Mon, 28 Jul 2008 17:12:52 -0700
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
Summary: Escape Analysis fixes.
Reviewed-by: never, rasbold
1 /*
2 * Copyright 1999-2008 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_loopopts.cpp.incl"
28 //=============================================================================
29 //------------------------------split_thru_phi---------------------------------
30 // Split Node 'n' through merge point if there is enough win.
31 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
32 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
33 // ConvI2L may have type information on it which is unsafe to push up
34 // so disable this for now
35 return NULL;
36 }
37 int wins = 0;
38 assert( !n->is_CFG(), "" );
39 assert( region->is_Region(), "" );
41 const Type* type = n->bottom_type();
42 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr();
43 Node *phi;
44 if( t_oop != NULL && t_oop->is_known_instance_field() ) {
45 int iid = t_oop->instance_id();
46 int index = C->get_alias_index(t_oop);
47 int offset = t_oop->offset();
48 phi = new (C,region->req()) PhiNode(region, type, NULL, iid, index, offset);
49 } else {
50 phi = new (C,region->req()) PhiNode(region, type);
51 }
52 uint old_unique = C->unique();
53 for( uint i = 1; i < region->req(); i++ ) {
54 Node *x;
55 Node* the_clone = NULL;
56 if( region->in(i) == C->top() ) {
57 x = C->top(); // Dead path? Use a dead data op
58 } else {
59 x = n->clone(); // Else clone up the data op
60 the_clone = x; // Remember for possible deletion.
61 // Alter data node to use pre-phi inputs
62 if( n->in(0) == region )
63 x->set_req( 0, region->in(i) );
64 for( uint j = 1; j < n->req(); j++ ) {
65 Node *in = n->in(j);
66 if( in->is_Phi() && in->in(0) == region )
67 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
68 }
69 }
70 // Check for a 'win' on some paths
71 const Type *t = x->Value(&_igvn);
73 bool singleton = t->singleton();
75 // A TOP singleton indicates that there are no possible values incoming
76 // along a particular edge. In most cases, this is OK, and the Phi will
77 // be eliminated later in an Ideal call. However, we can't allow this to
78 // happen if the singleton occurs on loop entry, as the elimination of
79 // the PhiNode may cause the resulting node to migrate back to a previous
80 // loop iteration.
81 if( singleton && t == Type::TOP ) {
82 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
83 // irreducible loop may not be indicated by an affirmative is_Loop());
84 // therefore, the only top we can split thru a phi is on a backedge of
85 // a loop.
86 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
87 }
89 if( singleton ) {
90 wins++;
91 x = ((PhaseGVN&)_igvn).makecon(t);
92 } else {
93 // We now call Identity to try to simplify the cloned node.
94 // Note that some Identity methods call phase->type(this).
95 // Make sure that the type array is big enough for
96 // our new node, even though we may throw the node away.
97 // (Note: This tweaking with igvn only works because x is a new node.)
98 _igvn.set_type(x, t);
99 Node *y = x->Identity(&_igvn);
100 if( y != x ) {
101 wins++;
102 x = y;
103 } else {
104 y = _igvn.hash_find(x);
105 if( y ) {
106 wins++;
107 x = y;
108 } else {
109 // Else x is a new node we are keeping
110 // We do not need register_new_node_with_optimizer
111 // because set_type has already been called.
112 _igvn._worklist.push(x);
113 }
114 }
115 }
116 if (x != the_clone && the_clone != NULL)
117 _igvn.remove_dead_node(the_clone);
118 phi->set_req( i, x );
119 }
120 // Too few wins?
121 if( wins <= policy ) {
122 _igvn.remove_dead_node(phi);
123 return NULL;
124 }
126 // Record Phi
127 register_new_node( phi, region );
129 for( uint i2 = 1; i2 < phi->req(); i2++ ) {
130 Node *x = phi->in(i2);
131 // If we commoned up the cloned 'x' with another existing Node,
132 // the existing Node picks up a new use. We need to make the
133 // existing Node occur higher up so it dominates its uses.
134 Node *old_ctrl;
135 IdealLoopTree *old_loop;
137 // The occasional new node
138 if( x->_idx >= old_unique ) { // Found a new, unplaced node?
139 old_ctrl = x->is_Con() ? C->root() : NULL;
140 old_loop = NULL; // Not in any prior loop
141 } else {
142 old_ctrl = x->is_Con() ? C->root() : get_ctrl(x);
143 old_loop = get_loop(old_ctrl); // Get prior loop
144 }
145 // New late point must dominate new use
146 Node *new_ctrl = dom_lca( old_ctrl, region->in(i2) );
147 // Set new location
148 set_ctrl(x, new_ctrl);
149 IdealLoopTree *new_loop = get_loop( new_ctrl );
150 // If changing loop bodies, see if we need to collect into new body
151 if( old_loop != new_loop ) {
152 if( old_loop && !old_loop->_child )
153 old_loop->_body.yank(x);
154 if( !new_loop->_child )
155 new_loop->_body.push(x); // Collect body info
156 }
157 }
159 return phi;
160 }
162 //------------------------------dominated_by------------------------------------
163 // Replace the dominated test with an obvious true or false. Place it on the
164 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
165 // live path up to the dominating control.
166 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff ) {
167 #ifndef PRODUCT
168 if( VerifyLoopOptimizations && PrintOpto ) tty->print_cr("dominating test");
169 #endif
172 // prevdom is the dominating projection of the dominating test.
173 assert( iff->is_If(), "" );
174 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
175 int pop = prevdom->Opcode();
176 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
177 // 'con' is set to true or false to kill the dominated test.
178 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
179 set_ctrl(con, C->root()); // Constant gets a new use
180 // Hack the dominated test
181 _igvn.hash_delete(iff);
182 iff->set_req(1, con);
183 _igvn._worklist.push(iff);
185 // If I dont have a reachable TRUE and FALSE path following the IfNode then
186 // I can assume this path reaches an infinite loop. In this case it's not
187 // important to optimize the data Nodes - either the whole compilation will
188 // be tossed or this path (and all data Nodes) will go dead.
189 if( iff->outcnt() != 2 ) return;
191 // Make control-dependent data Nodes on the live path (path that will remain
192 // once the dominated IF is removed) become control-dependent on the
193 // dominating projection.
194 Node* dp = ((IfNode*)iff)->proj_out(pop == Op_IfTrue);
195 IdealLoopTree *old_loop = get_loop(dp);
197 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
198 Node* cd = dp->fast_out(i); // Control-dependent node
199 if( cd->depends_only_on_test() ) {
200 assert( cd->in(0) == dp, "" );
201 _igvn.hash_delete( cd );
202 cd->set_req(0, prevdom);
203 set_early_ctrl( cd );
204 _igvn._worklist.push(cd);
205 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
206 if( old_loop != new_loop ) {
207 if( !old_loop->_child ) old_loop->_body.yank(cd);
208 if( !new_loop->_child ) new_loop->_body.push(cd);
209 }
210 --i;
211 --imax;
212 }
213 }
214 }
216 //------------------------------has_local_phi_input----------------------------
217 // Return TRUE if 'n' has Phi inputs from its local block and no other
218 // block-local inputs (all non-local-phi inputs come from earlier blocks)
219 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
220 Node *n_ctrl = get_ctrl(n);
221 // See if some inputs come from a Phi in this block, or from before
222 // this block.
223 uint i;
224 for( i = 1; i < n->req(); i++ ) {
225 Node *phi = n->in(i);
226 if( phi->is_Phi() && phi->in(0) == n_ctrl )
227 break;
228 }
229 if( i >= n->req() )
230 return NULL; // No Phi inputs; nowhere to clone thru
232 // Check for inputs created between 'n' and the Phi input. These
233 // must split as well; they have already been given the chance
234 // (courtesy of a post-order visit) and since they did not we must
235 // recover the 'cost' of splitting them by being very profitable
236 // when splitting 'n'. Since this is unlikely we simply give up.
237 for( i = 1; i < n->req(); i++ ) {
238 Node *m = n->in(i);
239 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
240 // We allow the special case of AddP's with no local inputs.
241 // This allows us to split-up address expressions.
242 if (m->is_AddP() &&
243 get_ctrl(m->in(2)) != n_ctrl &&
244 get_ctrl(m->in(3)) != n_ctrl) {
245 // Move the AddP up to dominating point
246 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl)));
247 continue;
248 }
249 return NULL;
250 }
251 }
253 return n_ctrl;
254 }
256 //------------------------------remix_address_expressions----------------------
257 // Rework addressing expressions to get the most loop-invariant stuff
258 // moved out. We'd like to do all associative operators, but it's especially
259 // important (common) to do address expressions.
260 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
261 if (!has_ctrl(n)) return NULL;
262 Node *n_ctrl = get_ctrl(n);
263 IdealLoopTree *n_loop = get_loop(n_ctrl);
265 // See if 'n' mixes loop-varying and loop-invariant inputs and
266 // itself is loop-varying.
268 // Only interested in binary ops (and AddP)
269 if( n->req() < 3 || n->req() > 4 ) return NULL;
271 Node *n1_ctrl = get_ctrl(n->in( 1));
272 Node *n2_ctrl = get_ctrl(n->in( 2));
273 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
274 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
275 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
276 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
278 // Does one of my inputs spin in a tighter loop than self?
279 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
280 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
281 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
282 return NULL; // Leave well enough alone
284 // Is at least one of my inputs loop-invariant?
285 if( n1_loop == n_loop &&
286 n2_loop == n_loop &&
287 n3_loop == n_loop )
288 return NULL; // No loop-invariant inputs
291 int n_op = n->Opcode();
293 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
294 if( n_op == Op_LShiftI ) {
295 // Scale is loop invariant
296 Node *scale = n->in(2);
297 Node *scale_ctrl = get_ctrl(scale);
298 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
299 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
300 return NULL;
301 const TypeInt *scale_t = scale->bottom_type()->isa_int();
302 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
303 return NULL; // Dont bother with byte/short masking
304 // Add must vary with loop (else shift would be loop-invariant)
305 Node *add = n->in(1);
306 Node *add_ctrl = get_ctrl(add);
307 IdealLoopTree *add_loop = get_loop(add_ctrl);
308 //assert( n_loop == add_loop, "" );
309 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
311 // Convert I-V into I+ (0-V); same for V-I
312 if( add->Opcode() == Op_SubI &&
313 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
314 Node *zero = _igvn.intcon(0);
315 set_ctrl(zero, C->root());
316 Node *neg = new (C, 3) SubINode( _igvn.intcon(0), add->in(2) );
317 register_new_node( neg, get_ctrl(add->in(2) ) );
318 add = new (C, 3) AddINode( add->in(1), neg );
319 register_new_node( add, add_ctrl );
320 }
321 if( add->Opcode() != Op_AddI ) return NULL;
322 // See if one add input is loop invariant
323 Node *add_var = add->in(1);
324 Node *add_var_ctrl = get_ctrl(add_var);
325 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
326 Node *add_invar = add->in(2);
327 Node *add_invar_ctrl = get_ctrl(add_invar);
328 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
329 if( add_var_loop == n_loop ) {
330 } else if( add_invar_loop == n_loop ) {
331 // Swap to find the invariant part
332 add_invar = add_var;
333 add_invar_ctrl = add_var_ctrl;
334 add_invar_loop = add_var_loop;
335 add_var = add->in(2);
336 Node *add_var_ctrl = get_ctrl(add_var);
337 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
338 } else // Else neither input is loop invariant
339 return NULL;
340 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
341 return NULL; // No invariant part of the add?
343 // Yes! Reshape address expression!
344 Node *inv_scale = new (C, 3) LShiftINode( add_invar, scale );
345 register_new_node( inv_scale, add_invar_ctrl );
346 Node *var_scale = new (C, 3) LShiftINode( add_var, scale );
347 register_new_node( var_scale, n_ctrl );
348 Node *var_add = new (C, 3) AddINode( var_scale, inv_scale );
349 register_new_node( var_add, n_ctrl );
350 _igvn.hash_delete( n );
351 _igvn.subsume_node( n, var_add );
352 return var_add;
353 }
355 // Replace (I+V) with (V+I)
356 if( n_op == Op_AddI ||
357 n_op == Op_AddL ||
358 n_op == Op_AddF ||
359 n_op == Op_AddD ||
360 n_op == Op_MulI ||
361 n_op == Op_MulL ||
362 n_op == Op_MulF ||
363 n_op == Op_MulD ) {
364 if( n2_loop == n_loop ) {
365 assert( n1_loop != n_loop, "" );
366 n->swap_edges(1, 2);
367 }
368 }
370 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
371 // but not if I2 is a constant.
372 if( n_op == Op_AddP ) {
373 if( n2_loop == n_loop && n3_loop != n_loop ) {
374 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
375 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
376 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
377 IdealLoopTree *n22loop = get_loop( n22_ctrl );
378 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
379 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
380 n23_loop == n_loop ) {
381 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
382 // Stuff new AddP in the loop preheader
383 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
384 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) );
385 register_new_node( add2, n_ctrl );
386 _igvn.hash_delete( n );
387 _igvn.subsume_node( n, add2 );
388 return add2;
389 }
390 }
391 }
393 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
394 if( n2_loop != n_loop && n3_loop == n_loop ) {
395 if( n->in(3)->Opcode() == Op_AddI ) {
396 Node *V = n->in(3)->in(1);
397 Node *I = n->in(3)->in(2);
398 if( is_member(n_loop,get_ctrl(V)) ) {
399 } else {
400 Node *tmp = V; V = I; I = tmp;
401 }
402 if( !is_member(n_loop,get_ctrl(I)) ) {
403 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I );
404 // Stuff new AddP in the loop preheader
405 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
406 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V );
407 register_new_node( add2, n_ctrl );
408 _igvn.hash_delete( n );
409 _igvn.subsume_node( n, add2 );
410 return add2;
411 }
412 }
413 }
414 }
416 return NULL;
417 }
419 //------------------------------conditional_move-------------------------------
420 // Attempt to replace a Phi with a conditional move. We have some pretty
421 // strict profitability requirements. All Phis at the merge point must
422 // be converted, so we can remove the control flow. We need to limit the
423 // number of c-moves to a small handful. All code that was in the side-arms
424 // of the CFG diamond is now speculatively executed. This code has to be
425 // "cheap enough". We are pretty much limited to CFG diamonds that merge
426 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
427 Node *PhaseIdealLoop::conditional_move( Node *region ) {
429 assert( region->is_Region(), "sanity check" );
430 if( region->req() != 3 ) return NULL;
432 // Check for CFG diamond
433 Node *lp = region->in(1);
434 Node *rp = region->in(2);
435 if( !lp || !rp ) return NULL;
436 Node *lp_c = lp->in(0);
437 if( lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If() ) return NULL;
438 IfNode *iff = lp_c->as_If();
440 // Check for highly predictable branch. No point in CMOV'ing if
441 // we are going to predict accurately all the time.
442 // %%% This hides patterns produced by utility methods like Math.min.
443 if( iff->_prob < PROB_UNLIKELY_MAG(3) ||
444 iff->_prob > PROB_LIKELY_MAG(3) )
445 return NULL;
447 // Check for ops pinned in an arm of the diamond.
448 // Can't remove the control flow in this case
449 if( lp->outcnt() > 1 ) return NULL;
450 if( rp->outcnt() > 1 ) return NULL;
452 // Check profitability
453 int cost = 0;
454 int phis = 0;
455 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
456 Node *out = region->fast_out(i);
457 if( !out->is_Phi() ) continue; // Ignore other control edges, etc
458 phis++;
459 PhiNode* phi = out->as_Phi();
460 switch (phi->type()->basic_type()) {
461 case T_LONG:
462 cost++; // Probably encodes as 2 CMOV's
463 case T_INT: // These all CMOV fine
464 case T_FLOAT:
465 case T_DOUBLE:
466 case T_ADDRESS: // (RawPtr)
467 case T_NARROWOOP:
468 cost++;
469 break;
470 case T_OBJECT: { // Base oops are OK, but not derived oops
471 const TypeOopPtr *tp = phi->type()->isa_oopptr();
472 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
473 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
474 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
475 // have a Phi for the base here that we convert to a CMOVE all is well
476 // and good. But if the base is dead, we'll not make a CMOVE. Later
477 // the allocator will have to produce a base by creating a CMOVE of the
478 // relevant bases. This puts the allocator in the business of
479 // manufacturing expensive instructions, generally a bad plan.
480 // Just Say No to Conditionally-Moved Derived Pointers.
481 if( tp && tp->offset() != 0 )
482 return NULL;
483 cost++;
484 break;
485 }
486 default:
487 return NULL; // In particular, can't do memory or I/O
488 }
489 // Add in cost any speculative ops
490 for( uint j = 1; j < region->req(); j++ ) {
491 Node *proj = region->in(j);
492 Node *inp = phi->in(j);
493 if (get_ctrl(inp) == proj) { // Found local op
494 cost++;
495 // Check for a chain of dependent ops; these will all become
496 // speculative in a CMOV.
497 for( uint k = 1; k < inp->req(); k++ )
498 if (get_ctrl(inp->in(k)) == proj)
499 return NULL; // Too much speculative goo
500 }
501 }
502 // See if the Phi is used by a Cmp. This will likely Split-If, a
503 // higher-payoff operation.
504 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
505 Node* use = phi->fast_out(k);
506 if( use->is_Cmp() )
507 return NULL;
508 }
509 }
510 if( cost >= ConditionalMoveLimit ) return NULL; // Too much goo
511 Node* bol = iff->in(1);
512 assert( bol->Opcode() == Op_Bool, "" );
513 int cmp_op = bol->in(1)->Opcode();
514 // It is expensive to generate flags from a float compare.
515 // Avoid duplicated float compare.
516 if( phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
518 // --------------
519 // Now replace all Phis with CMOV's
520 Node *cmov_ctrl = iff->in(0);
521 uint flip = (lp->Opcode() == Op_IfTrue);
522 while( 1 ) {
523 PhiNode* phi = NULL;
524 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
525 Node *out = region->fast_out(i);
526 if (out->is_Phi()) {
527 phi = out->as_Phi();
528 break;
529 }
530 }
531 if (phi == NULL) break;
532 #ifndef PRODUCT
533 if( PrintOpto && VerifyLoopOptimizations ) tty->print_cr("CMOV");
534 #endif
535 // Move speculative ops
536 for( uint j = 1; j < region->req(); j++ ) {
537 Node *proj = region->in(j);
538 Node *inp = phi->in(j);
539 if (get_ctrl(inp) == proj) { // Found local op
540 #ifndef PRODUCT
541 if( PrintOpto && VerifyLoopOptimizations ) {
542 tty->print(" speculate: ");
543 inp->dump();
544 }
545 #endif
546 set_ctrl(inp, cmov_ctrl);
547 }
548 }
549 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) );
550 register_new_node( cmov, cmov_ctrl );
551 _igvn.hash_delete(phi);
552 _igvn.subsume_node( phi, cmov );
553 #ifndef PRODUCT
554 if( VerifyLoopOptimizations ) verify();
555 #endif
556 }
558 // The useless CFG diamond will fold up later; see the optimization in
559 // RegionNode::Ideal.
560 _igvn._worklist.push(region);
562 return iff->in(1);
563 }
565 //------------------------------split_if_with_blocks_pre-----------------------
566 // Do the real work in a non-recursive function. Data nodes want to be
567 // cloned in the pre-order so they can feed each other nicely.
568 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
569 // Cloning these guys is unlikely to win
570 int n_op = n->Opcode();
571 if( n_op == Op_MergeMem ) return n;
572 if( n->is_Proj() ) return n;
573 // Do not clone-up CmpFXXX variations, as these are always
574 // followed by a CmpI
575 if( n->is_Cmp() ) return n;
576 // Attempt to use a conditional move instead of a phi/branch
577 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
578 Node *cmov = conditional_move( n );
579 if( cmov ) return cmov;
580 }
581 if( n->is_CFG() || n->is_LoadStore() )
582 return n;
583 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
584 n_op == Op_Opaque2 ) {
585 if( !C->major_progress() ) // If chance of no more loop opts...
586 _igvn._worklist.push(n); // maybe we'll remove them
587 return n;
588 }
590 if( n->is_Con() ) return n; // No cloning for Con nodes
592 Node *n_ctrl = get_ctrl(n);
593 if( !n_ctrl ) return n; // Dead node
595 // Attempt to remix address expressions for loop invariants
596 Node *m = remix_address_expressions( n );
597 if( m ) return m;
599 // Determine if the Node has inputs from some local Phi.
600 // Returns the block to clone thru.
601 Node *n_blk = has_local_phi_input( n );
602 if( !n_blk ) return n;
603 // Do not clone the trip counter through on a CountedLoop
604 // (messes up the canonical shape).
605 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
607 // Check for having no control input; not pinned. Allow
608 // dominating control.
609 if( n->in(0) ) {
610 Node *dom = idom(n_blk);
611 if( dom_lca( n->in(0), dom ) != n->in(0) )
612 return n;
613 }
614 // Policy: when is it profitable. You must get more wins than
615 // policy before it is considered profitable. Policy is usually 0,
616 // so 1 win is considered profitable. Big merges will require big
617 // cloning, so get a larger policy.
618 int policy = n_blk->req() >> 2;
620 // If the loop is a candidate for range check elimination,
621 // delay splitting through it's phi until a later loop optimization
622 if (n_blk->is_CountedLoop()) {
623 IdealLoopTree *lp = get_loop(n_blk);
624 if (lp && lp->_rce_candidate) {
625 return n;
626 }
627 }
629 // Use same limit as split_if_with_blocks_post
630 if( C->unique() > 35000 ) return n; // Method too big
632 // Split 'n' through the merge point if it is profitable
633 Node *phi = split_thru_phi( n, n_blk, policy );
634 if( !phi ) return n;
636 // Found a Phi to split thru!
637 // Replace 'n' with the new phi
638 _igvn.hash_delete(n);
639 _igvn.subsume_node( n, phi );
640 // Moved a load around the loop, 'en-registering' something.
641 if( n_blk->Opcode() == Op_Loop && n->is_Load() &&
642 !phi->in(LoopNode::LoopBackControl)->is_Load() )
643 C->set_major_progress();
645 return phi;
646 }
648 static bool merge_point_too_heavy(Compile* C, Node* region) {
649 // Bail out if the region and its phis have too many users.
650 int weight = 0;
651 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
652 weight += region->fast_out(i)->outcnt();
653 }
654 int nodes_left = MaxNodeLimit - C->unique();
655 if (weight * 8 > nodes_left) {
656 #ifndef PRODUCT
657 if (PrintOpto)
658 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
659 #endif
660 return true;
661 } else {
662 return false;
663 }
664 }
666 #ifdef _LP64
667 static bool merge_point_safe(Node* region) {
668 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
669 // having a PhiNode input. This sidesteps the dangerous case where the split
670 // ConvI2LNode may become TOP if the input Value() does not
671 // overlap the ConvI2L range, leaving a node which may not dominate its
672 // uses.
673 // A better fix for this problem can be found in the BugTraq entry, but
674 // expediency for Mantis demands this hack.
675 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
676 Node* n = region->fast_out(i);
677 if (n->is_Phi()) {
678 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
679 Node* m = n->fast_out(j);
680 if (m->Opcode() == Op_ConvI2L) {
681 return false;
682 }
683 }
684 }
685 }
686 return true;
687 }
688 #endif
691 //------------------------------place_near_use---------------------------------
692 // Place some computation next to use but not inside inner loops.
693 // For inner loop uses move it to the preheader area.
694 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const {
695 IdealLoopTree *u_loop = get_loop( useblock );
696 return (u_loop->_irreducible || u_loop->_child)
697 ? useblock
698 : u_loop->_head->in(LoopNode::EntryControl);
699 }
702 //------------------------------split_if_with_blocks_post----------------------
703 // Do the real work in a non-recursive function. CFG hackery wants to be
704 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
705 // info.
706 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) {
708 // Cloning Cmp through Phi's involves the split-if transform.
709 // FastLock is not used by an If
710 if( n->is_Cmp() && !n->is_FastLock() ) {
711 if( C->unique() > 35000 ) return; // Method too big
713 // Do not do 'split-if' if irreducible loops are present.
714 if( _has_irreducible_loops )
715 return;
717 Node *n_ctrl = get_ctrl(n);
718 // Determine if the Node has inputs from some local Phi.
719 // Returns the block to clone thru.
720 Node *n_blk = has_local_phi_input( n );
721 if( n_blk != n_ctrl ) return;
723 if( merge_point_too_heavy(C, n_ctrl) )
724 return;
726 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare?
727 Node *bol = n->unique_out();
728 assert( bol->is_Bool(), "expect a bool here" );
729 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare?
730 Node *iff = bol->unique_out();
732 // Check some safety conditions
733 if( iff->is_If() ) { // Classic split-if?
734 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if
735 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
736 if( get_ctrl(iff->in(2)) == n_ctrl ||
737 get_ctrl(iff->in(3)) == n_ctrl )
738 return; // Inputs not yet split-up
739 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) {
740 return; // Loop-invar test gates loop-varying CMOVE
741 }
742 } else {
743 return; // some other kind of node, such as an Allocate
744 }
746 // Do not do 'split-if' if some paths are dead. First do dead code
747 // elimination and then see if its still profitable.
748 for( uint i = 1; i < n_ctrl->req(); i++ )
749 if( n_ctrl->in(i) == C->top() )
750 return;
752 // When is split-if profitable? Every 'win' on means some control flow
753 // goes dead, so it's almost always a win.
754 int policy = 0;
755 // If trying to do a 'Split-If' at the loop head, it is only
756 // profitable if the cmp folds up on BOTH paths. Otherwise we
757 // risk peeling a loop forever.
759 // CNC - Disabled for now. Requires careful handling of loop
760 // body selection for the cloned code. Also, make sure we check
761 // for any input path not being in the same loop as n_ctrl. For
762 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
763 // because the alternative loop entry points won't be converted
764 // into LoopNodes.
765 IdealLoopTree *n_loop = get_loop(n_ctrl);
766 for( uint j = 1; j < n_ctrl->req(); j++ )
767 if( get_loop(n_ctrl->in(j)) != n_loop )
768 return;
770 #ifdef _LP64
771 // Check for safety of the merge point.
772 if( !merge_point_safe(n_ctrl) ) {
773 return;
774 }
775 #endif
777 // Split compare 'n' through the merge point if it is profitable
778 Node *phi = split_thru_phi( n, n_ctrl, policy );
779 if( !phi ) return;
781 // Found a Phi to split thru!
782 // Replace 'n' with the new phi
783 _igvn.hash_delete(n);
784 _igvn.subsume_node( n, phi );
786 // Now split the bool up thru the phi
787 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 );
788 _igvn.hash_delete(bol);
789 _igvn.subsume_node( bol, bolphi );
790 assert( iff->in(1) == bolphi, "" );
791 if( bolphi->Value(&_igvn)->singleton() )
792 return;
794 // Conditional-move? Must split up now
795 if( !iff->is_If() ) {
796 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 );
797 _igvn.hash_delete(iff);
798 _igvn.subsume_node( iff, cmovphi );
799 return;
800 }
802 // Now split the IF
803 do_split_if( iff );
804 return;
805 }
807 // Check for an IF ready to split; one that has its
808 // condition codes input coming from a Phi at the block start.
809 int n_op = n->Opcode();
811 // Check for an IF being dominated by another IF same test
812 if( n_op == Op_If ) {
813 Node *bol = n->in(1);
814 uint max = bol->outcnt();
815 // Check for same test used more than once?
816 if( n_op == Op_If && max > 1 && bol->is_Bool() ) {
817 // Search up IDOMs to see if this IF is dominated.
818 Node *cutoff = get_ctrl(bol);
820 // Now search up IDOMs till cutoff, looking for a dominating test
821 Node *prevdom = n;
822 Node *dom = idom(prevdom);
823 while( dom != cutoff ) {
824 if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) {
825 // Replace the dominated test with an obvious true or false.
826 // Place it on the IGVN worklist for later cleanup.
827 C->set_major_progress();
828 dominated_by( prevdom, n );
829 #ifndef PRODUCT
830 if( VerifyLoopOptimizations ) verify();
831 #endif
832 return;
833 }
834 prevdom = dom;
835 dom = idom(prevdom);
836 }
837 }
838 }
840 // See if a shared loop-varying computation has no loop-varying uses.
841 // Happens if something is only used for JVM state in uncommon trap exits,
842 // like various versions of induction variable+offset. Clone the
843 // computation per usage to allow it to sink out of the loop.
844 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
845 Node *n_ctrl = get_ctrl(n);
846 IdealLoopTree *n_loop = get_loop(n_ctrl);
847 if( n_loop != _ltree_root ) {
848 DUIterator_Fast imax, i = n->fast_outs(imax);
849 for (; i < imax; i++) {
850 Node* u = n->fast_out(i);
851 if( !has_ctrl(u) ) break; // Found control user
852 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
853 if( u_loop == n_loop ) break; // Found loop-varying use
854 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
855 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
856 }
857 bool did_break = (i < imax); // Did we break out of the previous loop?
858 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
859 Node *late_load_ctrl;
860 if (n->is_Load()) {
861 // If n is a load, get and save the result from get_late_ctrl(),
862 // to be later used in calculating the control for n's clones.
863 clear_dom_lca_tags();
864 late_load_ctrl = get_late_ctrl(n, n_ctrl);
865 }
866 // If n is a load, and the late control is the same as the current
867 // control, then the cloning of n is a pointless exercise, because
868 // GVN will ensure that we end up where we started.
869 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
870 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
871 Node *u = n->last_out(j); // Clone private computation per use
872 _igvn.hash_delete(u);
873 _igvn._worklist.push(u);
874 Node *x = n->clone(); // Clone computation
875 Node *x_ctrl = NULL;
876 if( u->is_Phi() ) {
877 // Replace all uses of normal nodes. Replace Phi uses
878 // individually, so the seperate Nodes can sink down
879 // different paths.
880 uint k = 1;
881 while( u->in(k) != n ) k++;
882 u->set_req( k, x );
883 // x goes next to Phi input path
884 x_ctrl = u->in(0)->in(k);
885 --j;
886 } else { // Normal use
887 // Replace all uses
888 for( uint k = 0; k < u->req(); k++ ) {
889 if( u->in(k) == n ) {
890 u->set_req( k, x );
891 --j;
892 }
893 }
894 x_ctrl = get_ctrl(u);
895 }
897 // Find control for 'x' next to use but not inside inner loops.
898 // For inner loop uses get the preheader area.
899 x_ctrl = place_near_use(x_ctrl);
901 if (n->is_Load()) {
902 // For loads, add a control edge to a CFG node outside of the loop
903 // to force them to not combine and return back inside the loop
904 // during GVN optimization (4641526).
905 //
906 // Because we are setting the actual control input, factor in
907 // the result from get_late_ctrl() so we respect any
908 // anti-dependences. (6233005).
909 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
911 // Don't allow the control input to be a CFG splitting node.
912 // Such nodes should only have ProjNodes as outs, e.g. IfNode
913 // should only have IfTrueNode and IfFalseNode (4985384).
914 x_ctrl = find_non_split_ctrl(x_ctrl);
915 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
917 x->set_req(0, x_ctrl);
918 }
919 register_new_node(x, x_ctrl);
921 // Some institutional knowledge is needed here: 'x' is
922 // yanked because if the optimizer runs GVN on it all the
923 // cloned x's will common up and undo this optimization and
924 // be forced back in the loop. This is annoying because it
925 // makes +VerifyOpto report false-positives on progress. I
926 // tried setting control edges on the x's to force them to
927 // not combine, but the matching gets worried when it tries
928 // to fold a StoreP and an AddP together (as part of an
929 // address expression) and the AddP and StoreP have
930 // different controls.
931 if( !x->is_Load() ) _igvn._worklist.yank(x);
932 }
933 _igvn.remove_dead_node(n);
934 }
935 }
936 }
937 }
939 // Check for Opaque2's who's loop has disappeared - who's input is in the
940 // same loop nest as their output. Remove 'em, they are no longer useful.
941 if( n_op == Op_Opaque2 &&
942 n->in(1) != NULL &&
943 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
944 _igvn.add_users_to_worklist(n);
945 _igvn.hash_delete(n);
946 _igvn.subsume_node( n, n->in(1) );
947 }
948 }
950 //------------------------------split_if_with_blocks---------------------------
951 // Check for aggressive application of 'split-if' optimization,
952 // using basic block level info.
953 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) {
954 Node *n = C->root();
955 visited.set(n->_idx); // first, mark node as visited
956 // Do pre-visit work for root
957 n = split_if_with_blocks_pre( n );
958 uint cnt = n->outcnt();
959 uint i = 0;
960 while (true) {
961 // Visit all children
962 if (i < cnt) {
963 Node* use = n->raw_out(i);
964 ++i;
965 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
966 // Now do pre-visit work for this use
967 use = split_if_with_blocks_pre( use );
968 nstack.push(n, i); // Save parent and next use's index.
969 n = use; // Process all children of current use.
970 cnt = use->outcnt();
971 i = 0;
972 }
973 }
974 else {
975 // All of n's children have been processed, complete post-processing.
976 if (cnt != 0 && !n->is_Con()) {
977 assert(has_node(n), "no dead nodes");
978 split_if_with_blocks_post( n );
979 }
980 if (nstack.is_empty()) {
981 // Finished all nodes on stack.
982 break;
983 }
984 // Get saved parent node and next use's index. Visit the rest of uses.
985 n = nstack.node();
986 cnt = n->outcnt();
987 i = nstack.index();
988 nstack.pop();
989 }
990 }
991 }
994 //=============================================================================
995 //
996 // C L O N E A L O O P B O D Y
997 //
999 //------------------------------clone_iff--------------------------------------
1000 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1001 // "Nearly" because all Nodes have been cloned from the original in the loop,
1002 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1003 // through the Phi recursively, and return a Bool.
1004 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) {
1006 // Convert this Phi into a Phi merging Bools
1007 uint i;
1008 for( i = 1; i < phi->req(); i++ ) {
1009 Node *b = phi->in(i);
1010 if( b->is_Phi() ) {
1011 _igvn.hash_delete(phi);
1012 _igvn._worklist.push(phi);
1013 phi->set_req(i, clone_iff( b->as_Phi(), loop ));
1014 } else {
1015 assert( b->is_Bool(), "" );
1016 }
1017 }
1019 Node *sample_bool = phi->in(1);
1020 Node *sample_cmp = sample_bool->in(1);
1022 // Make Phis to merge the Cmp's inputs.
1023 int size = phi->in(0)->req();
1024 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1025 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1026 for( i = 1; i < phi->req(); i++ ) {
1027 Node *n1 = phi->in(i)->in(1)->in(1);
1028 Node *n2 = phi->in(i)->in(1)->in(2);
1029 phi1->set_req( i, n1 );
1030 phi2->set_req( i, n2 );
1031 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1032 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1033 }
1034 // See if these Phis have been made before.
1035 // Register with optimizer
1036 Node *hit1 = _igvn.hash_find_insert(phi1);
1037 if( hit1 ) { // Hit, toss just made Phi
1038 _igvn.remove_dead_node(phi1); // Remove new phi
1039 assert( hit1->is_Phi(), "" );
1040 phi1 = (PhiNode*)hit1; // Use existing phi
1041 } else { // Miss
1042 _igvn.register_new_node_with_optimizer(phi1);
1043 }
1044 Node *hit2 = _igvn.hash_find_insert(phi2);
1045 if( hit2 ) { // Hit, toss just made Phi
1046 _igvn.remove_dead_node(phi2); // Remove new phi
1047 assert( hit2->is_Phi(), "" );
1048 phi2 = (PhiNode*)hit2; // Use existing phi
1049 } else { // Miss
1050 _igvn.register_new_node_with_optimizer(phi2);
1051 }
1052 // Register Phis with loop/block info
1053 set_ctrl(phi1, phi->in(0));
1054 set_ctrl(phi2, phi->in(0));
1055 // Make a new Cmp
1056 Node *cmp = sample_cmp->clone();
1057 cmp->set_req( 1, phi1 );
1058 cmp->set_req( 2, phi2 );
1059 _igvn.register_new_node_with_optimizer(cmp);
1060 set_ctrl(cmp, phi->in(0));
1062 // Make a new Bool
1063 Node *b = sample_bool->clone();
1064 b->set_req(1,cmp);
1065 _igvn.register_new_node_with_optimizer(b);
1066 set_ctrl(b, phi->in(0));
1068 assert( b->is_Bool(), "" );
1069 return (BoolNode*)b;
1070 }
1072 //------------------------------clone_bool-------------------------------------
1073 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1074 // "Nearly" because all Nodes have been cloned from the original in the loop,
1075 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1076 // through the Phi recursively, and return a Bool.
1077 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1078 uint i;
1079 // Convert this Phi into a Phi merging Bools
1080 for( i = 1; i < phi->req(); i++ ) {
1081 Node *b = phi->in(i);
1082 if( b->is_Phi() ) {
1083 _igvn.hash_delete(phi);
1084 _igvn._worklist.push(phi);
1085 phi->set_req(i, clone_bool( b->as_Phi(), loop ));
1086 } else {
1087 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1088 }
1089 }
1091 Node *sample_cmp = phi->in(1);
1093 // Make Phis to merge the Cmp's inputs.
1094 int size = phi->in(0)->req();
1095 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1096 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1097 for( uint j = 1; j < phi->req(); j++ ) {
1098 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1099 Node *n1, *n2;
1100 if( cmp_top->is_Cmp() ) {
1101 n1 = cmp_top->in(1);
1102 n2 = cmp_top->in(2);
1103 } else {
1104 n1 = n2 = cmp_top;
1105 }
1106 phi1->set_req( j, n1 );
1107 phi2->set_req( j, n2 );
1108 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1109 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1110 }
1112 // See if these Phis have been made before.
1113 // Register with optimizer
1114 Node *hit1 = _igvn.hash_find_insert(phi1);
1115 if( hit1 ) { // Hit, toss just made Phi
1116 _igvn.remove_dead_node(phi1); // Remove new phi
1117 assert( hit1->is_Phi(), "" );
1118 phi1 = (PhiNode*)hit1; // Use existing phi
1119 } else { // Miss
1120 _igvn.register_new_node_with_optimizer(phi1);
1121 }
1122 Node *hit2 = _igvn.hash_find_insert(phi2);
1123 if( hit2 ) { // Hit, toss just made Phi
1124 _igvn.remove_dead_node(phi2); // Remove new phi
1125 assert( hit2->is_Phi(), "" );
1126 phi2 = (PhiNode*)hit2; // Use existing phi
1127 } else { // Miss
1128 _igvn.register_new_node_with_optimizer(phi2);
1129 }
1130 // Register Phis with loop/block info
1131 set_ctrl(phi1, phi->in(0));
1132 set_ctrl(phi2, phi->in(0));
1133 // Make a new Cmp
1134 Node *cmp = sample_cmp->clone();
1135 cmp->set_req( 1, phi1 );
1136 cmp->set_req( 2, phi2 );
1137 _igvn.register_new_node_with_optimizer(cmp);
1138 set_ctrl(cmp, phi->in(0));
1140 assert( cmp->is_Cmp(), "" );
1141 return (CmpNode*)cmp;
1142 }
1144 //------------------------------sink_use---------------------------------------
1145 // If 'use' was in the loop-exit block, it now needs to be sunk
1146 // below the post-loop merge point.
1147 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1148 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1149 set_ctrl(use, post_loop);
1150 for (DUIterator j = use->outs(); use->has_out(j); j++)
1151 sink_use(use->out(j), post_loop);
1152 }
1153 }
1155 //------------------------------clone_loop-------------------------------------
1156 //
1157 // C L O N E A L O O P B O D Y
1158 //
1159 // This is the basic building block of the loop optimizations. It clones an
1160 // entire loop body. It makes an old_new loop body mapping; with this mapping
1161 // you can find the new-loop equivalent to an old-loop node. All new-loop
1162 // nodes are exactly equal to their old-loop counterparts, all edges are the
1163 // same. All exits from the old-loop now have a RegionNode that merges the
1164 // equivalent new-loop path. This is true even for the normal "loop-exit"
1165 // condition. All uses of loop-invariant old-loop values now come from (one
1166 // or more) Phis that merge their new-loop equivalents.
1167 //
1168 // This operation leaves the graph in an illegal state: there are two valid
1169 // control edges coming from the loop pre-header to both loop bodies. I'll
1170 // definitely have to hack the graph after running this transform.
1171 //
1172 // From this building block I will further edit edges to perform loop peeling
1173 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1174 //
1175 // Parameter side_by_size_idom:
1176 // When side_by_size_idom is NULL, the dominator tree is constructed for
1177 // the clone loop to dominate the original. Used in construction of
1178 // pre-main-post loop sequence.
1179 // When nonnull, the clone and original are side-by-side, both are
1180 // dominated by the side_by_side_idom node. Used in construction of
1181 // unswitched loops.
1182 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1183 Node* side_by_side_idom) {
1185 // Step 1: Clone the loop body. Make the old->new mapping.
1186 uint i;
1187 for( i = 0; i < loop->_body.size(); i++ ) {
1188 Node *old = loop->_body.at(i);
1189 Node *nnn = old->clone();
1190 old_new.map( old->_idx, nnn );
1191 _igvn.register_new_node_with_optimizer(nnn);
1192 }
1195 // Step 2: Fix the edges in the new body. If the old input is outside the
1196 // loop use it. If the old input is INside the loop, use the corresponding
1197 // new node instead.
1198 for( i = 0; i < loop->_body.size(); i++ ) {
1199 Node *old = loop->_body.at(i);
1200 Node *nnn = old_new[old->_idx];
1201 // Fix CFG/Loop controlling the new node
1202 if (has_ctrl(old)) {
1203 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
1204 } else {
1205 set_loop(nnn, loop->_parent);
1206 if (old->outcnt() > 0) {
1207 set_idom( nnn, old_new[idom(old)->_idx], dd );
1208 }
1209 }
1210 // Correct edges to the new node
1211 for( uint j = 0; j < nnn->req(); j++ ) {
1212 Node *n = nnn->in(j);
1213 if( n ) {
1214 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
1215 if( loop->is_member( old_in_loop ) )
1216 nnn->set_req(j, old_new[n->_idx]);
1217 }
1218 }
1219 _igvn.hash_find_insert(nnn);
1220 }
1221 Node *newhead = old_new[loop->_head->_idx];
1222 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1225 // Step 3: Now fix control uses. Loop varying control uses have already
1226 // been fixed up (as part of all input edges in Step 2). Loop invariant
1227 // control uses must be either an IfFalse or an IfTrue. Make a merge
1228 // point to merge the old and new IfFalse/IfTrue nodes; make the use
1229 // refer to this.
1230 ResourceArea *area = Thread::current()->resource_area();
1231 Node_List worklist(area);
1232 uint new_counter = C->unique();
1233 for( i = 0; i < loop->_body.size(); i++ ) {
1234 Node* old = loop->_body.at(i);
1235 if( !old->is_CFG() ) continue;
1236 Node* nnn = old_new[old->_idx];
1238 // Copy uses to a worklist, so I can munge the def-use info
1239 // with impunity.
1240 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1241 worklist.push(old->fast_out(j));
1243 while( worklist.size() ) { // Visit all uses
1244 Node *use = worklist.pop();
1245 if (!has_node(use)) continue; // Ignore dead nodes
1246 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1247 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
1248 // Both OLD and USE are CFG nodes here.
1249 assert( use->is_Proj(), "" );
1251 // Clone the loop exit control projection
1252 Node *newuse = use->clone();
1253 newuse->set_req(0,nnn);
1254 _igvn.register_new_node_with_optimizer(newuse);
1255 set_loop(newuse, use_loop);
1256 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
1258 // We need a Region to merge the exit from the peeled body and the
1259 // exit from the old loop body.
1260 RegionNode *r = new (C, 3) RegionNode(3);
1261 // Map the old use to the new merge point
1262 old_new.map( use->_idx, r );
1263 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
1264 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
1266 // The original user of 'use' uses 'r' instead.
1267 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
1268 Node* useuse = use->last_out(l);
1269 _igvn.hash_delete(useuse);
1270 _igvn._worklist.push(useuse);
1271 uint uses_found = 0;
1272 if( useuse->in(0) == use ) {
1273 useuse->set_req(0, r);
1274 uses_found++;
1275 if( useuse->is_CFG() ) {
1276 assert( dom_depth(useuse) > dd_r, "" );
1277 set_idom(useuse, r, dom_depth(useuse));
1278 }
1279 }
1280 for( uint k = 1; k < useuse->req(); k++ ) {
1281 if( useuse->in(k) == use ) {
1282 useuse->set_req(k, r);
1283 uses_found++;
1284 }
1285 }
1286 l -= uses_found; // we deleted 1 or more copies of this edge
1287 }
1289 // Now finish up 'r'
1290 r->set_req( 1, newuse );
1291 r->set_req( 2, use );
1292 _igvn.register_new_node_with_optimizer(r);
1293 set_loop(r, use_loop);
1294 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
1295 } // End of if a loop-exit test
1296 }
1297 }
1299 // Step 4: If loop-invariant use is not control, it must be dominated by a
1300 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
1301 // there if needed. Make a Phi there merging old and new used values.
1302 Node_List *split_if_set = NULL;
1303 Node_List *split_bool_set = NULL;
1304 Node_List *split_cex_set = NULL;
1305 for( i = 0; i < loop->_body.size(); i++ ) {
1306 Node* old = loop->_body.at(i);
1307 Node* nnn = old_new[old->_idx];
1308 // Copy uses to a worklist, so I can munge the def-use info
1309 // with impunity.
1310 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1311 worklist.push(old->fast_out(j));
1313 while( worklist.size() ) {
1314 Node *use = worklist.pop();
1315 if (!has_node(use)) continue; // Ignore dead nodes
1316 if (use->in(0) == C->top()) continue;
1317 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1318 // Check for data-use outside of loop - at least one of OLD or USE
1319 // must not be a CFG node.
1320 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) {
1322 // If the Data use is an IF, that means we have an IF outside of the
1323 // loop that is switching on a condition that is set inside of the
1324 // loop. Happens if people set a loop-exit flag; then test the flag
1325 // in the loop to break the loop, then test is again outside of the
1326 // loop to determine which way the loop exited.
1327 if( use->is_If() || use->is_CMove() ) {
1328 // Since this code is highly unlikely, we lazily build the worklist
1329 // of such Nodes to go split.
1330 if( !split_if_set )
1331 split_if_set = new Node_List(area);
1332 split_if_set->push(use);
1333 }
1334 if( use->is_Bool() ) {
1335 if( !split_bool_set )
1336 split_bool_set = new Node_List(area);
1337 split_bool_set->push(use);
1338 }
1339 if( use->Opcode() == Op_CreateEx ) {
1340 if( !split_cex_set )
1341 split_cex_set = new Node_List(area);
1342 split_cex_set->push(use);
1343 }
1346 // Get "block" use is in
1347 uint idx = 0;
1348 while( use->in(idx) != old ) idx++;
1349 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1350 assert( !loop->is_member( get_loop( prev ) ), "" );
1351 Node *cfg = prev->_idx >= new_counter
1352 ? prev->in(2)
1353 : idom(prev);
1354 if( use->is_Phi() ) // Phi use is in prior block
1355 cfg = prev->in(idx); // NOT in block of Phi itself
1356 if (cfg->is_top()) { // Use is dead?
1357 _igvn.hash_delete(use);
1358 _igvn._worklist.push(use);
1359 use->set_req(idx, C->top());
1360 continue;
1361 }
1363 while( !loop->is_member( get_loop( cfg ) ) ) {
1364 prev = cfg;
1365 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1366 }
1367 // If the use occurs after merging several exits from the loop, then
1368 // old value must have dominated all those exits. Since the same old
1369 // value was used on all those exits we did not need a Phi at this
1370 // merge point. NOW we do need a Phi here. Each loop exit value
1371 // is now merged with the peeled body exit; each exit gets its own
1372 // private Phi and those Phis need to be merged here.
1373 Node *phi;
1374 if( prev->is_Region() ) {
1375 if( idx == 0 ) { // Updating control edge?
1376 phi = prev; // Just use existing control
1377 } else { // Else need a new Phi
1378 phi = PhiNode::make( prev, old );
1379 // Now recursively fix up the new uses of old!
1380 for( uint i = 1; i < prev->req(); i++ ) {
1381 worklist.push(phi); // Onto worklist once for each 'old' input
1382 }
1383 }
1384 } else {
1385 // Get new RegionNode merging old and new loop exits
1386 prev = old_new[prev->_idx];
1387 assert( prev, "just made this in step 7" );
1388 if( idx == 0 ) { // Updating control edge?
1389 phi = prev; // Just use existing control
1390 } else { // Else need a new Phi
1391 // Make a new Phi merging data values properly
1392 phi = PhiNode::make( prev, old );
1393 phi->set_req( 1, nnn );
1394 }
1395 }
1396 // If inserting a new Phi, check for prior hits
1397 if( idx != 0 ) {
1398 Node *hit = _igvn.hash_find_insert(phi);
1399 if( hit == NULL ) {
1400 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1401 } else { // or
1402 // Remove the new phi from the graph and use the hit
1403 _igvn.remove_dead_node(phi);
1404 phi = hit; // Use existing phi
1405 }
1406 set_ctrl(phi, prev);
1407 }
1408 // Make 'use' use the Phi instead of the old loop body exit value
1409 _igvn.hash_delete(use);
1410 _igvn._worklist.push(use);
1411 use->set_req(idx, phi);
1412 if( use->_idx >= new_counter ) { // If updating new phis
1413 // Not needed for correctness, but prevents a weak assert
1414 // in AddPNode from tripping (when we end up with different
1415 // base & derived Phis that will become the same after
1416 // IGVN does CSE).
1417 Node *hit = _igvn.hash_find_insert(use);
1418 if( hit ) // Go ahead and re-hash for hits.
1419 _igvn.subsume_node( use, hit );
1420 }
1422 // If 'use' was in the loop-exit block, it now needs to be sunk
1423 // below the post-loop merge point.
1424 sink_use( use, prev );
1425 }
1426 }
1427 }
1429 // Check for IFs that need splitting/cloning. Happens if an IF outside of
1430 // the loop uses a condition set in the loop. The original IF probably
1431 // takes control from one or more OLD Regions (which in turn get from NEW
1432 // Regions). In any case, there will be a set of Phis for each merge point
1433 // from the IF up to where the original BOOL def exists the loop.
1434 if( split_if_set ) {
1435 while( split_if_set->size() ) {
1436 Node *iff = split_if_set->pop();
1437 if( iff->in(1)->is_Phi() ) {
1438 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop );
1439 _igvn.hash_delete(iff);
1440 _igvn._worklist.push(iff);
1441 iff->set_req(1, b);
1442 }
1443 }
1444 }
1445 if( split_bool_set ) {
1446 while( split_bool_set->size() ) {
1447 Node *b = split_bool_set->pop();
1448 Node *phi = b->in(1);
1449 assert( phi->is_Phi(), "" );
1450 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop );
1451 _igvn.hash_delete(b);
1452 _igvn._worklist.push(b);
1453 b->set_req(1, cmp);
1454 }
1455 }
1456 if( split_cex_set ) {
1457 while( split_cex_set->size() ) {
1458 Node *b = split_cex_set->pop();
1459 assert( b->in(0)->is_Region(), "" );
1460 assert( b->in(1)->is_Phi(), "" );
1461 assert( b->in(0)->in(0) == b->in(1)->in(0), "" );
1462 split_up( b, b->in(0), NULL );
1463 }
1464 }
1466 }
1469 //---------------------- stride_of_possible_iv -------------------------------------
1470 // Looks for an iff/bool/comp with one operand of the compare
1471 // being a cycle involving an add and a phi,
1472 // with an optional truncation (left-shift followed by a right-shift)
1473 // of the add. Returns zero if not an iv.
1474 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
1475 Node* trunc1 = NULL;
1476 Node* trunc2 = NULL;
1477 const TypeInt* ttype = NULL;
1478 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
1479 return 0;
1480 }
1481 BoolNode* bl = iff->in(1)->as_Bool();
1482 Node* cmp = bl->in(1);
1483 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) {
1484 return 0;
1485 }
1486 // Must have an invariant operand
1487 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
1488 return 0;
1489 }
1490 Node* add2 = NULL;
1491 Node* cmp1 = cmp->in(1);
1492 if (cmp1->is_Phi()) {
1493 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
1494 Node* phi = cmp1;
1495 for (uint i = 1; i < phi->req(); i++) {
1496 Node* in = phi->in(i);
1497 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
1498 &trunc1, &trunc2, &ttype);
1499 if (add && add->in(1) == phi) {
1500 add2 = add->in(2);
1501 break;
1502 }
1503 }
1504 } else {
1505 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
1506 Node* addtrunc = cmp1;
1507 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
1508 &trunc1, &trunc2, &ttype);
1509 if (add && add->in(1)->is_Phi()) {
1510 Node* phi = add->in(1);
1511 for (uint i = 1; i < phi->req(); i++) {
1512 if (phi->in(i) == addtrunc) {
1513 add2 = add->in(2);
1514 break;
1515 }
1516 }
1517 }
1518 }
1519 if (add2 != NULL) {
1520 const TypeInt* add2t = _igvn.type(add2)->is_int();
1521 if (add2t->is_con()) {
1522 return add2t->get_con();
1523 }
1524 }
1525 return 0;
1526 }
1529 //---------------------- stay_in_loop -------------------------------------
1530 // Return the (unique) control output node that's in the loop (if it exists.)
1531 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
1532 Node* unique = NULL;
1533 if (!n) return NULL;
1534 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1535 Node* use = n->fast_out(i);
1536 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
1537 if (unique != NULL) {
1538 return NULL;
1539 }
1540 unique = use;
1541 }
1542 }
1543 return unique;
1544 }
1546 //------------------------------ register_node -------------------------------------
1547 // Utility to register node "n" with PhaseIdealLoop
1548 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
1549 _igvn.register_new_node_with_optimizer(n);
1550 loop->_body.push(n);
1551 if (n->is_CFG()) {
1552 set_loop(n, loop);
1553 set_idom(n, pred, ddepth);
1554 } else {
1555 set_ctrl(n, pred);
1556 }
1557 }
1559 //------------------------------ proj_clone -------------------------------------
1560 // Utility to create an if-projection
1561 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
1562 ProjNode* c = p->clone()->as_Proj();
1563 c->set_req(0, iff);
1564 return c;
1565 }
1567 //------------------------------ short_circuit_if -------------------------------------
1568 // Force the iff control output to be the live_proj
1569 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
1570 int proj_con = live_proj->_con;
1571 assert(proj_con == 0 || proj_con == 1, "false or true projection");
1572 Node *con = _igvn.intcon(proj_con);
1573 set_ctrl(con, C->root());
1574 if (iff) {
1575 iff->set_req(1, con);
1576 }
1577 return con;
1578 }
1580 //------------------------------ insert_if_before_proj -------------------------------------
1581 // Insert a new if before an if projection (* - new node)
1582 //
1583 // before
1584 // if(test)
1585 // / \
1586 // v v
1587 // other-proj proj (arg)
1588 //
1589 // after
1590 // if(test)
1591 // / \
1592 // / v
1593 // | * proj-clone
1594 // v |
1595 // other-proj v
1596 // * new_if(relop(cmp[IU](left,right)))
1597 // / \
1598 // v v
1599 // * new-proj proj
1600 // (returned)
1601 //
1602 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
1603 IfNode* iff = proj->in(0)->as_If();
1604 IdealLoopTree *loop = get_loop(proj);
1605 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1606 int ddepth = dom_depth(proj);
1608 _igvn.hash_delete(iff);
1609 _igvn._worklist.push(iff);
1610 _igvn.hash_delete(proj);
1611 _igvn._worklist.push(proj);
1613 proj->set_req(0, NULL); // temporary disconnect
1614 ProjNode* proj2 = proj_clone(proj, iff);
1615 register_node(proj2, loop, iff, ddepth);
1617 Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right);
1618 register_node(cmp, loop, proj2, ddepth);
1620 BoolNode* bol = new (C,2)BoolNode(cmp, relop);
1621 register_node(bol, loop, proj2, ddepth);
1623 IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt);
1624 register_node(new_if, loop, proj2, ddepth);
1626 proj->set_req(0, new_if); // reattach
1627 set_idom(proj, new_if, ddepth);
1629 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
1630 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
1632 return new_exit;
1633 }
1635 //------------------------------ insert_region_before_proj -------------------------------------
1636 // Insert a region before an if projection (* - new node)
1637 //
1638 // before
1639 // if(test)
1640 // / |
1641 // v |
1642 // proj v
1643 // other-proj
1644 //
1645 // after
1646 // if(test)
1647 // / |
1648 // v |
1649 // * proj-clone v
1650 // | other-proj
1651 // v
1652 // * new-region
1653 // |
1654 // v
1655 // * dum_if
1656 // / \
1657 // v \
1658 // * dum-proj v
1659 // proj
1660 //
1661 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
1662 IfNode* iff = proj->in(0)->as_If();
1663 IdealLoopTree *loop = get_loop(proj);
1664 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1665 int ddepth = dom_depth(proj);
1667 _igvn.hash_delete(iff);
1668 _igvn._worklist.push(iff);
1669 _igvn.hash_delete(proj);
1670 _igvn._worklist.push(proj);
1672 proj->set_req(0, NULL); // temporary disconnect
1673 ProjNode* proj2 = proj_clone(proj, iff);
1674 register_node(proj2, loop, iff, ddepth);
1676 RegionNode* reg = new (C,2)RegionNode(2);
1677 reg->set_req(1, proj2);
1678 register_node(reg, loop, iff, ddepth);
1680 IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
1681 register_node(dum_if, loop, reg, ddepth);
1683 proj->set_req(0, dum_if); // reattach
1684 set_idom(proj, dum_if, ddepth);
1686 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
1687 register_node(dum_proj, loop, dum_if, ddepth);
1689 return reg;
1690 }
1692 //------------------------------ insert_cmpi_loop_exit -------------------------------------
1693 // Clone a signed compare loop exit from an unsigned compare and
1694 // insert it before the unsigned cmp on the stay-in-loop path.
1695 // All new nodes inserted in the dominator tree between the original
1696 // if and it's projections. The original if test is replaced with
1697 // a constant to force the stay-in-loop path.
1698 //
1699 // This is done to make sure that the original if and it's projections
1700 // still dominate the same set of control nodes, that the ctrl() relation
1701 // from data nodes to them is preserved, and that their loop nesting is
1702 // preserved.
1703 //
1704 // before
1705 // if(i <u limit) unsigned compare loop exit
1706 // / |
1707 // v v
1708 // exit-proj stay-in-loop-proj
1709 //
1710 // after
1711 // if(stay-in-loop-const) original if
1712 // / |
1713 // / v
1714 // / if(i < limit) new signed test
1715 // / / |
1716 // / / v
1717 // / / if(i <u limit) new cloned unsigned test
1718 // / / / |
1719 // v v v |
1720 // region |
1721 // | |
1722 // dum-if |
1723 // / | |
1724 // ether | |
1725 // v v
1726 // exit-proj stay-in-loop-proj
1727 //
1728 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
1729 const bool Signed = true;
1730 const bool Unsigned = false;
1732 BoolNode* bol = if_cmpu->in(1)->as_Bool();
1733 if (bol->_test._test != BoolTest::lt) return NULL;
1734 CmpNode* cmpu = bol->in(1)->as_Cmp();
1735 if (cmpu->Opcode() != Op_CmpU) return NULL;
1736 int stride = stride_of_possible_iv(if_cmpu);
1737 if (stride == 0) return NULL;
1739 ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj();
1740 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
1742 Node* limit = NULL;
1743 if (stride > 0) {
1744 limit = cmpu->in(2);
1745 } else {
1746 limit = _igvn.makecon(TypeInt::ZERO);
1747 set_ctrl(limit, C->root());
1748 }
1749 // Create a new region on the exit path
1750 RegionNode* reg = insert_region_before_proj(lp_exit);
1752 // Clone the if-cmpu-true-false using a signed compare
1753 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
1754 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
1755 reg->add_req(cmpi_exit);
1757 // Clone the if-cmpu-true-false
1758 BoolTest::mask rel_u = bol->_test._test;
1759 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
1760 reg->add_req(cmpu_exit);
1762 // Force original if to stay in loop.
1763 short_circuit_if(if_cmpu, lp_continue);
1765 return cmpi_exit->in(0)->as_If();
1766 }
1768 //------------------------------ remove_cmpi_loop_exit -------------------------------------
1769 // Remove a previously inserted signed compare loop exit.
1770 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
1771 Node* lp_proj = stay_in_loop(if_cmp, loop);
1772 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
1773 stay_in_loop(lp_proj, loop)->is_If() &&
1774 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
1775 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
1776 set_ctrl(con, C->root());
1777 if_cmp->set_req(1, con);
1778 }
1780 //------------------------------ scheduled_nodelist -------------------------------------
1781 // Create a post order schedule of nodes that are in the
1782 // "member" set. The list is returned in "sched".
1783 // The first node in "sched" is the loop head, followed by
1784 // nodes which have no inputs in the "member" set, and then
1785 // followed by the nodes that have an immediate input dependence
1786 // on a node in "sched".
1787 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
1789 assert(member.test(loop->_head->_idx), "loop head must be in member set");
1790 Arena *a = Thread::current()->resource_area();
1791 VectorSet visited(a);
1792 Node_Stack nstack(a, loop->_body.size());
1794 Node* n = loop->_head; // top of stack is cached in "n"
1795 uint idx = 0;
1796 visited.set(n->_idx);
1798 // Initially push all with no inputs from within member set
1799 for(uint i = 0; i < loop->_body.size(); i++ ) {
1800 Node *elt = loop->_body.at(i);
1801 if (member.test(elt->_idx)) {
1802 bool found = false;
1803 for (uint j = 0; j < elt->req(); j++) {
1804 Node* def = elt->in(j);
1805 if (def && member.test(def->_idx) && def != elt) {
1806 found = true;
1807 break;
1808 }
1809 }
1810 if (!found && elt != loop->_head) {
1811 nstack.push(n, idx);
1812 n = elt;
1813 assert(!visited.test(n->_idx), "not seen yet");
1814 visited.set(n->_idx);
1815 }
1816 }
1817 }
1819 // traverse out's that are in the member set
1820 while (true) {
1821 if (idx < n->outcnt()) {
1822 Node* use = n->raw_out(idx);
1823 idx++;
1824 if (!visited.test_set(use->_idx)) {
1825 if (member.test(use->_idx)) {
1826 nstack.push(n, idx);
1827 n = use;
1828 idx = 0;
1829 }
1830 }
1831 } else {
1832 // All outputs processed
1833 sched.push(n);
1834 if (nstack.is_empty()) break;
1835 n = nstack.node();
1836 idx = nstack.index();
1837 nstack.pop();
1838 }
1839 }
1840 }
1843 //------------------------------ has_use_in_set -------------------------------------
1844 // Has a use in the vector set
1845 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
1846 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1847 Node* use = n->fast_out(j);
1848 if (vset.test(use->_idx)) {
1849 return true;
1850 }
1851 }
1852 return false;
1853 }
1856 //------------------------------ has_use_internal_to_set -------------------------------------
1857 // Has use internal to the vector set (ie. not in a phi at the loop head)
1858 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
1859 Node* head = loop->_head;
1860 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1861 Node* use = n->fast_out(j);
1862 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
1863 return true;
1864 }
1865 }
1866 return false;
1867 }
1870 //------------------------------ clone_for_use_outside_loop -------------------------------------
1871 // clone "n" for uses that are outside of loop
1872 void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
1874 assert(worklist.size() == 0, "should be empty");
1875 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1876 Node* use = n->fast_out(j);
1877 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
1878 worklist.push(use);
1879 }
1880 }
1881 while( worklist.size() ) {
1882 Node *use = worklist.pop();
1883 if (!has_node(use) || use->in(0) == C->top()) continue;
1884 uint j;
1885 for (j = 0; j < use->req(); j++) {
1886 if (use->in(j) == n) break;
1887 }
1888 assert(j < use->req(), "must be there");
1890 // clone "n" and insert it between the inputs of "n" and the use outside the loop
1891 Node* n_clone = n->clone();
1892 _igvn.hash_delete(use);
1893 use->set_req(j, n_clone);
1894 _igvn._worklist.push(use);
1895 Node* use_c;
1896 if (!use->is_Phi()) {
1897 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
1898 } else {
1899 // Use in a phi is considered a use in the associated predecessor block
1900 use_c = use->in(0)->in(j);
1901 }
1902 if (use_c->is_CountedLoop()) {
1903 use_c = use_c->in(LoopNode::EntryControl);
1904 }
1905 set_ctrl(n_clone, use_c);
1906 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
1907 get_loop(use_c)->_body.push(n_clone);
1908 _igvn.register_new_node_with_optimizer(n_clone);
1909 #if !defined(PRODUCT)
1910 if (TracePartialPeeling) {
1911 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
1912 }
1913 #endif
1914 }
1915 }
1918 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
1919 // clone "n" for special uses that are in the not_peeled region.
1920 // If these def-uses occur in separate blocks, the code generator
1921 // marks the method as not compilable. For example, if a "BoolNode"
1922 // is in a different basic block than the "IfNode" that uses it, then
1923 // the compilation is aborted in the code generator.
1924 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1925 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
1926 if (n->is_Phi() || n->is_Load()) {
1927 return;
1928 }
1929 assert(worklist.size() == 0, "should be empty");
1930 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1931 Node* use = n->fast_out(j);
1932 if ( not_peel.test(use->_idx) &&
1933 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
1934 use->in(1) == n) {
1935 worklist.push(use);
1936 }
1937 }
1938 if (worklist.size() > 0) {
1939 // clone "n" and insert it between inputs of "n" and the use
1940 Node* n_clone = n->clone();
1941 loop->_body.push(n_clone);
1942 _igvn.register_new_node_with_optimizer(n_clone);
1943 set_ctrl(n_clone, get_ctrl(n));
1944 sink_list.push(n_clone);
1945 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
1946 #if !defined(PRODUCT)
1947 if (TracePartialPeeling) {
1948 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
1949 }
1950 #endif
1951 while( worklist.size() ) {
1952 Node *use = worklist.pop();
1953 _igvn.hash_delete(use);
1954 _igvn._worklist.push(use);
1955 for (uint j = 1; j < use->req(); j++) {
1956 if (use->in(j) == n) {
1957 use->set_req(j, n_clone);
1958 }
1959 }
1960 }
1961 }
1962 }
1965 //------------------------------ insert_phi_for_loop -------------------------------------
1966 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1967 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
1968 Node *phi = PhiNode::make(lp, back_edge_val);
1969 phi->set_req(LoopNode::EntryControl, lp_entry_val);
1970 // Use existing phi if it already exists
1971 Node *hit = _igvn.hash_find_insert(phi);
1972 if( hit == NULL ) {
1973 _igvn.register_new_node_with_optimizer(phi);
1974 set_ctrl(phi, lp);
1975 } else {
1976 // Remove the new phi from the graph and use the hit
1977 _igvn.remove_dead_node(phi);
1978 phi = hit;
1979 }
1980 _igvn.hash_delete(use);
1981 _igvn._worklist.push(use);
1982 use->set_req(idx, phi);
1983 }
1985 #ifdef ASSERT
1986 //------------------------------ is_valid_loop_partition -------------------------------------
1987 // Validate the loop partition sets: peel and not_peel
1988 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
1989 VectorSet& not_peel ) {
1990 uint i;
1991 // Check that peel_list entries are in the peel set
1992 for (i = 0; i < peel_list.size(); i++) {
1993 if (!peel.test(peel_list.at(i)->_idx)) {
1994 return false;
1995 }
1996 }
1997 // Check at loop members are in one of peel set or not_peel set
1998 for (i = 0; i < loop->_body.size(); i++ ) {
1999 Node *def = loop->_body.at(i);
2000 uint di = def->_idx;
2001 // Check that peel set elements are in peel_list
2002 if (peel.test(di)) {
2003 if (not_peel.test(di)) {
2004 return false;
2005 }
2006 // Must be in peel_list also
2007 bool found = false;
2008 for (uint j = 0; j < peel_list.size(); j++) {
2009 if (peel_list.at(j)->_idx == di) {
2010 found = true;
2011 break;
2012 }
2013 }
2014 if (!found) {
2015 return false;
2016 }
2017 } else if (not_peel.test(di)) {
2018 if (peel.test(di)) {
2019 return false;
2020 }
2021 } else {
2022 return false;
2023 }
2024 }
2025 return true;
2026 }
2028 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2029 // Ensure a use outside of loop is of the right form
2030 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2031 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2032 return (use->is_Phi() &&
2033 use_c->is_Region() && use_c->req() == 3 &&
2034 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2035 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2036 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2037 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2038 }
2040 //------------------------------ is_valid_clone_loop_form -------------------------------------
2041 // Ensure that all uses outside of loop are of the right form
2042 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2043 uint orig_exit_idx, uint clone_exit_idx) {
2044 uint len = peel_list.size();
2045 for (uint i = 0; i < len; i++) {
2046 Node *def = peel_list.at(i);
2048 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2049 Node *use = def->fast_out(j);
2050 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2051 if (!loop->is_member(get_loop(use_c))) {
2052 // use is not in the loop, check for correct structure
2053 if (use->in(0) == def) {
2054 // Okay
2055 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2056 return false;
2057 }
2058 }
2059 }
2060 }
2061 return true;
2062 }
2063 #endif
2065 //------------------------------ partial_peel -------------------------------------
2066 // Partially peel (aka loop rotation) the top portion of a loop (called
2067 // the peel section below) by cloning it and placing one copy just before
2068 // the new loop head and the other copy at the bottom of the new loop.
2069 //
2070 // before after where it came from
2071 //
2072 // stmt1 stmt1
2073 // loop: stmt2 clone
2074 // stmt2 if condA goto exitA clone
2075 // if condA goto exitA new_loop: new
2076 // stmt3 stmt3 clone
2077 // if !condB goto loop if condB goto exitB clone
2078 // exitB: stmt2 orig
2079 // stmt4 if !condA goto new_loop orig
2080 // exitA: goto exitA
2081 // exitB:
2082 // stmt4
2083 // exitA:
2084 //
2085 // Step 1: find the cut point: an exit test on probable
2086 // induction variable.
2087 // Step 2: schedule (with cloning) operations in the peel
2088 // section that can be executed after the cut into
2089 // the section that is not peeled. This may need
2090 // to clone operations into exit blocks. For
2091 // instance, a reference to A[i] in the not-peel
2092 // section and a reference to B[i] in an exit block
2093 // may cause a left-shift of i by 2 to be placed
2094 // in the peel block. This step will clone the left
2095 // shift into the exit block and sink the left shift
2096 // from the peel to the not-peel section.
2097 // Step 3: clone the loop, retarget the control, and insert
2098 // phis for values that are live across the new loop
2099 // head. This is very dependent on the graph structure
2100 // from clone_loop. It creates region nodes for
2101 // exit control and associated phi nodes for values
2102 // flow out of the loop through that exit. The region
2103 // node is dominated by the clone's control projection.
2104 // So the clone's peel section is placed before the
2105 // new loop head, and the clone's not-peel section is
2106 // forms the top part of the new loop. The original
2107 // peel section forms the tail of the new loop.
2108 // Step 4: update the dominator tree and recompute the
2109 // dominator depth.
2110 //
2111 // orig
2112 //
2113 // stmt1
2114 // |
2115 // v
2116 // loop<----+
2117 // | |
2118 // stmt2 |
2119 // | |
2120 // v |
2121 // ifA |
2122 // / | |
2123 // v v |
2124 // false true ^ <-- last_peel
2125 // / | |
2126 // / ===|==cut |
2127 // / stmt3 | <-- first_not_peel
2128 // / | |
2129 // | v |
2130 // v ifB |
2131 // exitA: / \ |
2132 // / \ |
2133 // v v |
2134 // false true |
2135 // / \ |
2136 // / ----+
2137 // |
2138 // v
2139 // exitB:
2140 // stmt4
2141 //
2142 //
2143 // after clone loop
2144 //
2145 // stmt1
2146 // / \
2147 // clone / \ orig
2148 // / \
2149 // / \
2150 // v v
2151 // +---->loop loop<----+
2152 // | | | |
2153 // | stmt2 stmt2 |
2154 // | | | |
2155 // | v v |
2156 // | ifA ifA |
2157 // | | \ / | |
2158 // | v v v v |
2159 // ^ true false false true ^ <-- last_peel
2160 // | | ^ \ / | |
2161 // | cut==|== \ \ / ===|==cut |
2162 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2163 // | | dom | | | |
2164 // | v \ 1v v2 v |
2165 // | ifB regionA ifB |
2166 // | / \ | / \ |
2167 // | / \ v / \ |
2168 // | v v exitA: v v |
2169 // | true false false true |
2170 // | / ^ \ / \ |
2171 // +---- \ \ / ----+
2172 // dom \ /
2173 // \ 1v v2
2174 // regionB
2175 // |
2176 // v
2177 // exitB:
2178 // stmt4
2179 //
2180 //
2181 // after partial peel
2182 //
2183 // stmt1
2184 // /
2185 // clone / orig
2186 // / TOP
2187 // / \
2188 // v v
2189 // TOP->region region----+
2190 // | | |
2191 // stmt2 stmt2 |
2192 // | | |
2193 // v v |
2194 // ifA ifA |
2195 // | \ / | |
2196 // v v v v |
2197 // true false false true | <-- last_peel
2198 // | ^ \ / +------|---+
2199 // +->newloop \ \ / === ==cut | |
2200 // | stmt3 \ \ / TOP | |
2201 // | | dom | | stmt3 | | <-- first_not_peel
2202 // | v \ 1v v2 v | |
2203 // | ifB regionA ifB ^ v
2204 // | / \ | / \ | |
2205 // | / \ v / \ | |
2206 // | v v exitA: v v | |
2207 // | true false false true | |
2208 // | / ^ \ / \ | |
2209 // | | \ \ / v | |
2210 // | | dom \ / TOP | |
2211 // | | \ 1v v2 | |
2212 // ^ v regionB | |
2213 // | | | | |
2214 // | | v ^ v
2215 // | | exitB: | |
2216 // | | stmt4 | |
2217 // | +------------>-----------------+ |
2218 // | |
2219 // +-----------------<---------------------+
2220 //
2221 //
2222 // final graph
2223 //
2224 // stmt1
2225 // |
2226 // v
2227 // ........> ifA clone
2228 // : / |
2229 // dom / |
2230 // : v v
2231 // : false true
2232 // : | |
2233 // : | stmt2 clone
2234 // : | |
2235 // : | v
2236 // : | newloop<-----+
2237 // : | | |
2238 // : | stmt3 clone |
2239 // : | | |
2240 // : | v |
2241 // : | ifB |
2242 // : | / \ |
2243 // : | v v |
2244 // : | false true |
2245 // : | | | |
2246 // : | v stmt2 |
2247 // : | exitB: | |
2248 // : | stmt4 v |
2249 // : | ifA orig |
2250 // : | / \ |
2251 // : | / \ |
2252 // : | v v |
2253 // : | false true |
2254 // : | / \ |
2255 // : v v -----+
2256 // RegionA
2257 // |
2258 // v
2259 // exitA
2260 //
2261 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
2263 if (!loop->_head->is_Loop()) {
2264 return false; }
2266 LoopNode *head = loop->_head->as_Loop();
2268 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
2269 return false;
2270 }
2272 // Check for complex exit control
2273 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
2274 Node *n = loop->_body.at(ii);
2275 int opc = n->Opcode();
2276 if (n->is_Call() ||
2277 opc == Op_Catch ||
2278 opc == Op_CatchProj ||
2279 opc == Op_Jump ||
2280 opc == Op_JumpProj) {
2281 #if !defined(PRODUCT)
2282 if (TracePartialPeeling) {
2283 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
2284 }
2285 #endif
2286 return false;
2287 }
2288 }
2290 int dd = dom_depth(head);
2292 // Step 1: find cut point
2294 // Walk up dominators to loop head looking for first loop exit
2295 // which is executed on every path thru loop.
2296 IfNode *peel_if = NULL;
2297 IfNode *peel_if_cmpu = NULL;
2299 Node *iff = loop->tail();
2300 while( iff != head ) {
2301 if( iff->is_If() ) {
2302 Node *ctrl = get_ctrl(iff->in(1));
2303 if (ctrl->is_top()) return false; // Dead test on live IF.
2304 // If loop-varying exit-test, check for induction variable
2305 if( loop->is_member(get_loop(ctrl)) &&
2306 loop->is_loop_exit(iff) &&
2307 is_possible_iv_test(iff)) {
2308 Node* cmp = iff->in(1)->in(1);
2309 if (cmp->Opcode() == Op_CmpI) {
2310 peel_if = iff->as_If();
2311 } else {
2312 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
2313 peel_if_cmpu = iff->as_If();
2314 }
2315 }
2316 }
2317 iff = idom(iff);
2318 }
2319 // Prefer signed compare over unsigned compare.
2320 IfNode* new_peel_if = NULL;
2321 if (peel_if == NULL) {
2322 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
2323 return false; // No peel point found
2324 }
2325 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
2326 if (new_peel_if == NULL) {
2327 return false; // No peel point found
2328 }
2329 peel_if = new_peel_if;
2330 }
2331 Node* last_peel = stay_in_loop(peel_if, loop);
2332 Node* first_not_peeled = stay_in_loop(last_peel, loop);
2333 if (first_not_peeled == NULL || first_not_peeled == head) {
2334 return false;
2335 }
2337 #if !defined(PRODUCT)
2338 if (TracePartialPeeling) {
2339 tty->print_cr("before partial peel one iteration");
2340 Node_List wl;
2341 Node* t = head->in(2);
2342 while (true) {
2343 wl.push(t);
2344 if (t == head) break;
2345 t = idom(t);
2346 }
2347 while (wl.size() > 0) {
2348 Node* tt = wl.pop();
2349 tt->dump();
2350 if (tt == last_peel) tty->print_cr("-- cut --");
2351 }
2352 }
2353 #endif
2354 ResourceArea *area = Thread::current()->resource_area();
2355 VectorSet peel(area);
2356 VectorSet not_peel(area);
2357 Node_List peel_list(area);
2358 Node_List worklist(area);
2359 Node_List sink_list(area);
2361 // Set of cfg nodes to peel are those that are executable from
2362 // the head through last_peel.
2363 assert(worklist.size() == 0, "should be empty");
2364 worklist.push(head);
2365 peel.set(head->_idx);
2366 while (worklist.size() > 0) {
2367 Node *n = worklist.pop();
2368 if (n != last_peel) {
2369 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2370 Node* use = n->fast_out(j);
2371 if (use->is_CFG() &&
2372 loop->is_member(get_loop(use)) &&
2373 !peel.test_set(use->_idx)) {
2374 worklist.push(use);
2375 }
2376 }
2377 }
2378 }
2380 // Set of non-cfg nodes to peel are those that are control
2381 // dependent on the cfg nodes.
2382 uint i;
2383 for(i = 0; i < loop->_body.size(); i++ ) {
2384 Node *n = loop->_body.at(i);
2385 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
2386 if (peel.test(n_c->_idx)) {
2387 peel.set(n->_idx);
2388 } else {
2389 not_peel.set(n->_idx);
2390 }
2391 }
2393 // Step 2: move operations from the peeled section down into the
2394 // not-peeled section
2396 // Get a post order schedule of nodes in the peel region
2397 // Result in right-most operand.
2398 scheduled_nodelist(loop, peel, peel_list );
2400 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2402 // For future check for too many new phis
2403 uint old_phi_cnt = 0;
2404 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2405 Node* use = head->fast_out(j);
2406 if (use->is_Phi()) old_phi_cnt++;
2407 }
2409 #if !defined(PRODUCT)
2410 if (TracePartialPeeling) {
2411 tty->print_cr("\npeeled list");
2412 }
2413 #endif
2415 // Evacuate nodes in peel region into the not_peeled region if possible
2416 uint new_phi_cnt = 0;
2417 for (i = 0; i < peel_list.size();) {
2418 Node* n = peel_list.at(i);
2419 #if !defined(PRODUCT)
2420 if (TracePartialPeeling) n->dump();
2421 #endif
2422 bool incr = true;
2423 if ( !n->is_CFG() ) {
2425 if ( has_use_in_set(n, not_peel) ) {
2427 // If not used internal to the peeled region,
2428 // move "n" from peeled to not_peeled region.
2430 if ( !has_use_internal_to_set(n, peel, loop) ) {
2432 // if not pinned and not a load (which maybe anti-dependent on a store)
2433 // and not a CMove (Matcher expects only bool->cmove).
2434 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
2435 clone_for_use_outside_loop( loop, n, worklist );
2437 sink_list.push(n);
2438 peel >>= n->_idx; // delete n from peel set.
2439 not_peel <<= n->_idx; // add n to not_peel set.
2440 peel_list.remove(i);
2441 incr = false;
2442 #if !defined(PRODUCT)
2443 if (TracePartialPeeling) {
2444 tty->print_cr("sink to not_peeled region: %d newbb: %d",
2445 n->_idx, get_ctrl(n)->_idx);
2446 }
2447 #endif
2448 }
2449 } else {
2450 // Otherwise check for special def-use cases that span
2451 // the peel/not_peel boundary such as bool->if
2452 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
2453 new_phi_cnt++;
2454 }
2455 }
2456 }
2457 if (incr) i++;
2458 }
2460 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
2461 #if !defined(PRODUCT)
2462 if (TracePartialPeeling) {
2463 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
2464 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
2465 }
2466 #endif
2467 if (new_peel_if != NULL) {
2468 remove_cmpi_loop_exit(new_peel_if, loop);
2469 }
2470 // Inhibit more partial peeling on this loop
2471 assert(!head->is_partial_peel_loop(), "not partial peeled");
2472 head->mark_partial_peel_failed();
2473 return false;
2474 }
2476 // Step 3: clone loop, retarget control, and insert new phis
2478 // Create new loop head for new phis and to hang
2479 // the nodes being moved (sinked) from the peel region.
2480 LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel);
2481 _igvn.register_new_node_with_optimizer(new_head);
2482 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
2483 first_not_peeled->set_req(0, new_head);
2484 set_loop(new_head, loop);
2485 loop->_body.push(new_head);
2486 not_peel.set(new_head->_idx);
2487 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
2488 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
2490 while (sink_list.size() > 0) {
2491 Node* n = sink_list.pop();
2492 set_ctrl(n, new_head);
2493 }
2495 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2497 clone_loop( loop, old_new, dd );
2499 const uint clone_exit_idx = 1;
2500 const uint orig_exit_idx = 2;
2501 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
2503 Node* head_clone = old_new[head->_idx];
2504 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
2505 Node* orig_tail_clone = head_clone->in(2);
2507 // Add phi if "def" node is in peel set and "use" is not
2509 for(i = 0; i < peel_list.size(); i++ ) {
2510 Node *def = peel_list.at(i);
2511 if (!def->is_CFG()) {
2512 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2513 Node *use = def->fast_out(j);
2514 if (has_node(use) && use->in(0) != C->top() &&
2515 (!peel.test(use->_idx) ||
2516 (use->is_Phi() && use->in(0) == head)) ) {
2517 worklist.push(use);
2518 }
2519 }
2520 while( worklist.size() ) {
2521 Node *use = worklist.pop();
2522 for (uint j = 1; j < use->req(); j++) {
2523 Node* n = use->in(j);
2524 if (n == def) {
2526 // "def" is in peel set, "use" is not in peel set
2527 // or "use" is in the entry boundary (a phi) of the peel set
2529 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
2531 if ( loop->is_member(get_loop( use_c )) ) {
2532 // use is in loop
2533 if (old_new[use->_idx] != NULL) { // null for dead code
2534 Node* use_clone = old_new[use->_idx];
2535 _igvn.hash_delete(use);
2536 use->set_req(j, C->top());
2537 _igvn._worklist.push(use);
2538 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
2539 }
2540 } else {
2541 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
2542 // use is not in the loop, check if the live range includes the cut
2543 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
2544 if (not_peel.test(lp_if->_idx)) {
2545 assert(j == orig_exit_idx, "use from original loop");
2546 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
2547 }
2548 }
2549 }
2550 }
2551 }
2552 }
2553 }
2555 // Step 3b: retarget control
2557 // Redirect control to the new loop head if a cloned node in
2558 // the not_peeled region has control that points into the peeled region.
2559 // This necessary because the cloned peeled region will be outside
2560 // the loop.
2561 // from to
2562 // cloned-peeled <---+
2563 // new_head_clone: | <--+
2564 // cloned-not_peeled in(0) in(0)
2565 // orig-peeled
2567 for(i = 0; i < loop->_body.size(); i++ ) {
2568 Node *n = loop->_body.at(i);
2569 if (!n->is_CFG() && n->in(0) != NULL &&
2570 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
2571 Node* n_clone = old_new[n->_idx];
2572 _igvn.hash_delete(n_clone);
2573 n_clone->set_req(0, new_head_clone);
2574 _igvn._worklist.push(n_clone);
2575 }
2576 }
2578 // Backedge of the surviving new_head (the clone) is original last_peel
2579 _igvn.hash_delete(new_head_clone);
2580 new_head_clone->set_req(LoopNode::LoopBackControl, last_peel);
2581 _igvn._worklist.push(new_head_clone);
2583 // Cut first node in original not_peel set
2584 _igvn.hash_delete(new_head);
2585 new_head->set_req(LoopNode::EntryControl, C->top());
2586 new_head->set_req(LoopNode::LoopBackControl, C->top());
2587 _igvn._worklist.push(new_head);
2589 // Copy head_clone back-branch info to original head
2590 // and remove original head's loop entry and
2591 // clone head's back-branch
2592 _igvn.hash_delete(head);
2593 _igvn.hash_delete(head_clone);
2594 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
2595 head->set_req(LoopNode::LoopBackControl, C->top());
2596 head_clone->set_req(LoopNode::LoopBackControl, C->top());
2597 _igvn._worklist.push(head);
2598 _igvn._worklist.push(head_clone);
2600 // Similarly modify the phis
2601 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
2602 Node* use = head->fast_out(k);
2603 if (use->is_Phi() && use->outcnt() > 0) {
2604 Node* use_clone = old_new[use->_idx];
2605 _igvn.hash_delete(use);
2606 _igvn.hash_delete(use_clone);
2607 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
2608 use->set_req(LoopNode::LoopBackControl, C->top());
2609 use_clone->set_req(LoopNode::LoopBackControl, C->top());
2610 _igvn._worklist.push(use);
2611 _igvn._worklist.push(use_clone);
2612 }
2613 }
2615 // Step 4: update dominator tree and dominator depth
2617 set_idom(head, orig_tail_clone, dd);
2618 recompute_dom_depth();
2620 // Inhibit more partial peeling on this loop
2621 new_head_clone->set_partial_peel_loop();
2622 C->set_major_progress();
2624 #if !defined(PRODUCT)
2625 if (TracePartialPeeling) {
2626 tty->print_cr("\nafter partial peel one iteration");
2627 Node_List wl(area);
2628 Node* t = last_peel;
2629 while (true) {
2630 wl.push(t);
2631 if (t == head_clone) break;
2632 t = idom(t);
2633 }
2634 while (wl.size() > 0) {
2635 Node* tt = wl.pop();
2636 if (tt == head) tty->print_cr("orig head");
2637 else if (tt == new_head_clone) tty->print_cr("new head");
2638 else if (tt == head_clone) tty->print_cr("clone head");
2639 tt->dump();
2640 }
2641 }
2642 #endif
2643 return true;
2644 }
2646 //------------------------------reorg_offsets----------------------------------
2647 // Reorganize offset computations to lower register pressure. Mostly
2648 // prevent loop-fallout uses of the pre-incremented trip counter (which are
2649 // then alive with the post-incremented trip counter forcing an extra
2650 // register move)
2651 void PhaseIdealLoop::reorg_offsets( IdealLoopTree *loop ) {
2653 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2654 CountedLoopEndNode *cle = cl->loopexit();
2655 if( !cle ) return; // The occasional dead loop
2656 // Find loop exit control
2657 Node *exit = cle->proj_out(false);
2658 assert( exit->Opcode() == Op_IfFalse, "" );
2660 // Check for the special case of folks using the pre-incremented
2661 // trip-counter on the fall-out path (forces the pre-incremented
2662 // and post-incremented trip counter to be live at the same time).
2663 // Fix this by adjusting to use the post-increment trip counter.
2664 Node *phi = cl->phi();
2665 if( !phi ) return; // Dead infinite loop
2666 bool progress = true;
2667 while (progress) {
2668 progress = false;
2669 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
2670 Node* use = phi->fast_out(i); // User of trip-counter
2671 if (!has_ctrl(use)) continue;
2672 Node *u_ctrl = get_ctrl(use);
2673 if( use->is_Phi() ) {
2674 u_ctrl = NULL;
2675 for( uint j = 1; j < use->req(); j++ )
2676 if( use->in(j) == phi )
2677 u_ctrl = dom_lca( u_ctrl, use->in(0)->in(j) );
2678 }
2679 IdealLoopTree *u_loop = get_loop(u_ctrl);
2680 // Look for loop-invariant use
2681 if( u_loop == loop ) continue;
2682 if( loop->is_member( u_loop ) ) continue;
2683 // Check that use is live out the bottom. Assuming the trip-counter
2684 // update is right at the bottom, uses of of the loop middle are ok.
2685 if( dom_lca( exit, u_ctrl ) != exit ) continue;
2686 // protect against stride not being a constant
2687 if( !cle->stride_is_con() ) continue;
2688 // Hit! Refactor use to use the post-incremented tripcounter.
2689 // Compute a post-increment tripcounter.
2690 Node *opaq = new (C, 2) Opaque2Node( C, cle->incr() );
2691 register_new_node( opaq, u_ctrl );
2692 Node *neg_stride = _igvn.intcon(-cle->stride_con());
2693 set_ctrl(neg_stride, C->root());
2694 Node *post = new (C, 3) AddINode( opaq, neg_stride);
2695 register_new_node( post, u_ctrl );
2696 _igvn.hash_delete(use);
2697 _igvn._worklist.push(use);
2698 for( uint j = 1; j < use->req(); j++ )
2699 if( use->in(j) == phi )
2700 use->set_req(j, post);
2701 // Since DU info changed, rerun loop
2702 progress = true;
2703 break;
2704 }
2705 }
2707 }