Thu, 13 Mar 2008 16:31:32 -0700
6673473: (Escape Analysis) Add the instance's field information to PhiNode
Summary: Avoid an infinite generation of instance's field values Phi nodes.
Reviewed-by: never
1 /*
2 * Copyright 1999-2006 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 int wins = 0;
33 assert( !n->is_CFG(), "" );
34 assert( region->is_Region(), "" );
36 const Type* type = n->bottom_type();
37 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr();
38 Node *phi;
39 if( t_oop != NULL && t_oop->is_instance_field() ) {
40 int iid = t_oop->instance_id();
41 int index = C->get_alias_index(t_oop);
42 int offset = t_oop->offset();
43 phi = new (C,region->req()) PhiNode(region, type, NULL, iid, index, offset);
44 } else {
45 phi = new (C,region->req()) PhiNode(region, type);
46 }
47 uint old_unique = C->unique();
48 for( uint i = 1; i < region->req(); i++ ) {
49 Node *x;
50 Node* the_clone = NULL;
51 if( region->in(i) == C->top() ) {
52 x = C->top(); // Dead path? Use a dead data op
53 } else {
54 x = n->clone(); // Else clone up the data op
55 the_clone = x; // Remember for possible deletion.
56 // Alter data node to use pre-phi inputs
57 if( n->in(0) == region )
58 x->set_req( 0, region->in(i) );
59 for( uint j = 1; j < n->req(); j++ ) {
60 Node *in = n->in(j);
61 if( in->is_Phi() && in->in(0) == region )
62 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
63 }
64 }
65 // Check for a 'win' on some paths
66 const Type *t = x->Value(&_igvn);
68 bool singleton = t->singleton();
70 // A TOP singleton indicates that there are no possible values incoming
71 // along a particular edge. In most cases, this is OK, and the Phi will
72 // be eliminated later in an Ideal call. However, we can't allow this to
73 // happen if the singleton occurs on loop entry, as the elimination of
74 // the PhiNode may cause the resulting node to migrate back to a previous
75 // loop iteration.
76 if( singleton && t == Type::TOP ) {
77 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
78 // irreducible loop may not be indicated by an affirmative is_Loop());
79 // therefore, the only top we can split thru a phi is on a backedge of
80 // a loop.
81 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
82 }
84 if( singleton ) {
85 wins++;
86 x = ((PhaseGVN&)_igvn).makecon(t);
87 } else {
88 // We now call Identity to try to simplify the cloned node.
89 // Note that some Identity methods call phase->type(this).
90 // Make sure that the type array is big enough for
91 // our new node, even though we may throw the node away.
92 // (Note: This tweaking with igvn only works because x is a new node.)
93 _igvn.set_type(x, t);
94 Node *y = x->Identity(&_igvn);
95 if( y != x ) {
96 wins++;
97 x = y;
98 } else {
99 y = _igvn.hash_find(x);
100 if( y ) {
101 wins++;
102 x = y;
103 } else {
104 // Else x is a new node we are keeping
105 // We do not need register_new_node_with_optimizer
106 // because set_type has already been called.
107 _igvn._worklist.push(x);
108 }
109 }
110 }
111 if (x != the_clone && the_clone != NULL)
112 _igvn.remove_dead_node(the_clone);
113 phi->set_req( i, x );
114 }
115 // Too few wins?
116 if( wins <= policy ) {
117 _igvn.remove_dead_node(phi);
118 return NULL;
119 }
121 // Record Phi
122 register_new_node( phi, region );
124 for( uint i2 = 1; i2 < phi->req(); i2++ ) {
125 Node *x = phi->in(i2);
126 // If we commoned up the cloned 'x' with another existing Node,
127 // the existing Node picks up a new use. We need to make the
128 // existing Node occur higher up so it dominates its uses.
129 Node *old_ctrl;
130 IdealLoopTree *old_loop;
132 // The occasional new node
133 if( x->_idx >= old_unique ) { // Found a new, unplaced node?
134 old_ctrl = x->is_Con() ? C->root() : NULL;
135 old_loop = NULL; // Not in any prior loop
136 } else {
137 old_ctrl = x->is_Con() ? C->root() : get_ctrl(x);
138 old_loop = get_loop(old_ctrl); // Get prior loop
139 }
140 // New late point must dominate new use
141 Node *new_ctrl = dom_lca( old_ctrl, region->in(i2) );
142 // Set new location
143 set_ctrl(x, new_ctrl);
144 IdealLoopTree *new_loop = get_loop( new_ctrl );
145 // If changing loop bodies, see if we need to collect into new body
146 if( old_loop != new_loop ) {
147 if( old_loop && !old_loop->_child )
148 old_loop->_body.yank(x);
149 if( !new_loop->_child )
150 new_loop->_body.push(x); // Collect body info
151 }
152 }
154 return phi;
155 }
157 //------------------------------dominated_by------------------------------------
158 // Replace the dominated test with an obvious true or false. Place it on the
159 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
160 // live path up to the dominating control.
161 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff ) {
162 #ifndef PRODUCT
163 if( VerifyLoopOptimizations && PrintOpto ) tty->print_cr("dominating test");
164 #endif
167 // prevdom is the dominating projection of the dominating test.
168 assert( iff->is_If(), "" );
169 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
170 int pop = prevdom->Opcode();
171 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
172 // 'con' is set to true or false to kill the dominated test.
173 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
174 set_ctrl(con, C->root()); // Constant gets a new use
175 // Hack the dominated test
176 _igvn.hash_delete(iff);
177 iff->set_req(1, con);
178 _igvn._worklist.push(iff);
180 // If I dont have a reachable TRUE and FALSE path following the IfNode then
181 // I can assume this path reaches an infinite loop. In this case it's not
182 // important to optimize the data Nodes - either the whole compilation will
183 // be tossed or this path (and all data Nodes) will go dead.
184 if( iff->outcnt() != 2 ) return;
186 // Make control-dependent data Nodes on the live path (path that will remain
187 // once the dominated IF is removed) become control-dependent on the
188 // dominating projection.
189 Node* dp = ((IfNode*)iff)->proj_out(pop == Op_IfTrue);
190 IdealLoopTree *old_loop = get_loop(dp);
192 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
193 Node* cd = dp->fast_out(i); // Control-dependent node
194 if( cd->depends_only_on_test() ) {
195 assert( cd->in(0) == dp, "" );
196 _igvn.hash_delete( cd );
197 cd->set_req(0, prevdom);
198 set_early_ctrl( cd );
199 _igvn._worklist.push(cd);
200 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
201 if( old_loop != new_loop ) {
202 if( !old_loop->_child ) old_loop->_body.yank(cd);
203 if( !new_loop->_child ) new_loop->_body.push(cd);
204 }
205 --i;
206 --imax;
207 }
208 }
209 }
211 //------------------------------has_local_phi_input----------------------------
212 // Return TRUE if 'n' has Phi inputs from its local block and no other
213 // block-local inputs (all non-local-phi inputs come from earlier blocks)
214 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
215 Node *n_ctrl = get_ctrl(n);
216 // See if some inputs come from a Phi in this block, or from before
217 // this block.
218 uint i;
219 for( i = 1; i < n->req(); i++ ) {
220 Node *phi = n->in(i);
221 if( phi->is_Phi() && phi->in(0) == n_ctrl )
222 break;
223 }
224 if( i >= n->req() )
225 return NULL; // No Phi inputs; nowhere to clone thru
227 // Check for inputs created between 'n' and the Phi input. These
228 // must split as well; they have already been given the chance
229 // (courtesy of a post-order visit) and since they did not we must
230 // recover the 'cost' of splitting them by being very profitable
231 // when splitting 'n'. Since this is unlikely we simply give up.
232 for( i = 1; i < n->req(); i++ ) {
233 Node *m = n->in(i);
234 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
235 // We allow the special case of AddP's with no local inputs.
236 // This allows us to split-up address expressions.
237 if (m->is_AddP() &&
238 get_ctrl(m->in(2)) != n_ctrl &&
239 get_ctrl(m->in(3)) != n_ctrl) {
240 // Move the AddP up to dominating point
241 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl)));
242 continue;
243 }
244 return NULL;
245 }
246 }
248 return n_ctrl;
249 }
251 //------------------------------remix_address_expressions----------------------
252 // Rework addressing expressions to get the most loop-invariant stuff
253 // moved out. We'd like to do all associative operators, but it's especially
254 // important (common) to do address expressions.
255 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
256 if (!has_ctrl(n)) return NULL;
257 Node *n_ctrl = get_ctrl(n);
258 IdealLoopTree *n_loop = get_loop(n_ctrl);
260 // See if 'n' mixes loop-varying and loop-invariant inputs and
261 // itself is loop-varying.
263 // Only interested in binary ops (and AddP)
264 if( n->req() < 3 || n->req() > 4 ) return NULL;
266 Node *n1_ctrl = get_ctrl(n->in( 1));
267 Node *n2_ctrl = get_ctrl(n->in( 2));
268 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
269 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
270 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
271 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
273 // Does one of my inputs spin in a tighter loop than self?
274 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
275 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
276 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
277 return NULL; // Leave well enough alone
279 // Is at least one of my inputs loop-invariant?
280 if( n1_loop == n_loop &&
281 n2_loop == n_loop &&
282 n3_loop == n_loop )
283 return NULL; // No loop-invariant inputs
286 int n_op = n->Opcode();
288 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
289 if( n_op == Op_LShiftI ) {
290 // Scale is loop invariant
291 Node *scale = n->in(2);
292 Node *scale_ctrl = get_ctrl(scale);
293 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
294 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
295 return NULL;
296 const TypeInt *scale_t = scale->bottom_type()->isa_int();
297 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
298 return NULL; // Dont bother with byte/short masking
299 // Add must vary with loop (else shift would be loop-invariant)
300 Node *add = n->in(1);
301 Node *add_ctrl = get_ctrl(add);
302 IdealLoopTree *add_loop = get_loop(add_ctrl);
303 //assert( n_loop == add_loop, "" );
304 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
306 // Convert I-V into I+ (0-V); same for V-I
307 if( add->Opcode() == Op_SubI &&
308 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
309 Node *zero = _igvn.intcon(0);
310 set_ctrl(zero, C->root());
311 Node *neg = new (C, 3) SubINode( _igvn.intcon(0), add->in(2) );
312 register_new_node( neg, get_ctrl(add->in(2) ) );
313 add = new (C, 3) AddINode( add->in(1), neg );
314 register_new_node( add, add_ctrl );
315 }
316 if( add->Opcode() != Op_AddI ) return NULL;
317 // See if one add input is loop invariant
318 Node *add_var = add->in(1);
319 Node *add_var_ctrl = get_ctrl(add_var);
320 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
321 Node *add_invar = add->in(2);
322 Node *add_invar_ctrl = get_ctrl(add_invar);
323 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
324 if( add_var_loop == n_loop ) {
325 } else if( add_invar_loop == n_loop ) {
326 // Swap to find the invariant part
327 add_invar = add_var;
328 add_invar_ctrl = add_var_ctrl;
329 add_invar_loop = add_var_loop;
330 add_var = add->in(2);
331 Node *add_var_ctrl = get_ctrl(add_var);
332 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
333 } else // Else neither input is loop invariant
334 return NULL;
335 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
336 return NULL; // No invariant part of the add?
338 // Yes! Reshape address expression!
339 Node *inv_scale = new (C, 3) LShiftINode( add_invar, scale );
340 register_new_node( inv_scale, add_invar_ctrl );
341 Node *var_scale = new (C, 3) LShiftINode( add_var, scale );
342 register_new_node( var_scale, n_ctrl );
343 Node *var_add = new (C, 3) AddINode( var_scale, inv_scale );
344 register_new_node( var_add, n_ctrl );
345 _igvn.hash_delete( n );
346 _igvn.subsume_node( n, var_add );
347 return var_add;
348 }
350 // Replace (I+V) with (V+I)
351 if( n_op == Op_AddI ||
352 n_op == Op_AddL ||
353 n_op == Op_AddF ||
354 n_op == Op_AddD ||
355 n_op == Op_MulI ||
356 n_op == Op_MulL ||
357 n_op == Op_MulF ||
358 n_op == Op_MulD ) {
359 if( n2_loop == n_loop ) {
360 assert( n1_loop != n_loop, "" );
361 n->swap_edges(1, 2);
362 }
363 }
365 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
366 // but not if I2 is a constant.
367 if( n_op == Op_AddP ) {
368 if( n2_loop == n_loop && n3_loop != n_loop ) {
369 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
370 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
371 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
372 IdealLoopTree *n22loop = get_loop( n22_ctrl );
373 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
374 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
375 n23_loop == n_loop ) {
376 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
377 // Stuff new AddP in the loop preheader
378 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
379 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) );
380 register_new_node( add2, n_ctrl );
381 _igvn.hash_delete( n );
382 _igvn.subsume_node( n, add2 );
383 return add2;
384 }
385 }
386 }
388 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
389 if( n2_loop != n_loop && n3_loop == n_loop ) {
390 if( n->in(3)->Opcode() == Op_AddI ) {
391 Node *V = n->in(3)->in(1);
392 Node *I = n->in(3)->in(2);
393 if( is_member(n_loop,get_ctrl(V)) ) {
394 } else {
395 Node *tmp = V; V = I; I = tmp;
396 }
397 if( !is_member(n_loop,get_ctrl(I)) ) {
398 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I );
399 // Stuff new AddP in the loop preheader
400 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
401 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V );
402 register_new_node( add2, n_ctrl );
403 _igvn.hash_delete( n );
404 _igvn.subsume_node( n, add2 );
405 return add2;
406 }
407 }
408 }
409 }
411 return NULL;
412 }
414 //------------------------------conditional_move-------------------------------
415 // Attempt to replace a Phi with a conditional move. We have some pretty
416 // strict profitability requirements. All Phis at the merge point must
417 // be converted, so we can remove the control flow. We need to limit the
418 // number of c-moves to a small handful. All code that was in the side-arms
419 // of the CFG diamond is now speculatively executed. This code has to be
420 // "cheap enough". We are pretty much limited to CFG diamonds that merge
421 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
422 Node *PhaseIdealLoop::conditional_move( Node *region ) {
424 assert( region->is_Region(), "sanity check" );
425 if( region->req() != 3 ) return NULL;
427 // Check for CFG diamond
428 Node *lp = region->in(1);
429 Node *rp = region->in(2);
430 if( !lp || !rp ) return NULL;
431 Node *lp_c = lp->in(0);
432 if( lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If() ) return NULL;
433 IfNode *iff = lp_c->as_If();
435 // Check for highly predictable branch. No point in CMOV'ing if
436 // we are going to predict accurately all the time.
437 // %%% This hides patterns produced by utility methods like Math.min.
438 if( iff->_prob < PROB_UNLIKELY_MAG(3) ||
439 iff->_prob > PROB_LIKELY_MAG(3) )
440 return NULL;
442 // Check for ops pinned in an arm of the diamond.
443 // Can't remove the control flow in this case
444 if( lp->outcnt() > 1 ) return NULL;
445 if( rp->outcnt() > 1 ) return NULL;
447 // Check profitability
448 int cost = 0;
449 int phis = 0;
450 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
451 Node *out = region->fast_out(i);
452 if( !out->is_Phi() ) continue; // Ignore other control edges, etc
453 phis++;
454 PhiNode* phi = out->as_Phi();
455 switch (phi->type()->basic_type()) {
456 case T_LONG:
457 cost++; // Probably encodes as 2 CMOV's
458 case T_INT: // These all CMOV fine
459 case T_FLOAT:
460 case T_DOUBLE:
461 case T_ADDRESS: // (RawPtr)
462 cost++;
463 break;
464 case T_OBJECT: { // Base oops are OK, but not derived oops
465 const TypeOopPtr *tp = phi->type()->isa_oopptr();
466 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
467 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
468 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
469 // have a Phi for the base here that we convert to a CMOVE all is well
470 // and good. But if the base is dead, we'll not make a CMOVE. Later
471 // the allocator will have to produce a base by creating a CMOVE of the
472 // relevant bases. This puts the allocator in the business of
473 // manufacturing expensive instructions, generally a bad plan.
474 // Just Say No to Conditionally-Moved Derived Pointers.
475 if( tp && tp->offset() != 0 )
476 return NULL;
477 cost++;
478 break;
479 }
480 default:
481 return NULL; // In particular, can't do memory or I/O
482 }
483 // Add in cost any speculative ops
484 for( uint j = 1; j < region->req(); j++ ) {
485 Node *proj = region->in(j);
486 Node *inp = phi->in(j);
487 if (get_ctrl(inp) == proj) { // Found local op
488 cost++;
489 // Check for a chain of dependent ops; these will all become
490 // speculative in a CMOV.
491 for( uint k = 1; k < inp->req(); k++ )
492 if (get_ctrl(inp->in(k)) == proj)
493 return NULL; // Too much speculative goo
494 }
495 }
496 // See if the Phi is used by a Cmp. This will likely Split-If, a
497 // higher-payoff operation.
498 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
499 Node* use = phi->fast_out(k);
500 if( use->is_Cmp() )
501 return NULL;
502 }
503 }
504 if( cost >= ConditionalMoveLimit ) return NULL; // Too much goo
505 Node* bol = iff->in(1);
506 assert( bol->Opcode() == Op_Bool, "" );
507 int cmp_op = bol->in(1)->Opcode();
508 // It is expensive to generate flags from a float compare.
509 // Avoid duplicated float compare.
510 if( phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
512 // --------------
513 // Now replace all Phis with CMOV's
514 Node *cmov_ctrl = iff->in(0);
515 uint flip = (lp->Opcode() == Op_IfTrue);
516 while( 1 ) {
517 PhiNode* phi = NULL;
518 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
519 Node *out = region->fast_out(i);
520 if (out->is_Phi()) {
521 phi = out->as_Phi();
522 break;
523 }
524 }
525 if (phi == NULL) break;
526 #ifndef PRODUCT
527 if( PrintOpto && VerifyLoopOptimizations ) tty->print_cr("CMOV");
528 #endif
529 // Move speculative ops
530 for( uint j = 1; j < region->req(); j++ ) {
531 Node *proj = region->in(j);
532 Node *inp = phi->in(j);
533 if (get_ctrl(inp) == proj) { // Found local op
534 #ifndef PRODUCT
535 if( PrintOpto && VerifyLoopOptimizations ) {
536 tty->print(" speculate: ");
537 inp->dump();
538 }
539 #endif
540 set_ctrl(inp, cmov_ctrl);
541 }
542 }
543 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) );
544 register_new_node( cmov, cmov_ctrl );
545 _igvn.hash_delete(phi);
546 _igvn.subsume_node( phi, cmov );
547 #ifndef PRODUCT
548 if( VerifyLoopOptimizations ) verify();
549 #endif
550 }
552 // The useless CFG diamond will fold up later; see the optimization in
553 // RegionNode::Ideal.
554 _igvn._worklist.push(region);
556 return iff->in(1);
557 }
559 //------------------------------split_if_with_blocks_pre-----------------------
560 // Do the real work in a non-recursive function. Data nodes want to be
561 // cloned in the pre-order so they can feed each other nicely.
562 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
563 // Cloning these guys is unlikely to win
564 int n_op = n->Opcode();
565 if( n_op == Op_MergeMem ) return n;
566 if( n->is_Proj() ) return n;
567 // Do not clone-up CmpFXXX variations, as these are always
568 // followed by a CmpI
569 if( n->is_Cmp() ) return n;
570 // Attempt to use a conditional move instead of a phi/branch
571 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
572 Node *cmov = conditional_move( n );
573 if( cmov ) return cmov;
574 }
575 if( n->is_CFG() || n_op == Op_StorePConditional || n_op == Op_StoreLConditional || n_op == Op_CompareAndSwapI || n_op == Op_CompareAndSwapL ||n_op == Op_CompareAndSwapP) return n;
576 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
577 n_op == Op_Opaque2 ) {
578 if( !C->major_progress() ) // If chance of no more loop opts...
579 _igvn._worklist.push(n); // maybe we'll remove them
580 return n;
581 }
583 if( n->is_Con() ) return n; // No cloning for Con nodes
585 Node *n_ctrl = get_ctrl(n);
586 if( !n_ctrl ) return n; // Dead node
588 // Attempt to remix address expressions for loop invariants
589 Node *m = remix_address_expressions( n );
590 if( m ) return m;
592 // Determine if the Node has inputs from some local Phi.
593 // Returns the block to clone thru.
594 Node *n_blk = has_local_phi_input( n );
595 if( !n_blk ) return n;
596 // Do not clone the trip counter through on a CountedLoop
597 // (messes up the canonical shape).
598 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
600 // Check for having no control input; not pinned. Allow
601 // dominating control.
602 if( n->in(0) ) {
603 Node *dom = idom(n_blk);
604 if( dom_lca( n->in(0), dom ) != n->in(0) )
605 return n;
606 }
607 // Policy: when is it profitable. You must get more wins than
608 // policy before it is considered profitable. Policy is usually 0,
609 // so 1 win is considered profitable. Big merges will require big
610 // cloning, so get a larger policy.
611 int policy = n_blk->req() >> 2;
613 // If the loop is a candidate for range check elimination,
614 // delay splitting through it's phi until a later loop optimization
615 if (n_blk->is_CountedLoop()) {
616 IdealLoopTree *lp = get_loop(n_blk);
617 if (lp && lp->_rce_candidate) {
618 return n;
619 }
620 }
622 // Use same limit as split_if_with_blocks_post
623 if( C->unique() > 35000 ) return n; // Method too big
625 // Split 'n' through the merge point if it is profitable
626 Node *phi = split_thru_phi( n, n_blk, policy );
627 if( !phi ) return n;
629 // Found a Phi to split thru!
630 // Replace 'n' with the new phi
631 _igvn.hash_delete(n);
632 _igvn.subsume_node( n, phi );
633 // Moved a load around the loop, 'en-registering' something.
634 if( n_blk->Opcode() == Op_Loop && n->is_Load() &&
635 !phi->in(LoopNode::LoopBackControl)->is_Load() )
636 C->set_major_progress();
638 return phi;
639 }
641 static bool merge_point_too_heavy(Compile* C, Node* region) {
642 // Bail out if the region and its phis have too many users.
643 int weight = 0;
644 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
645 weight += region->fast_out(i)->outcnt();
646 }
647 int nodes_left = MaxNodeLimit - C->unique();
648 if (weight * 8 > nodes_left) {
649 #ifndef PRODUCT
650 if (PrintOpto)
651 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
652 #endif
653 return true;
654 } else {
655 return false;
656 }
657 }
659 #ifdef _LP64
660 static bool merge_point_safe(Node* region) {
661 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
662 // having a PhiNode input. This sidesteps the dangerous case where the split
663 // ConvI2LNode may become TOP if the input Value() does not
664 // overlap the ConvI2L range, leaving a node which may not dominate its
665 // uses.
666 // A better fix for this problem can be found in the BugTraq entry, but
667 // expediency for Mantis demands this hack.
668 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
669 Node* n = region->fast_out(i);
670 if (n->is_Phi()) {
671 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
672 Node* m = n->fast_out(j);
673 if (m->Opcode() == Op_ConvI2L) {
674 return false;
675 }
676 }
677 }
678 }
679 return true;
680 }
681 #endif
684 //------------------------------place_near_use---------------------------------
685 // Place some computation next to use but not inside inner loops.
686 // For inner loop uses move it to the preheader area.
687 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const {
688 IdealLoopTree *u_loop = get_loop( useblock );
689 return (u_loop->_irreducible || u_loop->_child)
690 ? useblock
691 : u_loop->_head->in(LoopNode::EntryControl);
692 }
695 //------------------------------split_if_with_blocks_post----------------------
696 // Do the real work in a non-recursive function. CFG hackery wants to be
697 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
698 // info.
699 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) {
701 // Cloning Cmp through Phi's involves the split-if transform.
702 // FastLock is not used by an If
703 if( n->is_Cmp() && !n->is_FastLock() ) {
704 if( C->unique() > 35000 ) return; // Method too big
706 // Do not do 'split-if' if irreducible loops are present.
707 if( _has_irreducible_loops )
708 return;
710 Node *n_ctrl = get_ctrl(n);
711 // Determine if the Node has inputs from some local Phi.
712 // Returns the block to clone thru.
713 Node *n_blk = has_local_phi_input( n );
714 if( n_blk != n_ctrl ) return;
716 if( merge_point_too_heavy(C, n_ctrl) )
717 return;
719 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare?
720 Node *bol = n->unique_out();
721 assert( bol->is_Bool(), "expect a bool here" );
722 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare?
723 Node *iff = bol->unique_out();
725 // Check some safety conditions
726 if( iff->is_If() ) { // Classic split-if?
727 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if
728 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
729 if( get_ctrl(iff->in(2)) == n_ctrl ||
730 get_ctrl(iff->in(3)) == n_ctrl )
731 return; // Inputs not yet split-up
732 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) {
733 return; // Loop-invar test gates loop-varying CMOVE
734 }
735 } else {
736 return; // some other kind of node, such as an Allocate
737 }
739 // Do not do 'split-if' if some paths are dead. First do dead code
740 // elimination and then see if its still profitable.
741 for( uint i = 1; i < n_ctrl->req(); i++ )
742 if( n_ctrl->in(i) == C->top() )
743 return;
745 // When is split-if profitable? Every 'win' on means some control flow
746 // goes dead, so it's almost always a win.
747 int policy = 0;
748 // If trying to do a 'Split-If' at the loop head, it is only
749 // profitable if the cmp folds up on BOTH paths. Otherwise we
750 // risk peeling a loop forever.
752 // CNC - Disabled for now. Requires careful handling of loop
753 // body selection for the cloned code. Also, make sure we check
754 // for any input path not being in the same loop as n_ctrl. For
755 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
756 // because the alternative loop entry points won't be converted
757 // into LoopNodes.
758 IdealLoopTree *n_loop = get_loop(n_ctrl);
759 for( uint j = 1; j < n_ctrl->req(); j++ )
760 if( get_loop(n_ctrl->in(j)) != n_loop )
761 return;
763 #ifdef _LP64
764 // Check for safety of the merge point.
765 if( !merge_point_safe(n_ctrl) ) {
766 return;
767 }
768 #endif
770 // Split compare 'n' through the merge point if it is profitable
771 Node *phi = split_thru_phi( n, n_ctrl, policy );
772 if( !phi ) return;
774 // Found a Phi to split thru!
775 // Replace 'n' with the new phi
776 _igvn.hash_delete(n);
777 _igvn.subsume_node( n, phi );
779 // Now split the bool up thru the phi
780 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 );
781 _igvn.hash_delete(bol);
782 _igvn.subsume_node( bol, bolphi );
783 assert( iff->in(1) == bolphi, "" );
784 if( bolphi->Value(&_igvn)->singleton() )
785 return;
787 // Conditional-move? Must split up now
788 if( !iff->is_If() ) {
789 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 );
790 _igvn.hash_delete(iff);
791 _igvn.subsume_node( iff, cmovphi );
792 return;
793 }
795 // Now split the IF
796 do_split_if( iff );
797 return;
798 }
800 // Check for an IF ready to split; one that has its
801 // condition codes input coming from a Phi at the block start.
802 int n_op = n->Opcode();
804 // Check for an IF being dominated by another IF same test
805 if( n_op == Op_If ) {
806 Node *bol = n->in(1);
807 uint max = bol->outcnt();
808 // Check for same test used more than once?
809 if( n_op == Op_If && max > 1 && bol->is_Bool() ) {
810 // Search up IDOMs to see if this IF is dominated.
811 Node *cutoff = get_ctrl(bol);
813 // Now search up IDOMs till cutoff, looking for a dominating test
814 Node *prevdom = n;
815 Node *dom = idom(prevdom);
816 while( dom != cutoff ) {
817 if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) {
818 // Replace the dominated test with an obvious true or false.
819 // Place it on the IGVN worklist for later cleanup.
820 C->set_major_progress();
821 dominated_by( prevdom, n );
822 #ifndef PRODUCT
823 if( VerifyLoopOptimizations ) verify();
824 #endif
825 return;
826 }
827 prevdom = dom;
828 dom = idom(prevdom);
829 }
830 }
831 }
833 // See if a shared loop-varying computation has no loop-varying uses.
834 // Happens if something is only used for JVM state in uncommon trap exits,
835 // like various versions of induction variable+offset. Clone the
836 // computation per usage to allow it to sink out of the loop.
837 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
838 Node *n_ctrl = get_ctrl(n);
839 IdealLoopTree *n_loop = get_loop(n_ctrl);
840 if( n_loop != _ltree_root ) {
841 DUIterator_Fast imax, i = n->fast_outs(imax);
842 for (; i < imax; i++) {
843 Node* u = n->fast_out(i);
844 if( !has_ctrl(u) ) break; // Found control user
845 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
846 if( u_loop == n_loop ) break; // Found loop-varying use
847 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
848 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
849 }
850 bool did_break = (i < imax); // Did we break out of the previous loop?
851 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
852 Node *late_load_ctrl;
853 if (n->is_Load()) {
854 // If n is a load, get and save the result from get_late_ctrl(),
855 // to be later used in calculating the control for n's clones.
856 clear_dom_lca_tags();
857 late_load_ctrl = get_late_ctrl(n, n_ctrl);
858 }
859 // If n is a load, and the late control is the same as the current
860 // control, then the cloning of n is a pointless exercise, because
861 // GVN will ensure that we end up where we started.
862 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
863 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
864 Node *u = n->last_out(j); // Clone private computation per use
865 _igvn.hash_delete(u);
866 _igvn._worklist.push(u);
867 Node *x = n->clone(); // Clone computation
868 Node *x_ctrl = NULL;
869 if( u->is_Phi() ) {
870 // Replace all uses of normal nodes. Replace Phi uses
871 // individually, so the seperate Nodes can sink down
872 // different paths.
873 uint k = 1;
874 while( u->in(k) != n ) k++;
875 u->set_req( k, x );
876 // x goes next to Phi input path
877 x_ctrl = u->in(0)->in(k);
878 --j;
879 } else { // Normal use
880 // Replace all uses
881 for( uint k = 0; k < u->req(); k++ ) {
882 if( u->in(k) == n ) {
883 u->set_req( k, x );
884 --j;
885 }
886 }
887 x_ctrl = get_ctrl(u);
888 }
890 // Find control for 'x' next to use but not inside inner loops.
891 // For inner loop uses get the preheader area.
892 x_ctrl = place_near_use(x_ctrl);
894 if (n->is_Load()) {
895 // For loads, add a control edge to a CFG node outside of the loop
896 // to force them to not combine and return back inside the loop
897 // during GVN optimization (4641526).
898 //
899 // Because we are setting the actual control input, factor in
900 // the result from get_late_ctrl() so we respect any
901 // anti-dependences. (6233005).
902 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
904 // Don't allow the control input to be a CFG splitting node.
905 // Such nodes should only have ProjNodes as outs, e.g. IfNode
906 // should only have IfTrueNode and IfFalseNode (4985384).
907 x_ctrl = find_non_split_ctrl(x_ctrl);
908 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
910 x->set_req(0, x_ctrl);
911 }
912 register_new_node(x, x_ctrl);
914 // Some institutional knowledge is needed here: 'x' is
915 // yanked because if the optimizer runs GVN on it all the
916 // cloned x's will common up and undo this optimization and
917 // be forced back in the loop. This is annoying because it
918 // makes +VerifyOpto report false-positives on progress. I
919 // tried setting control edges on the x's to force them to
920 // not combine, but the matching gets worried when it tries
921 // to fold a StoreP and an AddP together (as part of an
922 // address expression) and the AddP and StoreP have
923 // different controls.
924 if( !x->is_Load() ) _igvn._worklist.yank(x);
925 }
926 _igvn.remove_dead_node(n);
927 }
928 }
929 }
930 }
932 // Check for Opaque2's who's loop has disappeared - who's input is in the
933 // same loop nest as their output. Remove 'em, they are no longer useful.
934 if( n_op == Op_Opaque2 &&
935 n->in(1) != NULL &&
936 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
937 _igvn.add_users_to_worklist(n);
938 _igvn.hash_delete(n);
939 _igvn.subsume_node( n, n->in(1) );
940 }
941 }
943 //------------------------------split_if_with_blocks---------------------------
944 // Check for aggressive application of 'split-if' optimization,
945 // using basic block level info.
946 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) {
947 Node *n = C->root();
948 visited.set(n->_idx); // first, mark node as visited
949 // Do pre-visit work for root
950 n = split_if_with_blocks_pre( n );
951 uint cnt = n->outcnt();
952 uint i = 0;
953 while (true) {
954 // Visit all children
955 if (i < cnt) {
956 Node* use = n->raw_out(i);
957 ++i;
958 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
959 // Now do pre-visit work for this use
960 use = split_if_with_blocks_pre( use );
961 nstack.push(n, i); // Save parent and next use's index.
962 n = use; // Process all children of current use.
963 cnt = use->outcnt();
964 i = 0;
965 }
966 }
967 else {
968 // All of n's children have been processed, complete post-processing.
969 if (cnt != 0 && !n->is_Con()) {
970 assert(has_node(n), "no dead nodes");
971 split_if_with_blocks_post( n );
972 }
973 if (nstack.is_empty()) {
974 // Finished all nodes on stack.
975 break;
976 }
977 // Get saved parent node and next use's index. Visit the rest of uses.
978 n = nstack.node();
979 cnt = n->outcnt();
980 i = nstack.index();
981 nstack.pop();
982 }
983 }
984 }
987 //=============================================================================
988 //
989 // C L O N E A L O O P B O D Y
990 //
992 //------------------------------clone_iff--------------------------------------
993 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
994 // "Nearly" because all Nodes have been cloned from the original in the loop,
995 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
996 // through the Phi recursively, and return a Bool.
997 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) {
999 // Convert this Phi into a Phi merging Bools
1000 uint i;
1001 for( i = 1; i < phi->req(); i++ ) {
1002 Node *b = phi->in(i);
1003 if( b->is_Phi() ) {
1004 _igvn.hash_delete(phi);
1005 _igvn._worklist.push(phi);
1006 phi->set_req(i, clone_iff( b->as_Phi(), loop ));
1007 } else {
1008 assert( b->is_Bool(), "" );
1009 }
1010 }
1012 Node *sample_bool = phi->in(1);
1013 Node *sample_cmp = sample_bool->in(1);
1015 // Make Phis to merge the Cmp's inputs.
1016 int size = phi->in(0)->req();
1017 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1018 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1019 for( i = 1; i < phi->req(); i++ ) {
1020 Node *n1 = phi->in(i)->in(1)->in(1);
1021 Node *n2 = phi->in(i)->in(1)->in(2);
1022 phi1->set_req( i, n1 );
1023 phi2->set_req( i, n2 );
1024 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1025 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1026 }
1027 // See if these Phis have been made before.
1028 // Register with optimizer
1029 Node *hit1 = _igvn.hash_find_insert(phi1);
1030 if( hit1 ) { // Hit, toss just made Phi
1031 _igvn.remove_dead_node(phi1); // Remove new phi
1032 assert( hit1->is_Phi(), "" );
1033 phi1 = (PhiNode*)hit1; // Use existing phi
1034 } else { // Miss
1035 _igvn.register_new_node_with_optimizer(phi1);
1036 }
1037 Node *hit2 = _igvn.hash_find_insert(phi2);
1038 if( hit2 ) { // Hit, toss just made Phi
1039 _igvn.remove_dead_node(phi2); // Remove new phi
1040 assert( hit2->is_Phi(), "" );
1041 phi2 = (PhiNode*)hit2; // Use existing phi
1042 } else { // Miss
1043 _igvn.register_new_node_with_optimizer(phi2);
1044 }
1045 // Register Phis with loop/block info
1046 set_ctrl(phi1, phi->in(0));
1047 set_ctrl(phi2, phi->in(0));
1048 // Make a new Cmp
1049 Node *cmp = sample_cmp->clone();
1050 cmp->set_req( 1, phi1 );
1051 cmp->set_req( 2, phi2 );
1052 _igvn.register_new_node_with_optimizer(cmp);
1053 set_ctrl(cmp, phi->in(0));
1055 // Make a new Bool
1056 Node *b = sample_bool->clone();
1057 b->set_req(1,cmp);
1058 _igvn.register_new_node_with_optimizer(b);
1059 set_ctrl(b, phi->in(0));
1061 assert( b->is_Bool(), "" );
1062 return (BoolNode*)b;
1063 }
1065 //------------------------------clone_bool-------------------------------------
1066 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1067 // "Nearly" because all Nodes have been cloned from the original in the loop,
1068 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1069 // through the Phi recursively, and return a Bool.
1070 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1071 uint i;
1072 // Convert this Phi into a Phi merging Bools
1073 for( i = 1; i < phi->req(); i++ ) {
1074 Node *b = phi->in(i);
1075 if( b->is_Phi() ) {
1076 _igvn.hash_delete(phi);
1077 _igvn._worklist.push(phi);
1078 phi->set_req(i, clone_bool( b->as_Phi(), loop ));
1079 } else {
1080 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1081 }
1082 }
1084 Node *sample_cmp = phi->in(1);
1086 // Make Phis to merge the Cmp's inputs.
1087 int size = phi->in(0)->req();
1088 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1089 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1090 for( uint j = 1; j < phi->req(); j++ ) {
1091 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1092 Node *n1, *n2;
1093 if( cmp_top->is_Cmp() ) {
1094 n1 = cmp_top->in(1);
1095 n2 = cmp_top->in(2);
1096 } else {
1097 n1 = n2 = cmp_top;
1098 }
1099 phi1->set_req( j, n1 );
1100 phi2->set_req( j, n2 );
1101 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1102 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1103 }
1105 // See if these Phis have been made before.
1106 // Register with optimizer
1107 Node *hit1 = _igvn.hash_find_insert(phi1);
1108 if( hit1 ) { // Hit, toss just made Phi
1109 _igvn.remove_dead_node(phi1); // Remove new phi
1110 assert( hit1->is_Phi(), "" );
1111 phi1 = (PhiNode*)hit1; // Use existing phi
1112 } else { // Miss
1113 _igvn.register_new_node_with_optimizer(phi1);
1114 }
1115 Node *hit2 = _igvn.hash_find_insert(phi2);
1116 if( hit2 ) { // Hit, toss just made Phi
1117 _igvn.remove_dead_node(phi2); // Remove new phi
1118 assert( hit2->is_Phi(), "" );
1119 phi2 = (PhiNode*)hit2; // Use existing phi
1120 } else { // Miss
1121 _igvn.register_new_node_with_optimizer(phi2);
1122 }
1123 // Register Phis with loop/block info
1124 set_ctrl(phi1, phi->in(0));
1125 set_ctrl(phi2, phi->in(0));
1126 // Make a new Cmp
1127 Node *cmp = sample_cmp->clone();
1128 cmp->set_req( 1, phi1 );
1129 cmp->set_req( 2, phi2 );
1130 _igvn.register_new_node_with_optimizer(cmp);
1131 set_ctrl(cmp, phi->in(0));
1133 assert( cmp->is_Cmp(), "" );
1134 return (CmpNode*)cmp;
1135 }
1137 //------------------------------sink_use---------------------------------------
1138 // If 'use' was in the loop-exit block, it now needs to be sunk
1139 // below the post-loop merge point.
1140 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1141 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1142 set_ctrl(use, post_loop);
1143 for (DUIterator j = use->outs(); use->has_out(j); j++)
1144 sink_use(use->out(j), post_loop);
1145 }
1146 }
1148 //------------------------------clone_loop-------------------------------------
1149 //
1150 // C L O N E A L O O P B O D Y
1151 //
1152 // This is the basic building block of the loop optimizations. It clones an
1153 // entire loop body. It makes an old_new loop body mapping; with this mapping
1154 // you can find the new-loop equivalent to an old-loop node. All new-loop
1155 // nodes are exactly equal to their old-loop counterparts, all edges are the
1156 // same. All exits from the old-loop now have a RegionNode that merges the
1157 // equivalent new-loop path. This is true even for the normal "loop-exit"
1158 // condition. All uses of loop-invariant old-loop values now come from (one
1159 // or more) Phis that merge their new-loop equivalents.
1160 //
1161 // This operation leaves the graph in an illegal state: there are two valid
1162 // control edges coming from the loop pre-header to both loop bodies. I'll
1163 // definitely have to hack the graph after running this transform.
1164 //
1165 // From this building block I will further edit edges to perform loop peeling
1166 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1167 //
1168 // Parameter side_by_size_idom:
1169 // When side_by_size_idom is NULL, the dominator tree is constructed for
1170 // the clone loop to dominate the original. Used in construction of
1171 // pre-main-post loop sequence.
1172 // When nonnull, the clone and original are side-by-side, both are
1173 // dominated by the side_by_side_idom node. Used in construction of
1174 // unswitched loops.
1175 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1176 Node* side_by_side_idom) {
1178 // Step 1: Clone the loop body. Make the old->new mapping.
1179 uint i;
1180 for( i = 0; i < loop->_body.size(); i++ ) {
1181 Node *old = loop->_body.at(i);
1182 Node *nnn = old->clone();
1183 old_new.map( old->_idx, nnn );
1184 _igvn.register_new_node_with_optimizer(nnn);
1185 }
1188 // Step 2: Fix the edges in the new body. If the old input is outside the
1189 // loop use it. If the old input is INside the loop, use the corresponding
1190 // new node instead.
1191 for( i = 0; i < loop->_body.size(); i++ ) {
1192 Node *old = loop->_body.at(i);
1193 Node *nnn = old_new[old->_idx];
1194 // Fix CFG/Loop controlling the new node
1195 if (has_ctrl(old)) {
1196 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
1197 } else {
1198 set_loop(nnn, loop->_parent);
1199 if (old->outcnt() > 0) {
1200 set_idom( nnn, old_new[idom(old)->_idx], dd );
1201 }
1202 }
1203 // Correct edges to the new node
1204 for( uint j = 0; j < nnn->req(); j++ ) {
1205 Node *n = nnn->in(j);
1206 if( n ) {
1207 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
1208 if( loop->is_member( old_in_loop ) )
1209 nnn->set_req(j, old_new[n->_idx]);
1210 }
1211 }
1212 _igvn.hash_find_insert(nnn);
1213 }
1214 Node *newhead = old_new[loop->_head->_idx];
1215 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1218 // Step 3: Now fix control uses. Loop varying control uses have already
1219 // been fixed up (as part of all input edges in Step 2). Loop invariant
1220 // control uses must be either an IfFalse or an IfTrue. Make a merge
1221 // point to merge the old and new IfFalse/IfTrue nodes; make the use
1222 // refer to this.
1223 ResourceArea *area = Thread::current()->resource_area();
1224 Node_List worklist(area);
1225 uint new_counter = C->unique();
1226 for( i = 0; i < loop->_body.size(); i++ ) {
1227 Node* old = loop->_body.at(i);
1228 if( !old->is_CFG() ) continue;
1229 Node* nnn = old_new[old->_idx];
1231 // Copy uses to a worklist, so I can munge the def-use info
1232 // with impunity.
1233 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1234 worklist.push(old->fast_out(j));
1236 while( worklist.size() ) { // Visit all uses
1237 Node *use = worklist.pop();
1238 if (!has_node(use)) continue; // Ignore dead nodes
1239 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1240 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
1241 // Both OLD and USE are CFG nodes here.
1242 assert( use->is_Proj(), "" );
1244 // Clone the loop exit control projection
1245 Node *newuse = use->clone();
1246 newuse->set_req(0,nnn);
1247 _igvn.register_new_node_with_optimizer(newuse);
1248 set_loop(newuse, use_loop);
1249 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
1251 // We need a Region to merge the exit from the peeled body and the
1252 // exit from the old loop body.
1253 RegionNode *r = new (C, 3) RegionNode(3);
1254 // Map the old use to the new merge point
1255 old_new.map( use->_idx, r );
1256 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
1257 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
1259 // The original user of 'use' uses 'r' instead.
1260 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
1261 Node* useuse = use->last_out(l);
1262 _igvn.hash_delete(useuse);
1263 _igvn._worklist.push(useuse);
1264 uint uses_found = 0;
1265 if( useuse->in(0) == use ) {
1266 useuse->set_req(0, r);
1267 uses_found++;
1268 if( useuse->is_CFG() ) {
1269 assert( dom_depth(useuse) > dd_r, "" );
1270 set_idom(useuse, r, dom_depth(useuse));
1271 }
1272 }
1273 for( uint k = 1; k < useuse->req(); k++ ) {
1274 if( useuse->in(k) == use ) {
1275 useuse->set_req(k, r);
1276 uses_found++;
1277 }
1278 }
1279 l -= uses_found; // we deleted 1 or more copies of this edge
1280 }
1282 // Now finish up 'r'
1283 r->set_req( 1, newuse );
1284 r->set_req( 2, use );
1285 _igvn.register_new_node_with_optimizer(r);
1286 set_loop(r, use_loop);
1287 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
1288 } // End of if a loop-exit test
1289 }
1290 }
1292 // Step 4: If loop-invariant use is not control, it must be dominated by a
1293 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
1294 // there if needed. Make a Phi there merging old and new used values.
1295 Node_List *split_if_set = NULL;
1296 Node_List *split_bool_set = NULL;
1297 Node_List *split_cex_set = NULL;
1298 for( i = 0; i < loop->_body.size(); i++ ) {
1299 Node* old = loop->_body.at(i);
1300 Node* nnn = old_new[old->_idx];
1301 // Copy uses to a worklist, so I can munge the def-use info
1302 // with impunity.
1303 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1304 worklist.push(old->fast_out(j));
1306 while( worklist.size() ) {
1307 Node *use = worklist.pop();
1308 if (!has_node(use)) continue; // Ignore dead nodes
1309 if (use->in(0) == C->top()) continue;
1310 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1311 // Check for data-use outside of loop - at least one of OLD or USE
1312 // must not be a CFG node.
1313 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) {
1315 // If the Data use is an IF, that means we have an IF outside of the
1316 // loop that is switching on a condition that is set inside of the
1317 // loop. Happens if people set a loop-exit flag; then test the flag
1318 // in the loop to break the loop, then test is again outside of the
1319 // loop to determine which way the loop exited.
1320 if( use->is_If() || use->is_CMove() ) {
1321 // Since this code is highly unlikely, we lazily build the worklist
1322 // of such Nodes to go split.
1323 if( !split_if_set )
1324 split_if_set = new Node_List(area);
1325 split_if_set->push(use);
1326 }
1327 if( use->is_Bool() ) {
1328 if( !split_bool_set )
1329 split_bool_set = new Node_List(area);
1330 split_bool_set->push(use);
1331 }
1332 if( use->Opcode() == Op_CreateEx ) {
1333 if( !split_cex_set )
1334 split_cex_set = new Node_List(area);
1335 split_cex_set->push(use);
1336 }
1339 // Get "block" use is in
1340 uint idx = 0;
1341 while( use->in(idx) != old ) idx++;
1342 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1343 assert( !loop->is_member( get_loop( prev ) ), "" );
1344 Node *cfg = prev->_idx >= new_counter
1345 ? prev->in(2)
1346 : idom(prev);
1347 if( use->is_Phi() ) // Phi use is in prior block
1348 cfg = prev->in(idx); // NOT in block of Phi itself
1349 if (cfg->is_top()) { // Use is dead?
1350 _igvn.hash_delete(use);
1351 _igvn._worklist.push(use);
1352 use->set_req(idx, C->top());
1353 continue;
1354 }
1356 while( !loop->is_member( get_loop( cfg ) ) ) {
1357 prev = cfg;
1358 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1359 }
1360 // If the use occurs after merging several exits from the loop, then
1361 // old value must have dominated all those exits. Since the same old
1362 // value was used on all those exits we did not need a Phi at this
1363 // merge point. NOW we do need a Phi here. Each loop exit value
1364 // is now merged with the peeled body exit; each exit gets its own
1365 // private Phi and those Phis need to be merged here.
1366 Node *phi;
1367 if( prev->is_Region() ) {
1368 if( idx == 0 ) { // Updating control edge?
1369 phi = prev; // Just use existing control
1370 } else { // Else need a new Phi
1371 phi = PhiNode::make( prev, old );
1372 // Now recursively fix up the new uses of old!
1373 for( uint i = 1; i < prev->req(); i++ ) {
1374 worklist.push(phi); // Onto worklist once for each 'old' input
1375 }
1376 }
1377 } else {
1378 // Get new RegionNode merging old and new loop exits
1379 prev = old_new[prev->_idx];
1380 assert( prev, "just made this in step 7" );
1381 if( idx == 0 ) { // Updating control edge?
1382 phi = prev; // Just use existing control
1383 } else { // Else need a new Phi
1384 // Make a new Phi merging data values properly
1385 phi = PhiNode::make( prev, old );
1386 phi->set_req( 1, nnn );
1387 }
1388 }
1389 // If inserting a new Phi, check for prior hits
1390 if( idx != 0 ) {
1391 Node *hit = _igvn.hash_find_insert(phi);
1392 if( hit == NULL ) {
1393 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1394 } else { // or
1395 // Remove the new phi from the graph and use the hit
1396 _igvn.remove_dead_node(phi);
1397 phi = hit; // Use existing phi
1398 }
1399 set_ctrl(phi, prev);
1400 }
1401 // Make 'use' use the Phi instead of the old loop body exit value
1402 _igvn.hash_delete(use);
1403 _igvn._worklist.push(use);
1404 use->set_req(idx, phi);
1405 if( use->_idx >= new_counter ) { // If updating new phis
1406 // Not needed for correctness, but prevents a weak assert
1407 // in AddPNode from tripping (when we end up with different
1408 // base & derived Phis that will become the same after
1409 // IGVN does CSE).
1410 Node *hit = _igvn.hash_find_insert(use);
1411 if( hit ) // Go ahead and re-hash for hits.
1412 _igvn.subsume_node( use, hit );
1413 }
1415 // If 'use' was in the loop-exit block, it now needs to be sunk
1416 // below the post-loop merge point.
1417 sink_use( use, prev );
1418 }
1419 }
1420 }
1422 // Check for IFs that need splitting/cloning. Happens if an IF outside of
1423 // the loop uses a condition set in the loop. The original IF probably
1424 // takes control from one or more OLD Regions (which in turn get from NEW
1425 // Regions). In any case, there will be a set of Phis for each merge point
1426 // from the IF up to where the original BOOL def exists the loop.
1427 if( split_if_set ) {
1428 while( split_if_set->size() ) {
1429 Node *iff = split_if_set->pop();
1430 if( iff->in(1)->is_Phi() ) {
1431 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop );
1432 _igvn.hash_delete(iff);
1433 _igvn._worklist.push(iff);
1434 iff->set_req(1, b);
1435 }
1436 }
1437 }
1438 if( split_bool_set ) {
1439 while( split_bool_set->size() ) {
1440 Node *b = split_bool_set->pop();
1441 Node *phi = b->in(1);
1442 assert( phi->is_Phi(), "" );
1443 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop );
1444 _igvn.hash_delete(b);
1445 _igvn._worklist.push(b);
1446 b->set_req(1, cmp);
1447 }
1448 }
1449 if( split_cex_set ) {
1450 while( split_cex_set->size() ) {
1451 Node *b = split_cex_set->pop();
1452 assert( b->in(0)->is_Region(), "" );
1453 assert( b->in(1)->is_Phi(), "" );
1454 assert( b->in(0)->in(0) == b->in(1)->in(0), "" );
1455 split_up( b, b->in(0), NULL );
1456 }
1457 }
1459 }
1462 //---------------------- stride_of_possible_iv -------------------------------------
1463 // Looks for an iff/bool/comp with one operand of the compare
1464 // being a cycle involving an add and a phi,
1465 // with an optional truncation (left-shift followed by a right-shift)
1466 // of the add. Returns zero if not an iv.
1467 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
1468 Node* trunc1 = NULL;
1469 Node* trunc2 = NULL;
1470 const TypeInt* ttype = NULL;
1471 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
1472 return 0;
1473 }
1474 BoolNode* bl = iff->in(1)->as_Bool();
1475 Node* cmp = bl->in(1);
1476 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) {
1477 return 0;
1478 }
1479 // Must have an invariant operand
1480 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
1481 return 0;
1482 }
1483 Node* add2 = NULL;
1484 Node* cmp1 = cmp->in(1);
1485 if (cmp1->is_Phi()) {
1486 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
1487 Node* phi = cmp1;
1488 for (uint i = 1; i < phi->req(); i++) {
1489 Node* in = phi->in(i);
1490 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
1491 &trunc1, &trunc2, &ttype);
1492 if (add && add->in(1) == phi) {
1493 add2 = add->in(2);
1494 break;
1495 }
1496 }
1497 } else {
1498 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
1499 Node* addtrunc = cmp1;
1500 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
1501 &trunc1, &trunc2, &ttype);
1502 if (add && add->in(1)->is_Phi()) {
1503 Node* phi = add->in(1);
1504 for (uint i = 1; i < phi->req(); i++) {
1505 if (phi->in(i) == addtrunc) {
1506 add2 = add->in(2);
1507 break;
1508 }
1509 }
1510 }
1511 }
1512 if (add2 != NULL) {
1513 const TypeInt* add2t = _igvn.type(add2)->is_int();
1514 if (add2t->is_con()) {
1515 return add2t->get_con();
1516 }
1517 }
1518 return 0;
1519 }
1522 //---------------------- stay_in_loop -------------------------------------
1523 // Return the (unique) control output node that's in the loop (if it exists.)
1524 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
1525 Node* unique = NULL;
1526 if (!n) return NULL;
1527 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1528 Node* use = n->fast_out(i);
1529 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
1530 if (unique != NULL) {
1531 return NULL;
1532 }
1533 unique = use;
1534 }
1535 }
1536 return unique;
1537 }
1539 //------------------------------ register_node -------------------------------------
1540 // Utility to register node "n" with PhaseIdealLoop
1541 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
1542 _igvn.register_new_node_with_optimizer(n);
1543 loop->_body.push(n);
1544 if (n->is_CFG()) {
1545 set_loop(n, loop);
1546 set_idom(n, pred, ddepth);
1547 } else {
1548 set_ctrl(n, pred);
1549 }
1550 }
1552 //------------------------------ proj_clone -------------------------------------
1553 // Utility to create an if-projection
1554 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
1555 ProjNode* c = p->clone()->as_Proj();
1556 c->set_req(0, iff);
1557 return c;
1558 }
1560 //------------------------------ short_circuit_if -------------------------------------
1561 // Force the iff control output to be the live_proj
1562 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
1563 int proj_con = live_proj->_con;
1564 assert(proj_con == 0 || proj_con == 1, "false or true projection");
1565 Node *con = _igvn.intcon(proj_con);
1566 set_ctrl(con, C->root());
1567 if (iff) {
1568 iff->set_req(1, con);
1569 }
1570 return con;
1571 }
1573 //------------------------------ insert_if_before_proj -------------------------------------
1574 // Insert a new if before an if projection (* - new node)
1575 //
1576 // before
1577 // if(test)
1578 // / \
1579 // v v
1580 // other-proj proj (arg)
1581 //
1582 // after
1583 // if(test)
1584 // / \
1585 // / v
1586 // | * proj-clone
1587 // v |
1588 // other-proj v
1589 // * new_if(relop(cmp[IU](left,right)))
1590 // / \
1591 // v v
1592 // * new-proj proj
1593 // (returned)
1594 //
1595 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
1596 IfNode* iff = proj->in(0)->as_If();
1597 IdealLoopTree *loop = get_loop(proj);
1598 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1599 int ddepth = dom_depth(proj);
1601 _igvn.hash_delete(iff);
1602 _igvn._worklist.push(iff);
1603 _igvn.hash_delete(proj);
1604 _igvn._worklist.push(proj);
1606 proj->set_req(0, NULL); // temporary disconnect
1607 ProjNode* proj2 = proj_clone(proj, iff);
1608 register_node(proj2, loop, iff, ddepth);
1610 Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right);
1611 register_node(cmp, loop, proj2, ddepth);
1613 BoolNode* bol = new (C,2)BoolNode(cmp, relop);
1614 register_node(bol, loop, proj2, ddepth);
1616 IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt);
1617 register_node(new_if, loop, proj2, ddepth);
1619 proj->set_req(0, new_if); // reattach
1620 set_idom(proj, new_if, ddepth);
1622 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
1623 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
1625 return new_exit;
1626 }
1628 //------------------------------ insert_region_before_proj -------------------------------------
1629 // Insert a region before an if projection (* - new node)
1630 //
1631 // before
1632 // if(test)
1633 // / |
1634 // v |
1635 // proj v
1636 // other-proj
1637 //
1638 // after
1639 // if(test)
1640 // / |
1641 // v |
1642 // * proj-clone v
1643 // | other-proj
1644 // v
1645 // * new-region
1646 // |
1647 // v
1648 // * dum_if
1649 // / \
1650 // v \
1651 // * dum-proj v
1652 // proj
1653 //
1654 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
1655 IfNode* iff = proj->in(0)->as_If();
1656 IdealLoopTree *loop = get_loop(proj);
1657 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1658 int ddepth = dom_depth(proj);
1660 _igvn.hash_delete(iff);
1661 _igvn._worklist.push(iff);
1662 _igvn.hash_delete(proj);
1663 _igvn._worklist.push(proj);
1665 proj->set_req(0, NULL); // temporary disconnect
1666 ProjNode* proj2 = proj_clone(proj, iff);
1667 register_node(proj2, loop, iff, ddepth);
1669 RegionNode* reg = new (C,2)RegionNode(2);
1670 reg->set_req(1, proj2);
1671 register_node(reg, loop, iff, ddepth);
1673 IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
1674 register_node(dum_if, loop, reg, ddepth);
1676 proj->set_req(0, dum_if); // reattach
1677 set_idom(proj, dum_if, ddepth);
1679 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
1680 register_node(dum_proj, loop, dum_if, ddepth);
1682 return reg;
1683 }
1685 //------------------------------ insert_cmpi_loop_exit -------------------------------------
1686 // Clone a signed compare loop exit from an unsigned compare and
1687 // insert it before the unsigned cmp on the stay-in-loop path.
1688 // All new nodes inserted in the dominator tree between the original
1689 // if and it's projections. The original if test is replaced with
1690 // a constant to force the stay-in-loop path.
1691 //
1692 // This is done to make sure that the original if and it's projections
1693 // still dominate the same set of control nodes, that the ctrl() relation
1694 // from data nodes to them is preserved, and that their loop nesting is
1695 // preserved.
1696 //
1697 // before
1698 // if(i <u limit) unsigned compare loop exit
1699 // / |
1700 // v v
1701 // exit-proj stay-in-loop-proj
1702 //
1703 // after
1704 // if(stay-in-loop-const) original if
1705 // / |
1706 // / v
1707 // / if(i < limit) new signed test
1708 // / / |
1709 // / / v
1710 // / / if(i <u limit) new cloned unsigned test
1711 // / / / |
1712 // v v v |
1713 // region |
1714 // | |
1715 // dum-if |
1716 // / | |
1717 // ether | |
1718 // v v
1719 // exit-proj stay-in-loop-proj
1720 //
1721 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
1722 const bool Signed = true;
1723 const bool Unsigned = false;
1725 BoolNode* bol = if_cmpu->in(1)->as_Bool();
1726 if (bol->_test._test != BoolTest::lt) return NULL;
1727 CmpNode* cmpu = bol->in(1)->as_Cmp();
1728 if (cmpu->Opcode() != Op_CmpU) return NULL;
1729 int stride = stride_of_possible_iv(if_cmpu);
1730 if (stride == 0) return NULL;
1732 ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj();
1733 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
1735 Node* limit = NULL;
1736 if (stride > 0) {
1737 limit = cmpu->in(2);
1738 } else {
1739 limit = _igvn.makecon(TypeInt::ZERO);
1740 set_ctrl(limit, C->root());
1741 }
1742 // Create a new region on the exit path
1743 RegionNode* reg = insert_region_before_proj(lp_exit);
1745 // Clone the if-cmpu-true-false using a signed compare
1746 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
1747 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
1748 reg->add_req(cmpi_exit);
1750 // Clone the if-cmpu-true-false
1751 BoolTest::mask rel_u = bol->_test._test;
1752 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
1753 reg->add_req(cmpu_exit);
1755 // Force original if to stay in loop.
1756 short_circuit_if(if_cmpu, lp_continue);
1758 return cmpi_exit->in(0)->as_If();
1759 }
1761 //------------------------------ remove_cmpi_loop_exit -------------------------------------
1762 // Remove a previously inserted signed compare loop exit.
1763 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
1764 Node* lp_proj = stay_in_loop(if_cmp, loop);
1765 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
1766 stay_in_loop(lp_proj, loop)->is_If() &&
1767 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
1768 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
1769 set_ctrl(con, C->root());
1770 if_cmp->set_req(1, con);
1771 }
1773 //------------------------------ scheduled_nodelist -------------------------------------
1774 // Create a post order schedule of nodes that are in the
1775 // "member" set. The list is returned in "sched".
1776 // The first node in "sched" is the loop head, followed by
1777 // nodes which have no inputs in the "member" set, and then
1778 // followed by the nodes that have an immediate input dependence
1779 // on a node in "sched".
1780 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
1782 assert(member.test(loop->_head->_idx), "loop head must be in member set");
1783 Arena *a = Thread::current()->resource_area();
1784 VectorSet visited(a);
1785 Node_Stack nstack(a, loop->_body.size());
1787 Node* n = loop->_head; // top of stack is cached in "n"
1788 uint idx = 0;
1789 visited.set(n->_idx);
1791 // Initially push all with no inputs from within member set
1792 for(uint i = 0; i < loop->_body.size(); i++ ) {
1793 Node *elt = loop->_body.at(i);
1794 if (member.test(elt->_idx)) {
1795 bool found = false;
1796 for (uint j = 0; j < elt->req(); j++) {
1797 Node* def = elt->in(j);
1798 if (def && member.test(def->_idx) && def != elt) {
1799 found = true;
1800 break;
1801 }
1802 }
1803 if (!found && elt != loop->_head) {
1804 nstack.push(n, idx);
1805 n = elt;
1806 assert(!visited.test(n->_idx), "not seen yet");
1807 visited.set(n->_idx);
1808 }
1809 }
1810 }
1812 // traverse out's that are in the member set
1813 while (true) {
1814 if (idx < n->outcnt()) {
1815 Node* use = n->raw_out(idx);
1816 idx++;
1817 if (!visited.test_set(use->_idx)) {
1818 if (member.test(use->_idx)) {
1819 nstack.push(n, idx);
1820 n = use;
1821 idx = 0;
1822 }
1823 }
1824 } else {
1825 // All outputs processed
1826 sched.push(n);
1827 if (nstack.is_empty()) break;
1828 n = nstack.node();
1829 idx = nstack.index();
1830 nstack.pop();
1831 }
1832 }
1833 }
1836 //------------------------------ has_use_in_set -------------------------------------
1837 // Has a use in the vector set
1838 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
1839 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1840 Node* use = n->fast_out(j);
1841 if (vset.test(use->_idx)) {
1842 return true;
1843 }
1844 }
1845 return false;
1846 }
1849 //------------------------------ has_use_internal_to_set -------------------------------------
1850 // Has use internal to the vector set (ie. not in a phi at the loop head)
1851 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
1852 Node* head = loop->_head;
1853 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1854 Node* use = n->fast_out(j);
1855 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
1856 return true;
1857 }
1858 }
1859 return false;
1860 }
1863 //------------------------------ clone_for_use_outside_loop -------------------------------------
1864 // clone "n" for uses that are outside of loop
1865 void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
1867 assert(worklist.size() == 0, "should be empty");
1868 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1869 Node* use = n->fast_out(j);
1870 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
1871 worklist.push(use);
1872 }
1873 }
1874 while( worklist.size() ) {
1875 Node *use = worklist.pop();
1876 if (!has_node(use) || use->in(0) == C->top()) continue;
1877 uint j;
1878 for (j = 0; j < use->req(); j++) {
1879 if (use->in(j) == n) break;
1880 }
1881 assert(j < use->req(), "must be there");
1883 // clone "n" and insert it between the inputs of "n" and the use outside the loop
1884 Node* n_clone = n->clone();
1885 _igvn.hash_delete(use);
1886 use->set_req(j, n_clone);
1887 _igvn._worklist.push(use);
1888 if (!use->is_Phi()) {
1889 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
1890 set_ctrl(n_clone, use_c);
1891 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
1892 get_loop(use_c)->_body.push(n_clone);
1893 } else {
1894 // Use in a phi is considered a use in the associated predecessor block
1895 Node *prevbb = use->in(0)->in(j);
1896 set_ctrl(n_clone, prevbb);
1897 assert(!loop->is_member(get_loop(prevbb)), "should be outside loop");
1898 get_loop(prevbb)->_body.push(n_clone);
1899 }
1900 _igvn.register_new_node_with_optimizer(n_clone);
1901 #if !defined(PRODUCT)
1902 if (TracePartialPeeling) {
1903 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
1904 }
1905 #endif
1906 }
1907 }
1910 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
1911 // clone "n" for special uses that are in the not_peeled region.
1912 // If these def-uses occur in separate blocks, the code generator
1913 // marks the method as not compilable. For example, if a "BoolNode"
1914 // is in a different basic block than the "IfNode" that uses it, then
1915 // the compilation is aborted in the code generator.
1916 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1917 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
1918 if (n->is_Phi() || n->is_Load()) {
1919 return;
1920 }
1921 assert(worklist.size() == 0, "should be empty");
1922 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1923 Node* use = n->fast_out(j);
1924 if ( not_peel.test(use->_idx) &&
1925 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
1926 use->in(1) == n) {
1927 worklist.push(use);
1928 }
1929 }
1930 if (worklist.size() > 0) {
1931 // clone "n" and insert it between inputs of "n" and the use
1932 Node* n_clone = n->clone();
1933 loop->_body.push(n_clone);
1934 _igvn.register_new_node_with_optimizer(n_clone);
1935 set_ctrl(n_clone, get_ctrl(n));
1936 sink_list.push(n_clone);
1937 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
1938 #if !defined(PRODUCT)
1939 if (TracePartialPeeling) {
1940 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
1941 }
1942 #endif
1943 while( worklist.size() ) {
1944 Node *use = worklist.pop();
1945 _igvn.hash_delete(use);
1946 _igvn._worklist.push(use);
1947 for (uint j = 1; j < use->req(); j++) {
1948 if (use->in(j) == n) {
1949 use->set_req(j, n_clone);
1950 }
1951 }
1952 }
1953 }
1954 }
1957 //------------------------------ insert_phi_for_loop -------------------------------------
1958 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1959 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
1960 Node *phi = PhiNode::make(lp, back_edge_val);
1961 phi->set_req(LoopNode::EntryControl, lp_entry_val);
1962 // Use existing phi if it already exists
1963 Node *hit = _igvn.hash_find_insert(phi);
1964 if( hit == NULL ) {
1965 _igvn.register_new_node_with_optimizer(phi);
1966 set_ctrl(phi, lp);
1967 } else {
1968 // Remove the new phi from the graph and use the hit
1969 _igvn.remove_dead_node(phi);
1970 phi = hit;
1971 }
1972 _igvn.hash_delete(use);
1973 _igvn._worklist.push(use);
1974 use->set_req(idx, phi);
1975 }
1977 #ifdef ASSERT
1978 //------------------------------ is_valid_loop_partition -------------------------------------
1979 // Validate the loop partition sets: peel and not_peel
1980 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
1981 VectorSet& not_peel ) {
1982 uint i;
1983 // Check that peel_list entries are in the peel set
1984 for (i = 0; i < peel_list.size(); i++) {
1985 if (!peel.test(peel_list.at(i)->_idx)) {
1986 return false;
1987 }
1988 }
1989 // Check at loop members are in one of peel set or not_peel set
1990 for (i = 0; i < loop->_body.size(); i++ ) {
1991 Node *def = loop->_body.at(i);
1992 uint di = def->_idx;
1993 // Check that peel set elements are in peel_list
1994 if (peel.test(di)) {
1995 if (not_peel.test(di)) {
1996 return false;
1997 }
1998 // Must be in peel_list also
1999 bool found = false;
2000 for (uint j = 0; j < peel_list.size(); j++) {
2001 if (peel_list.at(j)->_idx == di) {
2002 found = true;
2003 break;
2004 }
2005 }
2006 if (!found) {
2007 return false;
2008 }
2009 } else if (not_peel.test(di)) {
2010 if (peel.test(di)) {
2011 return false;
2012 }
2013 } else {
2014 return false;
2015 }
2016 }
2017 return true;
2018 }
2020 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2021 // Ensure a use outside of loop is of the right form
2022 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2023 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2024 return (use->is_Phi() &&
2025 use_c->is_Region() && use_c->req() == 3 &&
2026 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2027 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2028 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2029 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2030 }
2032 //------------------------------ is_valid_clone_loop_form -------------------------------------
2033 // Ensure that all uses outside of loop are of the right form
2034 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2035 uint orig_exit_idx, uint clone_exit_idx) {
2036 uint len = peel_list.size();
2037 for (uint i = 0; i < len; i++) {
2038 Node *def = peel_list.at(i);
2040 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2041 Node *use = def->fast_out(j);
2042 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2043 if (!loop->is_member(get_loop(use_c))) {
2044 // use is not in the loop, check for correct structure
2045 if (use->in(0) == def) {
2046 // Okay
2047 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2048 return false;
2049 }
2050 }
2051 }
2052 }
2053 return true;
2054 }
2055 #endif
2057 //------------------------------ partial_peel -------------------------------------
2058 // Partially peel (aka loop rotation) the top portion of a loop (called
2059 // the peel section below) by cloning it and placing one copy just before
2060 // the new loop head and the other copy at the bottom of the new loop.
2061 //
2062 // before after where it came from
2063 //
2064 // stmt1 stmt1
2065 // loop: stmt2 clone
2066 // stmt2 if condA goto exitA clone
2067 // if condA goto exitA new_loop: new
2068 // stmt3 stmt3 clone
2069 // if !condB goto loop if condB goto exitB clone
2070 // exitB: stmt2 orig
2071 // stmt4 if !condA goto new_loop orig
2072 // exitA: goto exitA
2073 // exitB:
2074 // stmt4
2075 // exitA:
2076 //
2077 // Step 1: find the cut point: an exit test on probable
2078 // induction variable.
2079 // Step 2: schedule (with cloning) operations in the peel
2080 // section that can be executed after the cut into
2081 // the section that is not peeled. This may need
2082 // to clone operations into exit blocks. For
2083 // instance, a reference to A[i] in the not-peel
2084 // section and a reference to B[i] in an exit block
2085 // may cause a left-shift of i by 2 to be placed
2086 // in the peel block. This step will clone the left
2087 // shift into the exit block and sink the left shift
2088 // from the peel to the not-peel section.
2089 // Step 3: clone the loop, retarget the control, and insert
2090 // phis for values that are live across the new loop
2091 // head. This is very dependent on the graph structure
2092 // from clone_loop. It creates region nodes for
2093 // exit control and associated phi nodes for values
2094 // flow out of the loop through that exit. The region
2095 // node is dominated by the clone's control projection.
2096 // So the clone's peel section is placed before the
2097 // new loop head, and the clone's not-peel section is
2098 // forms the top part of the new loop. The original
2099 // peel section forms the tail of the new loop.
2100 // Step 4: update the dominator tree and recompute the
2101 // dominator depth.
2102 //
2103 // orig
2104 //
2105 // stmt1
2106 // |
2107 // v
2108 // loop<----+
2109 // | |
2110 // stmt2 |
2111 // | |
2112 // v |
2113 // ifA |
2114 // / | |
2115 // v v |
2116 // false true ^ <-- last_peel
2117 // / | |
2118 // / ===|==cut |
2119 // / stmt3 | <-- first_not_peel
2120 // / | |
2121 // | v |
2122 // v ifB |
2123 // exitA: / \ |
2124 // / \ |
2125 // v v |
2126 // false true |
2127 // / \ |
2128 // / ----+
2129 // |
2130 // v
2131 // exitB:
2132 // stmt4
2133 //
2134 //
2135 // after clone loop
2136 //
2137 // stmt1
2138 // / \
2139 // clone / \ orig
2140 // / \
2141 // / \
2142 // v v
2143 // +---->loop loop<----+
2144 // | | | |
2145 // | stmt2 stmt2 |
2146 // | | | |
2147 // | v v |
2148 // | ifA ifA |
2149 // | | \ / | |
2150 // | v v v v |
2151 // ^ true false false true ^ <-- last_peel
2152 // | | ^ \ / | |
2153 // | cut==|== \ \ / ===|==cut |
2154 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2155 // | | dom | | | |
2156 // | v \ 1v v2 v |
2157 // | ifB regionA ifB |
2158 // | / \ | / \ |
2159 // | / \ v / \ |
2160 // | v v exitA: v v |
2161 // | true false false true |
2162 // | / ^ \ / \ |
2163 // +---- \ \ / ----+
2164 // dom \ /
2165 // \ 1v v2
2166 // regionB
2167 // |
2168 // v
2169 // exitB:
2170 // stmt4
2171 //
2172 //
2173 // after partial peel
2174 //
2175 // stmt1
2176 // /
2177 // clone / orig
2178 // / TOP
2179 // / \
2180 // v v
2181 // TOP->region region----+
2182 // | | |
2183 // stmt2 stmt2 |
2184 // | | |
2185 // v v |
2186 // ifA ifA |
2187 // | \ / | |
2188 // v v v v |
2189 // true false false true | <-- last_peel
2190 // | ^ \ / +------|---+
2191 // +->newloop \ \ / === ==cut | |
2192 // | stmt3 \ \ / TOP | |
2193 // | | dom | | stmt3 | | <-- first_not_peel
2194 // | v \ 1v v2 v | |
2195 // | ifB regionA ifB ^ v
2196 // | / \ | / \ | |
2197 // | / \ v / \ | |
2198 // | v v exitA: v v | |
2199 // | true false false true | |
2200 // | / ^ \ / \ | |
2201 // | | \ \ / v | |
2202 // | | dom \ / TOP | |
2203 // | | \ 1v v2 | |
2204 // ^ v regionB | |
2205 // | | | | |
2206 // | | v ^ v
2207 // | | exitB: | |
2208 // | | stmt4 | |
2209 // | +------------>-----------------+ |
2210 // | |
2211 // +-----------------<---------------------+
2212 //
2213 //
2214 // final graph
2215 //
2216 // stmt1
2217 // |
2218 // v
2219 // ........> ifA clone
2220 // : / |
2221 // dom / |
2222 // : v v
2223 // : false true
2224 // : | |
2225 // : | stmt2 clone
2226 // : | |
2227 // : | v
2228 // : | newloop<-----+
2229 // : | | |
2230 // : | stmt3 clone |
2231 // : | | |
2232 // : | v |
2233 // : | ifB |
2234 // : | / \ |
2235 // : | v v |
2236 // : | false true |
2237 // : | | | |
2238 // : | v stmt2 |
2239 // : | exitB: | |
2240 // : | stmt4 v |
2241 // : | ifA orig |
2242 // : | / \ |
2243 // : | / \ |
2244 // : | v v |
2245 // : | false true |
2246 // : | / \ |
2247 // : v v -----+
2248 // RegionA
2249 // |
2250 // v
2251 // exitA
2252 //
2253 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
2255 LoopNode *head = loop->_head->as_Loop();
2257 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
2258 return false;
2259 }
2261 // Check for complex exit control
2262 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
2263 Node *n = loop->_body.at(ii);
2264 int opc = n->Opcode();
2265 if (n->is_Call() ||
2266 opc == Op_Catch ||
2267 opc == Op_CatchProj ||
2268 opc == Op_Jump ||
2269 opc == Op_JumpProj) {
2270 #if !defined(PRODUCT)
2271 if (TracePartialPeeling) {
2272 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
2273 }
2274 #endif
2275 return false;
2276 }
2277 }
2279 int dd = dom_depth(head);
2281 // Step 1: find cut point
2283 // Walk up dominators to loop head looking for first loop exit
2284 // which is executed on every path thru loop.
2285 IfNode *peel_if = NULL;
2286 IfNode *peel_if_cmpu = NULL;
2288 Node *iff = loop->tail();
2289 while( iff != head ) {
2290 if( iff->is_If() ) {
2291 Node *ctrl = get_ctrl(iff->in(1));
2292 if (ctrl->is_top()) return false; // Dead test on live IF.
2293 // If loop-varying exit-test, check for induction variable
2294 if( loop->is_member(get_loop(ctrl)) &&
2295 loop->is_loop_exit(iff) &&
2296 is_possible_iv_test(iff)) {
2297 Node* cmp = iff->in(1)->in(1);
2298 if (cmp->Opcode() == Op_CmpI) {
2299 peel_if = iff->as_If();
2300 } else {
2301 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
2302 peel_if_cmpu = iff->as_If();
2303 }
2304 }
2305 }
2306 iff = idom(iff);
2307 }
2308 // Prefer signed compare over unsigned compare.
2309 IfNode* new_peel_if = NULL;
2310 if (peel_if == NULL) {
2311 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
2312 return false; // No peel point found
2313 }
2314 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
2315 if (new_peel_if == NULL) {
2316 return false; // No peel point found
2317 }
2318 peel_if = new_peel_if;
2319 }
2320 Node* last_peel = stay_in_loop(peel_if, loop);
2321 Node* first_not_peeled = stay_in_loop(last_peel, loop);
2322 if (first_not_peeled == NULL || first_not_peeled == head) {
2323 return false;
2324 }
2326 #if !defined(PRODUCT)
2327 if (TracePartialPeeling) {
2328 tty->print_cr("before partial peel one iteration");
2329 Node_List wl;
2330 Node* t = head->in(2);
2331 while (true) {
2332 wl.push(t);
2333 if (t == head) break;
2334 t = idom(t);
2335 }
2336 while (wl.size() > 0) {
2337 Node* tt = wl.pop();
2338 tt->dump();
2339 if (tt == last_peel) tty->print_cr("-- cut --");
2340 }
2341 }
2342 #endif
2343 ResourceArea *area = Thread::current()->resource_area();
2344 VectorSet peel(area);
2345 VectorSet not_peel(area);
2346 Node_List peel_list(area);
2347 Node_List worklist(area);
2348 Node_List sink_list(area);
2350 // Set of cfg nodes to peel are those that are executable from
2351 // the head through last_peel.
2352 assert(worklist.size() == 0, "should be empty");
2353 worklist.push(head);
2354 peel.set(head->_idx);
2355 while (worklist.size() > 0) {
2356 Node *n = worklist.pop();
2357 if (n != last_peel) {
2358 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2359 Node* use = n->fast_out(j);
2360 if (use->is_CFG() &&
2361 loop->is_member(get_loop(use)) &&
2362 !peel.test_set(use->_idx)) {
2363 worklist.push(use);
2364 }
2365 }
2366 }
2367 }
2369 // Set of non-cfg nodes to peel are those that are control
2370 // dependent on the cfg nodes.
2371 uint i;
2372 for(i = 0; i < loop->_body.size(); i++ ) {
2373 Node *n = loop->_body.at(i);
2374 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
2375 if (peel.test(n_c->_idx)) {
2376 peel.set(n->_idx);
2377 } else {
2378 not_peel.set(n->_idx);
2379 }
2380 }
2382 // Step 2: move operations from the peeled section down into the
2383 // not-peeled section
2385 // Get a post order schedule of nodes in the peel region
2386 // Result in right-most operand.
2387 scheduled_nodelist(loop, peel, peel_list );
2389 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2391 // For future check for too many new phis
2392 uint old_phi_cnt = 0;
2393 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2394 Node* use = head->fast_out(j);
2395 if (use->is_Phi()) old_phi_cnt++;
2396 }
2398 #if !defined(PRODUCT)
2399 if (TracePartialPeeling) {
2400 tty->print_cr("\npeeled list");
2401 }
2402 #endif
2404 // Evacuate nodes in peel region into the not_peeled region if possible
2405 uint new_phi_cnt = 0;
2406 for (i = 0; i < peel_list.size();) {
2407 Node* n = peel_list.at(i);
2408 #if !defined(PRODUCT)
2409 if (TracePartialPeeling) n->dump();
2410 #endif
2411 bool incr = true;
2412 if ( !n->is_CFG() ) {
2414 if ( has_use_in_set(n, not_peel) ) {
2416 // If not used internal to the peeled region,
2417 // move "n" from peeled to not_peeled region.
2419 if ( !has_use_internal_to_set(n, peel, loop) ) {
2421 // if not pinned and not a load (which maybe anti-dependent on a store)
2422 // and not a CMove (Matcher expects only bool->cmove).
2423 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
2424 clone_for_use_outside_loop( loop, n, worklist );
2426 sink_list.push(n);
2427 peel >>= n->_idx; // delete n from peel set.
2428 not_peel <<= n->_idx; // add n to not_peel set.
2429 peel_list.remove(i);
2430 incr = false;
2431 #if !defined(PRODUCT)
2432 if (TracePartialPeeling) {
2433 tty->print_cr("sink to not_peeled region: %d newbb: %d",
2434 n->_idx, get_ctrl(n)->_idx);
2435 }
2436 #endif
2437 }
2438 } else {
2439 // Otherwise check for special def-use cases that span
2440 // the peel/not_peel boundary such as bool->if
2441 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
2442 new_phi_cnt++;
2443 }
2444 }
2445 }
2446 if (incr) i++;
2447 }
2449 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
2450 #if !defined(PRODUCT)
2451 if (TracePartialPeeling) {
2452 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
2453 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
2454 }
2455 #endif
2456 if (new_peel_if != NULL) {
2457 remove_cmpi_loop_exit(new_peel_if, loop);
2458 }
2459 // Inhibit more partial peeling on this loop
2460 assert(!head->is_partial_peel_loop(), "not partial peeled");
2461 head->mark_partial_peel_failed();
2462 return false;
2463 }
2465 // Step 3: clone loop, retarget control, and insert new phis
2467 // Create new loop head for new phis and to hang
2468 // the nodes being moved (sinked) from the peel region.
2469 LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel);
2470 _igvn.register_new_node_with_optimizer(new_head);
2471 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
2472 first_not_peeled->set_req(0, new_head);
2473 set_loop(new_head, loop);
2474 loop->_body.push(new_head);
2475 not_peel.set(new_head->_idx);
2476 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
2477 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
2479 while (sink_list.size() > 0) {
2480 Node* n = sink_list.pop();
2481 set_ctrl(n, new_head);
2482 }
2484 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2486 clone_loop( loop, old_new, dd );
2488 const uint clone_exit_idx = 1;
2489 const uint orig_exit_idx = 2;
2490 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
2492 Node* head_clone = old_new[head->_idx];
2493 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
2494 Node* orig_tail_clone = head_clone->in(2);
2496 // Add phi if "def" node is in peel set and "use" is not
2498 for(i = 0; i < peel_list.size(); i++ ) {
2499 Node *def = peel_list.at(i);
2500 if (!def->is_CFG()) {
2501 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2502 Node *use = def->fast_out(j);
2503 if (has_node(use) && use->in(0) != C->top() &&
2504 (!peel.test(use->_idx) ||
2505 (use->is_Phi() && use->in(0) == head)) ) {
2506 worklist.push(use);
2507 }
2508 }
2509 while( worklist.size() ) {
2510 Node *use = worklist.pop();
2511 for (uint j = 1; j < use->req(); j++) {
2512 Node* n = use->in(j);
2513 if (n == def) {
2515 // "def" is in peel set, "use" is not in peel set
2516 // or "use" is in the entry boundary (a phi) of the peel set
2518 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
2520 if ( loop->is_member(get_loop( use_c )) ) {
2521 // use is in loop
2522 if (old_new[use->_idx] != NULL) { // null for dead code
2523 Node* use_clone = old_new[use->_idx];
2524 _igvn.hash_delete(use);
2525 use->set_req(j, C->top());
2526 _igvn._worklist.push(use);
2527 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
2528 }
2529 } else {
2530 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
2531 // use is not in the loop, check if the live range includes the cut
2532 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
2533 if (not_peel.test(lp_if->_idx)) {
2534 assert(j == orig_exit_idx, "use from original loop");
2535 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
2536 }
2537 }
2538 }
2539 }
2540 }
2541 }
2542 }
2544 // Step 3b: retarget control
2546 // Redirect control to the new loop head if a cloned node in
2547 // the not_peeled region has control that points into the peeled region.
2548 // This necessary because the cloned peeled region will be outside
2549 // the loop.
2550 // from to
2551 // cloned-peeled <---+
2552 // new_head_clone: | <--+
2553 // cloned-not_peeled in(0) in(0)
2554 // orig-peeled
2556 for(i = 0; i < loop->_body.size(); i++ ) {
2557 Node *n = loop->_body.at(i);
2558 if (!n->is_CFG() && n->in(0) != NULL &&
2559 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
2560 Node* n_clone = old_new[n->_idx];
2561 _igvn.hash_delete(n_clone);
2562 n_clone->set_req(0, new_head_clone);
2563 _igvn._worklist.push(n_clone);
2564 }
2565 }
2567 // Backedge of the surviving new_head (the clone) is original last_peel
2568 _igvn.hash_delete(new_head_clone);
2569 new_head_clone->set_req(LoopNode::LoopBackControl, last_peel);
2570 _igvn._worklist.push(new_head_clone);
2572 // Cut first node in original not_peel set
2573 _igvn.hash_delete(new_head);
2574 new_head->set_req(LoopNode::EntryControl, C->top());
2575 new_head->set_req(LoopNode::LoopBackControl, C->top());
2576 _igvn._worklist.push(new_head);
2578 // Copy head_clone back-branch info to original head
2579 // and remove original head's loop entry and
2580 // clone head's back-branch
2581 _igvn.hash_delete(head);
2582 _igvn.hash_delete(head_clone);
2583 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
2584 head->set_req(LoopNode::LoopBackControl, C->top());
2585 head_clone->set_req(LoopNode::LoopBackControl, C->top());
2586 _igvn._worklist.push(head);
2587 _igvn._worklist.push(head_clone);
2589 // Similarly modify the phis
2590 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
2591 Node* use = head->fast_out(k);
2592 if (use->is_Phi() && use->outcnt() > 0) {
2593 Node* use_clone = old_new[use->_idx];
2594 _igvn.hash_delete(use);
2595 _igvn.hash_delete(use_clone);
2596 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
2597 use->set_req(LoopNode::LoopBackControl, C->top());
2598 use_clone->set_req(LoopNode::LoopBackControl, C->top());
2599 _igvn._worklist.push(use);
2600 _igvn._worklist.push(use_clone);
2601 }
2602 }
2604 // Step 4: update dominator tree and dominator depth
2606 set_idom(head, orig_tail_clone, dd);
2607 recompute_dom_depth();
2609 // Inhibit more partial peeling on this loop
2610 new_head_clone->set_partial_peel_loop();
2611 C->set_major_progress();
2613 #if !defined(PRODUCT)
2614 if (TracePartialPeeling) {
2615 tty->print_cr("\nafter partial peel one iteration");
2616 Node_List wl(area);
2617 Node* t = last_peel;
2618 while (true) {
2619 wl.push(t);
2620 if (t == head_clone) break;
2621 t = idom(t);
2622 }
2623 while (wl.size() > 0) {
2624 Node* tt = wl.pop();
2625 if (tt == head) tty->print_cr("orig head");
2626 else if (tt == new_head_clone) tty->print_cr("new head");
2627 else if (tt == head_clone) tty->print_cr("clone head");
2628 tt->dump();
2629 }
2630 }
2631 #endif
2632 return true;
2633 }
2635 //------------------------------reorg_offsets----------------------------------
2636 // Reorganize offset computations to lower register pressure. Mostly
2637 // prevent loop-fallout uses of the pre-incremented trip counter (which are
2638 // then alive with the post-incremented trip counter forcing an extra
2639 // register move)
2640 void PhaseIdealLoop::reorg_offsets( IdealLoopTree *loop ) {
2642 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2643 CountedLoopEndNode *cle = cl->loopexit();
2644 if( !cle ) return; // The occasional dead loop
2645 // Find loop exit control
2646 Node *exit = cle->proj_out(false);
2647 assert( exit->Opcode() == Op_IfFalse, "" );
2649 // Check for the special case of folks using the pre-incremented
2650 // trip-counter on the fall-out path (forces the pre-incremented
2651 // and post-incremented trip counter to be live at the same time).
2652 // Fix this by adjusting to use the post-increment trip counter.
2653 Node *phi = cl->phi();
2654 if( !phi ) return; // Dead infinite loop
2655 bool progress = true;
2656 while (progress) {
2657 progress = false;
2658 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
2659 Node* use = phi->fast_out(i); // User of trip-counter
2660 if (!has_ctrl(use)) continue;
2661 Node *u_ctrl = get_ctrl(use);
2662 if( use->is_Phi() ) {
2663 u_ctrl = NULL;
2664 for( uint j = 1; j < use->req(); j++ )
2665 if( use->in(j) == phi )
2666 u_ctrl = dom_lca( u_ctrl, use->in(0)->in(j) );
2667 }
2668 IdealLoopTree *u_loop = get_loop(u_ctrl);
2669 // Look for loop-invariant use
2670 if( u_loop == loop ) continue;
2671 if( loop->is_member( u_loop ) ) continue;
2672 // Check that use is live out the bottom. Assuming the trip-counter
2673 // update is right at the bottom, uses of of the loop middle are ok.
2674 if( dom_lca( exit, u_ctrl ) != exit ) continue;
2675 // protect against stride not being a constant
2676 if( !cle->stride_is_con() ) continue;
2677 // Hit! Refactor use to use the post-incremented tripcounter.
2678 // Compute a post-increment tripcounter.
2679 Node *opaq = new (C, 2) Opaque2Node( cle->incr() );
2680 register_new_node( opaq, u_ctrl );
2681 Node *neg_stride = _igvn.intcon(-cle->stride_con());
2682 set_ctrl(neg_stride, C->root());
2683 Node *post = new (C, 3) AddINode( opaq, neg_stride);
2684 register_new_node( post, u_ctrl );
2685 _igvn.hash_delete(use);
2686 _igvn._worklist.push(use);
2687 for( uint j = 1; j < use->req(); j++ )
2688 if( use->in(j) == phi )
2689 use->set_req(j, post);
2690 // Since DU info changed, rerun loop
2691 progress = true;
2692 break;
2693 }
2694 }
2696 }