1 /*
2 * Copyright 1999-2006 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
24
25 #include "incls/_precompiled.incl"
26 #include "incls/_loopopts.cpp.incl"
27
28 //=============================================================================
29 //------------------------------split_thru_phi---------------------------------
30 // Split Node 'n' through merge point if there is enough win.
31 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
32 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
33 // ConvI2L may have type information on it which is unsafe to push up
34 // so disable this for now
35 return NULL;
36 }
37 int wins = 0;
38 assert( !n->is_CFG(), "" );
39 assert( region->is_Region(), "" );
40
41 const Type* type = n->bottom_type();
42 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr();
43 Node *phi;
44 if( t_oop != NULL && t_oop->is_instance_field() ) {
45 int iid = t_oop->instance_id();
46 int index = C->get_alias_index(t_oop);
47 int offset = t_oop->offset();
48 phi = new (C,region->req()) PhiNode(region, type, NULL, iid, index, offset);
49 } else {
50 phi = new (C,region->req()) PhiNode(region, type);
51 }
52 uint old_unique = C->unique();
53 for( uint i = 1; i < region->req(); i++ ) {
54 Node *x;
55 Node* the_clone = NULL;
56 if( region->in(i) == C->top() ) {
57 x = C->top(); // Dead path? Use a dead data op
58 } else {
59 x = n->clone(); // Else clone up the data op
60 the_clone = x; // Remember for possible deletion.
61 // Alter data node to use pre-phi inputs
62 if( n->in(0) == region )
63 x->set_req( 0, region->in(i) );
64 for( uint j = 1; j < n->req(); j++ ) {
65 Node *in = n->in(j);
66 if( in->is_Phi() && in->in(0) == region )
67 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
68 }
69 }
70 // Check for a 'win' on some paths
71 const Type *t = x->Value(&_igvn);
72
73 bool singleton = t->singleton();
74
75 // A TOP singleton indicates that there are no possible values incoming
76 // along a particular edge. In most cases, this is OK, and the Phi will
77 // be eliminated later in an Ideal call. However, we can't allow this to
78 // happen if the singleton occurs on loop entry, as the elimination of
79 // the PhiNode may cause the resulting node to migrate back to a previous
80 // loop iteration.
81 if( singleton && t == Type::TOP ) {
82 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
83 // irreducible loop may not be indicated by an affirmative is_Loop());
84 // therefore, the only top we can split thru a phi is on a backedge of
85 // a loop.
86 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
87 }
88
89 if( singleton ) {
90 wins++;
91 x = ((PhaseGVN&)_igvn).makecon(t);
92 } else {
93 // We now call Identity to try to simplify the cloned node.
94 // Note that some Identity methods call phase->type(this).
95 // Make sure that the type array is big enough for
96 // our new node, even though we may throw the node away.
97 // (Note: This tweaking with igvn only works because x is a new node.)
98 _igvn.set_type(x, t);
99 Node *y = x->Identity(&_igvn);
100 if( y != x ) {
101 wins++;
102 x = y;
103 } else {
104 y = _igvn.hash_find(x);
105 if( y ) {
106 wins++;
107 x = y;
108 } else {
109 // Else x is a new node we are keeping
110 // We do not need register_new_node_with_optimizer
111 // because set_type has already been called.
112 _igvn._worklist.push(x);
113 }
114 }
115 }
116 if (x != the_clone && the_clone != NULL)
117 _igvn.remove_dead_node(the_clone);
118 phi->set_req( i, x );
119 }
120 // Too few wins?
121 if( wins <= policy ) {
122 _igvn.remove_dead_node(phi);
123 return NULL;
124 }
125
126 // Record Phi
127 register_new_node( phi, region );
128
129 for( uint i2 = 1; i2 < phi->req(); i2++ ) {
130 Node *x = phi->in(i2);
131 // If we commoned up the cloned 'x' with another existing Node,
132 // the existing Node picks up a new use. We need to make the
133 // existing Node occur higher up so it dominates its uses.
134 Node *old_ctrl;
135 IdealLoopTree *old_loop;
136
137 // The occasional new node
138 if( x->_idx >= old_unique ) { // Found a new, unplaced node?
139 old_ctrl = x->is_Con() ? C->root() : NULL;
140 old_loop = NULL; // Not in any prior loop
141 } else {
142 old_ctrl = x->is_Con() ? C->root() : get_ctrl(x);
143 old_loop = get_loop(old_ctrl); // Get prior loop
144 }
145 // New late point must dominate new use
146 Node *new_ctrl = dom_lca( old_ctrl, region->in(i2) );
147 // Set new location
148 set_ctrl(x, new_ctrl);
149 IdealLoopTree *new_loop = get_loop( new_ctrl );
150 // If changing loop bodies, see if we need to collect into new body
151 if( old_loop != new_loop ) {
152 if( old_loop && !old_loop->_child )
153 old_loop->_body.yank(x);
154 if( !new_loop->_child )
155 new_loop->_body.push(x); // Collect body info
156 }
157 }
158
159 return phi;
160 }
161
162 //------------------------------dominated_by------------------------------------
163 // Replace the dominated test with an obvious true or false. Place it on the
164 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
165 // live path up to the dominating control.
166 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff ) {
167 #ifndef PRODUCT
168 if( VerifyLoopOptimizations && PrintOpto ) tty->print_cr("dominating test");
169 #endif
170
171
172 // prevdom is the dominating projection of the dominating test.
173 assert( iff->is_If(), "" );
174 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
175 int pop = prevdom->Opcode();
176 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
177 // 'con' is set to true or false to kill the dominated test.
178 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
179 set_ctrl(con, C->root()); // Constant gets a new use
180 // Hack the dominated test
181 _igvn.hash_delete(iff);
182 iff->set_req(1, con);
183 _igvn._worklist.push(iff);
184
185 // If I dont have a reachable TRUE and FALSE path following the IfNode then
186 // I can assume this path reaches an infinite loop. In this case it's not
187 // important to optimize the data Nodes - either the whole compilation will
188 // be tossed or this path (and all data Nodes) will go dead.
189 if( iff->outcnt() != 2 ) return;
190
191 // Make control-dependent data Nodes on the live path (path that will remain
192 // once the dominated IF is removed) become control-dependent on the
193 // dominating projection.
194 Node* dp = ((IfNode*)iff)->proj_out(pop == Op_IfTrue);
195 IdealLoopTree *old_loop = get_loop(dp);
196
197 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
198 Node* cd = dp->fast_out(i); // Control-dependent node
199 if( cd->depends_only_on_test() ) {
200 assert( cd->in(0) == dp, "" );
201 _igvn.hash_delete( cd );
202 cd->set_req(0, prevdom);
203 set_early_ctrl( cd );
204 _igvn._worklist.push(cd);
205 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
206 if( old_loop != new_loop ) {
207 if( !old_loop->_child ) old_loop->_body.yank(cd);
208 if( !new_loop->_child ) new_loop->_body.push(cd);
209 }
210 --i;
211 --imax;
212 }
213 }
214 }
215
216 //------------------------------has_local_phi_input----------------------------
217 // Return TRUE if 'n' has Phi inputs from its local block and no other
218 // block-local inputs (all non-local-phi inputs come from earlier blocks)
219 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
220 Node *n_ctrl = get_ctrl(n);
221 // See if some inputs come from a Phi in this block, or from before
222 // this block.
223 uint i;
224 for( i = 1; i < n->req(); i++ ) {
225 Node *phi = n->in(i);
226 if( phi->is_Phi() && phi->in(0) == n_ctrl )
227 break;
228 }
229 if( i >= n->req() )
230 return NULL; // No Phi inputs; nowhere to clone thru
231
232 // Check for inputs created between 'n' and the Phi input. These
233 // must split as well; they have already been given the chance
234 // (courtesy of a post-order visit) and since they did not we must
235 // recover the 'cost' of splitting them by being very profitable
236 // when splitting 'n'. Since this is unlikely we simply give up.
237 for( i = 1; i < n->req(); i++ ) {
238 Node *m = n->in(i);
239 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
240 // We allow the special case of AddP's with no local inputs.
241 // This allows us to split-up address expressions.
242 if (m->is_AddP() &&
243 get_ctrl(m->in(2)) != n_ctrl &&
244 get_ctrl(m->in(3)) != n_ctrl) {
245 // Move the AddP up to dominating point
246 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl)));
247 continue;
248 }
249 return NULL;
250 }
251 }
252
253 return n_ctrl;
254 }
255
256 //------------------------------remix_address_expressions----------------------
257 // Rework addressing expressions to get the most loop-invariant stuff
258 // moved out. We'd like to do all associative operators, but it's especially
259 // important (common) to do address expressions.
260 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
261 if (!has_ctrl(n)) return NULL;
262 Node *n_ctrl = get_ctrl(n);
263 IdealLoopTree *n_loop = get_loop(n_ctrl);
264
265 // See if 'n' mixes loop-varying and loop-invariant inputs and
266 // itself is loop-varying.
267
268 // Only interested in binary ops (and AddP)
269 if( n->req() < 3 || n->req() > 4 ) return NULL;
270
271 Node *n1_ctrl = get_ctrl(n->in( 1));
272 Node *n2_ctrl = get_ctrl(n->in( 2));
273 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
274 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
275 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
276 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
277
278 // Does one of my inputs spin in a tighter loop than self?
279 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
280 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
281 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
282 return NULL; // Leave well enough alone
283
284 // Is at least one of my inputs loop-invariant?
285 if( n1_loop == n_loop &&
286 n2_loop == n_loop &&
287 n3_loop == n_loop )
288 return NULL; // No loop-invariant inputs
289
290
291 int n_op = n->Opcode();
292
293 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
294 if( n_op == Op_LShiftI ) {
295 // Scale is loop invariant
296 Node *scale = n->in(2);
297 Node *scale_ctrl = get_ctrl(scale);
298 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
299 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
300 return NULL;
301 const TypeInt *scale_t = scale->bottom_type()->isa_int();
302 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
303 return NULL; // Dont bother with byte/short masking
304 // Add must vary with loop (else shift would be loop-invariant)
305 Node *add = n->in(1);
306 Node *add_ctrl = get_ctrl(add);
307 IdealLoopTree *add_loop = get_loop(add_ctrl);
308 //assert( n_loop == add_loop, "" );
309 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
310
311 // Convert I-V into I+ (0-V); same for V-I
312 if( add->Opcode() == Op_SubI &&
313 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
314 Node *zero = _igvn.intcon(0);
315 set_ctrl(zero, C->root());
316 Node *neg = new (C, 3) SubINode( _igvn.intcon(0), add->in(2) );
317 register_new_node( neg, get_ctrl(add->in(2) ) );
318 add = new (C, 3) AddINode( add->in(1), neg );
319 register_new_node( add, add_ctrl );
320 }
321 if( add->Opcode() != Op_AddI ) return NULL;
322 // See if one add input is loop invariant
323 Node *add_var = add->in(1);
324 Node *add_var_ctrl = get_ctrl(add_var);
325 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
326 Node *add_invar = add->in(2);
327 Node *add_invar_ctrl = get_ctrl(add_invar);
328 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
329 if( add_var_loop == n_loop ) {
330 } else if( add_invar_loop == n_loop ) {
331 // Swap to find the invariant part
332 add_invar = add_var;
333 add_invar_ctrl = add_var_ctrl;
334 add_invar_loop = add_var_loop;
335 add_var = add->in(2);
336 Node *add_var_ctrl = get_ctrl(add_var);
337 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
338 } else // Else neither input is loop invariant
339 return NULL;
340 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
341 return NULL; // No invariant part of the add?
342
343 // Yes! Reshape address expression!
344 Node *inv_scale = new (C, 3) LShiftINode( add_invar, scale );
345 register_new_node( inv_scale, add_invar_ctrl );
346 Node *var_scale = new (C, 3) LShiftINode( add_var, scale );
347 register_new_node( var_scale, n_ctrl );
348 Node *var_add = new (C, 3) AddINode( var_scale, inv_scale );
349 register_new_node( var_add, n_ctrl );
350 _igvn.hash_delete( n );
351 _igvn.subsume_node( n, var_add );
352 return var_add;
353 }
354
355 // Replace (I+V) with (V+I)
356 if( n_op == Op_AddI ||
357 n_op == Op_AddL ||
358 n_op == Op_AddF ||
359 n_op == Op_AddD ||
360 n_op == Op_MulI ||
361 n_op == Op_MulL ||
362 n_op == Op_MulF ||
363 n_op == Op_MulD ) {
364 if( n2_loop == n_loop ) {
365 assert( n1_loop != n_loop, "" );
366 n->swap_edges(1, 2);
367 }
368 }
369
370 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
371 // but not if I2 is a constant.
372 if( n_op == Op_AddP ) {
373 if( n2_loop == n_loop && n3_loop != n_loop ) {
374 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
375 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
376 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
377 IdealLoopTree *n22loop = get_loop( n22_ctrl );
378 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
379 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
380 n23_loop == n_loop ) {
381 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
382 // Stuff new AddP in the loop preheader
383 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
384 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) );
385 register_new_node( add2, n_ctrl );
386 _igvn.hash_delete( n );
387 _igvn.subsume_node( n, add2 );
388 return add2;
389 }
390 }
391 }
392
393 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
394 if( n2_loop != n_loop && n3_loop == n_loop ) {
395 if( n->in(3)->Opcode() == Op_AddI ) {
396 Node *V = n->in(3)->in(1);
397 Node *I = n->in(3)->in(2);
398 if( is_member(n_loop,get_ctrl(V)) ) {
399 } else {
400 Node *tmp = V; V = I; I = tmp;
401 }
402 if( !is_member(n_loop,get_ctrl(I)) ) {
403 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I );
404 // Stuff new AddP in the loop preheader
405 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
406 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V );
407 register_new_node( add2, n_ctrl );
408 _igvn.hash_delete( n );
409 _igvn.subsume_node( n, add2 );
410 return add2;
411 }
412 }
413 }
414 }
415
416 return NULL;
417 }
418
419 //------------------------------conditional_move-------------------------------
420 // Attempt to replace a Phi with a conditional move. We have some pretty
421 // strict profitability requirements. All Phis at the merge point must
422 // be converted, so we can remove the control flow. We need to limit the
423 // number of c-moves to a small handful. All code that was in the side-arms
424 // of the CFG diamond is now speculatively executed. This code has to be
425 // "cheap enough". We are pretty much limited to CFG diamonds that merge
426 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
427 Node *PhaseIdealLoop::conditional_move( Node *region ) {
428
429 assert( region->is_Region(), "sanity check" );
430 if( region->req() != 3 ) return NULL;
431
432 // Check for CFG diamond
433 Node *lp = region->in(1);
434 Node *rp = region->in(2);
435 if( !lp || !rp ) return NULL;
436 Node *lp_c = lp->in(0);
437 if( lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If() ) return NULL;
438 IfNode *iff = lp_c->as_If();
439
440 // Check for highly predictable branch. No point in CMOV'ing if
441 // we are going to predict accurately all the time.
442 // %%% This hides patterns produced by utility methods like Math.min.
443 if( iff->_prob < PROB_UNLIKELY_MAG(3) ||
444 iff->_prob > PROB_LIKELY_MAG(3) )
445 return NULL;
446
447 // Check for ops pinned in an arm of the diamond.
448 // Can't remove the control flow in this case
449 if( lp->outcnt() > 1 ) return NULL;
450 if( rp->outcnt() > 1 ) return NULL;
451
452 // Check profitability
453 int cost = 0;
454 int phis = 0;
455 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
456 Node *out = region->fast_out(i);
457 if( !out->is_Phi() ) continue; // Ignore other control edges, etc
458 phis++;
459 PhiNode* phi = out->as_Phi();
460 switch (phi->type()->basic_type()) {
461 case T_LONG:
462 cost++; // Probably encodes as 2 CMOV's
463 case T_INT: // These all CMOV fine
464 case T_FLOAT:
465 case T_DOUBLE:
466 case T_ADDRESS: // (RawPtr)
467 cost++;
468 break;
469 case T_OBJECT: { // Base oops are OK, but not derived oops
470 const TypeOopPtr *tp = phi->type()->isa_oopptr();
471 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
472 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
473 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
474 // have a Phi for the base here that we convert to a CMOVE all is well
475 // and good. But if the base is dead, we'll not make a CMOVE. Later
476 // the allocator will have to produce a base by creating a CMOVE of the
477 // relevant bases. This puts the allocator in the business of
478 // manufacturing expensive instructions, generally a bad plan.
479 // Just Say No to Conditionally-Moved Derived Pointers.
480 if( tp && tp->offset() != 0 )
481 return NULL;
482 cost++;
483 break;
484 }
485 default:
486 return NULL; // In particular, can't do memory or I/O
487 }
488 // Add in cost any speculative ops
489 for( uint j = 1; j < region->req(); j++ ) {
490 Node *proj = region->in(j);
491 Node *inp = phi->in(j);
492 if (get_ctrl(inp) == proj) { // Found local op
493 cost++;
494 // Check for a chain of dependent ops; these will all become
495 // speculative in a CMOV.
496 for( uint k = 1; k < inp->req(); k++ )
497 if (get_ctrl(inp->in(k)) == proj)
498 return NULL; // Too much speculative goo
499 }
500 }
501 // See if the Phi is used by a Cmp. This will likely Split-If, a
502 // higher-payoff operation.
503 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
504 Node* use = phi->fast_out(k);
505 if( use->is_Cmp() )
506 return NULL;
507 }
508 }
509 if( cost >= ConditionalMoveLimit ) return NULL; // Too much goo
510 Node* bol = iff->in(1);
511 assert( bol->Opcode() == Op_Bool, "" );
512 int cmp_op = bol->in(1)->Opcode();
513 // It is expensive to generate flags from a float compare.
514 // Avoid duplicated float compare.
515 if( phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
516
517 // --------------
518 // Now replace all Phis with CMOV's
519 Node *cmov_ctrl = iff->in(0);
520 uint flip = (lp->Opcode() == Op_IfTrue);
521 while( 1 ) {
522 PhiNode* phi = NULL;
523 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
524 Node *out = region->fast_out(i);
525 if (out->is_Phi()) {
526 phi = out->as_Phi();
527 break;
528 }
529 }
530 if (phi == NULL) break;
531 #ifndef PRODUCT
532 if( PrintOpto && VerifyLoopOptimizations ) tty->print_cr("CMOV");
533 #endif
534 // Move speculative ops
535 for( uint j = 1; j < region->req(); j++ ) {
536 Node *proj = region->in(j);
537 Node *inp = phi->in(j);
538 if (get_ctrl(inp) == proj) { // Found local op
539 #ifndef PRODUCT
540 if( PrintOpto && VerifyLoopOptimizations ) {
541 tty->print(" speculate: ");
542 inp->dump();
543 }
544 #endif
545 set_ctrl(inp, cmov_ctrl);
546 }
547 }
548 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) );
549 register_new_node( cmov, cmov_ctrl );
550 _igvn.hash_delete(phi);
551 _igvn.subsume_node( phi, cmov );
552 #ifndef PRODUCT
553 if( VerifyLoopOptimizations ) verify();
554 #endif
555 }
556
557 // The useless CFG diamond will fold up later; see the optimization in
558 // RegionNode::Ideal.
559 _igvn._worklist.push(region);
560
561 return iff->in(1);
562 }
563
564 //------------------------------split_if_with_blocks_pre-----------------------
565 // Do the real work in a non-recursive function. Data nodes want to be
566 // cloned in the pre-order so they can feed each other nicely.
567 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
568 // Cloning these guys is unlikely to win
569 int n_op = n->Opcode();
570 if( n_op == Op_MergeMem ) return n;
571 if( n->is_Proj() ) return n;
572 // Do not clone-up CmpFXXX variations, as these are always
573 // followed by a CmpI
574 if( n->is_Cmp() ) return n;
575 // Attempt to use a conditional move instead of a phi/branch
576 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
577 Node *cmov = conditional_move( n );
578 if( cmov ) return cmov;
579 }
580 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;
581 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
582 n_op == Op_Opaque2 ) {
583 if( !C->major_progress() ) // If chance of no more loop opts...
584 _igvn._worklist.push(n); // maybe we'll remove them
585 return n;
586 }
587
588 if( n->is_Con() ) return n; // No cloning for Con nodes
589
590 Node *n_ctrl = get_ctrl(n);
591 if( !n_ctrl ) return n; // Dead node
592
593 // Attempt to remix address expressions for loop invariants
594 Node *m = remix_address_expressions( n );
595 if( m ) return m;
596
597 // Determine if the Node has inputs from some local Phi.
598 // Returns the block to clone thru.
599 Node *n_blk = has_local_phi_input( n );
600 if( !n_blk ) return n;
601 // Do not clone the trip counter through on a CountedLoop
602 // (messes up the canonical shape).
603 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
604
605 // Check for having no control input; not pinned. Allow
606 // dominating control.
607 if( n->in(0) ) {
608 Node *dom = idom(n_blk);
609 if( dom_lca( n->in(0), dom ) != n->in(0) )
610 return n;
611 }
612 // Policy: when is it profitable. You must get more wins than
613 // policy before it is considered profitable. Policy is usually 0,
614 // so 1 win is considered profitable. Big merges will require big
615 // cloning, so get a larger policy.
616 int policy = n_blk->req() >> 2;
617
618 // If the loop is a candidate for range check elimination,
619 // delay splitting through it's phi until a later loop optimization
620 if (n_blk->is_CountedLoop()) {
621 IdealLoopTree *lp = get_loop(n_blk);
622 if (lp && lp->_rce_candidate) {
623 return n;
624 }
625 }
626
627 // Use same limit as split_if_with_blocks_post
628 if( C->unique() > 35000 ) return n; // Method too big
629
630 // Split 'n' through the merge point if it is profitable
631 Node *phi = split_thru_phi( n, n_blk, policy );
632 if( !phi ) return n;
633
634 // Found a Phi to split thru!
635 // Replace 'n' with the new phi
636 _igvn.hash_delete(n);
637 _igvn.subsume_node( n, phi );
638 // Moved a load around the loop, 'en-registering' something.
639 if( n_blk->Opcode() == Op_Loop && n->is_Load() &&
640 !phi->in(LoopNode::LoopBackControl)->is_Load() )
641 C->set_major_progress();
642
643 return phi;
644 }
645
646 static bool merge_point_too_heavy(Compile* C, Node* region) {
647 // Bail out if the region and its phis have too many users.
648 int weight = 0;
649 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
650 weight += region->fast_out(i)->outcnt();
651 }
652 int nodes_left = MaxNodeLimit - C->unique();
653 if (weight * 8 > nodes_left) {
654 #ifndef PRODUCT
655 if (PrintOpto)
656 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
657 #endif
658 return true;
659 } else {
660 return false;
661 }
662 }
663
664 #ifdef _LP64
665 static bool merge_point_safe(Node* region) {
666 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
667 // having a PhiNode input. This sidesteps the dangerous case where the split
668 // ConvI2LNode may become TOP if the input Value() does not
669 // overlap the ConvI2L range, leaving a node which may not dominate its
670 // uses.
671 // A better fix for this problem can be found in the BugTraq entry, but
672 // expediency for Mantis demands this hack.
673 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
674 Node* n = region->fast_out(i);
675 if (n->is_Phi()) {
676 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
677 Node* m = n->fast_out(j);
678 if (m->Opcode() == Op_ConvI2L) {
679 return false;
680 }
681 }
682 }
683 }
684 return true;
685 }
686 #endif
687
688
689 //------------------------------place_near_use---------------------------------
690 // Place some computation next to use but not inside inner loops.
691 // For inner loop uses move it to the preheader area.
692 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const {
693 IdealLoopTree *u_loop = get_loop( useblock );
694 return (u_loop->_irreducible || u_loop->_child)
695 ? useblock
696 : u_loop->_head->in(LoopNode::EntryControl);
697 }
698
699
700 //------------------------------split_if_with_blocks_post----------------------
701 // Do the real work in a non-recursive function. CFG hackery wants to be
702 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
703 // info.
704 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) {
705
706 // Cloning Cmp through Phi's involves the split-if transform.
707 // FastLock is not used by an If
708 if( n->is_Cmp() && !n->is_FastLock() ) {
709 if( C->unique() > 35000 ) return; // Method too big
710
711 // Do not do 'split-if' if irreducible loops are present.
712 if( _has_irreducible_loops )
713 return;
714
715 Node *n_ctrl = get_ctrl(n);
716 // Determine if the Node has inputs from some local Phi.
717 // Returns the block to clone thru.
718 Node *n_blk = has_local_phi_input( n );
719 if( n_blk != n_ctrl ) return;
720
721 if( merge_point_too_heavy(C, n_ctrl) )
722 return;
723
724 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare?
725 Node *bol = n->unique_out();
726 assert( bol->is_Bool(), "expect a bool here" );
727 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare?
728 Node *iff = bol->unique_out();
729
730 // Check some safety conditions
731 if( iff->is_If() ) { // Classic split-if?
732 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if
733 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
734 if( get_ctrl(iff->in(2)) == n_ctrl ||
735 get_ctrl(iff->in(3)) == n_ctrl )
736 return; // Inputs not yet split-up
737 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) {
738 return; // Loop-invar test gates loop-varying CMOVE
739 }
740 } else {
741 return; // some other kind of node, such as an Allocate
742 }
743
744 // Do not do 'split-if' if some paths are dead. First do dead code
745 // elimination and then see if its still profitable.
746 for( uint i = 1; i < n_ctrl->req(); i++ )
747 if( n_ctrl->in(i) == C->top() )
748 return;
749
750 // When is split-if profitable? Every 'win' on means some control flow
751 // goes dead, so it's almost always a win.
752 int policy = 0;
753 // If trying to do a 'Split-If' at the loop head, it is only
754 // profitable if the cmp folds up on BOTH paths. Otherwise we
755 // risk peeling a loop forever.
756
757 // CNC - Disabled for now. Requires careful handling of loop
758 // body selection for the cloned code. Also, make sure we check
759 // for any input path not being in the same loop as n_ctrl. For
760 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
761 // because the alternative loop entry points won't be converted
762 // into LoopNodes.
763 IdealLoopTree *n_loop = get_loop(n_ctrl);
764 for( uint j = 1; j < n_ctrl->req(); j++ )
765 if( get_loop(n_ctrl->in(j)) != n_loop )
766 return;
767
768 #ifdef _LP64
769 // Check for safety of the merge point.
770 if( !merge_point_safe(n_ctrl) ) {
771 return;
772 }
773 #endif
774
775 // Split compare 'n' through the merge point if it is profitable
776 Node *phi = split_thru_phi( n, n_ctrl, policy );
777 if( !phi ) return;
778
779 // Found a Phi to split thru!
780 // Replace 'n' with the new phi
781 _igvn.hash_delete(n);
782 _igvn.subsume_node( n, phi );
783
784 // Now split the bool up thru the phi
785 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 );
786 _igvn.hash_delete(bol);
787 _igvn.subsume_node( bol, bolphi );
788 assert( iff->in(1) == bolphi, "" );
789 if( bolphi->Value(&_igvn)->singleton() )
790 return;
791
792 // Conditional-move? Must split up now
793 if( !iff->is_If() ) {
794 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 );
795 _igvn.hash_delete(iff);
796 _igvn.subsume_node( iff, cmovphi );
797 return;
798 }
799
800 // Now split the IF
801 do_split_if( iff );
802 return;
803 }
804
805 // Check for an IF ready to split; one that has its
806 // condition codes input coming from a Phi at the block start.
807 int n_op = n->Opcode();
808
809 // Check for an IF being dominated by another IF same test
810 if( n_op == Op_If ) {
811 Node *bol = n->in(1);
812 uint max = bol->outcnt();
813 // Check for same test used more than once?
814 if( n_op == Op_If && max > 1 && bol->is_Bool() ) {
815 // Search up IDOMs to see if this IF is dominated.
816 Node *cutoff = get_ctrl(bol);
817
818 // Now search up IDOMs till cutoff, looking for a dominating test
819 Node *prevdom = n;
820 Node *dom = idom(prevdom);
821 while( dom != cutoff ) {
822 if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) {
823 // Replace the dominated test with an obvious true or false.
824 // Place it on the IGVN worklist for later cleanup.
825 C->set_major_progress();
826 dominated_by( prevdom, n );
827 #ifndef PRODUCT
828 if( VerifyLoopOptimizations ) verify();
829 #endif
830 return;
831 }
832 prevdom = dom;
833 dom = idom(prevdom);
834 }
835 }
836 }
837
838 // See if a shared loop-varying computation has no loop-varying uses.
839 // Happens if something is only used for JVM state in uncommon trap exits,
840 // like various versions of induction variable+offset. Clone the
841 // computation per usage to allow it to sink out of the loop.
842 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
843 Node *n_ctrl = get_ctrl(n);
844 IdealLoopTree *n_loop = get_loop(n_ctrl);
845 if( n_loop != _ltree_root ) {
846 DUIterator_Fast imax, i = n->fast_outs(imax);
847 for (; i < imax; i++) {
848 Node* u = n->fast_out(i);
849 if( !has_ctrl(u) ) break; // Found control user
850 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
851 if( u_loop == n_loop ) break; // Found loop-varying use
852 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
853 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
854 }
855 bool did_break = (i < imax); // Did we break out of the previous loop?
856 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
857 Node *late_load_ctrl;
858 if (n->is_Load()) {
859 // If n is a load, get and save the result from get_late_ctrl(),
860 // to be later used in calculating the control for n's clones.
861 clear_dom_lca_tags();
862 late_load_ctrl = get_late_ctrl(n, n_ctrl);
863 }
864 // If n is a load, and the late control is the same as the current
865 // control, then the cloning of n is a pointless exercise, because
866 // GVN will ensure that we end up where we started.
867 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
868 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
869 Node *u = n->last_out(j); // Clone private computation per use
870 _igvn.hash_delete(u);
871 _igvn._worklist.push(u);
872 Node *x = n->clone(); // Clone computation
873 Node *x_ctrl = NULL;
874 if( u->is_Phi() ) {
875 // Replace all uses of normal nodes. Replace Phi uses
876 // individually, so the seperate Nodes can sink down
877 // different paths.
878 uint k = 1;
879 while( u->in(k) != n ) k++;
880 u->set_req( k, x );
881 // x goes next to Phi input path
882 x_ctrl = u->in(0)->in(k);
883 --j;
884 } else { // Normal use
885 // Replace all uses
886 for( uint k = 0; k < u->req(); k++ ) {
887 if( u->in(k) == n ) {
888 u->set_req( k, x );
889 --j;
890 }
891 }
892 x_ctrl = get_ctrl(u);
893 }
894
895 // Find control for 'x' next to use but not inside inner loops.
896 // For inner loop uses get the preheader area.
897 x_ctrl = place_near_use(x_ctrl);
898
899 if (n->is_Load()) {
900 // For loads, add a control edge to a CFG node outside of the loop
901 // to force them to not combine and return back inside the loop
902 // during GVN optimization (4641526).
903 //
904 // Because we are setting the actual control input, factor in
905 // the result from get_late_ctrl() so we respect any
906 // anti-dependences. (6233005).
907 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
908
909 // Don't allow the control input to be a CFG splitting node.
910 // Such nodes should only have ProjNodes as outs, e.g. IfNode
911 // should only have IfTrueNode and IfFalseNode (4985384).
912 x_ctrl = find_non_split_ctrl(x_ctrl);
913 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
914
915 x->set_req(0, x_ctrl);
916 }
917 register_new_node(x, x_ctrl);
918
919 // Some institutional knowledge is needed here: 'x' is
920 // yanked because if the optimizer runs GVN on it all the
921 // cloned x's will common up and undo this optimization and
922 // be forced back in the loop. This is annoying because it
923 // makes +VerifyOpto report false-positives on progress. I
924 // tried setting control edges on the x's to force them to
925 // not combine, but the matching gets worried when it tries
926 // to fold a StoreP and an AddP together (as part of an
927 // address expression) and the AddP and StoreP have
928 // different controls.
929 if( !x->is_Load() ) _igvn._worklist.yank(x);
930 }
931 _igvn.remove_dead_node(n);
932 }
933 }
934 }
935 }
936
937 // Check for Opaque2's who's loop has disappeared - who's input is in the
938 // same loop nest as their output. Remove 'em, they are no longer useful.
939 if( n_op == Op_Opaque2 &&
940 n->in(1) != NULL &&
941 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
942 _igvn.add_users_to_worklist(n);
943 _igvn.hash_delete(n);
944 _igvn.subsume_node( n, n->in(1) );
945 }
946 }
947
948 //------------------------------split_if_with_blocks---------------------------
949 // Check for aggressive application of 'split-if' optimization,
950 // using basic block level info.
951 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) {
952 Node *n = C->root();
953 visited.set(n->_idx); // first, mark node as visited
954 // Do pre-visit work for root
955 n = split_if_with_blocks_pre( n );
956 uint cnt = n->outcnt();
957 uint i = 0;
958 while (true) {
959 // Visit all children
960 if (i < cnt) {
961 Node* use = n->raw_out(i);
962 ++i;
963 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
964 // Now do pre-visit work for this use
965 use = split_if_with_blocks_pre( use );
966 nstack.push(n, i); // Save parent and next use's index.
967 n = use; // Process all children of current use.
968 cnt = use->outcnt();
969 i = 0;
970 }
971 }
972 else {
973 // All of n's children have been processed, complete post-processing.
974 if (cnt != 0 && !n->is_Con()) {
975 assert(has_node(n), "no dead nodes");
976 split_if_with_blocks_post( n );
977 }
978 if (nstack.is_empty()) {
979 // Finished all nodes on stack.
980 break;
981 }
982 // Get saved parent node and next use's index. Visit the rest of uses.
983 n = nstack.node();
984 cnt = n->outcnt();
985 i = nstack.index();
986 nstack.pop();
987 }
988 }
989 }
990
991
992 //=============================================================================
993 //
994 // C L O N E A L O O P B O D Y
995 //
996
997 //------------------------------clone_iff--------------------------------------
998 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
999 // "Nearly" because all Nodes have been cloned from the original in the loop,
1000 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1001 // through the Phi recursively, and return a Bool.
1002 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) {
1003
1004 // Convert this Phi into a Phi merging Bools
1005 uint i;
1006 for( i = 1; i < phi->req(); i++ ) {
1007 Node *b = phi->in(i);
1008 if( b->is_Phi() ) {
1009 _igvn.hash_delete(phi);
1010 _igvn._worklist.push(phi);
1011 phi->set_req(i, clone_iff( b->as_Phi(), loop ));
1012 } else {
1013 assert( b->is_Bool(), "" );
1014 }
1015 }
1016
1017 Node *sample_bool = phi->in(1);
1018 Node *sample_cmp = sample_bool->in(1);
1019
1020 // Make Phis to merge the Cmp's inputs.
1021 int size = phi->in(0)->req();
1022 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1023 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1024 for( i = 1; i < phi->req(); i++ ) {
1025 Node *n1 = phi->in(i)->in(1)->in(1);
1026 Node *n2 = phi->in(i)->in(1)->in(2);
1027 phi1->set_req( i, n1 );
1028 phi2->set_req( i, n2 );
1029 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1030 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1031 }
1032 // See if these Phis have been made before.
1033 // Register with optimizer
1034 Node *hit1 = _igvn.hash_find_insert(phi1);
1035 if( hit1 ) { // Hit, toss just made Phi
1036 _igvn.remove_dead_node(phi1); // Remove new phi
1037 assert( hit1->is_Phi(), "" );
1038 phi1 = (PhiNode*)hit1; // Use existing phi
1039 } else { // Miss
1040 _igvn.register_new_node_with_optimizer(phi1);
1041 }
1042 Node *hit2 = _igvn.hash_find_insert(phi2);
1043 if( hit2 ) { // Hit, toss just made Phi
1044 _igvn.remove_dead_node(phi2); // Remove new phi
1045 assert( hit2->is_Phi(), "" );
1046 phi2 = (PhiNode*)hit2; // Use existing phi
1047 } else { // Miss
1048 _igvn.register_new_node_with_optimizer(phi2);
1049 }
1050 // Register Phis with loop/block info
1051 set_ctrl(phi1, phi->in(0));
1052 set_ctrl(phi2, phi->in(0));
1053 // Make a new Cmp
1054 Node *cmp = sample_cmp->clone();
1055 cmp->set_req( 1, phi1 );
1056 cmp->set_req( 2, phi2 );
1057 _igvn.register_new_node_with_optimizer(cmp);
1058 set_ctrl(cmp, phi->in(0));
1059
1060 // Make a new Bool
1061 Node *b = sample_bool->clone();
1062 b->set_req(1,cmp);
1063 _igvn.register_new_node_with_optimizer(b);
1064 set_ctrl(b, phi->in(0));
1065
1066 assert( b->is_Bool(), "" );
1067 return (BoolNode*)b;
1068 }
1069
1070 //------------------------------clone_bool-------------------------------------
1071 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1072 // "Nearly" because all Nodes have been cloned from the original in the loop,
1073 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1074 // through the Phi recursively, and return a Bool.
1075 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1076 uint i;
1077 // Convert this Phi into a Phi merging Bools
1078 for( i = 1; i < phi->req(); i++ ) {
1079 Node *b = phi->in(i);
1080 if( b->is_Phi() ) {
1081 _igvn.hash_delete(phi);
1082 _igvn._worklist.push(phi);
1083 phi->set_req(i, clone_bool( b->as_Phi(), loop ));
1084 } else {
1085 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1086 }
1087 }
1088
1089 Node *sample_cmp = phi->in(1);
1090
1091 // Make Phis to merge the Cmp's inputs.
1092 int size = phi->in(0)->req();
1093 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1094 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1095 for( uint j = 1; j < phi->req(); j++ ) {
1096 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1097 Node *n1, *n2;
1098 if( cmp_top->is_Cmp() ) {
1099 n1 = cmp_top->in(1);
1100 n2 = cmp_top->in(2);
1101 } else {
1102 n1 = n2 = cmp_top;
1103 }
1104 phi1->set_req( j, n1 );
1105 phi2->set_req( j, n2 );
1106 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1107 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1108 }
1109
1110 // See if these Phis have been made before.
1111 // Register with optimizer
1112 Node *hit1 = _igvn.hash_find_insert(phi1);
1113 if( hit1 ) { // Hit, toss just made Phi
1114 _igvn.remove_dead_node(phi1); // Remove new phi
1115 assert( hit1->is_Phi(), "" );
1116 phi1 = (PhiNode*)hit1; // Use existing phi
1117 } else { // Miss
1118 _igvn.register_new_node_with_optimizer(phi1);
1119 }
1120 Node *hit2 = _igvn.hash_find_insert(phi2);
1121 if( hit2 ) { // Hit, toss just made Phi
1122 _igvn.remove_dead_node(phi2); // Remove new phi
1123 assert( hit2->is_Phi(), "" );
1124 phi2 = (PhiNode*)hit2; // Use existing phi
1125 } else { // Miss
1126 _igvn.register_new_node_with_optimizer(phi2);
1127 }
1128 // Register Phis with loop/block info
1129 set_ctrl(phi1, phi->in(0));
1130 set_ctrl(phi2, phi->in(0));
1131 // Make a new Cmp
1132 Node *cmp = sample_cmp->clone();
1133 cmp->set_req( 1, phi1 );
1134 cmp->set_req( 2, phi2 );
1135 _igvn.register_new_node_with_optimizer(cmp);
1136 set_ctrl(cmp, phi->in(0));
1137
1138 assert( cmp->is_Cmp(), "" );
1139 return (CmpNode*)cmp;
1140 }
1141
1142 //------------------------------sink_use---------------------------------------
1143 // If 'use' was in the loop-exit block, it now needs to be sunk
1144 // below the post-loop merge point.
1145 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1146 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1147 set_ctrl(use, post_loop);
1148 for (DUIterator j = use->outs(); use->has_out(j); j++)
1149 sink_use(use->out(j), post_loop);
1150 }
1151 }
1152
1153 //------------------------------clone_loop-------------------------------------
1154 //
1155 // C L O N E A L O O P B O D Y
1156 //
1157 // This is the basic building block of the loop optimizations. It clones an
1158 // entire loop body. It makes an old_new loop body mapping; with this mapping
1159 // you can find the new-loop equivalent to an old-loop node. All new-loop
1160 // nodes are exactly equal to their old-loop counterparts, all edges are the
1161 // same. All exits from the old-loop now have a RegionNode that merges the
1162 // equivalent new-loop path. This is true even for the normal "loop-exit"
1163 // condition. All uses of loop-invariant old-loop values now come from (one
1164 // or more) Phis that merge their new-loop equivalents.
1165 //
1166 // This operation leaves the graph in an illegal state: there are two valid
1167 // control edges coming from the loop pre-header to both loop bodies. I'll
1168 // definitely have to hack the graph after running this transform.
1169 //
1170 // From this building block I will further edit edges to perform loop peeling
1171 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1172 //
1173 // Parameter side_by_size_idom:
1174 // When side_by_size_idom is NULL, the dominator tree is constructed for
1175 // the clone loop to dominate the original. Used in construction of
1176 // pre-main-post loop sequence.
1177 // When nonnull, the clone and original are side-by-side, both are
1178 // dominated by the side_by_side_idom node. Used in construction of
1179 // unswitched loops.
1180 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1181 Node* side_by_side_idom) {
1182
1183 // Step 1: Clone the loop body. Make the old->new mapping.
1184 uint i;
1185 for( i = 0; i < loop->_body.size(); i++ ) {
1186 Node *old = loop->_body.at(i);
1187 Node *nnn = old->clone();
1188 old_new.map( old->_idx, nnn );
1189 _igvn.register_new_node_with_optimizer(nnn);
1190 }
1191
1192
1193 // Step 2: Fix the edges in the new body. If the old input is outside the
1194 // loop use it. If the old input is INside the loop, use the corresponding
1195 // new node instead.
1196 for( i = 0; i < loop->_body.size(); i++ ) {
1197 Node *old = loop->_body.at(i);
1198 Node *nnn = old_new[old->_idx];
1199 // Fix CFG/Loop controlling the new node
1200 if (has_ctrl(old)) {
1201 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
1202 } else {
1203 set_loop(nnn, loop->_parent);
1204 if (old->outcnt() > 0) {
1205 set_idom( nnn, old_new[idom(old)->_idx], dd );
1206 }
1207 }
1208 // Correct edges to the new node
1209 for( uint j = 0; j < nnn->req(); j++ ) {
1210 Node *n = nnn->in(j);
1211 if( n ) {
1212 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
1213 if( loop->is_member( old_in_loop ) )
1214 nnn->set_req(j, old_new[n->_idx]);
1215 }
1216 }
1217 _igvn.hash_find_insert(nnn);
1218 }
1219 Node *newhead = old_new[loop->_head->_idx];
1220 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1221
1222
1223 // Step 3: Now fix control uses. Loop varying control uses have already
1224 // been fixed up (as part of all input edges in Step 2). Loop invariant
1225 // control uses must be either an IfFalse or an IfTrue. Make a merge
1226 // point to merge the old and new IfFalse/IfTrue nodes; make the use
1227 // refer to this.
1228 ResourceArea *area = Thread::current()->resource_area();
1229 Node_List worklist(area);
1230 uint new_counter = C->unique();
1231 for( i = 0; i < loop->_body.size(); i++ ) {
1232 Node* old = loop->_body.at(i);
1233 if( !old->is_CFG() ) continue;
1234 Node* nnn = old_new[old->_idx];
1235
1236 // Copy uses to a worklist, so I can munge the def-use info
1237 // with impunity.
1238 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1239 worklist.push(old->fast_out(j));
1240
1241 while( worklist.size() ) { // Visit all uses
1242 Node *use = worklist.pop();
1243 if (!has_node(use)) continue; // Ignore dead nodes
1244 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1245 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
1246 // Both OLD and USE are CFG nodes here.
1247 assert( use->is_Proj(), "" );
1248
1249 // Clone the loop exit control projection
1250 Node *newuse = use->clone();
1251 newuse->set_req(0,nnn);
1252 _igvn.register_new_node_with_optimizer(newuse);
1253 set_loop(newuse, use_loop);
1254 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
1255
1256 // We need a Region to merge the exit from the peeled body and the
1257 // exit from the old loop body.
1258 RegionNode *r = new (C, 3) RegionNode(3);
1259 // Map the old use to the new merge point
1260 old_new.map( use->_idx, r );
1261 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
1262 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
1263
1264 // The original user of 'use' uses 'r' instead.
1265 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
1266 Node* useuse = use->last_out(l);
1267 _igvn.hash_delete(useuse);
1268 _igvn._worklist.push(useuse);
1269 uint uses_found = 0;
1270 if( useuse->in(0) == use ) {
1271 useuse->set_req(0, r);
1272 uses_found++;
1273 if( useuse->is_CFG() ) {
1274 assert( dom_depth(useuse) > dd_r, "" );
1275 set_idom(useuse, r, dom_depth(useuse));
1276 }
1277 }
1278 for( uint k = 1; k < useuse->req(); k++ ) {
1279 if( useuse->in(k) == use ) {
1280 useuse->set_req(k, r);
1281 uses_found++;
1282 }
1283 }
1284 l -= uses_found; // we deleted 1 or more copies of this edge
1285 }
1286
1287 // Now finish up 'r'
1288 r->set_req( 1, newuse );
1289 r->set_req( 2, use );
1290 _igvn.register_new_node_with_optimizer(r);
1291 set_loop(r, use_loop);
1292 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
1293 } // End of if a loop-exit test
1294 }
1295 }
1296
1297 // Step 4: If loop-invariant use is not control, it must be dominated by a
1298 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
1299 // there if needed. Make a Phi there merging old and new used values.
1300 Node_List *split_if_set = NULL;
1301 Node_List *split_bool_set = NULL;
1302 Node_List *split_cex_set = NULL;
1303 for( i = 0; i < loop->_body.size(); i++ ) {
1304 Node* old = loop->_body.at(i);
1305 Node* nnn = old_new[old->_idx];
1306 // Copy uses to a worklist, so I can munge the def-use info
1307 // with impunity.
1308 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1309 worklist.push(old->fast_out(j));
1310
1311 while( worklist.size() ) {
1312 Node *use = worklist.pop();
1313 if (!has_node(use)) continue; // Ignore dead nodes
1314 if (use->in(0) == C->top()) continue;
1315 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1316 // Check for data-use outside of loop - at least one of OLD or USE
1317 // must not be a CFG node.
1318 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) {
1319
1320 // If the Data use is an IF, that means we have an IF outside of the
1321 // loop that is switching on a condition that is set inside of the
1322 // loop. Happens if people set a loop-exit flag; then test the flag
1323 // in the loop to break the loop, then test is again outside of the
1324 // loop to determine which way the loop exited.
1325 if( use->is_If() || use->is_CMove() ) {
1326 // Since this code is highly unlikely, we lazily build the worklist
1327 // of such Nodes to go split.
1328 if( !split_if_set )
1329 split_if_set = new Node_List(area);
1330 split_if_set->push(use);
1331 }
1332 if( use->is_Bool() ) {
1333 if( !split_bool_set )
1334 split_bool_set = new Node_List(area);
1335 split_bool_set->push(use);
1336 }
1337 if( use->Opcode() == Op_CreateEx ) {
1338 if( !split_cex_set )
1339 split_cex_set = new Node_List(area);
1340 split_cex_set->push(use);
1341 }
1342
1343
1344 // Get "block" use is in
1345 uint idx = 0;
1346 while( use->in(idx) != old ) idx++;
1347 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1348 assert( !loop->is_member( get_loop( prev ) ), "" );
1349 Node *cfg = prev->_idx >= new_counter
1350 ? prev->in(2)
1351 : idom(prev);
1352 if( use->is_Phi() ) // Phi use is in prior block
1353 cfg = prev->in(idx); // NOT in block of Phi itself
1354 if (cfg->is_top()) { // Use is dead?
1355 _igvn.hash_delete(use);
1356 _igvn._worklist.push(use);
1357 use->set_req(idx, C->top());
1358 continue;
1359 }
1360
1361 while( !loop->is_member( get_loop( cfg ) ) ) {
1362 prev = cfg;
1363 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1364 }
1365 // If the use occurs after merging several exits from the loop, then
1366 // old value must have dominated all those exits. Since the same old
1367 // value was used on all those exits we did not need a Phi at this
1368 // merge point. NOW we do need a Phi here. Each loop exit value
1369 // is now merged with the peeled body exit; each exit gets its own
1370 // private Phi and those Phis need to be merged here.
1371 Node *phi;
1372 if( prev->is_Region() ) {
1373 if( idx == 0 ) { // Updating control edge?
1374 phi = prev; // Just use existing control
1375 } else { // Else need a new Phi
1376 phi = PhiNode::make( prev, old );
1377 // Now recursively fix up the new uses of old!
1378 for( uint i = 1; i < prev->req(); i++ ) {
1379 worklist.push(phi); // Onto worklist once for each 'old' input
1380 }
1381 }
1382 } else {
1383 // Get new RegionNode merging old and new loop exits
1384 prev = old_new[prev->_idx];
1385 assert( prev, "just made this in step 7" );
1386 if( idx == 0 ) { // Updating control edge?
1387 phi = prev; // Just use existing control
1388 } else { // Else need a new Phi
1389 // Make a new Phi merging data values properly
1390 phi = PhiNode::make( prev, old );
1391 phi->set_req( 1, nnn );
1392 }
1393 }
1394 // If inserting a new Phi, check for prior hits
1395 if( idx != 0 ) {
1396 Node *hit = _igvn.hash_find_insert(phi);
1397 if( hit == NULL ) {
1398 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1399 } else { // or
1400 // Remove the new phi from the graph and use the hit
1401 _igvn.remove_dead_node(phi);
1402 phi = hit; // Use existing phi
1403 }
1404 set_ctrl(phi, prev);
1405 }
1406 // Make 'use' use the Phi instead of the old loop body exit value
1407 _igvn.hash_delete(use);
1408 _igvn._worklist.push(use);
1409 use->set_req(idx, phi);
1410 if( use->_idx >= new_counter ) { // If updating new phis
1411 // Not needed for correctness, but prevents a weak assert
1412 // in AddPNode from tripping (when we end up with different
1413 // base & derived Phis that will become the same after
1414 // IGVN does CSE).
1415 Node *hit = _igvn.hash_find_insert(use);
1416 if( hit ) // Go ahead and re-hash for hits.
1417 _igvn.subsume_node( use, hit );
1418 }
1419
1420 // If 'use' was in the loop-exit block, it now needs to be sunk
1421 // below the post-loop merge point.
1422 sink_use( use, prev );
1423 }
1424 }
1425 }
1426
1427 // Check for IFs that need splitting/cloning. Happens if an IF outside of
1428 // the loop uses a condition set in the loop. The original IF probably
1429 // takes control from one or more OLD Regions (which in turn get from NEW
1430 // Regions). In any case, there will be a set of Phis for each merge point
1431 // from the IF up to where the original BOOL def exists the loop.
1432 if( split_if_set ) {
1433 while( split_if_set->size() ) {
1434 Node *iff = split_if_set->pop();
1435 if( iff->in(1)->is_Phi() ) {
1436 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop );
1437 _igvn.hash_delete(iff);
1438 _igvn._worklist.push(iff);
1439 iff->set_req(1, b);
1440 }
1441 }
1442 }
1443 if( split_bool_set ) {
1444 while( split_bool_set->size() ) {
1445 Node *b = split_bool_set->pop();
1446 Node *phi = b->in(1);
1447 assert( phi->is_Phi(), "" );
1448 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop );
1449 _igvn.hash_delete(b);
1450 _igvn._worklist.push(b);
1451 b->set_req(1, cmp);
1452 }
1453 }
1454 if( split_cex_set ) {
1455 while( split_cex_set->size() ) {
1456 Node *b = split_cex_set->pop();
1457 assert( b->in(0)->is_Region(), "" );
1458 assert( b->in(1)->is_Phi(), "" );
1459 assert( b->in(0)->in(0) == b->in(1)->in(0), "" );
1460 split_up( b, b->in(0), NULL );
1461 }
1462 }
1463
1464 }
1465
1466
1467 //---------------------- stride_of_possible_iv -------------------------------------
1468 // Looks for an iff/bool/comp with one operand of the compare
1469 // being a cycle involving an add and a phi,
1470 // with an optional truncation (left-shift followed by a right-shift)
1471 // of the add. Returns zero if not an iv.
1472 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
1473 Node* trunc1 = NULL;
1474 Node* trunc2 = NULL;
1475 const TypeInt* ttype = NULL;
1476 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
1477 return 0;
1478 }
1479 BoolNode* bl = iff->in(1)->as_Bool();
1480 Node* cmp = bl->in(1);
1481 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) {
1482 return 0;
1483 }
1484 // Must have an invariant operand
1485 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
1486 return 0;
1487 }
1488 Node* add2 = NULL;
1489 Node* cmp1 = cmp->in(1);
1490 if (cmp1->is_Phi()) {
1491 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
1492 Node* phi = cmp1;
1493 for (uint i = 1; i < phi->req(); i++) {
1494 Node* in = phi->in(i);
1495 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
1496 &trunc1, &trunc2, &ttype);
1497 if (add && add->in(1) == phi) {
1498 add2 = add->in(2);
1499 break;
1500 }
1501 }
1502 } else {
1503 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
1504 Node* addtrunc = cmp1;
1505 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
1506 &trunc1, &trunc2, &ttype);
1507 if (add && add->in(1)->is_Phi()) {
1508 Node* phi = add->in(1);
1509 for (uint i = 1; i < phi->req(); i++) {
1510 if (phi->in(i) == addtrunc) {
1511 add2 = add->in(2);
1512 break;
1513 }
1514 }
1515 }
1516 }
1517 if (add2 != NULL) {
1518 const TypeInt* add2t = _igvn.type(add2)->is_int();
1519 if (add2t->is_con()) {
1520 return add2t->get_con();
1521 }
1522 }
1523 return 0;
1524 }
1525
1526
1527 //---------------------- stay_in_loop -------------------------------------
1528 // Return the (unique) control output node that's in the loop (if it exists.)
1529 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
1530 Node* unique = NULL;
1531 if (!n) return NULL;
1532 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1533 Node* use = n->fast_out(i);
1534 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
1535 if (unique != NULL) {
1536 return NULL;
1537 }
1538 unique = use;
1539 }
1540 }
1541 return unique;
1542 }
1543
1544 //------------------------------ register_node -------------------------------------
1545 // Utility to register node "n" with PhaseIdealLoop
1546 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
1547 _igvn.register_new_node_with_optimizer(n);
1548 loop->_body.push(n);
1549 if (n->is_CFG()) {
1550 set_loop(n, loop);
1551 set_idom(n, pred, ddepth);
1552 } else {
1553 set_ctrl(n, pred);
1554 }
1555 }
1556
1557 //------------------------------ proj_clone -------------------------------------
1558 // Utility to create an if-projection
1559 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
1560 ProjNode* c = p->clone()->as_Proj();
1561 c->set_req(0, iff);
1562 return c;
1563 }
1564
1565 //------------------------------ short_circuit_if -------------------------------------
1566 // Force the iff control output to be the live_proj
1567 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
1568 int proj_con = live_proj->_con;
1569 assert(proj_con == 0 || proj_con == 1, "false or true projection");
1570 Node *con = _igvn.intcon(proj_con);
1571 set_ctrl(con, C->root());
1572 if (iff) {
1573 iff->set_req(1, con);
1574 }
1575 return con;
1576 }
1577
1578 //------------------------------ insert_if_before_proj -------------------------------------
1579 // Insert a new if before an if projection (* - new node)
1580 //
1581 // before
1582 // if(test)
1583 // / \
1584 // v v
1585 // other-proj proj (arg)
1586 //
1587 // after
1588 // if(test)
1589 // / \
1590 // / v
1591 // | * proj-clone
1592 // v |
1593 // other-proj v
1594 // * new_if(relop(cmp[IU](left,right)))
1595 // / \
1596 // v v
1597 // * new-proj proj
1598 // (returned)
1599 //
1600 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
1601 IfNode* iff = proj->in(0)->as_If();
1602 IdealLoopTree *loop = get_loop(proj);
1603 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1604 int ddepth = dom_depth(proj);
1605
1606 _igvn.hash_delete(iff);
1607 _igvn._worklist.push(iff);
1608 _igvn.hash_delete(proj);
1609 _igvn._worklist.push(proj);
1610
1611 proj->set_req(0, NULL); // temporary disconnect
1612 ProjNode* proj2 = proj_clone(proj, iff);
1613 register_node(proj2, loop, iff, ddepth);
1614
1615 Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right);
1616 register_node(cmp, loop, proj2, ddepth);
1617
1618 BoolNode* bol = new (C,2)BoolNode(cmp, relop);
1619 register_node(bol, loop, proj2, ddepth);
1620
1621 IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt);
1622 register_node(new_if, loop, proj2, ddepth);
1623
1624 proj->set_req(0, new_if); // reattach
1625 set_idom(proj, new_if, ddepth);
1626
1627 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
1628 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
1629
1630 return new_exit;
1631 }
1632
1633 //------------------------------ insert_region_before_proj -------------------------------------
1634 // Insert a region before an if projection (* - new node)
1635 //
1636 // before
1637 // if(test)
1638 // / |
1639 // v |
1640 // proj v
1641 // other-proj
1642 //
1643 // after
1644 // if(test)
1645 // / |
1646 // v |
1647 // * proj-clone v
1648 // | other-proj
1649 // v
1650 // * new-region
1651 // |
1652 // v
1653 // * dum_if
1654 // / \
1655 // v \
1656 // * dum-proj v
1657 // proj
1658 //
1659 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
1660 IfNode* iff = proj->in(0)->as_If();
1661 IdealLoopTree *loop = get_loop(proj);
1662 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1663 int ddepth = dom_depth(proj);
1664
1665 _igvn.hash_delete(iff);
1666 _igvn._worklist.push(iff);
1667 _igvn.hash_delete(proj);
1668 _igvn._worklist.push(proj);
1669
1670 proj->set_req(0, NULL); // temporary disconnect
1671 ProjNode* proj2 = proj_clone(proj, iff);
1672 register_node(proj2, loop, iff, ddepth);
1673
1674 RegionNode* reg = new (C,2)RegionNode(2);
1675 reg->set_req(1, proj2);
1676 register_node(reg, loop, iff, ddepth);
1677
1678 IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
1679 register_node(dum_if, loop, reg, ddepth);
1680
1681 proj->set_req(0, dum_if); // reattach
1682 set_idom(proj, dum_if, ddepth);
1683
1684 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
1685 register_node(dum_proj, loop, dum_if, ddepth);
1686
1687 return reg;
1688 }
1689
1690 //------------------------------ insert_cmpi_loop_exit -------------------------------------
1691 // Clone a signed compare loop exit from an unsigned compare and
1692 // insert it before the unsigned cmp on the stay-in-loop path.
1693 // All new nodes inserted in the dominator tree between the original
1694 // if and it's projections. The original if test is replaced with
1695 // a constant to force the stay-in-loop path.
1696 //
1697 // This is done to make sure that the original if and it's projections
1698 // still dominate the same set of control nodes, that the ctrl() relation
1699 // from data nodes to them is preserved, and that their loop nesting is
1700 // preserved.
1701 //
1702 // before
1703 // if(i <u limit) unsigned compare loop exit
1704 // / |
1705 // v v
1706 // exit-proj stay-in-loop-proj
1707 //
1708 // after
1709 // if(stay-in-loop-const) original if
1710 // / |
1711 // / v
1712 // / if(i < limit) new signed test
1713 // / / |
1714 // / / v
1715 // / / if(i <u limit) new cloned unsigned test
1716 // / / / |
1717 // v v v |
1718 // region |
1719 // | |
1720 // dum-if |
1721 // / | |
1722 // ether | |
1723 // v v
1724 // exit-proj stay-in-loop-proj
1725 //
1726 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
1727 const bool Signed = true;
1728 const bool Unsigned = false;
1729
1730 BoolNode* bol = if_cmpu->in(1)->as_Bool();
1731 if (bol->_test._test != BoolTest::lt) return NULL;
1732 CmpNode* cmpu = bol->in(1)->as_Cmp();
1733 if (cmpu->Opcode() != Op_CmpU) return NULL;
1734 int stride = stride_of_possible_iv(if_cmpu);
1735 if (stride == 0) return NULL;
1736
1737 ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj();
1738 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
1739
1740 Node* limit = NULL;
1741 if (stride > 0) {
1742 limit = cmpu->in(2);
1743 } else {
1744 limit = _igvn.makecon(TypeInt::ZERO);
1745 set_ctrl(limit, C->root());
1746 }
1747 // Create a new region on the exit path
1748 RegionNode* reg = insert_region_before_proj(lp_exit);
1749
1750 // Clone the if-cmpu-true-false using a signed compare
1751 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
1752 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
1753 reg->add_req(cmpi_exit);
1754
1755 // Clone the if-cmpu-true-false
1756 BoolTest::mask rel_u = bol->_test._test;
1757 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
1758 reg->add_req(cmpu_exit);
1759
1760 // Force original if to stay in loop.
1761 short_circuit_if(if_cmpu, lp_continue);
1762
1763 return cmpi_exit->in(0)->as_If();
1764 }
1765
1766 //------------------------------ remove_cmpi_loop_exit -------------------------------------
1767 // Remove a previously inserted signed compare loop exit.
1768 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
1769 Node* lp_proj = stay_in_loop(if_cmp, loop);
1770 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
1771 stay_in_loop(lp_proj, loop)->is_If() &&
1772 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
1773 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
1774 set_ctrl(con, C->root());
1775 if_cmp->set_req(1, con);
1776 }
1777
1778 //------------------------------ scheduled_nodelist -------------------------------------
1779 // Create a post order schedule of nodes that are in the
1780 // "member" set. The list is returned in "sched".
1781 // The first node in "sched" is the loop head, followed by
1782 // nodes which have no inputs in the "member" set, and then
1783 // followed by the nodes that have an immediate input dependence
1784 // on a node in "sched".
1785 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
1786
1787 assert(member.test(loop->_head->_idx), "loop head must be in member set");
1788 Arena *a = Thread::current()->resource_area();
1789 VectorSet visited(a);
1790 Node_Stack nstack(a, loop->_body.size());
1791
1792 Node* n = loop->_head; // top of stack is cached in "n"
1793 uint idx = 0;
1794 visited.set(n->_idx);
1795
1796 // Initially push all with no inputs from within member set
1797 for(uint i = 0; i < loop->_body.size(); i++ ) {
1798 Node *elt = loop->_body.at(i);
1799 if (member.test(elt->_idx)) {
1800 bool found = false;
1801 for (uint j = 0; j < elt->req(); j++) {
1802 Node* def = elt->in(j);
1803 if (def && member.test(def->_idx) && def != elt) {
1804 found = true;
1805 break;
1806 }
1807 }
1808 if (!found && elt != loop->_head) {
1809 nstack.push(n, idx);
1810 n = elt;
1811 assert(!visited.test(n->_idx), "not seen yet");
1812 visited.set(n->_idx);
1813 }
1814 }
1815 }
1816
1817 // traverse out's that are in the member set
1818 while (true) {
1819 if (idx < n->outcnt()) {
1820 Node* use = n->raw_out(idx);
1821 idx++;
1822 if (!visited.test_set(use->_idx)) {
1823 if (member.test(use->_idx)) {
1824 nstack.push(n, idx);
1825 n = use;
1826 idx = 0;
1827 }
1828 }
1829 } else {
1830 // All outputs processed
1831 sched.push(n);
1832 if (nstack.is_empty()) break;
1833 n = nstack.node();
1834 idx = nstack.index();
1835 nstack.pop();
1836 }
1837 }
1838 }
1839
1840
1841 //------------------------------ has_use_in_set -------------------------------------
1842 // Has a use in the vector set
1843 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
1844 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1845 Node* use = n->fast_out(j);
1846 if (vset.test(use->_idx)) {
1847 return true;
1848 }
1849 }
1850 return false;
1851 }
1852
1853
1854 //------------------------------ has_use_internal_to_set -------------------------------------
1855 // Has use internal to the vector set (ie. not in a phi at the loop head)
1856 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
1857 Node* head = loop->_head;
1858 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1859 Node* use = n->fast_out(j);
1860 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
1861 return true;
1862 }
1863 }
1864 return false;
1865 }
1866
1867
1868 //------------------------------ clone_for_use_outside_loop -------------------------------------
1869 // clone "n" for uses that are outside of loop
1870 void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
1871
1872 assert(worklist.size() == 0, "should be empty");
1873 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1874 Node* use = n->fast_out(j);
1875 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
1876 worklist.push(use);
1877 }
1878 }
1879 while( worklist.size() ) {
1880 Node *use = worklist.pop();
1881 if (!has_node(use) || use->in(0) == C->top()) continue;
1882 uint j;
1883 for (j = 0; j < use->req(); j++) {
1884 if (use->in(j) == n) break;
1885 }
1886 assert(j < use->req(), "must be there");
1887
1888 // clone "n" and insert it between the inputs of "n" and the use outside the loop
1889 Node* n_clone = n->clone();
1890 _igvn.hash_delete(use);
1891 use->set_req(j, n_clone);
1892 _igvn._worklist.push(use);
1893 if (!use->is_Phi()) {
1894 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
1895 set_ctrl(n_clone, use_c);
1896 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
1897 get_loop(use_c)->_body.push(n_clone);
1898 } else {
1899 // Use in a phi is considered a use in the associated predecessor block
1900 Node *prevbb = use->in(0)->in(j);
1901 set_ctrl(n_clone, prevbb);
1902 assert(!loop->is_member(get_loop(prevbb)), "should be outside loop");
1903 get_loop(prevbb)->_body.push(n_clone);
1904 }
1905 _igvn.register_new_node_with_optimizer(n_clone);
1906 #if !defined(PRODUCT)
1907 if (TracePartialPeeling) {
1908 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
1909 }
1910 #endif
1911 }
1912 }
1913
1914
1915 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
1916 // clone "n" for special uses that are in the not_peeled region.
1917 // If these def-uses occur in separate blocks, the code generator
1918 // marks the method as not compilable. For example, if a "BoolNode"
1919 // is in a different basic block than the "IfNode" that uses it, then
1920 // the compilation is aborted in the code generator.
1921 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1922 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
1923 if (n->is_Phi() || n->is_Load()) {
1924 return;
1925 }
1926 assert(worklist.size() == 0, "should be empty");
1927 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1928 Node* use = n->fast_out(j);
1929 if ( not_peel.test(use->_idx) &&
1930 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
1931 use->in(1) == n) {
1932 worklist.push(use);
1933 }
1934 }
1935 if (worklist.size() > 0) {
1936 // clone "n" and insert it between inputs of "n" and the use
1937 Node* n_clone = n->clone();
1938 loop->_body.push(n_clone);
1939 _igvn.register_new_node_with_optimizer(n_clone);
1940 set_ctrl(n_clone, get_ctrl(n));
1941 sink_list.push(n_clone);
1942 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
1943 #if !defined(PRODUCT)
1944 if (TracePartialPeeling) {
1945 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
1946 }
1947 #endif
1948 while( worklist.size() ) {
1949 Node *use = worklist.pop();
1950 _igvn.hash_delete(use);
1951 _igvn._worklist.push(use);
1952 for (uint j = 1; j < use->req(); j++) {
1953 if (use->in(j) == n) {
1954 use->set_req(j, n_clone);
1955 }
1956 }
1957 }
1958 }
1959 }
1960
1961
1962 //------------------------------ insert_phi_for_loop -------------------------------------
1963 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1964 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
1965 Node *phi = PhiNode::make(lp, back_edge_val);
1966 phi->set_req(LoopNode::EntryControl, lp_entry_val);
1967 // Use existing phi if it already exists
1968 Node *hit = _igvn.hash_find_insert(phi);
1969 if( hit == NULL ) {
1970 _igvn.register_new_node_with_optimizer(phi);
1971 set_ctrl(phi, lp);
1972 } else {
1973 // Remove the new phi from the graph and use the hit
1974 _igvn.remove_dead_node(phi);
1975 phi = hit;
1976 }
1977 _igvn.hash_delete(use);
1978 _igvn._worklist.push(use);
1979 use->set_req(idx, phi);
1980 }
1981
1982 #ifdef ASSERT
1983 //------------------------------ is_valid_loop_partition -------------------------------------
1984 // Validate the loop partition sets: peel and not_peel
1985 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
1986 VectorSet& not_peel ) {
1987 uint i;
1988 // Check that peel_list entries are in the peel set
1989 for (i = 0; i < peel_list.size(); i++) {
1990 if (!peel.test(peel_list.at(i)->_idx)) {
1991 return false;
1992 }
1993 }
1994 // Check at loop members are in one of peel set or not_peel set
1995 for (i = 0; i < loop->_body.size(); i++ ) {
1996 Node *def = loop->_body.at(i);
1997 uint di = def->_idx;
1998 // Check that peel set elements are in peel_list
1999 if (peel.test(di)) {
2000 if (not_peel.test(di)) {
2001 return false;
2002 }
2003 // Must be in peel_list also
2004 bool found = false;
2005 for (uint j = 0; j < peel_list.size(); j++) {
2006 if (peel_list.at(j)->_idx == di) {
2007 found = true;
2008 break;
2009 }
2010 }
2011 if (!found) {
2012 return false;
2013 }
2014 } else if (not_peel.test(di)) {
2015 if (peel.test(di)) {
2016 return false;
2017 }
2018 } else {
2019 return false;
2020 }
2021 }
2022 return true;
2023 }
2024
2025 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2026 // Ensure a use outside of loop is of the right form
2027 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2028 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2029 return (use->is_Phi() &&
2030 use_c->is_Region() && use_c->req() == 3 &&
2031 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2032 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2033 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2034 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2035 }
2036
2037 //------------------------------ is_valid_clone_loop_form -------------------------------------
2038 // Ensure that all uses outside of loop are of the right form
2039 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2040 uint orig_exit_idx, uint clone_exit_idx) {
2041 uint len = peel_list.size();
2042 for (uint i = 0; i < len; i++) {
2043 Node *def = peel_list.at(i);
2044
2045 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2046 Node *use = def->fast_out(j);
2047 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2048 if (!loop->is_member(get_loop(use_c))) {
2049 // use is not in the loop, check for correct structure
2050 if (use->in(0) == def) {
2051 // Okay
2052 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2053 return false;
2054 }
2055 }
2056 }
2057 }
2058 return true;
2059 }
2060 #endif
2061
2062 //------------------------------ partial_peel -------------------------------------
2063 // Partially peel (aka loop rotation) the top portion of a loop (called
2064 // the peel section below) by cloning it and placing one copy just before
2065 // the new loop head and the other copy at the bottom of the new loop.
2066 //
2067 // before after where it came from
2068 //
2069 // stmt1 stmt1
2070 // loop: stmt2 clone
2071 // stmt2 if condA goto exitA clone
2072 // if condA goto exitA new_loop: new
2073 // stmt3 stmt3 clone
2074 // if !condB goto loop if condB goto exitB clone
2075 // exitB: stmt2 orig
2076 // stmt4 if !condA goto new_loop orig
2077 // exitA: goto exitA
2078 // exitB:
2079 // stmt4
2080 // exitA:
2081 //
2082 // Step 1: find the cut point: an exit test on probable
2083 // induction variable.
2084 // Step 2: schedule (with cloning) operations in the peel
2085 // section that can be executed after the cut into
2086 // the section that is not peeled. This may need
2087 // to clone operations into exit blocks. For
2088 // instance, a reference to A[i] in the not-peel
2089 // section and a reference to B[i] in an exit block
2090 // may cause a left-shift of i by 2 to be placed
2091 // in the peel block. This step will clone the left
2092 // shift into the exit block and sink the left shift
2093 // from the peel to the not-peel section.
2094 // Step 3: clone the loop, retarget the control, and insert
2095 // phis for values that are live across the new loop
2096 // head. This is very dependent on the graph structure
2097 // from clone_loop. It creates region nodes for
2098 // exit control and associated phi nodes for values
2099 // flow out of the loop through that exit. The region
2100 // node is dominated by the clone's control projection.
2101 // So the clone's peel section is placed before the
2102 // new loop head, and the clone's not-peel section is
2103 // forms the top part of the new loop. The original
2104 // peel section forms the tail of the new loop.
2105 // Step 4: update the dominator tree and recompute the
2106 // dominator depth.
2107 //
2108 // orig
2109 //
2110 // stmt1
2111 // |
2112 // v
2113 // loop<----+
2114 // | |
2115 // stmt2 |
2116 // | |
2117 // v |
2118 // ifA |
2119 // / | |
2120 // v v |
2121 // false true ^ <-- last_peel
2122 // / | |
2123 // / ===|==cut |
2124 // / stmt3 | <-- first_not_peel
2125 // / | |
2126 // | v |
2127 // v ifB |
2128 // exitA: / \ |
2129 // / \ |
2130 // v v |
2131 // false true |
2132 // / \ |
2133 // / ----+
2134 // |
2135 // v
2136 // exitB:
2137 // stmt4
2138 //
2139 //
2140 // after clone loop
2141 //
2142 // stmt1
2143 // / \
2144 // clone / \ orig
2145 // / \
2146 // / \
2147 // v v
2148 // +---->loop loop<----+
2149 // | | | |
2150 // | stmt2 stmt2 |
2151 // | | | |
2152 // | v v |
2153 // | ifA ifA |
2154 // | | \ / | |
2155 // | v v v v |
2156 // ^ true false false true ^ <-- last_peel
2157 // | | ^ \ / | |
2158 // | cut==|== \ \ / ===|==cut |
2159 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2160 // | | dom | | | |
2161 // | v \ 1v v2 v |
2162 // | ifB regionA ifB |
2163 // | / \ | / \ |
2164 // | / \ v / \ |
2165 // | v v exitA: v v |
2166 // | true false false true |
2167 // | / ^ \ / \ |
2168 // +---- \ \ / ----+
2169 // dom \ /
2170 // \ 1v v2
2171 // regionB
2172 // |
2173 // v
2174 // exitB:
2175 // stmt4
2176 //
2177 //
2178 // after partial peel
2179 //
2180 // stmt1
2181 // /
2182 // clone / orig
2183 // / TOP
2184 // / \
2185 // v v
2186 // TOP->region region----+
2187 // | | |
2188 // stmt2 stmt2 |
2189 // | | |
2190 // v v |
2191 // ifA ifA |
2192 // | \ / | |
2193 // v v v v |
2194 // true false false true | <-- last_peel
2195 // | ^ \ / +------|---+
2196 // +->newloop \ \ / === ==cut | |
2197 // | stmt3 \ \ / TOP | |
2198 // | | dom | | stmt3 | | <-- first_not_peel
2199 // | v \ 1v v2 v | |
2200 // | ifB regionA ifB ^ v
2201 // | / \ | / \ | |
2202 // | / \ v / \ | |
2203 // | v v exitA: v v | |
2204 // | true false false true | |
2205 // | / ^ \ / \ | |
2206 // | | \ \ / v | |
2207 // | | dom \ / TOP | |
2208 // | | \ 1v v2 | |
2209 // ^ v regionB | |
2210 // | | | | |
2211 // | | v ^ v
2212 // | | exitB: | |
2213 // | | stmt4 | |
2214 // | +------------>-----------------+ |
2215 // | |
2216 // +-----------------<---------------------+
2217 //
2218 //
2219 // final graph
2220 //
2221 // stmt1
2222 // |
2223 // v
2224 // ........> ifA clone
2225 // : / |
2226 // dom / |
2227 // : v v
2228 // : false true
2229 // : | |
2230 // : | stmt2 clone
2231 // : | |
2232 // : | v
2233 // : | newloop<-----+
2234 // : | | |
2235 // : | stmt3 clone |
2236 // : | | |
2237 // : | v |
2238 // : | ifB |
2239 // : | / \ |
2240 // : | v v |
2241 // : | false true |
2242 // : | | | |
2243 // : | v stmt2 |
2244 // : | exitB: | |
2245 // : | stmt4 v |
2246 // : | ifA orig |
2247 // : | / \ |
2248 // : | / \ |
2249 // : | v v |
2250 // : | false true |
2251 // : | / \ |
2252 // : v v -----+
2253 // RegionA
2254 // |
2255 // v
2256 // exitA
2257 //
2258 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
2259
2260 LoopNode *head = loop->_head->as_Loop();
2261
2262 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
2263 return false;
2264 }
2265
2266 // Check for complex exit control
2267 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
2268 Node *n = loop->_body.at(ii);
2269 int opc = n->Opcode();
2270 if (n->is_Call() ||
2271 opc == Op_Catch ||
2272 opc == Op_CatchProj ||
2273 opc == Op_Jump ||
2274 opc == Op_JumpProj) {
2275 #if !defined(PRODUCT)
2276 if (TracePartialPeeling) {
2277 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
2278 }
2279 #endif
2280 return false;
2281 }
2282 }
2283
2284 int dd = dom_depth(head);
2285
2286 // Step 1: find cut point
2287
2288 // Walk up dominators to loop head looking for first loop exit
2289 // which is executed on every path thru loop.
2290 IfNode *peel_if = NULL;
2291 IfNode *peel_if_cmpu = NULL;
2292
2293 Node *iff = loop->tail();
2294 while( iff != head ) {
2295 if( iff->is_If() ) {
2296 Node *ctrl = get_ctrl(iff->in(1));
2297 if (ctrl->is_top()) return false; // Dead test on live IF.
2298 // If loop-varying exit-test, check for induction variable
2299 if( loop->is_member(get_loop(ctrl)) &&
2300 loop->is_loop_exit(iff) &&
2301 is_possible_iv_test(iff)) {
2302 Node* cmp = iff->in(1)->in(1);
2303 if (cmp->Opcode() == Op_CmpI) {
2304 peel_if = iff->as_If();
2305 } else {
2306 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
2307 peel_if_cmpu = iff->as_If();
2308 }
2309 }
2310 }
2311 iff = idom(iff);
2312 }
2313 // Prefer signed compare over unsigned compare.
2314 IfNode* new_peel_if = NULL;
2315 if (peel_if == NULL) {
2316 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
2317 return false; // No peel point found
2318 }
2319 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
2320 if (new_peel_if == NULL) {
2321 return false; // No peel point found
2322 }
2323 peel_if = new_peel_if;
2324 }
2325 Node* last_peel = stay_in_loop(peel_if, loop);
2326 Node* first_not_peeled = stay_in_loop(last_peel, loop);
2327 if (first_not_peeled == NULL || first_not_peeled == head) {
2328 return false;
2329 }
2330
2331 #if !defined(PRODUCT)
2332 if (TracePartialPeeling) {
2333 tty->print_cr("before partial peel one iteration");
2334 Node_List wl;
2335 Node* t = head->in(2);
2336 while (true) {
2337 wl.push(t);
2338 if (t == head) break;
2339 t = idom(t);
2340 }
2341 while (wl.size() > 0) {
2342 Node* tt = wl.pop();
2343 tt->dump();
2344 if (tt == last_peel) tty->print_cr("-- cut --");
2345 }
2346 }
2347 #endif
2348 ResourceArea *area = Thread::current()->resource_area();
2349 VectorSet peel(area);
2350 VectorSet not_peel(area);
2351 Node_List peel_list(area);
2352 Node_List worklist(area);
2353 Node_List sink_list(area);
2354
2355 // Set of cfg nodes to peel are those that are executable from
2356 // the head through last_peel.
2357 assert(worklist.size() == 0, "should be empty");
2358 worklist.push(head);
2359 peel.set(head->_idx);
2360 while (worklist.size() > 0) {
2361 Node *n = worklist.pop();
2362 if (n != last_peel) {
2363 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2364 Node* use = n->fast_out(j);
2365 if (use->is_CFG() &&
2366 loop->is_member(get_loop(use)) &&
2367 !peel.test_set(use->_idx)) {
2368 worklist.push(use);
2369 }
2370 }
2371 }
2372 }
2373
2374 // Set of non-cfg nodes to peel are those that are control
2375 // dependent on the cfg nodes.
2376 uint i;
2377 for(i = 0; i < loop->_body.size(); i++ ) {
2378 Node *n = loop->_body.at(i);
2379 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
2380 if (peel.test(n_c->_idx)) {
2381 peel.set(n->_idx);
2382 } else {
2383 not_peel.set(n->_idx);
2384 }
2385 }
2386
2387 // Step 2: move operations from the peeled section down into the
2388 // not-peeled section
2389
2390 // Get a post order schedule of nodes in the peel region
2391 // Result in right-most operand.
2392 scheduled_nodelist(loop, peel, peel_list );
2393
2394 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2395
2396 // For future check for too many new phis
2397 uint old_phi_cnt = 0;
2398 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2399 Node* use = head->fast_out(j);
2400 if (use->is_Phi()) old_phi_cnt++;
2401 }
2402
2403 #if !defined(PRODUCT)
2404 if (TracePartialPeeling) {
2405 tty->print_cr("\npeeled list");
2406 }
2407 #endif
2408
2409 // Evacuate nodes in peel region into the not_peeled region if possible
2410 uint new_phi_cnt = 0;
2411 for (i = 0; i < peel_list.size();) {
2412 Node* n = peel_list.at(i);
2413 #if !defined(PRODUCT)
2414 if (TracePartialPeeling) n->dump();
2415 #endif
2416 bool incr = true;
2417 if ( !n->is_CFG() ) {
2418
2419 if ( has_use_in_set(n, not_peel) ) {
2420
2421 // If not used internal to the peeled region,
2422 // move "n" from peeled to not_peeled region.
2423
2424 if ( !has_use_internal_to_set(n, peel, loop) ) {
2425
2426 // if not pinned and not a load (which maybe anti-dependent on a store)
2427 // and not a CMove (Matcher expects only bool->cmove).
2428 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
2429 clone_for_use_outside_loop( loop, n, worklist );
2430
2431 sink_list.push(n);
2432 peel >>= n->_idx; // delete n from peel set.
2433 not_peel <<= n->_idx; // add n to not_peel set.
2434 peel_list.remove(i);
2435 incr = false;
2436 #if !defined(PRODUCT)
2437 if (TracePartialPeeling) {
2438 tty->print_cr("sink to not_peeled region: %d newbb: %d",
2439 n->_idx, get_ctrl(n)->_idx);
2440 }
2441 #endif
2442 }
2443 } else {
2444 // Otherwise check for special def-use cases that span
2445 // the peel/not_peel boundary such as bool->if
2446 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
2447 new_phi_cnt++;
2448 }
2449 }
2450 }
2451 if (incr) i++;
2452 }
2453
2454 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
2455 #if !defined(PRODUCT)
2456 if (TracePartialPeeling) {
2457 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
2458 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
2459 }
2460 #endif
2461 if (new_peel_if != NULL) {
2462 remove_cmpi_loop_exit(new_peel_if, loop);
2463 }
2464 // Inhibit more partial peeling on this loop
2465 assert(!head->is_partial_peel_loop(), "not partial peeled");
2466 head->mark_partial_peel_failed();
2467 return false;
2468 }
2469
2470 // Step 3: clone loop, retarget control, and insert new phis
2471
2472 // Create new loop head for new phis and to hang
2473 // the nodes being moved (sinked) from the peel region.
2474 LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel);
2475 _igvn.register_new_node_with_optimizer(new_head);
2476 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
2477 first_not_peeled->set_req(0, new_head);
2478 set_loop(new_head, loop);
2479 loop->_body.push(new_head);
2480 not_peel.set(new_head->_idx);
2481 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
2482 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
2483
2484 while (sink_list.size() > 0) {
2485 Node* n = sink_list.pop();
2486 set_ctrl(n, new_head);
2487 }
2488
2489 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2490
2491 clone_loop( loop, old_new, dd );
2492
2493 const uint clone_exit_idx = 1;
2494 const uint orig_exit_idx = 2;
2495 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
2496
2497 Node* head_clone = old_new[head->_idx];
2498 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
2499 Node* orig_tail_clone = head_clone->in(2);
2500
2501 // Add phi if "def" node is in peel set and "use" is not
2502
2503 for(i = 0; i < peel_list.size(); i++ ) {
2504 Node *def = peel_list.at(i);
2505 if (!def->is_CFG()) {
2506 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2507 Node *use = def->fast_out(j);
2508 if (has_node(use) && use->in(0) != C->top() &&
2509 (!peel.test(use->_idx) ||
2510 (use->is_Phi() && use->in(0) == head)) ) {
2511 worklist.push(use);
2512 }
2513 }
2514 while( worklist.size() ) {
2515 Node *use = worklist.pop();
2516 for (uint j = 1; j < use->req(); j++) {
2517 Node* n = use->in(j);
2518 if (n == def) {
2519
2520 // "def" is in peel set, "use" is not in peel set
2521 // or "use" is in the entry boundary (a phi) of the peel set
2522
2523 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
2524
2525 if ( loop->is_member(get_loop( use_c )) ) {
2526 // use is in loop
2527 if (old_new[use->_idx] != NULL) { // null for dead code
2528 Node* use_clone = old_new[use->_idx];
2529 _igvn.hash_delete(use);
2530 use->set_req(j, C->top());
2531 _igvn._worklist.push(use);
2532 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
2533 }
2534 } else {
2535 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
2536 // use is not in the loop, check if the live range includes the cut
2537 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
2538 if (not_peel.test(lp_if->_idx)) {
2539 assert(j == orig_exit_idx, "use from original loop");
2540 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
2541 }
2542 }
2543 }
2544 }
2545 }
2546 }
2547 }
2548
2549 // Step 3b: retarget control
2550
2551 // Redirect control to the new loop head if a cloned node in
2552 // the not_peeled region has control that points into the peeled region.
2553 // This necessary because the cloned peeled region will be outside
2554 // the loop.
2555 // from to
2556 // cloned-peeled <---+
2557 // new_head_clone: | <--+
2558 // cloned-not_peeled in(0) in(0)
2559 // orig-peeled
2560
2561 for(i = 0; i < loop->_body.size(); i++ ) {
2562 Node *n = loop->_body.at(i);
2563 if (!n->is_CFG() && n->in(0) != NULL &&
2564 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
2565 Node* n_clone = old_new[n->_idx];
2566 _igvn.hash_delete(n_clone);
2567 n_clone->set_req(0, new_head_clone);
2568 _igvn._worklist.push(n_clone);
2569 }
2570 }
2571
2572 // Backedge of the surviving new_head (the clone) is original last_peel
2573 _igvn.hash_delete(new_head_clone);
2574 new_head_clone->set_req(LoopNode::LoopBackControl, last_peel);
2575 _igvn._worklist.push(new_head_clone);
2576
2577 // Cut first node in original not_peel set
2578 _igvn.hash_delete(new_head);
2579 new_head->set_req(LoopNode::EntryControl, C->top());
2580 new_head->set_req(LoopNode::LoopBackControl, C->top());
2581 _igvn._worklist.push(new_head);
2582
2583 // Copy head_clone back-branch info to original head
2584 // and remove original head's loop entry and
2585 // clone head's back-branch
2586 _igvn.hash_delete(head);
2587 _igvn.hash_delete(head_clone);
2588 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
2589 head->set_req(LoopNode::LoopBackControl, C->top());
2590 head_clone->set_req(LoopNode::LoopBackControl, C->top());
2591 _igvn._worklist.push(head);
2592 _igvn._worklist.push(head_clone);
2593
2594 // Similarly modify the phis
2595 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
2596 Node* use = head->fast_out(k);
2597 if (use->is_Phi() && use->outcnt() > 0) {
2598 Node* use_clone = old_new[use->_idx];
2599 _igvn.hash_delete(use);
2600 _igvn.hash_delete(use_clone);
2601 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
2602 use->set_req(LoopNode::LoopBackControl, C->top());
2603 use_clone->set_req(LoopNode::LoopBackControl, C->top());
2604 _igvn._worklist.push(use);
2605 _igvn._worklist.push(use_clone);
2606 }
2607 }
2608
2609 // Step 4: update dominator tree and dominator depth
2610
2611 set_idom(head, orig_tail_clone, dd);
2612 recompute_dom_depth();
2613
2614 // Inhibit more partial peeling on this loop
2615 new_head_clone->set_partial_peel_loop();
2616 C->set_major_progress();
2617
2618 #if !defined(PRODUCT)
2619 if (TracePartialPeeling) {
2620 tty->print_cr("\nafter partial peel one iteration");
2621 Node_List wl(area);
2622 Node* t = last_peel;
2623 while (true) {
2624 wl.push(t);
2625 if (t == head_clone) break;
2626 t = idom(t);
2627 }
2628 while (wl.size() > 0) {
2629 Node* tt = wl.pop();
2630 if (tt == head) tty->print_cr("orig head");
2631 else if (tt == new_head_clone) tty->print_cr("new head");
2632 else if (tt == head_clone) tty->print_cr("clone head");
2633 tt->dump();
2634 }
2635 }
2636 #endif
2637 return true;
2638 }
2639
2640 //------------------------------reorg_offsets----------------------------------
2641 // Reorganize offset computations to lower register pressure. Mostly
2642 // prevent loop-fallout uses of the pre-incremented trip counter (which are
2643 // then alive with the post-incremented trip counter forcing an extra
2644 // register move)
2645 void PhaseIdealLoop::reorg_offsets( IdealLoopTree *loop ) {
2646
2647 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2648 CountedLoopEndNode *cle = cl->loopexit();
2649 if( !cle ) return; // The occasional dead loop
2650 // Find loop exit control
2651 Node *exit = cle->proj_out(false);
2652 assert( exit->Opcode() == Op_IfFalse, "" );
2653
2654 // Check for the special case of folks using the pre-incremented
2655 // trip-counter on the fall-out path (forces the pre-incremented
2656 // and post-incremented trip counter to be live at the same time).
2657 // Fix this by adjusting to use the post-increment trip counter.
2658 Node *phi = cl->phi();
2659 if( !phi ) return; // Dead infinite loop
2660 bool progress = true;
2661 while (progress) {
2662 progress = false;
2663 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
2664 Node* use = phi->fast_out(i); // User of trip-counter
2665 if (!has_ctrl(use)) continue;
2666 Node *u_ctrl = get_ctrl(use);
2667 if( use->is_Phi() ) {
2668 u_ctrl = NULL;
2669 for( uint j = 1; j < use->req(); j++ )
2670 if( use->in(j) == phi )
2671 u_ctrl = dom_lca( u_ctrl, use->in(0)->in(j) );
2672 }
2673 IdealLoopTree *u_loop = get_loop(u_ctrl);
2674 // Look for loop-invariant use
2675 if( u_loop == loop ) continue;
2676 if( loop->is_member( u_loop ) ) continue;
2677 // Check that use is live out the bottom. Assuming the trip-counter
2678 // update is right at the bottom, uses of of the loop middle are ok.
2679 if( dom_lca( exit, u_ctrl ) != exit ) continue;
2680 // protect against stride not being a constant
2681 if( !cle->stride_is_con() ) continue;
2682 // Hit! Refactor use to use the post-incremented tripcounter.
2683 // Compute a post-increment tripcounter.
2684 Node *opaq = new (C, 2) Opaque2Node( cle->incr() );
2685 register_new_node( opaq, u_ctrl );
2686 Node *neg_stride = _igvn.intcon(-cle->stride_con());
2687 set_ctrl(neg_stride, C->root());
2688 Node *post = new (C, 3) AddINode( opaq, neg_stride);
2689 register_new_node( post, u_ctrl );
2690 _igvn.hash_delete(use);
2691 _igvn._worklist.push(use);
2692 for( uint j = 1; j < use->req(); j++ )
2693 if( use->in(j) == phi )
2694 use->set_req(j, post);
2695 // Since DU info changed, rerun loop
2696 progress = true;
2697 break;
2698 }
2699 }
2700
2701 }