grow_traces();
1681
1682 // Merge one trace into another, but only at fall-through points.
1683 // This may make diamonds and other related shapes in a trace.
1684 merge_traces(true);
1685
1686 // Run merge again, allowing two traces to be catenated, even if
1687 // one does not fall through into the other. This appends loosely
1688 // related traces to be near each other.
1689 merge_traces(false);
1690
1691 // Re-order all the remaining traces by frequency
1692 reorder_traces(size);
1693
1694 assert(_cfg.number_of_blocks() >= (uint) (size - 1), "number of blocks can not shrink");
1695}
1696
1697
1698// Edge e completes a loop in a trace. If the target block is head of the
1699// loop, rotate the loop block so that the loop ends in a conditional branch.
1700bool Trace::backedge(CFGEdge *e) {
1701 bool loop_rotated = false;
1702 Block *src_block = e->from();
1703 Block *targ_block = e->to();
1704
1705 assert(last_block() == src_block, "loop discovery at back branch");
1706 if (first_block() == targ_block) {
1707 if (BlockLayoutRotateLoops && last_block()->num_fall_throughs() < 2) {
1708 // Find the last block in the trace that has a conditional
1709 // branch.
1710 Block *b;
1711 for (b = last_block(); b != NULL; b = prev(b)) {
1712 if (b->num_fall_throughs() == 2) {
1713 break;
1714 }
1715 }
1716
1717 if (b != last_block() && b != NULL) {
1718 loop_rotated = true;
1719
1720 // Rotate the loop by doing two-part linked-list surgery.
1721 append(first_block());
1722 break_loop_after(b);
1723 }
1724 }
1725
1726 // Backbranch to the top of a trace
1727 // Scroll forward through the trace from the targ_block. If we find
1728 // a loop head before another loop top, use the the loop head alignment.
1729 for (Block *b = targ_block; b != NULL; b = next(b)) {
1730 if (b->has_loop_alignment()) {
1731 break;
1732 }
1733 if (b->head()->is_Loop()) {
1734 targ_block = b;
1735 break;
1736 }
1737 }
1738
1739 first_block()->set_loop_alignment(targ_block);
1740
1741 } else {
1742 // That loop may already have a loop top (we're reaching it again
1743 // through the backedge of an outer loop)
1744 Block* b = prev(targ_block);
1745 bool has_top = targ_block->head()->is_Loop() && b->has_loop_alignment() && !b->head()->is_Loop();
1746 if (!has_top) {
1747 // Backbranch into the middle of a trace
1748 targ_block->set_loop_alignment(targ_block);
1749 }
1750 }
1751
1752 return loop_rotated;
1753}
1754
1755// push blocks onto the CFG list
1756// ensure that blocks have the correct two-way branch sense
1757void Trace::fixup_blocks(PhaseCFG &cfg) {
1758 Block *last = last_block();
1759 for (Block *b = first_block(); b != NULL; b = next(b)) {
1760 cfg.add_block(b);
1761 if (!b->is_connector()) {
1762 int nfallthru = b->num_fall_throughs();
1763 if (b != last) {
1764 if (nfallthru == 2) {
1765 // Ensure that the sense of the branch is correct
1766 Block *bnext = next(b);
1767 Block *bs0 = b->non_connector_successor(0);
1768
1769 MachNode *iff = b->get_node(b->number_of_nodes() - 3)->as_Mach();
1770 ProjNode *proj0 = b->get_node(b->number_of_nodes() - 2)->as_Proj();
1771 ProjNode *proj1 = b->get_node(b->number_of_nodes() - 1)->as_Proj();
1772
1773 if (bnext == bs0) {
1774 // Fall-thru case in succs[0], should be in succs[1]
1775
1776 // Flip targets in _succs map
1777 Block *tbs0 = b->_succs[0];
1778 Block *tbs1 = b->_succs[1];
1779 b->_succs.map( 0, tbs1 );
1780 b->_succs.map( 1, tbs0 );
1781
1782 // Flip projections to match targets
1783 b->map_node(proj1, b->number_of_nodes() - 2);
1784 b->map_node(proj0, b->number_of_nodes() - 1);
1785 }
1786 }
1787 }
1788 }
1789 }
1790}
1791