1 #ifdef USE_PRAGMA_IDENT_SRC
2 #pragma ident "@(#)loopopts.cpp 1.222 08/11/24 12:23:09 JVM"
3 #endif
4 /*
5 * Copyright 1999-2006 Sun Microsystems, Inc. All Rights Reserved.
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7 *
8 * This code is free software; you can redistribute it and/or modify it
9 * under the terms of the GNU General Public License version 2 only, as
10 * published by the Free Software Foundation.
11 *
12 * This code is distributed in the hope that it will be useful, but WITHOUT
13 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 * version 2 for more details (a copy is included in the LICENSE file that
16 * accompanied this code).
17 *
18 * You should have received a copy of the GNU General Public License version
19 * 2 along with this work; if not, write to the Free Software Foundation,
20 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
21 *
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
24 * have any questions.
25 *
26 */
27
28 #include "incls/_precompiled.incl"
29 #include "incls/_loopopts.cpp.incl"
30
31 //=============================================================================
32 //------------------------------split_thru_phi---------------------------------
33 // Split Node 'n' through merge point if there is enough win.
34 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
35 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::BOTTOM) {
36 // ConvI2L may have type information on it which is unsafe to push up
37 // so disable this for now
38 return NULL;
39 }
40 int wins = 0;
41 assert( !n->is_CFG(), "" );
42 assert( region->is_Region(), "" );
43 Node *phi = new (C, region->req()) PhiNode( region, n->bottom_type() );
44 uint old_unique = C->unique();
45 for( uint i = 1; i < region->req(); i++ ) {
46 Node *x;
47 Node* the_clone = NULL;
48 if( region->in(i) == C->top() ) {
49 x = C->top(); // Dead path? Use a dead data op
50 } else {
51 x = n->clone(); // Else clone up the data op
52 the_clone = x; // Remember for possible deletion.
53 // Alter data node to use pre-phi inputs
54 if( n->in(0) == region )
55 x->set_req( 0, region->in(i) );
56 for( uint j = 1; j < n->req(); j++ ) {
57 Node *in = n->in(j);
58 if( in->is_Phi() && in->in(0) == region )
59 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
60 }
61 }
62 // Check for a 'win' on some paths
63 const Type *t = x->Value(&_igvn);
64
65 bool singleton = t->singleton();
66
67 // A TOP singleton indicates that there are no possible values incoming
68 // along a particular edge. In most cases, this is OK, and the Phi will
69 // be eliminated later in an Ideal call. However, we can't allow this to
70 // happen if the singleton occurs on loop entry, as the elimination of
71 // the PhiNode may cause the resulting node to migrate back to a previous
72 // loop iteration.
73 if( singleton && t == Type::TOP ) {
74 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
75 // irreducible loop may not be indicated by an affirmative is_Loop());
76 // therefore, the only top we can split thru a phi is on a backedge of
77 // a loop.
78 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
79 }
80
81 if( singleton ) {
82 wins++;
83 x = ((PhaseGVN&)_igvn).makecon(t);
84 } else {
85 // We now call Identity to try to simplify the cloned node.
86 // Note that some Identity methods call phase->type(this).
87 // Make sure that the type array is big enough for
88 // our new node, even though we may throw the node away.
89 // (Note: This tweaking with igvn only works because x is a new node.)
90 _igvn.set_type(x, t);
91 Node *y = x->Identity(&_igvn);
92 if( y != x ) {
93 wins++;
94 x = y;
95 } else {
96 y = _igvn.hash_find(x);
97 if( y ) {
98 wins++;
99 x = y;
100 } else {
101 // Else x is a new node we are keeping
102 // We do not need register_new_node_with_optimizer
103 // because set_type has already been called.
104 _igvn._worklist.push(x);
105 }
106 }
107 }
108 if (x != the_clone && the_clone != NULL)
109 _igvn.remove_dead_node(the_clone);
110 phi->set_req( i, x );
111 }
112 // Too few wins?
113 if( wins <= policy ) {
114 _igvn.remove_dead_node(phi);
115 return NULL;
116 }
117
118 // Record Phi
119 register_new_node( phi, region );
120
121 for( uint i2 = 1; i2 < phi->req(); i2++ ) {
122 Node *x = phi->in(i2);
123 // If we commoned up the cloned 'x' with another existing Node,
124 // the existing Node picks up a new use. We need to make the
125 // existing Node occur higher up so it dominates its uses.
126 Node *old_ctrl;
127 IdealLoopTree *old_loop;
128
129 // The occasional new node
130 if( x->_idx >= old_unique ) { // Found a new, unplaced node?
131 old_ctrl = x->is_Con() ? C->root() : NULL;
132 old_loop = NULL; // Not in any prior loop
133 } else {
134 old_ctrl = x->is_Con() ? C->root() : get_ctrl(x);
135 old_loop = get_loop(old_ctrl); // Get prior loop
136 }
137 // New late point must dominate new use
138 Node *new_ctrl = dom_lca( old_ctrl, region->in(i2) );
139 // Set new location
140 set_ctrl(x, new_ctrl);
141 IdealLoopTree *new_loop = get_loop( new_ctrl );
142 // If changing loop bodies, see if we need to collect into new body
143 if( old_loop != new_loop ) {
144 if( old_loop && !old_loop->_child )
145 old_loop->_body.yank(x);
146 if( !new_loop->_child )
147 new_loop->_body.push(x); // Collect body info
148 }
149 }
150
151 return phi;
152 }
153
154 //------------------------------dominated_by------------------------------------
155 // Replace the dominated test with an obvious true or false. Place it on the
156 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
157 // live path up to the dominating control.
158 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff ) {
159 #ifndef PRODUCT
160 if( VerifyLoopOptimizations && PrintOpto ) tty->print_cr("dominating test");
161 #endif
162
163
164 // prevdom is the dominating projection of the dominating test.
165 assert( iff->is_If(), "" );
166 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
167 int pop = prevdom->Opcode();
168 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
169 // 'con' is set to true or false to kill the dominated test.
170 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
171 set_ctrl(con, C->root()); // Constant gets a new use
172 // Hack the dominated test
173 _igvn.hash_delete(iff);
174 iff->set_req(1, con);
175 _igvn._worklist.push(iff);
176
177 // If I dont have a reachable TRUE and FALSE path following the IfNode then
178 // I can assume this path reaches an infinite loop. In this case it's not
179 // important to optimize the data Nodes - either the whole compilation will
180 // be tossed or this path (and all data Nodes) will go dead.
181 if( iff->outcnt() != 2 ) return;
182
183 // Make control-dependent data Nodes on the live path (path that will remain
184 // once the dominated IF is removed) become control-dependent on the
185 // dominating projection.
186 Node* dp = ((IfNode*)iff)->proj_out(pop == Op_IfTrue);
187 IdealLoopTree *old_loop = get_loop(dp);
188
189 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
190 Node* cd = dp->fast_out(i); // Control-dependent node
191 if( cd->depends_only_on_test() ) {
192 assert( cd->in(0) == dp, "" );
193 _igvn.hash_delete( cd );
194 cd->set_req(0, prevdom);
195 set_early_ctrl( cd );
196 _igvn._worklist.push(cd);
197 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
198 if( old_loop != new_loop ) {
199 if( !old_loop->_child ) old_loop->_body.yank(cd);
200 if( !new_loop->_child ) new_loop->_body.push(cd);
201 }
202 --i;
203 --imax;
204 }
205 }
206 }
207
208 //------------------------------has_local_phi_input----------------------------
209 // Return TRUE if 'n' has Phi inputs from its local block and no other
210 // block-local inputs (all non-local-phi inputs come from earlier blocks)
211 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
212 Node *n_ctrl = get_ctrl(n);
213 // See if some inputs come from a Phi in this block, or from before
214 // this block.
215 uint i;
216 for( i = 1; i < n->req(); i++ ) {
217 Node *phi = n->in(i);
218 if( phi->is_Phi() && phi->in(0) == n_ctrl )
219 break;
220 }
221 if( i >= n->req() )
222 return NULL; // No Phi inputs; nowhere to clone thru
223
224 // Check for inputs created between 'n' and the Phi input. These
225 // must split as well; they have already been given the chance
226 // (courtesy of a post-order visit) and since they did not we must
227 // recover the 'cost' of splitting them by being very profitable
228 // when splitting 'n'. Since this is unlikely we simply give up.
229 for( i = 1; i < n->req(); i++ ) {
230 Node *m = n->in(i);
231 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
232 // We allow the special case of AddP's with no local inputs.
233 // This allows us to split-up address expressions.
234 if (m->is_AddP() &&
235 get_ctrl(m->in(2)) != n_ctrl &&
236 get_ctrl(m->in(3)) != n_ctrl) {
237 // Move the AddP up to dominating point
238 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl)));
239 continue;
240 }
241 return NULL;
242 }
243 }
244
245 return n_ctrl;
246 }
247
248 //------------------------------remix_address_expressions----------------------
249 // Rework addressing expressions to get the most loop-invariant stuff
250 // moved out. We'd like to do all associative operators, but it's especially
251 // important (common) to do address expressions.
252 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
253 if (!has_ctrl(n)) return NULL;
254 Node *n_ctrl = get_ctrl(n);
255 IdealLoopTree *n_loop = get_loop(n_ctrl);
256
257 // See if 'n' mixes loop-varying and loop-invariant inputs and
258 // itself is loop-varying.
259
260 // Only interested in binary ops (and AddP)
261 if( n->req() < 3 || n->req() > 4 ) return NULL;
262
263 Node *n1_ctrl = get_ctrl(n->in( 1));
264 Node *n2_ctrl = get_ctrl(n->in( 2));
265 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
266 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
267 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
268 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
269
270 // Does one of my inputs spin in a tighter loop than self?
271 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
272 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
273 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
274 return NULL; // Leave well enough alone
275
276 // Is at least one of my inputs loop-invariant?
277 if( n1_loop == n_loop &&
278 n2_loop == n_loop &&
279 n3_loop == n_loop )
280 return NULL; // No loop-invariant inputs
281
282
283 int n_op = n->Opcode();
284
285 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
286 if( n_op == Op_LShiftI ) {
287 // Scale is loop invariant
288 Node *scale = n->in(2);
289 Node *scale_ctrl = get_ctrl(scale);
290 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
291 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
292 return NULL;
293 const TypeInt *scale_t = scale->bottom_type()->isa_int();
294 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
295 return NULL; // Dont bother with byte/short masking
296 // Add must vary with loop (else shift would be loop-invariant)
297 Node *add = n->in(1);
298 Node *add_ctrl = get_ctrl(add);
299 IdealLoopTree *add_loop = get_loop(add_ctrl);
300 //assert( n_loop == add_loop, "" );
301 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
302
303 // Convert I-V into I+ (0-V); same for V-I
304 if( add->Opcode() == Op_SubI &&
305 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
306 Node *zero = _igvn.intcon(0);
307 set_ctrl(zero, C->root());
308 Node *neg = new (C, 3) SubINode( _igvn.intcon(0), add->in(2) );
309 register_new_node( neg, get_ctrl(add->in(2) ) );
310 add = new (C, 3) AddINode( add->in(1), neg );
311 register_new_node( add, add_ctrl );
312 }
313 if( add->Opcode() != Op_AddI ) return NULL;
314 // See if one add input is loop invariant
315 Node *add_var = add->in(1);
316 Node *add_var_ctrl = get_ctrl(add_var);
317 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
318 Node *add_invar = add->in(2);
319 Node *add_invar_ctrl = get_ctrl(add_invar);
320 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
321 if( add_var_loop == n_loop ) {
322 } else if( add_invar_loop == n_loop ) {
323 // Swap to find the invariant part
324 add_invar = add_var;
325 add_invar_ctrl = add_var_ctrl;
326 add_invar_loop = add_var_loop;
327 add_var = add->in(2);
328 Node *add_var_ctrl = get_ctrl(add_var);
329 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
330 } else // Else neither input is loop invariant
331 return NULL;
332 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
333 return NULL; // No invariant part of the add?
334
335 // Yes! Reshape address expression!
336 Node *inv_scale = new (C, 3) LShiftINode( add_invar, scale );
337 register_new_node( inv_scale, add_invar_ctrl );
338 Node *var_scale = new (C, 3) LShiftINode( add_var, scale );
339 register_new_node( var_scale, n_ctrl );
340 Node *var_add = new (C, 3) AddINode( var_scale, inv_scale );
341 register_new_node( var_add, n_ctrl );
342 _igvn.hash_delete( n );
343 _igvn.subsume_node( n, var_add );
344 return var_add;
345 }
346
347 // Replace (I+V) with (V+I)
348 if( n_op == Op_AddI ||
349 n_op == Op_AddL ||
350 n_op == Op_AddF ||
351 n_op == Op_AddD ||
352 n_op == Op_MulI ||
353 n_op == Op_MulL ||
354 n_op == Op_MulF ||
355 n_op == Op_MulD ) {
356 if( n2_loop == n_loop ) {
357 assert( n1_loop != n_loop, "" );
358 n->swap_edges(1, 2);
359 }
360 }
361
362 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
363 // but not if I2 is a constant.
364 if( n_op == Op_AddP ) {
365 if( n2_loop == n_loop && n3_loop != n_loop ) {
366 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
367 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
368 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
369 IdealLoopTree *n22loop = get_loop( n22_ctrl );
370 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
371 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
372 n23_loop == n_loop ) {
373 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
374 // Stuff new AddP in the loop preheader
375 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
376 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) );
377 register_new_node( add2, n_ctrl );
378 _igvn.hash_delete( n );
379 _igvn.subsume_node( n, add2 );
380 return add2;
381 }
382 }
383 }
384
385 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
386 if( n2_loop != n_loop && n3_loop == n_loop ) {
387 if( n->in(3)->Opcode() == Op_AddI ) {
388 Node *V = n->in(3)->in(1);
389 Node *I = n->in(3)->in(2);
390 if( is_member(n_loop,get_ctrl(V)) ) {
391 } else {
392 Node *tmp = V; V = I; I = tmp;
393 }
394 if( !is_member(n_loop,get_ctrl(I)) ) {
395 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I );
396 // Stuff new AddP in the loop preheader
397 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
398 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V );
399 register_new_node( add2, n_ctrl );
400 _igvn.hash_delete( n );
401 _igvn.subsume_node( n, add2 );
402 return add2;
403 }
404 }
405 }
406 }
407
408 return NULL;
409 }
410
411 //------------------------------conditional_move-------------------------------
412 // Attempt to replace a Phi with a conditional move. We have some pretty
413 // strict profitability requirements. All Phis at the merge point must
414 // be converted, so we can remove the control flow. We need to limit the
415 // number of c-moves to a small handful. All code that was in the side-arms
416 // of the CFG diamond is now speculatively executed. This code has to be
417 // "cheap enough". We are pretty much limited to CFG diamonds that merge
418 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
419 Node *PhaseIdealLoop::conditional_move( Node *region ) {
420
421 assert( region->is_Region(), "sanity check" );
422 if( region->req() != 3 ) return NULL;
423
424 // Check for CFG diamond
425 Node *lp = region->in(1);
426 Node *rp = region->in(2);
427 if( !lp || !rp ) return NULL;
428 Node *lp_c = lp->in(0);
429 if( lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If() ) return NULL;
430 IfNode *iff = lp_c->as_If();
431
432 // Check for highly predictable branch. No point in CMOV'ing if
433 // we are going to predict accurately all the time.
434 // %%% This hides patterns produced by utility methods like Math.min.
435 if( iff->_prob < PROB_UNLIKELY_MAG(3) ||
436 iff->_prob > PROB_LIKELY_MAG(3) )
437 return NULL;
438
439 // Check for ops pinned in an arm of the diamond.
440 // Can't remove the control flow in this case
441 if( lp->outcnt() > 1 ) return NULL;
442 if( rp->outcnt() > 1 ) return NULL;
443
444 // Check profitability
445 int cost = 0;
446 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
447 Node *out = region->fast_out(i);
448 if( !out->is_Phi() ) continue; // Ignore other control edges, etc
449 PhiNode* phi = out->as_Phi();
450 switch (phi->type()->basic_type()) {
451 case T_LONG:
452 cost++; // Probably encodes as 2 CMOV's
453 case T_INT: // These all CMOV fine
454 case T_FLOAT:
455 case T_DOUBLE:
456 case T_ADDRESS: // (RawPtr)
457 cost++;
458 break;
459 case T_OBJECT: { // Base oops are OK, but not derived oops
460 const TypeOopPtr *tp = phi->type()->isa_oopptr();
461 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
462 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
463 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
464 // have a Phi for the base here that we convert to a CMOVE all is well
465 // and good. But if the base is dead, we'll not make a CMOVE. Later
466 // the allocator will have to produce a base by creating a CMOVE of the
467 // relevant bases. This puts the allocator in the business of
468 // manufacturing expensive instructions, generally a bad plan.
469 // Just Say No to Conditionally-Moved Derived Pointers.
470 if( tp && tp->offset() != 0 )
471 return NULL;
472 cost++;
473 break;
474 }
475 default:
476 return NULL; // In particular, can't do memory or I/O
477 }
478 // Add in cost any speculative ops
479 for( uint j = 1; j < region->req(); j++ ) {
480 Node *proj = region->in(j);
481 Node *inp = phi->in(j);
482 if (get_ctrl(inp) == proj) { // Found local op
483 cost++;
484 // Check for a chain of dependent ops; these will all become
485 // speculative in a CMOV.
486 for( uint k = 1; k < inp->req(); k++ )
487 if (get_ctrl(inp->in(k)) == proj)
488 return NULL; // Too much speculative goo
489 }
490 }
491 // See if the Phi is used by a Cmp. This will likely Split-If, a
492 // higher-payoff operation.
493 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
494 Node* use = phi->fast_out(k);
495 if( use->is_Cmp() )
496 return NULL;
497 }
498 }
499 if( cost >= ConditionalMoveLimit ) return NULL; // Too much goo
500
501 // --------------
502 // Now replace all Phis with CMOV's
503 Node *cmov_ctrl = iff->in(0);
504 uint flip = (lp->Opcode() == Op_IfTrue);
505 while( 1 ) {
506 PhiNode* phi = NULL;
507 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
508 Node *out = region->fast_out(i);
509 if (out->is_Phi()) {
510 phi = out->as_Phi();
511 break;
512 }
513 }
514 if (phi == NULL) break;
515 #ifndef PRODUCT
516 if( PrintOpto && VerifyLoopOptimizations ) tty->print_cr("CMOV");
517 #endif
518 // Move speculative ops
519 for( uint j = 1; j < region->req(); j++ ) {
520 Node *proj = region->in(j);
521 Node *inp = phi->in(j);
522 if (get_ctrl(inp) == proj) { // Found local op
523 #ifndef PRODUCT
524 if( PrintOpto && VerifyLoopOptimizations ) {
525 tty->print(" speculate: ");
526 inp->dump();
527 }
528 #endif
529 set_ctrl(inp, cmov_ctrl);
530 }
531 }
532 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) );
533 register_new_node( cmov, cmov_ctrl );
534 _igvn.hash_delete(phi);
535 _igvn.subsume_node( phi, cmov );
536 #ifndef PRODUCT
537 if( VerifyLoopOptimizations ) verify();
538 #endif
539 }
540
541 // The useless CFG diamond will fold up later; see the optimization in
542 // RegionNode::Ideal.
543 _igvn._worklist.push(region);
544
545 return iff->in(1);
546 }
547
548 //------------------------------split_if_with_blocks_pre-----------------------
549 // Do the real work in a non-recursive function. Data nodes want to be
550 // cloned in the pre-order so they can feed each other nicely.
551 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
552 // Cloning these guys is unlikely to win
553 int n_op = n->Opcode();
554 if( n_op == Op_MergeMem ) return n;
555 if( n->is_Proj() ) return n;
556 // Do not clone-up CmpFXXX variations, as these are always
557 // followed by a CmpI
558 if( n->is_Cmp() ) return n;
559 // Attempt to use a conditional move instead of a phi/branch
560 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
561 Node *cmov = conditional_move( n );
562 if( cmov ) return cmov;
563 }
564 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;
565 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
566 n_op == Op_Opaque2 ) {
567 if( !C->major_progress() ) // If chance of no more loop opts...
568 _igvn._worklist.push(n); // maybe we'll remove them
569 return n;
570 }
571
572 if( n->is_Con() ) return n; // No cloning for Con nodes
573
574 Node *n_ctrl = get_ctrl(n);
575 if( !n_ctrl ) return n; // Dead node
576
577 // Attempt to remix address expressions for loop invariants
578 Node *m = remix_address_expressions( n );
579 if( m ) return m;
580
581 // Determine if the Node has inputs from some local Phi.
582 // Returns the block to clone thru.
583 Node *n_blk = has_local_phi_input( n );
584 if( !n_blk ) return n;
585 // Do not clone the trip counter through on a CountedLoop
586 // (messes up the canonical shape).
587 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
588
589 // Check for having no control input; not pinned. Allow
590 // dominating control.
591 if( n->in(0) ) {
592 Node *dom = idom(n_blk);
593 if( dom_lca( n->in(0), dom ) != n->in(0) )
594 return n;
595 }
596 // Policy: when is it profitable. You must get more wins than
597 // policy before it is considered profitable. Policy is usually 0,
598 // so 1 win is considered profitable. Big merges will require big
599 // cloning, so get a larger policy.
600 int policy = n_blk->req() >> 2;
601
602 // If the loop is a candidate for range check elimination,
603 // delay splitting through it's phi until a later loop optimization
604 if (n_blk->is_CountedLoop()) {
605 IdealLoopTree *lp = get_loop(n_blk);
606 if (lp && lp->_rce_candidate) {
607 return n;
608 }
609 }
610
611 // Use same limit as split_if_with_blocks_post
612 if( C->unique() > 35000 ) return n; // Method too big
613
614 // Split 'n' through the merge point if it is profitable
615 Node *phi = split_thru_phi( n, n_blk, policy );
616 if( !phi ) return n;
617
618 // Found a Phi to split thru!
619 // Replace 'n' with the new phi
620 _igvn.hash_delete(n);
621 _igvn.subsume_node( n, phi );
622 // Moved a load around the loop, 'en-registering' something.
623 if( n_blk->Opcode() == Op_Loop && n->is_Load() &&
624 !phi->in(LoopNode::LoopBackControl)->is_Load() )
625 C->set_major_progress();
626
627 return phi;
628 }
629
630 static bool merge_point_too_heavy(Compile* C, Node* region) {
631 // Bail out if the region and its phis have too many users.
632 int weight = 0;
633 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
634 weight += region->fast_out(i)->outcnt();
635 }
636 int nodes_left = MaxNodeLimit - C->unique();
637 if (weight * 8 > nodes_left) {
638 #ifndef PRODUCT
639 if (PrintOpto)
640 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
641 #endif
642 return true;
643 } else {
644 return false;
645 }
646 }
647
648 #ifdef _LP64
649 static bool merge_point_safe(Node* region) {
650 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
651 // having a PhiNode input. This sidesteps the dangerous case where the split
652 // ConvI2LNode may become TOP if the input Value() does not
653 // overlap the ConvI2L range, leaving a node which may not dominate its
654 // uses.
655 // A better fix for this problem can be found in the BugTraq entry, but
656 // expediency for Mantis demands this hack.
657 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
658 Node* n = region->fast_out(i);
659 if (n->is_Phi()) {
660 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
661 Node* m = n->fast_out(j);
662 if (m->Opcode() == Op_ConvI2L) {
663 return false;
664 }
665 }
666 }
667 }
668 return true;
669 }
670 #endif
671
672
673 //------------------------------place_near_use---------------------------------
674 // Place some computation next to use but not inside inner loops.
675 // For inner loop uses move it to the preheader area.
676 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const {
677 IdealLoopTree *u_loop = get_loop( useblock );
678 return (u_loop->_irreducible || u_loop->_child)
679 ? useblock
680 : u_loop->_head->in(LoopNode::EntryControl);
681 }
682
683
684 //------------------------------split_if_with_blocks_post----------------------
685 // Do the real work in a non-recursive function. CFG hackery wants to be
686 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
687 // info.
688 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) {
689
690 // Cloning Cmp through Phi's involves the split-if transform.
691 // FastLock is not used by an If
692 if( n->is_Cmp() && !n->is_FastLock() ) {
693 if( C->unique() > 35000 ) return; // Method too big
694
695 // Do not do 'split-if' if irreducible loops are present.
696 if( _has_irreducible_loops )
697 return;
698
699 Node *n_ctrl = get_ctrl(n);
700 // Determine if the Node has inputs from some local Phi.
701 // Returns the block to clone thru.
702 Node *n_blk = has_local_phi_input( n );
703 if( n_blk != n_ctrl ) return;
704
705 if( merge_point_too_heavy(C, n_ctrl) )
706 return;
707
708 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare?
709 Node *bol = n->unique_out();
710 assert( bol->is_Bool(), "expect a bool here" );
711 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare?
712 Node *iff = bol->unique_out();
713
714 // Check some safety conditions
715 if( iff->is_If() ) { // Classic split-if?
716 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if
717 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
718 if( get_ctrl(iff->in(2)) == n_ctrl ||
719 get_ctrl(iff->in(3)) == n_ctrl )
720 return; // Inputs not yet split-up
721 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) {
722 return; // Loop-invar test gates loop-varying CMOVE
723 }
724 } else {
725 return; // some other kind of node, such as an Allocate
726 }
727
728 // Do not do 'split-if' if some paths are dead. First do dead code
729 // elimination and then see if its still profitable.
730 for( uint i = 1; i < n_ctrl->req(); i++ )
731 if( n_ctrl->in(i) == C->top() )
732 return;
733
734 // When is split-if profitable? Every 'win' on means some control flow
735 // goes dead, so it's almost always a win.
736 int policy = 0;
737 // If trying to do a 'Split-If' at the loop head, it is only
738 // profitable if the cmp folds up on BOTH paths. Otherwise we
739 // risk peeling a loop forever.
740
741 // CNC - Disabled for now. Requires careful handling of loop
742 // body selection for the cloned code. Also, make sure we check
743 // for any input path not being in the same loop as n_ctrl. For
744 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
745 // because the alternative loop entry points won't be converted
746 // into LoopNodes.
747 IdealLoopTree *n_loop = get_loop(n_ctrl);
748 for( uint j = 1; j < n_ctrl->req(); j++ )
749 if( get_loop(n_ctrl->in(j)) != n_loop )
750 return;
751
752 #ifdef _LP64
753 // Check for safety of the merge point.
754 if( !merge_point_safe(n_ctrl) ) {
755 return;
756 }
757 #endif
758
759 // Split compare 'n' through the merge point if it is profitable
760 Node *phi = split_thru_phi( n, n_ctrl, policy );
761 if( !phi ) return;
762
763 // Found a Phi to split thru!
764 // Replace 'n' with the new phi
765 _igvn.hash_delete(n);
766 _igvn.subsume_node( n, phi );
767
768 // Now split the bool up thru the phi
769 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 );
770 _igvn.hash_delete(bol);
771 _igvn.subsume_node( bol, bolphi );
772 assert( iff->in(1) == bolphi, "" );
773 if( bolphi->Value(&_igvn)->singleton() )
774 return;
775
776 // Conditional-move? Must split up now
777 if( !iff->is_If() ) {
778 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 );
779 _igvn.hash_delete(iff);
780 _igvn.subsume_node( iff, cmovphi );
781 return;
782 }
783
784 // Now split the IF
785 do_split_if( iff );
786 return;
787 }
788
789 // Check for an IF ready to split; one that has its
790 // condition codes input coming from a Phi at the block start.
791 int n_op = n->Opcode();
792
793 // Check for an IF being dominated by another IF same test
794 if( n_op == Op_If ) {
795 Node *bol = n->in(1);
796 uint max = bol->outcnt();
797 // Check for same test used more than once?
798 if( n_op == Op_If && max > 1 && bol->is_Bool() ) {
799 // Search up IDOMs to see if this IF is dominated.
800 Node *cutoff = get_ctrl(bol);
801
802 // Now search up IDOMs till cutoff, looking for a dominating test
803 Node *prevdom = n;
804 Node *dom = idom(prevdom);
805 while( dom != cutoff ) {
806 if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) {
807 // Replace the dominated test with an obvious true or false.
808 // Place it on the IGVN worklist for later cleanup.
809 C->set_major_progress();
810 dominated_by( prevdom, n );
811 #ifndef PRODUCT
812 if( VerifyLoopOptimizations ) verify();
813 #endif
814 return;
815 }
816 prevdom = dom;
817 dom = idom(prevdom);
818 }
819 }
820 }
821
822 // See if a shared loop-varying computation has no loop-varying uses.
823 // Happens if something is only used for JVM state in uncommon trap exits,
824 // like various versions of induction variable+offset. Clone the
825 // computation per usage to allow it to sink out of the loop.
826 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
827 Node *n_ctrl = get_ctrl(n);
828 IdealLoopTree *n_loop = get_loop(n_ctrl);
829 if( n_loop != _ltree_root ) {
830 DUIterator_Fast imax, i = n->fast_outs(imax);
831 for (; i < imax; i++) {
832 Node* u = n->fast_out(i);
833 if( !has_ctrl(u) ) break; // Found control user
834 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
835 if( u_loop == n_loop ) break; // Found loop-varying use
836 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
837 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
838 }
839 bool did_break = (i < imax); // Did we break out of the previous loop?
840 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
841 Node *late_load_ctrl;
842 if (n->is_Load()) {
843 // If n is a load, get and save the result from get_late_ctrl(),
844 // to be later used in calculating the control for n's clones.
845 clear_dom_lca_tags();
846 late_load_ctrl = get_late_ctrl(n, n_ctrl);
847 }
848 // If n is a load, and the late control is the same as the current
849 // control, then the cloning of n is a pointless exercise, because
850 // GVN will ensure that we end up where we started.
851 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
852 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
853 Node *u = n->last_out(j); // Clone private computation per use
854 _igvn.hash_delete(u);
855 _igvn._worklist.push(u);
856 Node *x = n->clone(); // Clone computation
857 Node *x_ctrl = NULL;
858 if( u->is_Phi() ) {
859 // Replace all uses of normal nodes. Replace Phi uses
860 // individually, so the seperate Nodes can sink down
861 // different paths.
862 uint k = 1;
863 while( u->in(k) != n ) k++;
864 u->set_req( k, x );
865 // x goes next to Phi input path
866 x_ctrl = u->in(0)->in(k);
867 --j;
868 } else { // Normal use
869 // Replace all uses
870 for( uint k = 0; k < u->req(); k++ ) {
871 if( u->in(k) == n ) {
872 u->set_req( k, x );
873 --j;
874 }
875 }
876 x_ctrl = get_ctrl(u);
877 }
878
879 // Find control for 'x' next to use but not inside inner loops.
880 // For inner loop uses get the preheader area.
881 x_ctrl = place_near_use(x_ctrl);
882
883 if (n->is_Load()) {
884 // For loads, add a control edge to a CFG node outside of the loop
885 // to force them to not combine and return back inside the loop
886 // during GVN optimization (4641526).
887 //
888 // Because we are setting the actual control input, factor in
889 // the result from get_late_ctrl() so we respect any
890 // anti-dependences. (6233005).
891 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
892
893 // Don't allow the control input to be a CFG splitting node.
894 // Such nodes should only have ProjNodes as outs, e.g. IfNode
895 // should only have IfTrueNode and IfFalseNode (4985384).
896 x_ctrl = find_non_split_ctrl(x_ctrl);
897 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
898
899 x->set_req(0, x_ctrl);
900 }
901 register_new_node(x, x_ctrl);
902
903 // Some institutional knowledge is needed here: 'x' is
904 // yanked because if the optimizer runs GVN on it all the
905 // cloned x's will common up and undo this optimization and
906 // be forced back in the loop. This is annoying because it
907 // makes +VerifyOpto report false-positives on progress. I
908 // tried setting control edges on the x's to force them to
909 // not combine, but the matching gets worried when it tries
910 // to fold a StoreP and an AddP together (as part of an
911 // address expression) and the AddP and StoreP have
912 // different controls.
913 if( !x->is_Load() ) _igvn._worklist.yank(x);
914 }
915 _igvn.remove_dead_node(n);
916 }
917 }
918 }
919 }
920
921 // Check for Opaque2's who's loop has disappeared - who's input is in the
922 // same loop nest as their output. Remove 'em, they are no longer useful.
923 if( n_op == Op_Opaque2 &&
924 n->in(1) != NULL &&
925 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
926 _igvn.add_users_to_worklist(n);
927 _igvn.hash_delete(n);
928 _igvn.subsume_node( n, n->in(1) );
929 }
930 }
931
932 //------------------------------split_if_with_blocks---------------------------
933 // Check for aggressive application of 'split-if' optimization,
934 // using basic block level info.
935 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) {
936 Node *n = C->root();
937 visited.set(n->_idx); // first, mark node as visited
938 // Do pre-visit work for root
939 n = split_if_with_blocks_pre( n );
940 uint cnt = n->outcnt();
941 uint i = 0;
942 while (true) {
943 // Visit all children
944 if (i < cnt) {
945 Node* use = n->raw_out(i);
946 ++i;
947 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
948 // Now do pre-visit work for this use
949 use = split_if_with_blocks_pre( use );
950 nstack.push(n, i); // Save parent and next use's index.
951 n = use; // Process all children of current use.
952 cnt = use->outcnt();
953 i = 0;
954 }
955 }
956 else {
957 // All of n's children have been processed, complete post-processing.
958 if (cnt != 0 && !n->is_Con()) {
959 assert(has_node(n), "no dead nodes");
960 split_if_with_blocks_post( n );
961 }
962 if (nstack.is_empty()) {
963 // Finished all nodes on stack.
964 break;
965 }
966 // Get saved parent node and next use's index. Visit the rest of uses.
967 n = nstack.node();
968 cnt = n->outcnt();
969 i = nstack.index();
970 nstack.pop();
971 }
972 }
973 }
974
975
976 //=============================================================================
977 //
978 // C L O N E A L O O P B O D Y
979 //
980
981 //------------------------------clone_iff--------------------------------------
982 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
983 // "Nearly" because all Nodes have been cloned from the original in the loop,
984 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
985 // through the Phi recursively, and return a Bool.
986 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) {
987
988 // Convert this Phi into a Phi merging Bools
989 uint i;
990 for( i = 1; i < phi->req(); i++ ) {
991 Node *b = phi->in(i);
992 if( b->is_Phi() ) {
993 _igvn.hash_delete(phi);
994 _igvn._worklist.push(phi);
995 phi->set_req(i, clone_iff( b->as_Phi(), loop ));
996 } else {
997 assert( b->is_Bool(), "" );
998 }
999 }
1000
1001 Node *sample_bool = phi->in(1);
1002 Node *sample_cmp = sample_bool->in(1);
1003
1004 // Make Phis to merge the Cmp's inputs.
1005 int size = phi->in(0)->req();
1006 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1007 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1008 for( i = 1; i < phi->req(); i++ ) {
1009 Node *n1 = phi->in(i)->in(1)->in(1);
1010 Node *n2 = phi->in(i)->in(1)->in(2);
1011 phi1->set_req( i, n1 );
1012 phi2->set_req( i, n2 );
1013 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1014 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1015 }
1016 // See if these Phis have been made before.
1017 // Register with optimizer
1018 Node *hit1 = _igvn.hash_find_insert(phi1);
1019 if( hit1 ) { // Hit, toss just made Phi
1020 _igvn.remove_dead_node(phi1); // Remove new phi
1021 assert( hit1->is_Phi(), "" );
1022 phi1 = (PhiNode*)hit1; // Use existing phi
1023 } else { // Miss
1024 _igvn.register_new_node_with_optimizer(phi1);
1025 }
1026 Node *hit2 = _igvn.hash_find_insert(phi2);
1027 if( hit2 ) { // Hit, toss just made Phi
1028 _igvn.remove_dead_node(phi2); // Remove new phi
1029 assert( hit2->is_Phi(), "" );
1030 phi2 = (PhiNode*)hit2; // Use existing phi
1031 } else { // Miss
1032 _igvn.register_new_node_with_optimizer(phi2);
1033 }
1034 // Register Phis with loop/block info
1035 set_ctrl(phi1, phi->in(0));
1036 set_ctrl(phi2, phi->in(0));
1037 // Make a new Cmp
1038 Node *cmp = sample_cmp->clone();
1039 cmp->set_req( 1, phi1 );
1040 cmp->set_req( 2, phi2 );
1041 _igvn.register_new_node_with_optimizer(cmp);
1042 set_ctrl(cmp, phi->in(0));
1043
1044 // Make a new Bool
1045 Node *b = sample_bool->clone();
1046 b->set_req(1,cmp);
1047 _igvn.register_new_node_with_optimizer(b);
1048 set_ctrl(b, phi->in(0));
1049
1050 assert( b->is_Bool(), "" );
1051 return (BoolNode*)b;
1052 }
1053
1054 //------------------------------clone_bool-------------------------------------
1055 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1056 // "Nearly" because all Nodes have been cloned from the original in the loop,
1057 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1058 // through the Phi recursively, and return a Bool.
1059 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1060 uint i;
1061 // Convert this Phi into a Phi merging Bools
1062 for( i = 1; i < phi->req(); i++ ) {
1063 Node *b = phi->in(i);
1064 if( b->is_Phi() ) {
1065 _igvn.hash_delete(phi);
1066 _igvn._worklist.push(phi);
1067 phi->set_req(i, clone_bool( b->as_Phi(), loop ));
1068 } else {
1069 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1070 }
1071 }
1072
1073 Node *sample_cmp = phi->in(1);
1074
1075 // Make Phis to merge the Cmp's inputs.
1076 int size = phi->in(0)->req();
1077 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1078 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1079 for( uint j = 1; j < phi->req(); j++ ) {
1080 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1081 Node *n1, *n2;
1082 if( cmp_top->is_Cmp() ) {
1083 n1 = cmp_top->in(1);
1084 n2 = cmp_top->in(2);
1085 } else {
1086 n1 = n2 = cmp_top;
1087 }
1088 phi1->set_req( j, n1 );
1089 phi2->set_req( j, n2 );
1090 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1091 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1092 }
1093
1094 // See if these Phis have been made before.
1095 // Register with optimizer
1096 Node *hit1 = _igvn.hash_find_insert(phi1);
1097 if( hit1 ) { // Hit, toss just made Phi
1098 _igvn.remove_dead_node(phi1); // Remove new phi
1099 assert( hit1->is_Phi(), "" );
1100 phi1 = (PhiNode*)hit1; // Use existing phi
1101 } else { // Miss
1102 _igvn.register_new_node_with_optimizer(phi1);
1103 }
1104 Node *hit2 = _igvn.hash_find_insert(phi2);
1105 if( hit2 ) { // Hit, toss just made Phi
1106 _igvn.remove_dead_node(phi2); // Remove new phi
1107 assert( hit2->is_Phi(), "" );
1108 phi2 = (PhiNode*)hit2; // Use existing phi
1109 } else { // Miss
1110 _igvn.register_new_node_with_optimizer(phi2);
1111 }
1112 // Register Phis with loop/block info
1113 set_ctrl(phi1, phi->in(0));
1114 set_ctrl(phi2, phi->in(0));
1115 // Make a new Cmp
1116 Node *cmp = sample_cmp->clone();
1117 cmp->set_req( 1, phi1 );
1118 cmp->set_req( 2, phi2 );
1119 _igvn.register_new_node_with_optimizer(cmp);
1120 set_ctrl(cmp, phi->in(0));
1121
1122 assert( cmp->is_Cmp(), "" );
1123 return (CmpNode*)cmp;
1124 }
1125
1126 //------------------------------sink_use---------------------------------------
1127 // If 'use' was in the loop-exit block, it now needs to be sunk
1128 // below the post-loop merge point.
1129 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1130 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1131 set_ctrl(use, post_loop);
1132 for (DUIterator j = use->outs(); use->has_out(j); j++)
1133 sink_use(use->out(j), post_loop);
1134 }
1135 }
1136
1137 //------------------------------clone_loop-------------------------------------
1138 //
1139 // C L O N E A L O O P B O D Y
1140 //
1141 // This is the basic building block of the loop optimizations. It clones an
1142 // entire loop body. It makes an old_new loop body mapping; with this mapping
1143 // you can find the new-loop equivalent to an old-loop node. All new-loop
1144 // nodes are exactly equal to their old-loop counterparts, all edges are the
1145 // same. All exits from the old-loop now have a RegionNode that merges the
1146 // equivalent new-loop path. This is true even for the normal "loop-exit"
1147 // condition. All uses of loop-invariant old-loop values now come from (one
1148 // or more) Phis that merge their new-loop equivalents.
1149 //
1150 // This operation leaves the graph in an illegal state: there are two valid
1151 // control edges coming from the loop pre-header to both loop bodies. I'll
1152 // definitely have to hack the graph after running this transform.
1153 //
1154 // From this building block I will further edit edges to perform loop peeling
1155 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1156 //
1157 // Parameter side_by_size_idom:
1158 // When side_by_size_idom is NULL, the dominator tree is constructed for
1159 // the clone loop to dominate the original. Used in construction of
1160 // pre-main-post loop sequence.
1161 // When nonnull, the clone and original are side-by-side, both are
1162 // dominated by the side_by_side_idom node. Used in construction of
1163 // unswitched loops.
1164 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1165 Node* side_by_side_idom) {
1166
1167 // Step 1: Clone the loop body. Make the old->new mapping.
1168 uint i;
1169 for( i = 0; i < loop->_body.size(); i++ ) {
1170 Node *old = loop->_body.at(i);
1171 Node *nnn = old->clone();
1172 old_new.map( old->_idx, nnn );
1173 _igvn.register_new_node_with_optimizer(nnn);
1174 }
1175
1176
1177 // Step 2: Fix the edges in the new body. If the old input is outside the
1178 // loop use it. If the old input is INside the loop, use the corresponding
1179 // new node instead.
1180 for( i = 0; i < loop->_body.size(); i++ ) {
1181 Node *old = loop->_body.at(i);
1182 Node *nnn = old_new[old->_idx];
1183 // Fix CFG/Loop controlling the new node
1184 if (has_ctrl(old)) {
1185 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
1186 } else {
1187 set_loop(nnn, loop->_parent);
1188 if (old->outcnt() > 0) {
1189 set_idom( nnn, old_new[idom(old)->_idx], dd );
1190 }
1191 }
1192 // Correct edges to the new node
1193 for( uint j = 0; j < nnn->req(); j++ ) {
1194 Node *n = nnn->in(j);
1195 if( n ) {
1196 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
1197 if( loop->is_member( old_in_loop ) )
1198 nnn->set_req(j, old_new[n->_idx]);
1199 }
1200 }
1201 _igvn.hash_find_insert(nnn);
1202 }
1203 Node *newhead = old_new[loop->_head->_idx];
1204 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1205
1206
1207 // Step 3: Now fix control uses. Loop varying control uses have already
1208 // been fixed up (as part of all input edges in Step 2). Loop invariant
1209 // control uses must be either an IfFalse or an IfTrue. Make a merge
1210 // point to merge the old and new IfFalse/IfTrue nodes; make the use
1211 // refer to this.
1212 ResourceArea *area = Thread::current()->resource_area();
1213 Node_List worklist(area);
1214 uint new_counter = C->unique();
1215 for( i = 0; i < loop->_body.size(); i++ ) {
1216 Node* old = loop->_body.at(i);
1217 if( !old->is_CFG() ) continue;
1218 Node* nnn = old_new[old->_idx];
1219
1220 // Copy uses to a worklist, so I can munge the def-use info
1221 // with impunity.
1222 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1223 worklist.push(old->fast_out(j));
1224
1225 while( worklist.size() ) { // Visit all uses
1226 Node *use = worklist.pop();
1227 if (!has_node(use)) continue; // Ignore dead nodes
1228 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1229 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
1230 // Both OLD and USE are CFG nodes here.
1231 assert( use->is_Proj(), "" );
1232
1233 // Clone the loop exit control projection
1234 Node *newuse = use->clone();
1235 newuse->set_req(0,nnn);
1236 _igvn.register_new_node_with_optimizer(newuse);
1237 set_loop(newuse, use_loop);
1238 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
1239
1240 // We need a Region to merge the exit from the peeled body and the
1241 // exit from the old loop body.
1242 RegionNode *r = new (C, 3) RegionNode(3);
1243 // Map the old use to the new merge point
1244 old_new.map( use->_idx, r );
1245 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
1246 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
1247
1248 // The original user of 'use' uses 'r' instead.
1249 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
1250 Node* useuse = use->last_out(l);
1251 _igvn.hash_delete(useuse);
1252 _igvn._worklist.push(useuse);
1253 uint uses_found = 0;
1254 if( useuse->in(0) == use ) {
1255 useuse->set_req(0, r);
1256 uses_found++;
1257 if( useuse->is_CFG() ) {
1258 assert( dom_depth(useuse) > dd_r, "" );
1259 set_idom(useuse, r, dom_depth(useuse));
1260 }
1261 }
1262 for( uint k = 1; k < useuse->req(); k++ ) {
1263 if( useuse->in(k) == use ) {
1264 useuse->set_req(k, r);
1265 uses_found++;
1266 }
1267 }
1268 l -= uses_found; // we deleted 1 or more copies of this edge
1269 }
1270
1271 // Now finish up 'r'
1272 r->set_req( 1, newuse );
1273 r->set_req( 2, use );
1274 _igvn.register_new_node_with_optimizer(r);
1275 set_loop(r, use_loop);
1276 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
1277 } // End of if a loop-exit test
1278 }
1279 }
1280
1281 // Step 4: If loop-invariant use is not control, it must be dominated by a
1282 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
1283 // there if needed. Make a Phi there merging old and new used values.
1284 Node_List *split_if_set = NULL;
1285 Node_List *split_bool_set = NULL;
1286 Node_List *split_cex_set = NULL;
1287 for( i = 0; i < loop->_body.size(); i++ ) {
1288 Node* old = loop->_body.at(i);
1289 Node* nnn = old_new[old->_idx];
1290 // Copy uses to a worklist, so I can munge the def-use info
1291 // with impunity.
1292 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1293 worklist.push(old->fast_out(j));
1294
1295 while( worklist.size() ) {
1296 Node *use = worklist.pop();
1297 if (!has_node(use)) continue; // Ignore dead nodes
1298 if (use->in(0) == C->top()) continue;
1299 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1300 // Check for data-use outside of loop - at least one of OLD or USE
1301 // must not be a CFG node.
1302 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) {
1303
1304 // If the Data use is an IF, that means we have an IF outside of the
1305 // loop that is switching on a condition that is set inside of the
1306 // loop. Happens if people set a loop-exit flag; then test the flag
1307 // in the loop to break the loop, then test is again outside of the
1308 // loop to determine which way the loop exited.
1309 if( use->is_If() || use->is_CMove() ) {
1310 // Since this code is highly unlikely, we lazily build the worklist
1311 // of such Nodes to go split.
1312 if( !split_if_set )
1313 split_if_set = new Node_List(area);
1314 split_if_set->push(use);
1315 }
1316 if( use->is_Bool() ) {
1317 if( !split_bool_set )
1318 split_bool_set = new Node_List(area);
1319 split_bool_set->push(use);
1320 }
1321 if( use->Opcode() == Op_CreateEx ) {
1322 if( !split_cex_set )
1323 split_cex_set = new Node_List(area);
1324 split_cex_set->push(use);
1325 }
1326
1327
1328 // Get "block" use is in
1329 uint idx = 0;
1330 while( use->in(idx) != old ) idx++;
1331 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1332 assert( !loop->is_member( get_loop( prev ) ), "" );
1333 Node *cfg = prev->_idx >= new_counter
1334 ? prev->in(2)
1335 : idom(prev);
1336 if( use->is_Phi() ) // Phi use is in prior block
1337 cfg = prev->in(idx); // NOT in block of Phi itself
1338 if (cfg->is_top()) { // Use is dead?
1339 _igvn.hash_delete(use);
1340 _igvn._worklist.push(use);
1341 use->set_req(idx, C->top());
1342 continue;
1343 }
1344
1345 while( !loop->is_member( get_loop( cfg ) ) ) {
1346 prev = cfg;
1347 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1348 }
1349 // If the use occurs after merging several exits from the loop, then
1350 // old value must have dominated all those exits. Since the same old
1351 // value was used on all those exits we did not need a Phi at this
1352 // merge point. NOW we do need a Phi here. Each loop exit value
1353 // is now merged with the peeled body exit; each exit gets its own
1354 // private Phi and those Phis need to be merged here.
1355 Node *phi;
1356 if( prev->is_Region() ) {
1357 if( idx == 0 ) { // Updating control edge?
1358 phi = prev; // Just use existing control
1359 } else { // Else need a new Phi
1360 phi = PhiNode::make( prev, old );
1361 // Now recursively fix up the new uses of old!
1362 for( uint i = 1; i < prev->req(); i++ ) {
1363 worklist.push(phi); // Onto worklist once for each 'old' input
1364 }
1365 }
1366 } else {
1367 // Get new RegionNode merging old and new loop exits
1368 prev = old_new[prev->_idx];
1369 assert( prev, "just made this in step 7" );
1370 if( idx == 0 ) { // Updating control edge?
1371 phi = prev; // Just use existing control
1372 } else { // Else need a new Phi
1373 // Make a new Phi merging data values properly
1374 phi = PhiNode::make( prev, old );
1375 phi->set_req( 1, nnn );
1376 }
1377 }
1378 // If inserting a new Phi, check for prior hits
1379 if( idx != 0 ) {
1380 Node *hit = _igvn.hash_find_insert(phi);
1381 if( hit == NULL ) {
1382 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1383 } else { // or
1384 // Remove the new phi from the graph and use the hit
1385 _igvn.remove_dead_node(phi);
1386 phi = hit; // Use existing phi
1387 }
1388 set_ctrl(phi, prev);
1389 }
1390 // Make 'use' use the Phi instead of the old loop body exit value
1391 _igvn.hash_delete(use);
1392 _igvn._worklist.push(use);
1393 use->set_req(idx, phi);
1394 if( use->_idx >= new_counter ) { // If updating new phis
1395 // Not needed for correctness, but prevents a weak assert
1396 // in AddPNode from tripping (when we end up with different
1397 // base & derived Phis that will become the same after
1398 // IGVN does CSE).
1399 Node *hit = _igvn.hash_find_insert(use);
1400 if( hit ) // Go ahead and re-hash for hits.
1401 _igvn.subsume_node( use, hit );
1402 }
1403
1404 // If 'use' was in the loop-exit block, it now needs to be sunk
1405 // below the post-loop merge point.
1406 sink_use( use, prev );
1407 }
1408 }
1409 }
1410
1411 // Check for IFs that need splitting/cloning. Happens if an IF outside of
1412 // the loop uses a condition set in the loop. The original IF probably
1413 // takes control from one or more OLD Regions (which in turn get from NEW
1414 // Regions). In any case, there will be a set of Phis for each merge point
1415 // from the IF up to where the original BOOL def exists the loop.
1416 if( split_if_set ) {
1417 while( split_if_set->size() ) {
1418 Node *iff = split_if_set->pop();
1419 if( iff->in(1)->is_Phi() ) {
1420 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop );
1421 _igvn.hash_delete(iff);
1422 _igvn._worklist.push(iff);
1423 iff->set_req(1, b);
1424 }
1425 }
1426 }
1427 if( split_bool_set ) {
1428 while( split_bool_set->size() ) {
1429 Node *b = split_bool_set->pop();
1430 Node *phi = b->in(1);
1431 assert( phi->is_Phi(), "" );
1432 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop );
1433 _igvn.hash_delete(b);
1434 _igvn._worklist.push(b);
1435 b->set_req(1, cmp);
1436 }
1437 }
1438 if( split_cex_set ) {
1439 while( split_cex_set->size() ) {
1440 Node *b = split_cex_set->pop();
1441 assert( b->in(0)->is_Region(), "" );
1442 assert( b->in(1)->is_Phi(), "" );
1443 assert( b->in(0)->in(0) == b->in(1)->in(0), "" );
1444 split_up( b, b->in(0), NULL );
1445 }
1446 }
1447
1448 }
1449
1450
1451 //---------------------- stride_of_possible_iv -------------------------------------
1452 // Looks for an iff/bool/comp with one operand of the compare
1453 // being a cycle involving an add and a phi,
1454 // with an optional truncation (left-shift followed by a right-shift)
1455 // of the add. Returns zero if not an iv.
1456 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
1457 Node* trunc1 = NULL;
1458 Node* trunc2 = NULL;
1459 const TypeInt* ttype = NULL;
1460 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
1461 return 0;
1462 }
1463 BoolNode* bl = iff->in(1)->as_Bool();
1464 Node* cmp = bl->in(1);
1465 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) {
1466 return 0;
1467 }
1468 // Must have an invariant operand
1469 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
1470 return 0;
1471 }
1472 Node* add2 = NULL;
1473 Node* cmp1 = cmp->in(1);
1474 if (cmp1->is_Phi()) {
1475 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
1476 Node* phi = cmp1;
1477 for (uint i = 1; i < phi->req(); i++) {
1478 Node* in = phi->in(i);
1479 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
1480 &trunc1, &trunc2, &ttype);
1481 if (add && add->in(1) == phi) {
1482 add2 = add->in(2);
1483 break;
1484 }
1485 }
1486 } else {
1487 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
1488 Node* addtrunc = cmp1;
1489 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
1490 &trunc1, &trunc2, &ttype);
1491 if (add && add->in(1)->is_Phi()) {
1492 Node* phi = add->in(1);
1493 for (uint i = 1; i < phi->req(); i++) {
1494 if (phi->in(i) == addtrunc) {
1495 add2 = add->in(2);
1496 break;
1497 }
1498 }
1499 }
1500 }
1501 if (add2 != NULL) {
1502 const TypeInt* add2t = _igvn.type(add2)->is_int();
1503 if (add2t->is_con()) {
1504 return add2t->get_con();
1505 }
1506 }
1507 return 0;
1508 }
1509
1510
1511 //---------------------- stay_in_loop -------------------------------------
1512 // Return the (unique) control output node that's in the loop (if it exists.)
1513 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
1514 Node* unique = NULL;
1515 if (!n) return NULL;
1516 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1517 Node* use = n->fast_out(i);
1518 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
1519 if (unique != NULL) {
1520 return NULL;
1521 }
1522 unique = use;
1523 }
1524 }
1525 return unique;
1526 }
1527
1528 //------------------------------ register_node -------------------------------------
1529 // Utility to register node "n" with PhaseIdealLoop
1530 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
1531 _igvn.register_new_node_with_optimizer(n);
1532 loop->_body.push(n);
1533 if (n->is_CFG()) {
1534 set_loop(n, loop);
1535 set_idom(n, pred, ddepth);
1536 } else {
1537 set_ctrl(n, pred);
1538 }
1539 }
1540
1541 //------------------------------ proj_clone -------------------------------------
1542 // Utility to create an if-projection
1543 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
1544 ProjNode* c = p->clone()->as_Proj();
1545 c->set_req(0, iff);
1546 return c;
1547 }
1548
1549 //------------------------------ short_circuit_if -------------------------------------
1550 // Force the iff control output to be the live_proj
1551 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
1552 int proj_con = live_proj->_con;
1553 assert(proj_con == 0 || proj_con == 1, "false or true projection");
1554 Node *con = _igvn.intcon(proj_con);
1555 set_ctrl(con, C->root());
1556 if (iff) {
1557 iff->set_req(1, con);
1558 }
1559 return con;
1560 }
1561
1562 //------------------------------ insert_if_before_proj -------------------------------------
1563 // Insert a new if before an if projection (* - new node)
1564 //
1565 // before
1566 // if(test)
1567 // / \
1568 // v v
1569 // other-proj proj (arg)
1570 //
1571 // after
1572 // if(test)
1573 // / \
1574 // / v
1575 // | * proj-clone
1576 // v |
1577 // other-proj v
1578 // * new_if(relop(cmp[IU](left,right)))
1579 // / \
1580 // v v
1581 // * new-proj proj
1582 // (returned)
1583 //
1584 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
1585 IfNode* iff = proj->in(0)->as_If();
1586 IdealLoopTree *loop = get_loop(proj);
1587 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1588 int ddepth = dom_depth(proj);
1589
1590 _igvn.hash_delete(iff);
1591 _igvn._worklist.push(iff);
1592 _igvn.hash_delete(proj);
1593 _igvn._worklist.push(proj);
1594
1595 proj->set_req(0, NULL); // temporary disconnect
1596 ProjNode* proj2 = proj_clone(proj, iff);
1597 register_node(proj2, loop, iff, ddepth);
1598
1599 Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right);
1600 register_node(cmp, loop, proj2, ddepth);
1601
1602 BoolNode* bol = new (C,2)BoolNode(cmp, relop);
1603 register_node(bol, loop, proj2, ddepth);
1604
1605 IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt);
1606 register_node(new_if, loop, proj2, ddepth);
1607
1608 proj->set_req(0, new_if); // reattach
1609 set_idom(proj, new_if, ddepth);
1610
1611 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
1612 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
1613
1614 return new_exit;
1615 }
1616
1617 //------------------------------ insert_region_before_proj -------------------------------------
1618 // Insert a region before an if projection (* - new node)
1619 //
1620 // before
1621 // if(test)
1622 // / |
1623 // v |
1624 // proj v
1625 // other-proj
1626 //
1627 // after
1628 // if(test)
1629 // / |
1630 // v |
1631 // * proj-clone v
1632 // | other-proj
1633 // v
1634 // * new-region
1635 // |
1636 // v
1637 // * dum_if
1638 // / \
1639 // v \
1640 // * dum-proj v
1641 // proj
1642 //
1643 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
1644 IfNode* iff = proj->in(0)->as_If();
1645 IdealLoopTree *loop = get_loop(proj);
1646 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1647 int ddepth = dom_depth(proj);
1648
1649 _igvn.hash_delete(iff);
1650 _igvn._worklist.push(iff);
1651 _igvn.hash_delete(proj);
1652 _igvn._worklist.push(proj);
1653
1654 proj->set_req(0, NULL); // temporary disconnect
1655 ProjNode* proj2 = proj_clone(proj, iff);
1656 register_node(proj2, loop, iff, ddepth);
1657
1658 RegionNode* reg = new (C,2)RegionNode(2);
1659 reg->set_req(1, proj2);
1660 register_node(reg, loop, iff, ddepth);
1661
1662 IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
1663 register_node(dum_if, loop, reg, ddepth);
1664
1665 proj->set_req(0, dum_if); // reattach
1666 set_idom(proj, dum_if, ddepth);
1667
1668 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
1669 register_node(dum_proj, loop, dum_if, ddepth);
1670
1671 return reg;
1672 }
1673
1674 //------------------------------ insert_cmpi_loop_exit -------------------------------------
1675 // Clone a signed compare loop exit from an unsigned compare and
1676 // insert it before the unsigned cmp on the stay-in-loop path.
1677 // All new nodes inserted in the dominator tree between the original
1678 // if and it's projections. The original if test is replaced with
1679 // a constant to force the stay-in-loop path.
1680 //
1681 // This is done to make sure that the original if and it's projections
1682 // still dominate the same set of control nodes, that the ctrl() relation
1683 // from data nodes to them is preserved, and that their loop nesting is
1684 // preserved.
1685 //
1686 // before
1687 // if(i <u limit) unsigned compare loop exit
1688 // / |
1689 // v v
1690 // exit-proj stay-in-loop-proj
1691 //
1692 // after
1693 // if(stay-in-loop-const) original if
1694 // / |
1695 // / v
1696 // / if(i < limit) new signed test
1697 // / / |
1698 // / / v
1699 // / / if(i <u limit) new cloned unsigned test
1700 // / / / |
1701 // v v v |
1702 // region |
1703 // | |
1704 // dum-if |
1705 // / | |
1706 // ether | |
1707 // v v
1708 // exit-proj stay-in-loop-proj
1709 //
1710 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
1711 const bool Signed = true;
1712 const bool Unsigned = false;
1713
1714 BoolNode* bol = if_cmpu->in(1)->as_Bool();
1715 if (bol->_test._test != BoolTest::lt) return NULL;
1716 CmpNode* cmpu = bol->in(1)->as_Cmp();
1717 if (cmpu->Opcode() != Op_CmpU) return NULL;
1718 int stride = stride_of_possible_iv(if_cmpu);
1719 if (stride == 0) return NULL;
1720
1721 ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj();
1722 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
1723
1724 Node* limit = NULL;
1725 if (stride > 0) {
1726 limit = cmpu->in(2);
1727 } else {
1728 limit = _igvn.makecon(TypeInt::ZERO);
1729 set_ctrl(limit, C->root());
1730 }
1731 // Create a new region on the exit path
1732 RegionNode* reg = insert_region_before_proj(lp_exit);
1733
1734 // Clone the if-cmpu-true-false using a signed compare
1735 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
1736 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
1737 reg->add_req(cmpi_exit);
1738
1739 // Clone the if-cmpu-true-false
1740 BoolTest::mask rel_u = bol->_test._test;
1741 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
1742 reg->add_req(cmpu_exit);
1743
1744 // Force original if to stay in loop.
1745 short_circuit_if(if_cmpu, lp_continue);
1746
1747 return cmpi_exit->in(0)->as_If();
1748 }
1749
1750 //------------------------------ remove_cmpi_loop_exit -------------------------------------
1751 // Remove a previously inserted signed compare loop exit.
1752 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
1753 Node* lp_proj = stay_in_loop(if_cmp, loop);
1754 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
1755 stay_in_loop(lp_proj, loop)->is_If() &&
1756 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
1757 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
1758 set_ctrl(con, C->root());
1759 if_cmp->set_req(1, con);
1760 }
1761
1762 //------------------------------ scheduled_nodelist -------------------------------------
1763 // Create a post order schedule of nodes that are in the
1764 // "member" set. The list is returned in "sched".
1765 // The first node in "sched" is the loop head, followed by
1766 // nodes which have no inputs in the "member" set, and then
1767 // followed by the nodes that have an immediate input dependence
1768 // on a node in "sched".
1769 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
1770
1771 assert(member.test(loop->_head->_idx), "loop head must be in member set");
1772 Arena *a = Thread::current()->resource_area();
1773 VectorSet visited(a);
1774 Node_Stack nstack(a, loop->_body.size());
1775
1776 Node* n = loop->_head; // top of stack is cached in "n"
1777 uint idx = 0;
1778 visited.set(n->_idx);
1779
1780 // Initially push all with no inputs from within member set
1781 for(uint i = 0; i < loop->_body.size(); i++ ) {
1782 Node *elt = loop->_body.at(i);
1783 if (member.test(elt->_idx)) {
1784 bool found = false;
1785 for (uint j = 0; j < elt->req(); j++) {
1786 Node* def = elt->in(j);
1787 if (def && member.test(def->_idx) && def != elt) {
1788 found = true;
1789 break;
1790 }
1791 }
1792 if (!found && elt != loop->_head) {
1793 nstack.push(n, idx);
1794 n = elt;
1795 assert(!visited.test(n->_idx), "not seen yet");
1796 visited.set(n->_idx);
1797 }
1798 }
1799 }
1800
1801 // traverse out's that are in the member set
1802 while (true) {
1803 if (idx < n->outcnt()) {
1804 Node* use = n->raw_out(idx);
1805 idx++;
1806 if (!visited.test_set(use->_idx)) {
1807 if (member.test(use->_idx)) {
1808 nstack.push(n, idx);
1809 n = use;
1810 idx = 0;
1811 }
1812 }
1813 } else {
1814 // All outputs processed
1815 sched.push(n);
1816 if (nstack.is_empty()) break;
1817 n = nstack.node();
1818 idx = nstack.index();
1819 nstack.pop();
1820 }
1821 }
1822 }
1823
1824
1825 //------------------------------ has_use_in_set -------------------------------------
1826 // Has a use in the vector set
1827 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
1828 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1829 Node* use = n->fast_out(j);
1830 if (vset.test(use->_idx)) {
1831 return true;
1832 }
1833 }
1834 return false;
1835 }
1836
1837
1838 //------------------------------ has_use_internal_to_set -------------------------------------
1839 // Has use internal to the vector set (ie. not in a phi at the loop head)
1840 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
1841 Node* head = loop->_head;
1842 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1843 Node* use = n->fast_out(j);
1844 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
1845 return true;
1846 }
1847 }
1848 return false;
1849 }
1850
1851
1852 //------------------------------ clone_for_use_outside_loop -------------------------------------
1853 // clone "n" for uses that are outside of loop
1854 void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
1855
1856 assert(worklist.size() == 0, "should be empty");
1857 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1858 Node* use = n->fast_out(j);
1859 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
1860 worklist.push(use);
1861 }
1862 }
1863 while( worklist.size() ) {
1864 Node *use = worklist.pop();
1865 if (!has_node(use) || use->in(0) == C->top()) continue;
1866 uint j;
1867 for (j = 0; j < use->req(); j++) {
1868 if (use->in(j) == n) break;
1869 }
1870 assert(j < use->req(), "must be there");
1871
1872 // clone "n" and insert it between the inputs of "n" and the use outside the loop
1873 Node* n_clone = n->clone();
1874 _igvn.hash_delete(use);
1875 use->set_req(j, n_clone);
1876 _igvn._worklist.push(use);
1877 if (!use->is_Phi()) {
1878 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
1879 set_ctrl(n_clone, use_c);
1880 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
1881 get_loop(use_c)->_body.push(n_clone);
1882 } else {
1883 // Use in a phi is considered a use in the associated predecessor block
1884 Node *prevbb = use->in(0)->in(j);
1885 set_ctrl(n_clone, prevbb);
1886 assert(!loop->is_member(get_loop(prevbb)), "should be outside loop");
1887 get_loop(prevbb)->_body.push(n_clone);
1888 }
1889 _igvn.register_new_node_with_optimizer(n_clone);
1890 #if !defined(PRODUCT)
1891 if (TracePartialPeeling) {
1892 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
1893 }
1894 #endif
1895 }
1896 }
1897
1898
1899 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
1900 // clone "n" for special uses that are in the not_peeled region.
1901 // If these def-uses occur in separate blocks, the code generator
1902 // marks the method as not compilable. For example, if a "BoolNode"
1903 // is in a different basic block than the "IfNode" that uses it, then
1904 // the compilation is aborted in the code generator.
1905 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1906 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
1907 if (n->is_Phi() || n->is_Load()) {
1908 return;
1909 }
1910 assert(worklist.size() == 0, "should be empty");
1911 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1912 Node* use = n->fast_out(j);
1913 if ( not_peel.test(use->_idx) &&
1914 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
1915 use->in(1) == n) {
1916 worklist.push(use);
1917 }
1918 }
1919 if (worklist.size() > 0) {
1920 // clone "n" and insert it between inputs of "n" and the use
1921 Node* n_clone = n->clone();
1922 loop->_body.push(n_clone);
1923 _igvn.register_new_node_with_optimizer(n_clone);
1924 set_ctrl(n_clone, get_ctrl(n));
1925 sink_list.push(n_clone);
1926 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
1927 #if !defined(PRODUCT)
1928 if (TracePartialPeeling) {
1929 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
1930 }
1931 #endif
1932 while( worklist.size() ) {
1933 Node *use = worklist.pop();
1934 _igvn.hash_delete(use);
1935 _igvn._worklist.push(use);
1936 for (uint j = 1; j < use->req(); j++) {
1937 if (use->in(j) == n) {
1938 use->set_req(j, n_clone);
1939 }
1940 }
1941 }
1942 }
1943 }
1944
1945
1946 //------------------------------ insert_phi_for_loop -------------------------------------
1947 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1948 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
1949 Node *phi = PhiNode::make(lp, back_edge_val);
1950 phi->set_req(LoopNode::EntryControl, lp_entry_val);
1951 // Use existing phi if it already exists
1952 Node *hit = _igvn.hash_find_insert(phi);
1953 if( hit == NULL ) {
1954 _igvn.register_new_node_with_optimizer(phi);
1955 set_ctrl(phi, lp);
1956 } else {
1957 // Remove the new phi from the graph and use the hit
1958 _igvn.remove_dead_node(phi);
1959 phi = hit;
1960 }
1961 _igvn.hash_delete(use);
1962 _igvn._worklist.push(use);
1963 use->set_req(idx, phi);
1964 }
1965
1966 #ifdef ASSERT
1967 //------------------------------ is_valid_loop_partition -------------------------------------
1968 // Validate the loop partition sets: peel and not_peel
1969 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
1970 VectorSet& not_peel ) {
1971 uint i;
1972 // Check that peel_list entries are in the peel set
1973 for (i = 0; i < peel_list.size(); i++) {
1974 if (!peel.test(peel_list.at(i)->_idx)) {
1975 return false;
1976 }
1977 }
1978 // Check at loop members are in one of peel set or not_peel set
1979 for (i = 0; i < loop->_body.size(); i++ ) {
1980 Node *def = loop->_body.at(i);
1981 uint di = def->_idx;
1982 // Check that peel set elements are in peel_list
1983 if (peel.test(di)) {
1984 if (not_peel.test(di)) {
1985 return false;
1986 }
1987 // Must be in peel_list also
1988 bool found = false;
1989 for (uint j = 0; j < peel_list.size(); j++) {
1990 if (peel_list.at(j)->_idx == di) {
1991 found = true;
1992 break;
1993 }
1994 }
1995 if (!found) {
1996 return false;
1997 }
1998 } else if (not_peel.test(di)) {
1999 if (peel.test(di)) {
2000 return false;
2001 }
2002 } else {
2003 return false;
2004 }
2005 }
2006 return true;
2007 }
2008
2009 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2010 // Ensure a use outside of loop is of the right form
2011 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2012 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2013 return (use->is_Phi() &&
2014 use_c->is_Region() && use_c->req() == 3 &&
2015 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2016 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2017 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2018 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2019 }
2020
2021 //------------------------------ is_valid_clone_loop_form -------------------------------------
2022 // Ensure that all uses outside of loop are of the right form
2023 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2024 uint orig_exit_idx, uint clone_exit_idx) {
2025 uint len = peel_list.size();
2026 for (uint i = 0; i < len; i++) {
2027 Node *def = peel_list.at(i);
2028
2029 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2030 Node *use = def->fast_out(j);
2031 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2032 if (!loop->is_member(get_loop(use_c))) {
2033 // use is not in the loop, check for correct structure
2034 if (use->in(0) == def) {
2035 // Okay
2036 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2037 return false;
2038 }
2039 }
2040 }
2041 }
2042 return true;
2043 }
2044 #endif
2045
2046 //------------------------------ partial_peel -------------------------------------
2047 // Partially peel (aka loop rotation) the top portion of a loop (called
2048 // the peel section below) by cloning it and placing one copy just before
2049 // the new loop head and the other copy at the bottom of the new loop.
2050 //
2051 // before after where it came from
2052 //
2053 // stmt1 stmt1
2054 // loop: stmt2 clone
2055 // stmt2 if condA goto exitA clone
2056 // if condA goto exitA new_loop: new
2057 // stmt3 stmt3 clone
2058 // if !condB goto loop if condB goto exitB clone
2059 // exitB: stmt2 orig
2060 // stmt4 if !condA goto new_loop orig
2061 // exitA: goto exitA
2062 // exitB:
2063 // stmt4
2064 // exitA:
2065 //
2066 // Step 1: find the cut point: an exit test on probable
2067 // induction variable.
2068 // Step 2: schedule (with cloning) operations in the peel
2069 // section that can be executed after the cut into
2070 // the section that is not peeled. This may need
2071 // to clone operations into exit blocks. For
2072 // instance, a reference to A[i] in the not-peel
2073 // section and a reference to B[i] in an exit block
2074 // may cause a left-shift of i by 2 to be placed
2075 // in the peel block. This step will clone the left
2076 // shift into the exit block and sink the left shift
2077 // from the peel to the not-peel section.
2078 // Step 3: clone the loop, retarget the control, and insert
2079 // phis for values that are live across the new loop
2080 // head. This is very dependent on the graph structure
2081 // from clone_loop. It creates region nodes for
2082 // exit control and associated phi nodes for values
2083 // flow out of the loop through that exit. The region
2084 // node is dominated by the clone's control projection.
2085 // So the clone's peel section is placed before the
2086 // new loop head, and the clone's not-peel section is
2087 // forms the top part of the new loop. The original
2088 // peel section forms the tail of the new loop.
2089 // Step 4: update the dominator tree and recompute the
2090 // dominator depth.
2091 //
2092 // orig
2093 //
2094 // stmt1
2095 // |
2096 // v
2097 // loop<----+
2098 // | |
2099 // stmt2 |
2100 // | |
2101 // v |
2102 // ifA |
2103 // / | |
2104 // v v |
2105 // false true ^ <-- last_peel
2106 // / | |
2107 // / ===|==cut |
2108 // / stmt3 | <-- first_not_peel
2109 // / | |
2110 // | v |
2111 // v ifB |
2112 // exitA: / \ |
2113 // / \ |
2114 // v v |
2115 // false true |
2116 // / \ |
2117 // / ----+
2118 // |
2119 // v
2120 // exitB:
2121 // stmt4
2122 //
2123 //
2124 // after clone loop
2125 //
2126 // stmt1
2127 // / \
2128 // clone / \ orig
2129 // / \
2130 // / \
2131 // v v
2132 // +---->loop loop<----+
2133 // | | | |
2134 // | stmt2 stmt2 |
2135 // | | | |
2136 // | v v |
2137 // | ifA ifA |
2138 // | | \ / | |
2139 // | v v v v |
2140 // ^ true false false true ^ <-- last_peel
2141 // | | ^ \ / | |
2142 // | cut==|== \ \ / ===|==cut |
2143 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2144 // | | dom | | | |
2145 // | v \ 1v v2 v |
2146 // | ifB regionA ifB |
2147 // | / \ | / \ |
2148 // | / \ v / \ |
2149 // | v v exitA: v v |
2150 // | true false false true |
2151 // | / ^ \ / \ |
2152 // +---- \ \ / ----+
2153 // dom \ /
2154 // \ 1v v2
2155 // regionB
2156 // |
2157 // v
2158 // exitB:
2159 // stmt4
2160 //
2161 //
2162 // after partial peel
2163 //
2164 // stmt1
2165 // /
2166 // clone / orig
2167 // / TOP
2168 // / \
2169 // v v
2170 // TOP->region region----+
2171 // | | |
2172 // stmt2 stmt2 |
2173 // | | |
2174 // v v |
2175 // ifA ifA |
2176 // | \ / | |
2177 // v v v v |
2178 // true false false true | <-- last_peel
2179 // | ^ \ / +------|---+
2180 // +->newloop \ \ / === ==cut | |
2181 // | stmt3 \ \ / TOP | |
2182 // | | dom | | stmt3 | | <-- first_not_peel
2183 // | v \ 1v v2 v | |
2184 // | ifB regionA ifB ^ v
2185 // | / \ | / \ | |
2186 // | / \ v / \ | |
2187 // | v v exitA: v v | |
2188 // | true false false true | |
2189 // | / ^ \ / \ | |
2190 // | | \ \ / v | |
2191 // | | dom \ / TOP | |
2192 // | | \ 1v v2 | |
2193 // ^ v regionB | |
2194 // | | | | |
2195 // | | v ^ v
2196 // | | exitB: | |
2197 // | | stmt4 | |
2198 // | +------------>-----------------+ |
2199 // | |
2200 // +-----------------<---------------------+
2201 //
2202 //
2203 // final graph
2204 //
2205 // stmt1
2206 // |
2207 // v
2208 // ........> ifA clone
2209 // : / |
2210 // dom / |
2211 // : v v
2212 // : false true
2213 // : | |
2214 // : | stmt2 clone
2215 // : | |
2216 // : | v
2217 // : | newloop<-----+
2218 // : | | |
2219 // : | stmt3 clone |
2220 // : | | |
2221 // : | v |
2222 // : | ifB |
2223 // : | / \ |
2224 // : | v v |
2225 // : | false true |
2226 // : | | | |
2227 // : | v stmt2 |
2228 // : | exitB: | |
2229 // : | stmt4 v |
2230 // : | ifA orig |
2231 // : | / \ |
2232 // : | / \ |
2233 // : | v v |
2234 // : | false true |
2235 // : | / \ |
2236 // : v v -----+
2237 // RegionA
2238 // |
2239 // v
2240 // exitA
2241 //
2242 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
2243
2244 LoopNode *head = loop->_head->as_Loop();
2245
2246 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
2247 return false;
2248 }
2249
2250 // Check for complex exit control
2251 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
2252 Node *n = loop->_body.at(ii);
2253 int opc = n->Opcode();
2254 if (n->is_Call() ||
2255 opc == Op_Catch ||
2256 opc == Op_CatchProj ||
2257 opc == Op_Jump ||
2258 opc == Op_JumpProj) {
2259 #if !defined(PRODUCT)
2260 if (TracePartialPeeling) {
2261 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
2262 }
2263 #endif
2264 return false;
2265 }
2266 }
2267
2268 int dd = dom_depth(head);
2269
2270 // Step 1: find cut point
2271
2272 // Walk up dominators to loop head looking for first loop exit
2273 // which is executed on every path thru loop.
2274 IfNode *peel_if = NULL;
2275 IfNode *peel_if_cmpu = NULL;
2276
2277 Node *iff = loop->tail();
2278 while( iff != head ) {
2279 if( iff->is_If() ) {
2280 Node *ctrl = get_ctrl(iff->in(1));
2281 if (ctrl->is_top()) return false; // Dead test on live IF.
2282 // If loop-varying exit-test, check for induction variable
2283 if( loop->is_member(get_loop(ctrl)) &&
2284 loop->is_loop_exit(iff) &&
2285 is_possible_iv_test(iff)) {
2286 Node* cmp = iff->in(1)->in(1);
2287 if (cmp->Opcode() == Op_CmpI) {
2288 peel_if = iff->as_If();
2289 } else {
2290 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
2291 peel_if_cmpu = iff->as_If();
2292 }
2293 }
2294 }
2295 iff = idom(iff);
2296 }
2297 // Prefer signed compare over unsigned compare.
2298 IfNode* new_peel_if = NULL;
2299 if (peel_if == NULL) {
2300 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
2301 return false; // No peel point found
2302 }
2303 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
2304 if (new_peel_if == NULL) {
2305 return false; // No peel point found
2306 }
2307 peel_if = new_peel_if;
2308 }
2309 Node* last_peel = stay_in_loop(peel_if, loop);
2310 Node* first_not_peeled = stay_in_loop(last_peel, loop);
2311 if (first_not_peeled == NULL || first_not_peeled == head) {
2312 return false;
2313 }
2314
2315 #if !defined(PRODUCT)
2316 if (TracePartialPeeling) {
2317 tty->print_cr("before partial peel one iteration");
2318 Node_List wl;
2319 Node* t = head->in(2);
2320 while (true) {
2321 wl.push(t);
2322 if (t == head) break;
2323 t = idom(t);
2324 }
2325 while (wl.size() > 0) {
2326 Node* tt = wl.pop();
2327 tt->dump();
2328 if (tt == last_peel) tty->print_cr("-- cut --");
2329 }
2330 }
2331 #endif
2332 ResourceArea *area = Thread::current()->resource_area();
2333 VectorSet peel(area);
2334 VectorSet not_peel(area);
2335 Node_List peel_list(area);
2336 Node_List worklist(area);
2337 Node_List sink_list(area);
2338
2339 // Set of cfg nodes to peel are those that are executable from
2340 // the head through last_peel.
2341 assert(worklist.size() == 0, "should be empty");
2342 worklist.push(head);
2343 peel.set(head->_idx);
2344 while (worklist.size() > 0) {
2345 Node *n = worklist.pop();
2346 if (n != last_peel) {
2347 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2348 Node* use = n->fast_out(j);
2349 if (use->is_CFG() &&
2350 loop->is_member(get_loop(use)) &&
2351 !peel.test_set(use->_idx)) {
2352 worklist.push(use);
2353 }
2354 }
2355 }
2356 }
2357
2358 // Set of non-cfg nodes to peel are those that are control
2359 // dependent on the cfg nodes.
2360 uint i;
2361 for(i = 0; i < loop->_body.size(); i++ ) {
2362 Node *n = loop->_body.at(i);
2363 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
2364 if (peel.test(n_c->_idx)) {
2365 peel.set(n->_idx);
2366 } else {
2367 not_peel.set(n->_idx);
2368 }
2369 }
2370
2371 // Step 2: move operations from the peeled section down into the
2372 // not-peeled section
2373
2374 // Get a post order schedule of nodes in the peel region
2375 // Result in right-most operand.
2376 scheduled_nodelist(loop, peel, peel_list );
2377
2378 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2379
2380 // For future check for too many new phis
2381 uint old_phi_cnt = 0;
2382 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2383 Node* use = head->fast_out(j);
2384 if (use->is_Phi()) old_phi_cnt++;
2385 }
2386
2387 #if !defined(PRODUCT)
2388 if (TracePartialPeeling) {
2389 tty->print_cr("\npeeled list");
2390 }
2391 #endif
2392
2393 // Evacuate nodes in peel region into the not_peeled region if possible
2394 uint new_phi_cnt = 0;
2395 for (i = 0; i < peel_list.size();) {
2396 Node* n = peel_list.at(i);
2397 #if !defined(PRODUCT)
2398 if (TracePartialPeeling) n->dump();
2399 #endif
2400 bool incr = true;
2401 if ( !n->is_CFG() ) {
2402
2403 if ( has_use_in_set(n, not_peel) ) {
2404
2405 // If not used internal to the peeled region,
2406 // move "n" from peeled to not_peeled region.
2407
2408 if ( !has_use_internal_to_set(n, peel, loop) ) {
2409
2410 // if not pinned and not a load (which maybe anti-dependent on a store)
2411 // and not a CMove (Matcher expects only bool->cmove).
2412 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
2413 clone_for_use_outside_loop( loop, n, worklist );
2414
2415 sink_list.push(n);
2416 peel >>= n->_idx; // delete n from peel set.
2417 not_peel <<= n->_idx; // add n to not_peel set.
2418 peel_list.remove(i);
2419 incr = false;
2420 #if !defined(PRODUCT)
2421 if (TracePartialPeeling) {
2422 tty->print_cr("sink to not_peeled region: %d newbb: %d",
2423 n->_idx, get_ctrl(n)->_idx);
2424 }
2425 #endif
2426 }
2427 } else {
2428 // Otherwise check for special def-use cases that span
2429 // the peel/not_peel boundary such as bool->if
2430 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
2431 new_phi_cnt++;
2432 }
2433 }
2434 }
2435 if (incr) i++;
2436 }
2437
2438 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
2439 #if !defined(PRODUCT)
2440 if (TracePartialPeeling) {
2441 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
2442 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
2443 }
2444 #endif
2445 if (new_peel_if != NULL) {
2446 remove_cmpi_loop_exit(new_peel_if, loop);
2447 }
2448 // Inhibit more partial peeling on this loop
2449 assert(!head->is_partial_peel_loop(), "not partial peeled");
2450 head->mark_partial_peel_failed();
2451 return false;
2452 }
2453
2454 // Step 3: clone loop, retarget control, and insert new phis
2455
2456 // Create new loop head for new phis and to hang
2457 // the nodes being moved (sinked) from the peel region.
2458 LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel);
2459 _igvn.register_new_node_with_optimizer(new_head);
2460 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
2461 first_not_peeled->set_req(0, new_head);
2462 set_loop(new_head, loop);
2463 loop->_body.push(new_head);
2464 not_peel.set(new_head->_idx);
2465 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
2466 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
2467
2468 while (sink_list.size() > 0) {
2469 Node* n = sink_list.pop();
2470 set_ctrl(n, new_head);
2471 }
2472
2473 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2474
2475 clone_loop( loop, old_new, dd );
2476
2477 const uint clone_exit_idx = 1;
2478 const uint orig_exit_idx = 2;
2479 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
2480
2481 Node* head_clone = old_new[head->_idx];
2482 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
2483 Node* orig_tail_clone = head_clone->in(2);
2484
2485 // Add phi if "def" node is in peel set and "use" is not
2486
2487 for(i = 0; i < peel_list.size(); i++ ) {
2488 Node *def = peel_list.at(i);
2489 if (!def->is_CFG()) {
2490 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2491 Node *use = def->fast_out(j);
2492 if (has_node(use) && use->in(0) != C->top() &&
2493 (!peel.test(use->_idx) ||
2494 (use->is_Phi() && use->in(0) == head)) ) {
2495 worklist.push(use);
2496 }
2497 }
2498 while( worklist.size() ) {
2499 Node *use = worklist.pop();
2500 for (uint j = 1; j < use->req(); j++) {
2501 Node* n = use->in(j);
2502 if (n == def) {
2503
2504 // "def" is in peel set, "use" is not in peel set
2505 // or "use" is in the entry boundary (a phi) of the peel set
2506
2507 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
2508
2509 if ( loop->is_member(get_loop( use_c )) ) {
2510 // use is in loop
2511 if (old_new[use->_idx] != NULL) { // null for dead code
2512 Node* use_clone = old_new[use->_idx];
2513 _igvn.hash_delete(use);
2514 use->set_req(j, C->top());
2515 _igvn._worklist.push(use);
2516 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
2517 }
2518 } else {
2519 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
2520 // use is not in the loop, check if the live range includes the cut
2521 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
2522 if (not_peel.test(lp_if->_idx)) {
2523 assert(j == orig_exit_idx, "use from original loop");
2524 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
2525 }
2526 }
2527 }
2528 }
2529 }
2530 }
2531 }
2532
2533 // Step 3b: retarget control
2534
2535 // Redirect control to the new loop head if a cloned node in
2536 // the not_peeled region has control that points into the peeled region.
2537 // This necessary because the cloned peeled region will be outside
2538 // the loop.
2539 // from to
2540 // cloned-peeled <---+
2541 // new_head_clone: | <--+
2542 // cloned-not_peeled in(0) in(0)
2543 // orig-peeled
2544
2545 for(i = 0; i < loop->_body.size(); i++ ) {
2546 Node *n = loop->_body.at(i);
2547 if (!n->is_CFG() && n->in(0) != NULL &&
2548 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
2549 Node* n_clone = old_new[n->_idx];
2550 _igvn.hash_delete(n_clone);
2551 n_clone->set_req(0, new_head_clone);
2552 _igvn._worklist.push(n_clone);
2553 }
2554 }
2555
2556 // Backedge of the surviving new_head (the clone) is original last_peel
2557 _igvn.hash_delete(new_head_clone);
2558 new_head_clone->set_req(LoopNode::LoopBackControl, last_peel);
2559 _igvn._worklist.push(new_head_clone);
2560
2561 // Cut first node in original not_peel set
2562 _igvn.hash_delete(new_head);
2563 new_head->set_req(LoopNode::EntryControl, C->top());
2564 new_head->set_req(LoopNode::LoopBackControl, C->top());
2565 _igvn._worklist.push(new_head);
2566
2567 // Copy head_clone back-branch info to original head
2568 // and remove original head's loop entry and
2569 // clone head's back-branch
2570 _igvn.hash_delete(head);
2571 _igvn.hash_delete(head_clone);
2572 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
2573 head->set_req(LoopNode::LoopBackControl, C->top());
2574 head_clone->set_req(LoopNode::LoopBackControl, C->top());
2575 _igvn._worklist.push(head);
2576 _igvn._worklist.push(head_clone);
2577
2578 // Similarly modify the phis
2579 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
2580 Node* use = head->fast_out(k);
2581 if (use->is_Phi() && use->outcnt() > 0) {
2582 Node* use_clone = old_new[use->_idx];
2583 _igvn.hash_delete(use);
2584 _igvn.hash_delete(use_clone);
2585 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
2586 use->set_req(LoopNode::LoopBackControl, C->top());
2587 use_clone->set_req(LoopNode::LoopBackControl, C->top());
2588 _igvn._worklist.push(use);
2589 _igvn._worklist.push(use_clone);
2590 }
2591 }
2592
2593 // Step 4: update dominator tree and dominator depth
2594
2595 set_idom(head, orig_tail_clone, dd);
2596 recompute_dom_depth();
2597
2598 // Inhibit more partial peeling on this loop
2599 new_head_clone->set_partial_peel_loop();
2600 C->set_major_progress();
2601
2602 #if !defined(PRODUCT)
2603 if (TracePartialPeeling) {
2604 tty->print_cr("\nafter partial peel one iteration");
2605 Node_List wl(area);
2606 Node* t = last_peel;
2607 while (true) {
2608 wl.push(t);
2609 if (t == head_clone) break;
2610 t = idom(t);
2611 }
2612 while (wl.size() > 0) {
2613 Node* tt = wl.pop();
2614 if (tt == head) tty->print_cr("orig head");
2615 else if (tt == new_head_clone) tty->print_cr("new head");
2616 else if (tt == head_clone) tty->print_cr("clone head");
2617 tt->dump();
2618 }
2619 }
2620 #endif
2621 return true;
2622 }
2623
2624 //------------------------------reorg_offsets----------------------------------
2625 // Reorganize offset computations to lower register pressure. Mostly
2626 // prevent loop-fallout uses of the pre-incremented trip counter (which are
2627 // then alive with the post-incremented trip counter forcing an extra
2628 // register move)
2629 void PhaseIdealLoop::reorg_offsets( IdealLoopTree *loop ) {
2630
2631 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2632 CountedLoopEndNode *cle = cl->loopexit();
2633 if( !cle ) return; // The occasional dead loop
2634 // Find loop exit control
2635 Node *exit = cle->proj_out(false);
2636 assert( exit->Opcode() == Op_IfFalse, "" );
2637
2638 // Check for the special case of folks using the pre-incremented
2639 // trip-counter on the fall-out path (forces the pre-incremented
2640 // and post-incremented trip counter to be live at the same time).
2641 // Fix this by adjusting to use the post-increment trip counter.
2642 Node *phi = cl->phi();
2643 if( !phi ) return; // Dead infinite loop
2644 bool progress = true;
2645 while (progress) {
2646 progress = false;
2647 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
2648 Node* use = phi->fast_out(i); // User of trip-counter
2649 if (!has_ctrl(use)) continue;
2650 Node *u_ctrl = get_ctrl(use);
2651 if( use->is_Phi() ) {
2652 u_ctrl = NULL;
2653 for( uint j = 1; j < use->req(); j++ )
2654 if( use->in(j) == phi )
2655 u_ctrl = dom_lca( u_ctrl, use->in(0)->in(j) );
2656 }
2657 IdealLoopTree *u_loop = get_loop(u_ctrl);
2658 // Look for loop-invariant use
2659 if( u_loop == loop ) continue;
2660 if( loop->is_member( u_loop ) ) continue;
2661 // Check that use is live out the bottom. Assuming the trip-counter
2662 // update is right at the bottom, uses of of the loop middle are ok.
2663 if( dom_lca( exit, u_ctrl ) != exit ) continue;
2664 // protect against stride not being a constant
2665 if( !cle->stride_is_con() ) continue;
2666 // Hit! Refactor use to use the post-incremented tripcounter.
2667 // Compute a post-increment tripcounter.
2668 Node *opaq = new (C, 2) Opaque2Node( cle->incr() );
2669 register_new_node( opaq, u_ctrl );
2670 Node *neg_stride = _igvn.intcon(-cle->stride_con());
2671 set_ctrl(neg_stride, C->root());
2672 Node *post = new (C, 3) AddINode( opaq, neg_stride);
2673 register_new_node( post, u_ctrl );
2674 _igvn.hash_delete(use);
2675 _igvn._worklist.push(use);
2676 for( uint j = 1; j < use->req(); j++ )
2677 if( use->in(j) == phi )
2678 use->set_req(j, post);
2679 // Since DU info changed, rerun loop
2680 progress = true;
2681 break;
2682 }
2683 }
2684
2685 }
2686