1 /*
2 * Copyright 1999-2008 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
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_known_instance_field() ) {
45 int iid = t_oop->instance_id();
46 int index = C->get_alias_index(t_oop);
47 int offset = t_oop->offset();
48 phi = new (C,region->req()) PhiNode(region, type, NULL, iid, index, offset);
49 } else {
50 phi = new (C,region->req()) PhiNode(region, type);
51 }
52 uint old_unique = C->unique();
53 for( uint i = 1; i < region->req(); i++ ) {
54 Node *x;
55 Node* the_clone = NULL;
56 if( region->in(i) == C->top() ) {
57 x = C->top(); // Dead path? Use a dead data op
58 } else {
59 x = n->clone(); // Else clone up the data op
60 the_clone = x; // Remember for possible deletion.
61 // Alter data node to use pre-phi inputs
62 if( n->in(0) == region )
63 x->set_req( 0, region->in(i) );
64 for( uint j = 1; j < n->req(); j++ ) {
65 Node *in = n->in(j);
66 if( in->is_Phi() && in->in(0) == region )
67 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
68 }
69 }
70 // Check for a 'win' on some paths
71 const Type *t = x->Value(&_igvn);
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 // If x is a TypeNode, capture any more-precise type permanently into Node
100 // othewise it will be not updated during igvn->transform since
101 // igvn->type(x) is set to x->Value() already.
102 x->raise_bottom_type(t);
103 Node *y = x->Identity(&_igvn);
104 if( y != x ) {
105 wins++;
106 x = y;
107 } else {
108 y = _igvn.hash_find(x);
109 if( y ) {
110 wins++;
111 x = y;
112 } else {
113 // Else x is a new node we are keeping
114 // We do not need register_new_node_with_optimizer
115 // because set_type has already been called.
116 _igvn._worklist.push(x);
117 }
118 }
119 }
120 if (x != the_clone && the_clone != NULL)
121 _igvn.remove_dead_node(the_clone);
122 phi->set_req( i, x );
123 }
124 // Too few wins?
125 if( wins <= policy ) {
126 _igvn.remove_dead_node(phi);
127 return NULL;
128 }
129
130 // Record Phi
131 register_new_node( phi, region );
132
133 for( uint i2 = 1; i2 < phi->req(); i2++ ) {
134 Node *x = phi->in(i2);
135 // If we commoned up the cloned 'x' with another existing Node,
136 // the existing Node picks up a new use. We need to make the
137 // existing Node occur higher up so it dominates its uses.
138 Node *old_ctrl;
139 IdealLoopTree *old_loop;
140
141 // The occasional new node
142 if( x->_idx >= old_unique ) { // Found a new, unplaced node?
143 old_ctrl = x->is_Con() ? C->root() : NULL;
144 old_loop = NULL; // Not in any prior loop
145 } else {
146 old_ctrl = x->is_Con() ? C->root() : get_ctrl(x);
147 old_loop = get_loop(old_ctrl); // Get prior loop
148 }
149 // New late point must dominate new use
150 Node *new_ctrl = dom_lca( old_ctrl, region->in(i2) );
151 // Set new location
152 set_ctrl(x, new_ctrl);
153 IdealLoopTree *new_loop = get_loop( new_ctrl );
154 // If changing loop bodies, see if we need to collect into new body
155 if( old_loop != new_loop ) {
156 if( old_loop && !old_loop->_child )
157 old_loop->_body.yank(x);
158 if( !new_loop->_child )
159 new_loop->_body.push(x); // Collect body info
160 }
161 }
162
163 return phi;
164 }
165
166 //------------------------------dominated_by------------------------------------
167 // Replace the dominated test with an obvious true or false. Place it on the
168 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
169 // live path up to the dominating control.
170 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff ) {
171 #ifndef PRODUCT
172 if( VerifyLoopOptimizations && PrintOpto ) tty->print_cr("dominating test");
173 #endif
174
175
176 // prevdom is the dominating projection of the dominating test.
177 assert( iff->is_If(), "" );
178 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
179 int pop = prevdom->Opcode();
180 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
181 // 'con' is set to true or false to kill the dominated test.
182 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
183 set_ctrl(con, C->root()); // Constant gets a new use
184 // Hack the dominated test
185 _igvn.hash_delete(iff);
186 iff->set_req(1, con);
187 _igvn._worklist.push(iff);
188
189 // If I dont have a reachable TRUE and FALSE path following the IfNode then
190 // I can assume this path reaches an infinite loop. In this case it's not
191 // important to optimize the data Nodes - either the whole compilation will
192 // be tossed or this path (and all data Nodes) will go dead.
193 if( iff->outcnt() != 2 ) return;
194
195 // Make control-dependent data Nodes on the live path (path that will remain
196 // once the dominated IF is removed) become control-dependent on the
197 // dominating projection.
198 Node* dp = ((IfNode*)iff)->proj_out(pop == Op_IfTrue);
199 IdealLoopTree *old_loop = get_loop(dp);
200
201 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
202 Node* cd = dp->fast_out(i); // Control-dependent node
203 if( cd->depends_only_on_test() ) {
204 assert( cd->in(0) == dp, "" );
205 _igvn.hash_delete( cd );
206 cd->set_req(0, prevdom);
207 set_early_ctrl( cd );
208 _igvn._worklist.push(cd);
209 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
210 if( old_loop != new_loop ) {
211 if( !old_loop->_child ) old_loop->_body.yank(cd);
212 if( !new_loop->_child ) new_loop->_body.push(cd);
213 }
214 --i;
215 --imax;
216 }
217 }
218 }
219
220 //------------------------------has_local_phi_input----------------------------
221 // Return TRUE if 'n' has Phi inputs from its local block and no other
222 // block-local inputs (all non-local-phi inputs come from earlier blocks)
223 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
224 Node *n_ctrl = get_ctrl(n);
225 // See if some inputs come from a Phi in this block, or from before
226 // this block.
227 uint i;
228 for( i = 1; i < n->req(); i++ ) {
229 Node *phi = n->in(i);
230 if( phi->is_Phi() && phi->in(0) == n_ctrl )
231 break;
232 }
233 if( i >= n->req() )
234 return NULL; // No Phi inputs; nowhere to clone thru
235
236 // Check for inputs created between 'n' and the Phi input. These
237 // must split as well; they have already been given the chance
238 // (courtesy of a post-order visit) and since they did not we must
239 // recover the 'cost' of splitting them by being very profitable
240 // when splitting 'n'. Since this is unlikely we simply give up.
241 for( i = 1; i < n->req(); i++ ) {
242 Node *m = n->in(i);
243 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
244 // We allow the special case of AddP's with no local inputs.
245 // This allows us to split-up address expressions.
246 if (m->is_AddP() &&
247 get_ctrl(m->in(2)) != n_ctrl &&
248 get_ctrl(m->in(3)) != n_ctrl) {
249 // Move the AddP up to dominating point
250 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl)));
251 continue;
252 }
253 return NULL;
254 }
255 }
256
257 return n_ctrl;
258 }
259
260 //------------------------------remix_address_expressions----------------------
261 // Rework addressing expressions to get the most loop-invariant stuff
262 // moved out. We'd like to do all associative operators, but it's especially
263 // important (common) to do address expressions.
264 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
265 if (!has_ctrl(n)) return NULL;
266 Node *n_ctrl = get_ctrl(n);
267 IdealLoopTree *n_loop = get_loop(n_ctrl);
268
269 // See if 'n' mixes loop-varying and loop-invariant inputs and
270 // itself is loop-varying.
271
272 // Only interested in binary ops (and AddP)
273 if( n->req() < 3 || n->req() > 4 ) return NULL;
274
275 Node *n1_ctrl = get_ctrl(n->in( 1));
276 Node *n2_ctrl = get_ctrl(n->in( 2));
277 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
278 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
279 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
280 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
281
282 // Does one of my inputs spin in a tighter loop than self?
283 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
284 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
285 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
286 return NULL; // Leave well enough alone
287
288 // Is at least one of my inputs loop-invariant?
289 if( n1_loop == n_loop &&
290 n2_loop == n_loop &&
291 n3_loop == n_loop )
292 return NULL; // No loop-invariant inputs
293
294
295 int n_op = n->Opcode();
296
297 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
298 if( n_op == Op_LShiftI ) {
299 // Scale is loop invariant
300 Node *scale = n->in(2);
301 Node *scale_ctrl = get_ctrl(scale);
302 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
303 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
304 return NULL;
305 const TypeInt *scale_t = scale->bottom_type()->isa_int();
306 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
307 return NULL; // Dont bother with byte/short masking
308 // Add must vary with loop (else shift would be loop-invariant)
309 Node *add = n->in(1);
310 Node *add_ctrl = get_ctrl(add);
311 IdealLoopTree *add_loop = get_loop(add_ctrl);
312 //assert( n_loop == add_loop, "" );
313 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
314
315 // Convert I-V into I+ (0-V); same for V-I
316 if( add->Opcode() == Op_SubI &&
317 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
318 Node *zero = _igvn.intcon(0);
319 set_ctrl(zero, C->root());
320 Node *neg = new (C, 3) SubINode( _igvn.intcon(0), add->in(2) );
321 register_new_node( neg, get_ctrl(add->in(2) ) );
322 add = new (C, 3) AddINode( add->in(1), neg );
323 register_new_node( add, add_ctrl );
324 }
325 if( add->Opcode() != Op_AddI ) return NULL;
326 // See if one add input is loop invariant
327 Node *add_var = add->in(1);
328 Node *add_var_ctrl = get_ctrl(add_var);
329 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
330 Node *add_invar = add->in(2);
331 Node *add_invar_ctrl = get_ctrl(add_invar);
332 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
333 if( add_var_loop == n_loop ) {
334 } else if( add_invar_loop == n_loop ) {
335 // Swap to find the invariant part
336 add_invar = add_var;
337 add_invar_ctrl = add_var_ctrl;
338 add_invar_loop = add_var_loop;
339 add_var = add->in(2);
340 Node *add_var_ctrl = get_ctrl(add_var);
341 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
342 } else // Else neither input is loop invariant
343 return NULL;
344 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
345 return NULL; // No invariant part of the add?
346
347 // Yes! Reshape address expression!
348 Node *inv_scale = new (C, 3) LShiftINode( add_invar, scale );
349 register_new_node( inv_scale, add_invar_ctrl );
350 Node *var_scale = new (C, 3) LShiftINode( add_var, scale );
351 register_new_node( var_scale, n_ctrl );
352 Node *var_add = new (C, 3) AddINode( var_scale, inv_scale );
353 register_new_node( var_add, n_ctrl );
354 _igvn.hash_delete( n );
355 _igvn.subsume_node( n, var_add );
356 return var_add;
357 }
358
359 // Replace (I+V) with (V+I)
360 if( n_op == Op_AddI ||
361 n_op == Op_AddL ||
362 n_op == Op_AddF ||
363 n_op == Op_AddD ||
364 n_op == Op_MulI ||
365 n_op == Op_MulL ||
366 n_op == Op_MulF ||
367 n_op == Op_MulD ) {
368 if( n2_loop == n_loop ) {
369 assert( n1_loop != n_loop, "" );
370 n->swap_edges(1, 2);
371 }
372 }
373
374 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
375 // but not if I2 is a constant.
376 if( n_op == Op_AddP ) {
377 if( n2_loop == n_loop && n3_loop != n_loop ) {
378 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
379 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
380 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
381 IdealLoopTree *n22loop = get_loop( n22_ctrl );
382 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
383 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
384 n23_loop == n_loop ) {
385 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
386 // Stuff new AddP in the loop preheader
387 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
388 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) );
389 register_new_node( add2, n_ctrl );
390 _igvn.hash_delete( n );
391 _igvn.subsume_node( n, add2 );
392 return add2;
393 }
394 }
395 }
396
397 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
398 if( n2_loop != n_loop && n3_loop == n_loop ) {
399 if( n->in(3)->Opcode() == Op_AddI ) {
400 Node *V = n->in(3)->in(1);
401 Node *I = n->in(3)->in(2);
402 if( is_member(n_loop,get_ctrl(V)) ) {
403 } else {
404 Node *tmp = V; V = I; I = tmp;
405 }
406 if( !is_member(n_loop,get_ctrl(I)) ) {
407 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I );
408 // Stuff new AddP in the loop preheader
409 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
410 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V );
411 register_new_node( add2, n_ctrl );
412 _igvn.hash_delete( n );
413 _igvn.subsume_node( n, add2 );
414 return add2;
415 }
416 }
417 }
418 }
419
420 return NULL;
421 }
422
423 //------------------------------conditional_move-------------------------------
424 // Attempt to replace a Phi with a conditional move. We have some pretty
425 // strict profitability requirements. All Phis at the merge point must
426 // be converted, so we can remove the control flow. We need to limit the
427 // number of c-moves to a small handful. All code that was in the side-arms
428 // of the CFG diamond is now speculatively executed. This code has to be
429 // "cheap enough". We are pretty much limited to CFG diamonds that merge
430 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
431 Node *PhaseIdealLoop::conditional_move( Node *region ) {
432
433 assert( region->is_Region(), "sanity check" );
434 if( region->req() != 3 ) return NULL;
435
436 // Check for CFG diamond
437 Node *lp = region->in(1);
438 Node *rp = region->in(2);
439 if( !lp || !rp ) return NULL;
440 Node *lp_c = lp->in(0);
441 if( lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If() ) return NULL;
442 IfNode *iff = lp_c->as_If();
443
444 // Check for highly predictable branch. No point in CMOV'ing if
445 // we are going to predict accurately all the time.
446 // %%% This hides patterns produced by utility methods like Math.min.
447 if( iff->_prob < PROB_UNLIKELY_MAG(3) ||
448 iff->_prob > PROB_LIKELY_MAG(3) )
449 return NULL;
450
451 // Check for ops pinned in an arm of the diamond.
452 // Can't remove the control flow in this case
453 if( lp->outcnt() > 1 ) return NULL;
454 if( rp->outcnt() > 1 ) return NULL;
455
456 // Check profitability
457 int cost = 0;
458 int phis = 0;
459 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
460 Node *out = region->fast_out(i);
461 if( !out->is_Phi() ) continue; // Ignore other control edges, etc
462 phis++;
463 PhiNode* phi = out->as_Phi();
464 switch (phi->type()->basic_type()) {
465 case T_LONG:
466 cost++; // Probably encodes as 2 CMOV's
467 case T_INT: // These all CMOV fine
468 case T_FLOAT:
469 case T_DOUBLE:
470 case T_ADDRESS: // (RawPtr)
471 cost++;
472 break;
473 case T_NARROWOOP: // Fall through
474 case T_OBJECT: { // Base oops are OK, but not derived oops
475 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
476 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
477 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
478 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
479 // have a Phi for the base here that we convert to a CMOVE all is well
480 // and good. But if the base is dead, we'll not make a CMOVE. Later
481 // the allocator will have to produce a base by creating a CMOVE of the
482 // relevant bases. This puts the allocator in the business of
483 // manufacturing expensive instructions, generally a bad plan.
484 // Just Say No to Conditionally-Moved Derived Pointers.
485 if( tp && tp->offset() != 0 )
486 return NULL;
487 cost++;
488 break;
489 }
490 default:
491 return NULL; // In particular, can't do memory or I/O
492 }
493 // Add in cost any speculative ops
494 for( uint j = 1; j < region->req(); j++ ) {
495 Node *proj = region->in(j);
496 Node *inp = phi->in(j);
497 if (get_ctrl(inp) == proj) { // Found local op
498 cost++;
499 // Check for a chain of dependent ops; these will all become
500 // speculative in a CMOV.
501 for( uint k = 1; k < inp->req(); k++ )
502 if (get_ctrl(inp->in(k)) == proj)
503 return NULL; // Too much speculative goo
504 }
505 }
506 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
507 // This will likely Split-If, a higher-payoff operation.
508 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
509 Node* use = phi->fast_out(k);
510 if( use->is_Cmp() || use->is_DecodeN() || use->is_EncodeP() )
511 return NULL;
512 }
513 }
514 if( cost >= ConditionalMoveLimit ) return NULL; // Too much goo
515 Node* bol = iff->in(1);
516 assert( bol->Opcode() == Op_Bool, "" );
517 int cmp_op = bol->in(1)->Opcode();
518 // It is expensive to generate flags from a float compare.
519 // Avoid duplicated float compare.
520 if( phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
521
522 // --------------
523 // Now replace all Phis with CMOV's
524 Node *cmov_ctrl = iff->in(0);
525 uint flip = (lp->Opcode() == Op_IfTrue);
526 while( 1 ) {
527 PhiNode* phi = NULL;
528 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
529 Node *out = region->fast_out(i);
530 if (out->is_Phi()) {
531 phi = out->as_Phi();
532 break;
533 }
534 }
535 if (phi == NULL) break;
536 #ifndef PRODUCT
537 if( PrintOpto && VerifyLoopOptimizations ) tty->print_cr("CMOV");
538 #endif
539 // Move speculative ops
540 for( uint j = 1; j < region->req(); j++ ) {
541 Node *proj = region->in(j);
542 Node *inp = phi->in(j);
543 if (get_ctrl(inp) == proj) { // Found local op
544 #ifndef PRODUCT
545 if( PrintOpto && VerifyLoopOptimizations ) {
546 tty->print(" speculate: ");
547 inp->dump();
548 }
549 #endif
550 set_ctrl(inp, cmov_ctrl);
551 }
552 }
553 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) );
554 register_new_node( cmov, cmov_ctrl );
555 _igvn.hash_delete(phi);
556 _igvn.subsume_node( phi, cmov );
557 #ifndef PRODUCT
558 if( VerifyLoopOptimizations ) verify();
559 #endif
560 }
561
562 // The useless CFG diamond will fold up later; see the optimization in
563 // RegionNode::Ideal.
564 _igvn._worklist.push(region);
565
566 return iff->in(1);
567 }
568
569 //------------------------------split_if_with_blocks_pre-----------------------
570 // Do the real work in a non-recursive function. Data nodes want to be
571 // cloned in the pre-order so they can feed each other nicely.
572 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
573 // Cloning these guys is unlikely to win
574 int n_op = n->Opcode();
575 if( n_op == Op_MergeMem ) return n;
576 if( n->is_Proj() ) return n;
577 // Do not clone-up CmpFXXX variations, as these are always
578 // followed by a CmpI
579 if( n->is_Cmp() ) return n;
580 // Attempt to use a conditional move instead of a phi/branch
581 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
582 Node *cmov = conditional_move( n );
583 if( cmov ) return cmov;
584 }
585 if( n->is_CFG() || n->is_LoadStore() )
586 return n;
587 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
588 n_op == Op_Opaque2 ) {
589 if( !C->major_progress() ) // If chance of no more loop opts...
590 _igvn._worklist.push(n); // maybe we'll remove them
591 return n;
592 }
593
594 if( n->is_Con() ) return n; // No cloning for Con nodes
595
596 Node *n_ctrl = get_ctrl(n);
597 if( !n_ctrl ) return n; // Dead node
598
599 // Attempt to remix address expressions for loop invariants
600 Node *m = remix_address_expressions( n );
601 if( m ) return m;
602
603 // Determine if the Node has inputs from some local Phi.
604 // Returns the block to clone thru.
605 Node *n_blk = has_local_phi_input( n );
606 if( !n_blk ) return n;
607 // Do not clone the trip counter through on a CountedLoop
608 // (messes up the canonical shape).
609 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
610
611 // Check for having no control input; not pinned. Allow
612 // dominating control.
613 if( n->in(0) ) {
614 Node *dom = idom(n_blk);
615 if( dom_lca( n->in(0), dom ) != n->in(0) )
616 return n;
617 }
618 // Policy: when is it profitable. You must get more wins than
619 // policy before it is considered profitable. Policy is usually 0,
620 // so 1 win is considered profitable. Big merges will require big
621 // cloning, so get a larger policy.
622 int policy = n_blk->req() >> 2;
623
624 // If the loop is a candidate for range check elimination,
625 // delay splitting through it's phi until a later loop optimization
626 if (n_blk->is_CountedLoop()) {
627 IdealLoopTree *lp = get_loop(n_blk);
628 if (lp && lp->_rce_candidate) {
629 return n;
630 }
631 }
632
633 // Use same limit as split_if_with_blocks_post
634 if( C->unique() > 35000 ) return n; // Method too big
635
636 // Split 'n' through the merge point if it is profitable
637 Node *phi = split_thru_phi( n, n_blk, policy );
638 if( !phi ) return n;
639
640 // Found a Phi to split thru!
641 // Replace 'n' with the new phi
642 _igvn.hash_delete(n);
643 _igvn.subsume_node( n, phi );
644 // Moved a load around the loop, 'en-registering' something.
645 if( n_blk->Opcode() == Op_Loop && n->is_Load() &&
646 !phi->in(LoopNode::LoopBackControl)->is_Load() )
647 C->set_major_progress();
648
649 return phi;
650 }
651
652 static bool merge_point_too_heavy(Compile* C, Node* region) {
653 // Bail out if the region and its phis have too many users.
654 int weight = 0;
655 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
656 weight += region->fast_out(i)->outcnt();
657 }
658 int nodes_left = MaxNodeLimit - C->unique();
659 if (weight * 8 > nodes_left) {
660 #ifndef PRODUCT
661 if (PrintOpto)
662 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
663 #endif
664 return true;
665 } else {
666 return false;
667 }
668 }
669
670 #ifdef _LP64
671 static bool merge_point_safe(Node* region) {
672 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
673 // having a PhiNode input. This sidesteps the dangerous case where the split
674 // ConvI2LNode may become TOP if the input Value() does not
675 // overlap the ConvI2L range, leaving a node which may not dominate its
676 // uses.
677 // A better fix for this problem can be found in the BugTraq entry, but
678 // expediency for Mantis demands this hack.
679 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
680 Node* n = region->fast_out(i);
681 if (n->is_Phi()) {
682 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
683 Node* m = n->fast_out(j);
684 if (m->Opcode() == Op_ConvI2L) {
685 return false;
686 }
687 }
688 }
689 }
690 return true;
691 }
692 #endif
693
694
695 //------------------------------place_near_use---------------------------------
696 // Place some computation next to use but not inside inner loops.
697 // For inner loop uses move it to the preheader area.
698 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const {
699 IdealLoopTree *u_loop = get_loop( useblock );
700 return (u_loop->_irreducible || u_loop->_child)
701 ? useblock
702 : u_loop->_head->in(LoopNode::EntryControl);
703 }
704
705
706 //------------------------------split_if_with_blocks_post----------------------
707 // Do the real work in a non-recursive function. CFG hackery wants to be
708 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
709 // info.
710 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) {
711
712 // Cloning Cmp through Phi's involves the split-if transform.
713 // FastLock is not used by an If
714 if( n->is_Cmp() && !n->is_FastLock() ) {
715 if( C->unique() > 35000 ) return; // Method too big
716
717 // Do not do 'split-if' if irreducible loops are present.
718 if( _has_irreducible_loops )
719 return;
720
721 Node *n_ctrl = get_ctrl(n);
722 // Determine if the Node has inputs from some local Phi.
723 // Returns the block to clone thru.
724 Node *n_blk = has_local_phi_input( n );
725 if( n_blk != n_ctrl ) return;
726
727 if( merge_point_too_heavy(C, n_ctrl) )
728 return;
729
730 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare?
731 Node *bol = n->unique_out();
732 assert( bol->is_Bool(), "expect a bool here" );
733 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare?
734 Node *iff = bol->unique_out();
735
736 // Check some safety conditions
737 if( iff->is_If() ) { // Classic split-if?
738 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if
739 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
740 if( get_ctrl(iff->in(2)) == n_ctrl ||
741 get_ctrl(iff->in(3)) == n_ctrl )
742 return; // Inputs not yet split-up
743 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) {
744 return; // Loop-invar test gates loop-varying CMOVE
745 }
746 } else {
747 return; // some other kind of node, such as an Allocate
748 }
749
750 // Do not do 'split-if' if some paths are dead. First do dead code
751 // elimination and then see if its still profitable.
752 for( uint i = 1; i < n_ctrl->req(); i++ )
753 if( n_ctrl->in(i) == C->top() )
754 return;
755
756 // When is split-if profitable? Every 'win' on means some control flow
757 // goes dead, so it's almost always a win.
758 int policy = 0;
759 // If trying to do a 'Split-If' at the loop head, it is only
760 // profitable if the cmp folds up on BOTH paths. Otherwise we
761 // risk peeling a loop forever.
762
763 // CNC - Disabled for now. Requires careful handling of loop
764 // body selection for the cloned code. Also, make sure we check
765 // for any input path not being in the same loop as n_ctrl. For
766 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
767 // because the alternative loop entry points won't be converted
768 // into LoopNodes.
769 IdealLoopTree *n_loop = get_loop(n_ctrl);
770 for( uint j = 1; j < n_ctrl->req(); j++ )
771 if( get_loop(n_ctrl->in(j)) != n_loop )
772 return;
773
774 #ifdef _LP64
775 // Check for safety of the merge point.
776 if( !merge_point_safe(n_ctrl) ) {
777 return;
778 }
779 #endif
780
781 // Split compare 'n' through the merge point if it is profitable
782 Node *phi = split_thru_phi( n, n_ctrl, policy );
783 if( !phi ) return;
784
785 // Found a Phi to split thru!
786 // Replace 'n' with the new phi
787 _igvn.hash_delete(n);
788 _igvn.subsume_node( n, phi );
789
790 // Now split the bool up thru the phi
791 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 );
792 _igvn.hash_delete(bol);
793 _igvn.subsume_node( bol, bolphi );
794 assert( iff->in(1) == bolphi, "" );
795 if( bolphi->Value(&_igvn)->singleton() )
796 return;
797
798 // Conditional-move? Must split up now
799 if( !iff->is_If() ) {
800 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 );
801 _igvn.hash_delete(iff);
802 _igvn.subsume_node( iff, cmovphi );
803 return;
804 }
805
806 // Now split the IF
807 do_split_if( iff );
808 return;
809 }
810
811 // Check for an IF ready to split; one that has its
812 // condition codes input coming from a Phi at the block start.
813 int n_op = n->Opcode();
814
815 // Check for an IF being dominated by another IF same test
816 if( n_op == Op_If ) {
817 Node *bol = n->in(1);
818 uint max = bol->outcnt();
819 // Check for same test used more than once?
820 if( n_op == Op_If && max > 1 && bol->is_Bool() ) {
821 // Search up IDOMs to see if this IF is dominated.
822 Node *cutoff = get_ctrl(bol);
823
824 // Now search up IDOMs till cutoff, looking for a dominating test
825 Node *prevdom = n;
826 Node *dom = idom(prevdom);
827 while( dom != cutoff ) {
828 if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) {
829 // Replace the dominated test with an obvious true or false.
830 // Place it on the IGVN worklist for later cleanup.
831 C->set_major_progress();
832 dominated_by( prevdom, n );
833 #ifndef PRODUCT
834 if( VerifyLoopOptimizations ) verify();
835 #endif
836 return;
837 }
838 prevdom = dom;
839 dom = idom(prevdom);
840 }
841 }
842 }
843
844 // See if a shared loop-varying computation has no loop-varying uses.
845 // Happens if something is only used for JVM state in uncommon trap exits,
846 // like various versions of induction variable+offset. Clone the
847 // computation per usage to allow it to sink out of the loop.
848 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
849 Node *n_ctrl = get_ctrl(n);
850 IdealLoopTree *n_loop = get_loop(n_ctrl);
851 if( n_loop != _ltree_root ) {
852 DUIterator_Fast imax, i = n->fast_outs(imax);
853 for (; i < imax; i++) {
854 Node* u = n->fast_out(i);
855 if( !has_ctrl(u) ) break; // Found control user
856 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
857 if( u_loop == n_loop ) break; // Found loop-varying use
858 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
859 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
860 }
861 bool did_break = (i < imax); // Did we break out of the previous loop?
862 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
863 Node *late_load_ctrl;
864 if (n->is_Load()) {
865 // If n is a load, get and save the result from get_late_ctrl(),
866 // to be later used in calculating the control for n's clones.
867 clear_dom_lca_tags();
868 late_load_ctrl = get_late_ctrl(n, n_ctrl);
869 }
870 // If n is a load, and the late control is the same as the current
871 // control, then the cloning of n is a pointless exercise, because
872 // GVN will ensure that we end up where we started.
873 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
874 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
875 Node *u = n->last_out(j); // Clone private computation per use
876 _igvn.hash_delete(u);
877 _igvn._worklist.push(u);
878 Node *x = n->clone(); // Clone computation
879 Node *x_ctrl = NULL;
880 if( u->is_Phi() ) {
881 // Replace all uses of normal nodes. Replace Phi uses
882 // individually, so the seperate Nodes can sink down
883 // different paths.
884 uint k = 1;
885 while( u->in(k) != n ) k++;
886 u->set_req( k, x );
887 // x goes next to Phi input path
888 x_ctrl = u->in(0)->in(k);
889 --j;
890 } else { // Normal use
891 // Replace all uses
892 for( uint k = 0; k < u->req(); k++ ) {
893 if( u->in(k) == n ) {
894 u->set_req( k, x );
895 --j;
896 }
897 }
898 x_ctrl = get_ctrl(u);
899 }
900
901 // Find control for 'x' next to use but not inside inner loops.
902 // For inner loop uses get the preheader area.
903 x_ctrl = place_near_use(x_ctrl);
904
905 if (n->is_Load()) {
906 // For loads, add a control edge to a CFG node outside of the loop
907 // to force them to not combine and return back inside the loop
908 // during GVN optimization (4641526).
909 //
910 // Because we are setting the actual control input, factor in
911 // the result from get_late_ctrl() so we respect any
912 // anti-dependences. (6233005).
913 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
914
915 // Don't allow the control input to be a CFG splitting node.
916 // Such nodes should only have ProjNodes as outs, e.g. IfNode
917 // should only have IfTrueNode and IfFalseNode (4985384).
918 x_ctrl = find_non_split_ctrl(x_ctrl);
919 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
920
921 x->set_req(0, x_ctrl);
922 }
923 register_new_node(x, x_ctrl);
924
925 // Some institutional knowledge is needed here: 'x' is
926 // yanked because if the optimizer runs GVN on it all the
927 // cloned x's will common up and undo this optimization and
928 // be forced back in the loop. This is annoying because it
929 // makes +VerifyOpto report false-positives on progress. I
930 // tried setting control edges on the x's to force them to
931 // not combine, but the matching gets worried when it tries
932 // to fold a StoreP and an AddP together (as part of an
933 // address expression) and the AddP and StoreP have
934 // different controls.
935 if( !x->is_Load() && !x->is_DecodeN() ) _igvn._worklist.yank(x);
936 }
937 _igvn.remove_dead_node(n);
938 }
939 }
940 }
941 }
942
943 // Check for Opaque2's who's loop has disappeared - who's input is in the
944 // same loop nest as their output. Remove 'em, they are no longer useful.
945 if( n_op == Op_Opaque2 &&
946 n->in(1) != NULL &&
947 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
948 _igvn.add_users_to_worklist(n);
949 _igvn.hash_delete(n);
950 _igvn.subsume_node( n, n->in(1) );
951 }
952 }
953
954 //------------------------------split_if_with_blocks---------------------------
955 // Check for aggressive application of 'split-if' optimization,
956 // using basic block level info.
957 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) {
958 Node *n = C->root();
959 visited.set(n->_idx); // first, mark node as visited
960 // Do pre-visit work for root
961 n = split_if_with_blocks_pre( n );
962 uint cnt = n->outcnt();
963 uint i = 0;
964 while (true) {
965 // Visit all children
966 if (i < cnt) {
967 Node* use = n->raw_out(i);
968 ++i;
969 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
970 // Now do pre-visit work for this use
971 use = split_if_with_blocks_pre( use );
972 nstack.push(n, i); // Save parent and next use's index.
973 n = use; // Process all children of current use.
974 cnt = use->outcnt();
975 i = 0;
976 }
977 }
978 else {
979 // All of n's children have been processed, complete post-processing.
980 if (cnt != 0 && !n->is_Con()) {
981 assert(has_node(n), "no dead nodes");
982 split_if_with_blocks_post( n );
983 }
984 if (nstack.is_empty()) {
985 // Finished all nodes on stack.
986 break;
987 }
988 // Get saved parent node and next use's index. Visit the rest of uses.
989 n = nstack.node();
990 cnt = n->outcnt();
991 i = nstack.index();
992 nstack.pop();
993 }
994 }
995 }
996
997
998 //=============================================================================
999 //
1000 // C L O N E A L O O P B O D Y
1001 //
1002
1003 //------------------------------clone_iff--------------------------------------
1004 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1005 // "Nearly" because all Nodes have been cloned from the original in the loop,
1006 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1007 // through the Phi recursively, and return a Bool.
1008 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) {
1009
1010 // Convert this Phi into a Phi merging Bools
1011 uint i;
1012 for( i = 1; i < phi->req(); i++ ) {
1013 Node *b = phi->in(i);
1014 if( b->is_Phi() ) {
1015 _igvn.hash_delete(phi);
1016 _igvn._worklist.push(phi);
1017 phi->set_req(i, clone_iff( b->as_Phi(), loop ));
1018 } else {
1019 assert( b->is_Bool(), "" );
1020 }
1021 }
1022
1023 Node *sample_bool = phi->in(1);
1024 Node *sample_cmp = sample_bool->in(1);
1025
1026 // Make Phis to merge the Cmp's inputs.
1027 int size = phi->in(0)->req();
1028 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1029 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1030 for( i = 1; i < phi->req(); i++ ) {
1031 Node *n1 = phi->in(i)->in(1)->in(1);
1032 Node *n2 = phi->in(i)->in(1)->in(2);
1033 phi1->set_req( i, n1 );
1034 phi2->set_req( i, n2 );
1035 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1036 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1037 }
1038 // See if these Phis have been made before.
1039 // Register with optimizer
1040 Node *hit1 = _igvn.hash_find_insert(phi1);
1041 if( hit1 ) { // Hit, toss just made Phi
1042 _igvn.remove_dead_node(phi1); // Remove new phi
1043 assert( hit1->is_Phi(), "" );
1044 phi1 = (PhiNode*)hit1; // Use existing phi
1045 } else { // Miss
1046 _igvn.register_new_node_with_optimizer(phi1);
1047 }
1048 Node *hit2 = _igvn.hash_find_insert(phi2);
1049 if( hit2 ) { // Hit, toss just made Phi
1050 _igvn.remove_dead_node(phi2); // Remove new phi
1051 assert( hit2->is_Phi(), "" );
1052 phi2 = (PhiNode*)hit2; // Use existing phi
1053 } else { // Miss
1054 _igvn.register_new_node_with_optimizer(phi2);
1055 }
1056 // Register Phis with loop/block info
1057 set_ctrl(phi1, phi->in(0));
1058 set_ctrl(phi2, phi->in(0));
1059 // Make a new Cmp
1060 Node *cmp = sample_cmp->clone();
1061 cmp->set_req( 1, phi1 );
1062 cmp->set_req( 2, phi2 );
1063 _igvn.register_new_node_with_optimizer(cmp);
1064 set_ctrl(cmp, phi->in(0));
1065
1066 // Make a new Bool
1067 Node *b = sample_bool->clone();
1068 b->set_req(1,cmp);
1069 _igvn.register_new_node_with_optimizer(b);
1070 set_ctrl(b, phi->in(0));
1071
1072 assert( b->is_Bool(), "" );
1073 return (BoolNode*)b;
1074 }
1075
1076 //------------------------------clone_bool-------------------------------------
1077 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1078 // "Nearly" because all Nodes have been cloned from the original in the loop,
1079 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1080 // through the Phi recursively, and return a Bool.
1081 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1082 uint i;
1083 // Convert this Phi into a Phi merging Bools
1084 for( i = 1; i < phi->req(); i++ ) {
1085 Node *b = phi->in(i);
1086 if( b->is_Phi() ) {
1087 _igvn.hash_delete(phi);
1088 _igvn._worklist.push(phi);
1089 phi->set_req(i, clone_bool( b->as_Phi(), loop ));
1090 } else {
1091 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1092 }
1093 }
1094
1095 Node *sample_cmp = phi->in(1);
1096
1097 // Make Phis to merge the Cmp's inputs.
1098 int size = phi->in(0)->req();
1099 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1100 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1101 for( uint j = 1; j < phi->req(); j++ ) {
1102 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1103 Node *n1, *n2;
1104 if( cmp_top->is_Cmp() ) {
1105 n1 = cmp_top->in(1);
1106 n2 = cmp_top->in(2);
1107 } else {
1108 n1 = n2 = cmp_top;
1109 }
1110 phi1->set_req( j, n1 );
1111 phi2->set_req( j, n2 );
1112 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1113 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1114 }
1115
1116 // See if these Phis have been made before.
1117 // Register with optimizer
1118 Node *hit1 = _igvn.hash_find_insert(phi1);
1119 if( hit1 ) { // Hit, toss just made Phi
1120 _igvn.remove_dead_node(phi1); // Remove new phi
1121 assert( hit1->is_Phi(), "" );
1122 phi1 = (PhiNode*)hit1; // Use existing phi
1123 } else { // Miss
1124 _igvn.register_new_node_with_optimizer(phi1);
1125 }
1126 Node *hit2 = _igvn.hash_find_insert(phi2);
1127 if( hit2 ) { // Hit, toss just made Phi
1128 _igvn.remove_dead_node(phi2); // Remove new phi
1129 assert( hit2->is_Phi(), "" );
1130 phi2 = (PhiNode*)hit2; // Use existing phi
1131 } else { // Miss
1132 _igvn.register_new_node_with_optimizer(phi2);
1133 }
1134 // Register Phis with loop/block info
1135 set_ctrl(phi1, phi->in(0));
1136 set_ctrl(phi2, phi->in(0));
1137 // Make a new Cmp
1138 Node *cmp = sample_cmp->clone();
1139 cmp->set_req( 1, phi1 );
1140 cmp->set_req( 2, phi2 );
1141 _igvn.register_new_node_with_optimizer(cmp);
1142 set_ctrl(cmp, phi->in(0));
1143
1144 assert( cmp->is_Cmp(), "" );
1145 return (CmpNode*)cmp;
1146 }
1147
1148 //------------------------------sink_use---------------------------------------
1149 // If 'use' was in the loop-exit block, it now needs to be sunk
1150 // below the post-loop merge point.
1151 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1152 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1153 set_ctrl(use, post_loop);
1154 for (DUIterator j = use->outs(); use->has_out(j); j++)
1155 sink_use(use->out(j), post_loop);
1156 }
1157 }
1158
1159 //------------------------------clone_loop-------------------------------------
1160 //
1161 // C L O N E A L O O P B O D Y
1162 //
1163 // This is the basic building block of the loop optimizations. It clones an
1164 // entire loop body. It makes an old_new loop body mapping; with this mapping
1165 // you can find the new-loop equivalent to an old-loop node. All new-loop
1166 // nodes are exactly equal to their old-loop counterparts, all edges are the
1167 // same. All exits from the old-loop now have a RegionNode that merges the
1168 // equivalent new-loop path. This is true even for the normal "loop-exit"
1169 // condition. All uses of loop-invariant old-loop values now come from (one
1170 // or more) Phis that merge their new-loop equivalents.
1171 //
1172 // This operation leaves the graph in an illegal state: there are two valid
1173 // control edges coming from the loop pre-header to both loop bodies. I'll
1174 // definitely have to hack the graph after running this transform.
1175 //
1176 // From this building block I will further edit edges to perform loop peeling
1177 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1178 //
1179 // Parameter side_by_size_idom:
1180 // When side_by_size_idom is NULL, the dominator tree is constructed for
1181 // the clone loop to dominate the original. Used in construction of
1182 // pre-main-post loop sequence.
1183 // When nonnull, the clone and original are side-by-side, both are
1184 // dominated by the side_by_side_idom node. Used in construction of
1185 // unswitched loops.
1186 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1187 Node* side_by_side_idom) {
1188
1189 // Step 1: Clone the loop body. Make the old->new mapping.
1190 uint i;
1191 for( i = 0; i < loop->_body.size(); i++ ) {
1192 Node *old = loop->_body.at(i);
1193 Node *nnn = old->clone();
1194 old_new.map( old->_idx, nnn );
1195 _igvn.register_new_node_with_optimizer(nnn);
1196 }
1197
1198
1199 // Step 2: Fix the edges in the new body. If the old input is outside the
1200 // loop use it. If the old input is INside the loop, use the corresponding
1201 // new node instead.
1202 for( i = 0; i < loop->_body.size(); i++ ) {
1203 Node *old = loop->_body.at(i);
1204 Node *nnn = old_new[old->_idx];
1205 // Fix CFG/Loop controlling the new node
1206 if (has_ctrl(old)) {
1207 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
1208 } else {
1209 set_loop(nnn, loop->_parent);
1210 if (old->outcnt() > 0) {
1211 set_idom( nnn, old_new[idom(old)->_idx], dd );
1212 }
1213 }
1214 // Correct edges to the new node
1215 for( uint j = 0; j < nnn->req(); j++ ) {
1216 Node *n = nnn->in(j);
1217 if( n ) {
1218 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
1219 if( loop->is_member( old_in_loop ) )
1220 nnn->set_req(j, old_new[n->_idx]);
1221 }
1222 }
1223 _igvn.hash_find_insert(nnn);
1224 }
1225 Node *newhead = old_new[loop->_head->_idx];
1226 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1227
1228
1229 // Step 3: Now fix control uses. Loop varying control uses have already
1230 // been fixed up (as part of all input edges in Step 2). Loop invariant
1231 // control uses must be either an IfFalse or an IfTrue. Make a merge
1232 // point to merge the old and new IfFalse/IfTrue nodes; make the use
1233 // refer to this.
1234 ResourceArea *area = Thread::current()->resource_area();
1235 Node_List worklist(area);
1236 uint new_counter = C->unique();
1237 for( i = 0; i < loop->_body.size(); i++ ) {
1238 Node* old = loop->_body.at(i);
1239 if( !old->is_CFG() ) continue;
1240 Node* nnn = old_new[old->_idx];
1241
1242 // Copy uses to a worklist, so I can munge the def-use info
1243 // with impunity.
1244 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1245 worklist.push(old->fast_out(j));
1246
1247 while( worklist.size() ) { // Visit all uses
1248 Node *use = worklist.pop();
1249 if (!has_node(use)) continue; // Ignore dead nodes
1250 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1251 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
1252 // Both OLD and USE are CFG nodes here.
1253 assert( use->is_Proj(), "" );
1254
1255 // Clone the loop exit control projection
1256 Node *newuse = use->clone();
1257 newuse->set_req(0,nnn);
1258 _igvn.register_new_node_with_optimizer(newuse);
1259 set_loop(newuse, use_loop);
1260 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
1261
1262 // We need a Region to merge the exit from the peeled body and the
1263 // exit from the old loop body.
1264 RegionNode *r = new (C, 3) RegionNode(3);
1265 // Map the old use to the new merge point
1266 old_new.map( use->_idx, r );
1267 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
1268 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
1269
1270 // The original user of 'use' uses 'r' instead.
1271 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
1272 Node* useuse = use->last_out(l);
1273 _igvn.hash_delete(useuse);
1274 _igvn._worklist.push(useuse);
1275 uint uses_found = 0;
1276 if( useuse->in(0) == use ) {
1277 useuse->set_req(0, r);
1278 uses_found++;
1279 if( useuse->is_CFG() ) {
1280 assert( dom_depth(useuse) > dd_r, "" );
1281 set_idom(useuse, r, dom_depth(useuse));
1282 }
1283 }
1284 for( uint k = 1; k < useuse->req(); k++ ) {
1285 if( useuse->in(k) == use ) {
1286 useuse->set_req(k, r);
1287 uses_found++;
1288 }
1289 }
1290 l -= uses_found; // we deleted 1 or more copies of this edge
1291 }
1292
1293 // Now finish up 'r'
1294 r->set_req( 1, newuse );
1295 r->set_req( 2, use );
1296 _igvn.register_new_node_with_optimizer(r);
1297 set_loop(r, use_loop);
1298 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
1299 } // End of if a loop-exit test
1300 }
1301 }
1302
1303 // Step 4: If loop-invariant use is not control, it must be dominated by a
1304 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
1305 // there if needed. Make a Phi there merging old and new used values.
1306 Node_List *split_if_set = NULL;
1307 Node_List *split_bool_set = NULL;
1308 Node_List *split_cex_set = NULL;
1309 for( i = 0; i < loop->_body.size(); i++ ) {
1310 Node* old = loop->_body.at(i);
1311 Node* nnn = old_new[old->_idx];
1312 // Copy uses to a worklist, so I can munge the def-use info
1313 // with impunity.
1314 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1315 worklist.push(old->fast_out(j));
1316
1317 while( worklist.size() ) {
1318 Node *use = worklist.pop();
1319 if (!has_node(use)) continue; // Ignore dead nodes
1320 if (use->in(0) == C->top()) continue;
1321 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1322 // Check for data-use outside of loop - at least one of OLD or USE
1323 // must not be a CFG node.
1324 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) {
1325
1326 // If the Data use is an IF, that means we have an IF outside of the
1327 // loop that is switching on a condition that is set inside of the
1328 // loop. Happens if people set a loop-exit flag; then test the flag
1329 // in the loop to break the loop, then test is again outside of the
1330 // loop to determine which way the loop exited.
1331 if( use->is_If() || use->is_CMove() ) {
1332 // Since this code is highly unlikely, we lazily build the worklist
1333 // of such Nodes to go split.
1334 if( !split_if_set )
1335 split_if_set = new Node_List(area);
1336 split_if_set->push(use);
1337 }
1338 if( use->is_Bool() ) {
1339 if( !split_bool_set )
1340 split_bool_set = new Node_List(area);
1341 split_bool_set->push(use);
1342 }
1343 if( use->Opcode() == Op_CreateEx ) {
1344 if( !split_cex_set )
1345 split_cex_set = new Node_List(area);
1346 split_cex_set->push(use);
1347 }
1348
1349
1350 // Get "block" use is in
1351 uint idx = 0;
1352 while( use->in(idx) != old ) idx++;
1353 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1354 assert( !loop->is_member( get_loop( prev ) ), "" );
1355 Node *cfg = prev->_idx >= new_counter
1356 ? prev->in(2)
1357 : idom(prev);
1358 if( use->is_Phi() ) // Phi use is in prior block
1359 cfg = prev->in(idx); // NOT in block of Phi itself
1360 if (cfg->is_top()) { // Use is dead?
1361 _igvn.hash_delete(use);
1362 _igvn._worklist.push(use);
1363 use->set_req(idx, C->top());
1364 continue;
1365 }
1366
1367 while( !loop->is_member( get_loop( cfg ) ) ) {
1368 prev = cfg;
1369 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1370 }
1371 // If the use occurs after merging several exits from the loop, then
1372 // old value must have dominated all those exits. Since the same old
1373 // value was used on all those exits we did not need a Phi at this
1374 // merge point. NOW we do need a Phi here. Each loop exit value
1375 // is now merged with the peeled body exit; each exit gets its own
1376 // private Phi and those Phis need to be merged here.
1377 Node *phi;
1378 if( prev->is_Region() ) {
1379 if( idx == 0 ) { // Updating control edge?
1380 phi = prev; // Just use existing control
1381 } else { // Else need a new Phi
1382 phi = PhiNode::make( prev, old );
1383 // Now recursively fix up the new uses of old!
1384 for( uint i = 1; i < prev->req(); i++ ) {
1385 worklist.push(phi); // Onto worklist once for each 'old' input
1386 }
1387 }
1388 } else {
1389 // Get new RegionNode merging old and new loop exits
1390 prev = old_new[prev->_idx];
1391 assert( prev, "just made this in step 7" );
1392 if( idx == 0 ) { // Updating control edge?
1393 phi = prev; // Just use existing control
1394 } else { // Else need a new Phi
1395 // Make a new Phi merging data values properly
1396 phi = PhiNode::make( prev, old );
1397 phi->set_req( 1, nnn );
1398 }
1399 }
1400 // If inserting a new Phi, check for prior hits
1401 if( idx != 0 ) {
1402 Node *hit = _igvn.hash_find_insert(phi);
1403 if( hit == NULL ) {
1404 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1405 } else { // or
1406 // Remove the new phi from the graph and use the hit
1407 _igvn.remove_dead_node(phi);
1408 phi = hit; // Use existing phi
1409 }
1410 set_ctrl(phi, prev);
1411 }
1412 // Make 'use' use the Phi instead of the old loop body exit value
1413 _igvn.hash_delete(use);
1414 _igvn._worklist.push(use);
1415 use->set_req(idx, phi);
1416 if( use->_idx >= new_counter ) { // If updating new phis
1417 // Not needed for correctness, but prevents a weak assert
1418 // in AddPNode from tripping (when we end up with different
1419 // base & derived Phis that will become the same after
1420 // IGVN does CSE).
1421 Node *hit = _igvn.hash_find_insert(use);
1422 if( hit ) // Go ahead and re-hash for hits.
1423 _igvn.subsume_node( use, hit );
1424 }
1425
1426 // If 'use' was in the loop-exit block, it now needs to be sunk
1427 // below the post-loop merge point.
1428 sink_use( use, prev );
1429 }
1430 }
1431 }
1432
1433 // Check for IFs that need splitting/cloning. Happens if an IF outside of
1434 // the loop uses a condition set in the loop. The original IF probably
1435 // takes control from one or more OLD Regions (which in turn get from NEW
1436 // Regions). In any case, there will be a set of Phis for each merge point
1437 // from the IF up to where the original BOOL def exists the loop.
1438 if( split_if_set ) {
1439 while( split_if_set->size() ) {
1440 Node *iff = split_if_set->pop();
1441 if( iff->in(1)->is_Phi() ) {
1442 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop );
1443 _igvn.hash_delete(iff);
1444 _igvn._worklist.push(iff);
1445 iff->set_req(1, b);
1446 }
1447 }
1448 }
1449 if( split_bool_set ) {
1450 while( split_bool_set->size() ) {
1451 Node *b = split_bool_set->pop();
1452 Node *phi = b->in(1);
1453 assert( phi->is_Phi(), "" );
1454 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop );
1455 _igvn.hash_delete(b);
1456 _igvn._worklist.push(b);
1457 b->set_req(1, cmp);
1458 }
1459 }
1460 if( split_cex_set ) {
1461 while( split_cex_set->size() ) {
1462 Node *b = split_cex_set->pop();
1463 assert( b->in(0)->is_Region(), "" );
1464 assert( b->in(1)->is_Phi(), "" );
1465 assert( b->in(0)->in(0) == b->in(1)->in(0), "" );
1466 split_up( b, b->in(0), NULL );
1467 }
1468 }
1469
1470 }
1471
1472
1473 //---------------------- stride_of_possible_iv -------------------------------------
1474 // Looks for an iff/bool/comp with one operand of the compare
1475 // being a cycle involving an add and a phi,
1476 // with an optional truncation (left-shift followed by a right-shift)
1477 // of the add. Returns zero if not an iv.
1478 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
1479 Node* trunc1 = NULL;
1480 Node* trunc2 = NULL;
1481 const TypeInt* ttype = NULL;
1482 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
1483 return 0;
1484 }
1485 BoolNode* bl = iff->in(1)->as_Bool();
1486 Node* cmp = bl->in(1);
1487 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) {
1488 return 0;
1489 }
1490 // Must have an invariant operand
1491 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
1492 return 0;
1493 }
1494 Node* add2 = NULL;
1495 Node* cmp1 = cmp->in(1);
1496 if (cmp1->is_Phi()) {
1497 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
1498 Node* phi = cmp1;
1499 for (uint i = 1; i < phi->req(); i++) {
1500 Node* in = phi->in(i);
1501 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
1502 &trunc1, &trunc2, &ttype);
1503 if (add && add->in(1) == phi) {
1504 add2 = add->in(2);
1505 break;
1506 }
1507 }
1508 } else {
1509 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
1510 Node* addtrunc = cmp1;
1511 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
1512 &trunc1, &trunc2, &ttype);
1513 if (add && add->in(1)->is_Phi()) {
1514 Node* phi = add->in(1);
1515 for (uint i = 1; i < phi->req(); i++) {
1516 if (phi->in(i) == addtrunc) {
1517 add2 = add->in(2);
1518 break;
1519 }
1520 }
1521 }
1522 }
1523 if (add2 != NULL) {
1524 const TypeInt* add2t = _igvn.type(add2)->is_int();
1525 if (add2t->is_con()) {
1526 return add2t->get_con();
1527 }
1528 }
1529 return 0;
1530 }
1531
1532
1533 //---------------------- stay_in_loop -------------------------------------
1534 // Return the (unique) control output node that's in the loop (if it exists.)
1535 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
1536 Node* unique = NULL;
1537 if (!n) return NULL;
1538 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1539 Node* use = n->fast_out(i);
1540 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
1541 if (unique != NULL) {
1542 return NULL;
1543 }
1544 unique = use;
1545 }
1546 }
1547 return unique;
1548 }
1549
1550 //------------------------------ register_node -------------------------------------
1551 // Utility to register node "n" with PhaseIdealLoop
1552 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
1553 _igvn.register_new_node_with_optimizer(n);
1554 loop->_body.push(n);
1555 if (n->is_CFG()) {
1556 set_loop(n, loop);
1557 set_idom(n, pred, ddepth);
1558 } else {
1559 set_ctrl(n, pred);
1560 }
1561 }
1562
1563 //------------------------------ proj_clone -------------------------------------
1564 // Utility to create an if-projection
1565 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
1566 ProjNode* c = p->clone()->as_Proj();
1567 c->set_req(0, iff);
1568 return c;
1569 }
1570
1571 //------------------------------ short_circuit_if -------------------------------------
1572 // Force the iff control output to be the live_proj
1573 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
1574 int proj_con = live_proj->_con;
1575 assert(proj_con == 0 || proj_con == 1, "false or true projection");
1576 Node *con = _igvn.intcon(proj_con);
1577 set_ctrl(con, C->root());
1578 if (iff) {
1579 iff->set_req(1, con);
1580 }
1581 return con;
1582 }
1583
1584 //------------------------------ insert_if_before_proj -------------------------------------
1585 // Insert a new if before an if projection (* - new node)
1586 //
1587 // before
1588 // if(test)
1589 // / \
1590 // v v
1591 // other-proj proj (arg)
1592 //
1593 // after
1594 // if(test)
1595 // / \
1596 // / v
1597 // | * proj-clone
1598 // v |
1599 // other-proj v
1600 // * new_if(relop(cmp[IU](left,right)))
1601 // / \
1602 // v v
1603 // * new-proj proj
1604 // (returned)
1605 //
1606 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
1607 IfNode* iff = proj->in(0)->as_If();
1608 IdealLoopTree *loop = get_loop(proj);
1609 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1610 int ddepth = dom_depth(proj);
1611
1612 _igvn.hash_delete(iff);
1613 _igvn._worklist.push(iff);
1614 _igvn.hash_delete(proj);
1615 _igvn._worklist.push(proj);
1616
1617 proj->set_req(0, NULL); // temporary disconnect
1618 ProjNode* proj2 = proj_clone(proj, iff);
1619 register_node(proj2, loop, iff, ddepth);
1620
1621 Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right);
1622 register_node(cmp, loop, proj2, ddepth);
1623
1624 BoolNode* bol = new (C,2)BoolNode(cmp, relop);
1625 register_node(bol, loop, proj2, ddepth);
1626
1627 IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt);
1628 register_node(new_if, loop, proj2, ddepth);
1629
1630 proj->set_req(0, new_if); // reattach
1631 set_idom(proj, new_if, ddepth);
1632
1633 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
1634 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
1635
1636 return new_exit;
1637 }
1638
1639 //------------------------------ insert_region_before_proj -------------------------------------
1640 // Insert a region before an if projection (* - new node)
1641 //
1642 // before
1643 // if(test)
1644 // / |
1645 // v |
1646 // proj v
1647 // other-proj
1648 //
1649 // after
1650 // if(test)
1651 // / |
1652 // v |
1653 // * proj-clone v
1654 // | other-proj
1655 // v
1656 // * new-region
1657 // |
1658 // v
1659 // * dum_if
1660 // / \
1661 // v \
1662 // * dum-proj v
1663 // proj
1664 //
1665 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
1666 IfNode* iff = proj->in(0)->as_If();
1667 IdealLoopTree *loop = get_loop(proj);
1668 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1669 int ddepth = dom_depth(proj);
1670
1671 _igvn.hash_delete(iff);
1672 _igvn._worklist.push(iff);
1673 _igvn.hash_delete(proj);
1674 _igvn._worklist.push(proj);
1675
1676 proj->set_req(0, NULL); // temporary disconnect
1677 ProjNode* proj2 = proj_clone(proj, iff);
1678 register_node(proj2, loop, iff, ddepth);
1679
1680 RegionNode* reg = new (C,2)RegionNode(2);
1681 reg->set_req(1, proj2);
1682 register_node(reg, loop, iff, ddepth);
1683
1684 IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
1685 register_node(dum_if, loop, reg, ddepth);
1686
1687 proj->set_req(0, dum_if); // reattach
1688 set_idom(proj, dum_if, ddepth);
1689
1690 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
1691 register_node(dum_proj, loop, dum_if, ddepth);
1692
1693 return reg;
1694 }
1695
1696 //------------------------------ insert_cmpi_loop_exit -------------------------------------
1697 // Clone a signed compare loop exit from an unsigned compare and
1698 // insert it before the unsigned cmp on the stay-in-loop path.
1699 // All new nodes inserted in the dominator tree between the original
1700 // if and it's projections. The original if test is replaced with
1701 // a constant to force the stay-in-loop path.
1702 //
1703 // This is done to make sure that the original if and it's projections
1704 // still dominate the same set of control nodes, that the ctrl() relation
1705 // from data nodes to them is preserved, and that their loop nesting is
1706 // preserved.
1707 //
1708 // before
1709 // if(i <u limit) unsigned compare loop exit
1710 // / |
1711 // v v
1712 // exit-proj stay-in-loop-proj
1713 //
1714 // after
1715 // if(stay-in-loop-const) original if
1716 // / |
1717 // / v
1718 // / if(i < limit) new signed test
1719 // / / |
1720 // / / v
1721 // / / if(i <u limit) new cloned unsigned test
1722 // / / / |
1723 // v v v |
1724 // region |
1725 // | |
1726 // dum-if |
1727 // / | |
1728 // ether | |
1729 // v v
1730 // exit-proj stay-in-loop-proj
1731 //
1732 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
1733 const bool Signed = true;
1734 const bool Unsigned = false;
1735
1736 BoolNode* bol = if_cmpu->in(1)->as_Bool();
1737 if (bol->_test._test != BoolTest::lt) return NULL;
1738 CmpNode* cmpu = bol->in(1)->as_Cmp();
1739 if (cmpu->Opcode() != Op_CmpU) return NULL;
1740 int stride = stride_of_possible_iv(if_cmpu);
1741 if (stride == 0) return NULL;
1742
1743 ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj();
1744 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
1745
1746 Node* limit = NULL;
1747 if (stride > 0) {
1748 limit = cmpu->in(2);
1749 } else {
1750 limit = _igvn.makecon(TypeInt::ZERO);
1751 set_ctrl(limit, C->root());
1752 }
1753 // Create a new region on the exit path
1754 RegionNode* reg = insert_region_before_proj(lp_exit);
1755
1756 // Clone the if-cmpu-true-false using a signed compare
1757 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
1758 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
1759 reg->add_req(cmpi_exit);
1760
1761 // Clone the if-cmpu-true-false
1762 BoolTest::mask rel_u = bol->_test._test;
1763 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
1764 reg->add_req(cmpu_exit);
1765
1766 // Force original if to stay in loop.
1767 short_circuit_if(if_cmpu, lp_continue);
1768
1769 return cmpi_exit->in(0)->as_If();
1770 }
1771
1772 //------------------------------ remove_cmpi_loop_exit -------------------------------------
1773 // Remove a previously inserted signed compare loop exit.
1774 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
1775 Node* lp_proj = stay_in_loop(if_cmp, loop);
1776 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
1777 stay_in_loop(lp_proj, loop)->is_If() &&
1778 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
1779 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
1780 set_ctrl(con, C->root());
1781 if_cmp->set_req(1, con);
1782 }
1783
1784 //------------------------------ scheduled_nodelist -------------------------------------
1785 // Create a post order schedule of nodes that are in the
1786 // "member" set. The list is returned in "sched".
1787 // The first node in "sched" is the loop head, followed by
1788 // nodes which have no inputs in the "member" set, and then
1789 // followed by the nodes that have an immediate input dependence
1790 // on a node in "sched".
1791 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
1792
1793 assert(member.test(loop->_head->_idx), "loop head must be in member set");
1794 Arena *a = Thread::current()->resource_area();
1795 VectorSet visited(a);
1796 Node_Stack nstack(a, loop->_body.size());
1797
1798 Node* n = loop->_head; // top of stack is cached in "n"
1799 uint idx = 0;
1800 visited.set(n->_idx);
1801
1802 // Initially push all with no inputs from within member set
1803 for(uint i = 0; i < loop->_body.size(); i++ ) {
1804 Node *elt = loop->_body.at(i);
1805 if (member.test(elt->_idx)) {
1806 bool found = false;
1807 for (uint j = 0; j < elt->req(); j++) {
1808 Node* def = elt->in(j);
1809 if (def && member.test(def->_idx) && def != elt) {
1810 found = true;
1811 break;
1812 }
1813 }
1814 if (!found && elt != loop->_head) {
1815 nstack.push(n, idx);
1816 n = elt;
1817 assert(!visited.test(n->_idx), "not seen yet");
1818 visited.set(n->_idx);
1819 }
1820 }
1821 }
1822
1823 // traverse out's that are in the member set
1824 while (true) {
1825 if (idx < n->outcnt()) {
1826 Node* use = n->raw_out(idx);
1827 idx++;
1828 if (!visited.test_set(use->_idx)) {
1829 if (member.test(use->_idx)) {
1830 nstack.push(n, idx);
1831 n = use;
1832 idx = 0;
1833 }
1834 }
1835 } else {
1836 // All outputs processed
1837 sched.push(n);
1838 if (nstack.is_empty()) break;
1839 n = nstack.node();
1840 idx = nstack.index();
1841 nstack.pop();
1842 }
1843 }
1844 }
1845
1846
1847 //------------------------------ has_use_in_set -------------------------------------
1848 // Has a use in the vector set
1849 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
1850 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1851 Node* use = n->fast_out(j);
1852 if (vset.test(use->_idx)) {
1853 return true;
1854 }
1855 }
1856 return false;
1857 }
1858
1859
1860 //------------------------------ has_use_internal_to_set -------------------------------------
1861 // Has use internal to the vector set (ie. not in a phi at the loop head)
1862 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
1863 Node* head = loop->_head;
1864 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1865 Node* use = n->fast_out(j);
1866 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
1867 return true;
1868 }
1869 }
1870 return false;
1871 }
1872
1873
1874 //------------------------------ clone_for_use_outside_loop -------------------------------------
1875 // clone "n" for uses that are outside of loop
1876 void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
1877
1878 assert(worklist.size() == 0, "should be empty");
1879 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1880 Node* use = n->fast_out(j);
1881 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
1882 worklist.push(use);
1883 }
1884 }
1885 while( worklist.size() ) {
1886 Node *use = worklist.pop();
1887 if (!has_node(use) || use->in(0) == C->top()) continue;
1888 uint j;
1889 for (j = 0; j < use->req(); j++) {
1890 if (use->in(j) == n) break;
1891 }
1892 assert(j < use->req(), "must be there");
1893
1894 // clone "n" and insert it between the inputs of "n" and the use outside the loop
1895 Node* n_clone = n->clone();
1896 _igvn.hash_delete(use);
1897 use->set_req(j, n_clone);
1898 _igvn._worklist.push(use);
1899 Node* use_c;
1900 if (!use->is_Phi()) {
1901 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
1902 } else {
1903 // Use in a phi is considered a use in the associated predecessor block
1904 use_c = use->in(0)->in(j);
1905 }
1906 if (use_c->is_CountedLoop()) {
1907 use_c = use_c->in(LoopNode::EntryControl);
1908 }
1909 set_ctrl(n_clone, use_c);
1910 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
1911 get_loop(use_c)->_body.push(n_clone);
1912 _igvn.register_new_node_with_optimizer(n_clone);
1913 #if !defined(PRODUCT)
1914 if (TracePartialPeeling) {
1915 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
1916 }
1917 #endif
1918 }
1919 }
1920
1921
1922 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
1923 // clone "n" for special uses that are in the not_peeled region.
1924 // If these def-uses occur in separate blocks, the code generator
1925 // marks the method as not compilable. For example, if a "BoolNode"
1926 // is in a different basic block than the "IfNode" that uses it, then
1927 // the compilation is aborted in the code generator.
1928 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1929 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
1930 if (n->is_Phi() || n->is_Load()) {
1931 return;
1932 }
1933 assert(worklist.size() == 0, "should be empty");
1934 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1935 Node* use = n->fast_out(j);
1936 if ( not_peel.test(use->_idx) &&
1937 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
1938 use->in(1) == n) {
1939 worklist.push(use);
1940 }
1941 }
1942 if (worklist.size() > 0) {
1943 // clone "n" and insert it between inputs of "n" and the use
1944 Node* n_clone = n->clone();
1945 loop->_body.push(n_clone);
1946 _igvn.register_new_node_with_optimizer(n_clone);
1947 set_ctrl(n_clone, get_ctrl(n));
1948 sink_list.push(n_clone);
1949 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
1950 #if !defined(PRODUCT)
1951 if (TracePartialPeeling) {
1952 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
1953 }
1954 #endif
1955 while( worklist.size() ) {
1956 Node *use = worklist.pop();
1957 _igvn.hash_delete(use);
1958 _igvn._worklist.push(use);
1959 for (uint j = 1; j < use->req(); j++) {
1960 if (use->in(j) == n) {
1961 use->set_req(j, n_clone);
1962 }
1963 }
1964 }
1965 }
1966 }
1967
1968
1969 //------------------------------ insert_phi_for_loop -------------------------------------
1970 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1971 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
1972 Node *phi = PhiNode::make(lp, back_edge_val);
1973 phi->set_req(LoopNode::EntryControl, lp_entry_val);
1974 // Use existing phi if it already exists
1975 Node *hit = _igvn.hash_find_insert(phi);
1976 if( hit == NULL ) {
1977 _igvn.register_new_node_with_optimizer(phi);
1978 set_ctrl(phi, lp);
1979 } else {
1980 // Remove the new phi from the graph and use the hit
1981 _igvn.remove_dead_node(phi);
1982 phi = hit;
1983 }
1984 _igvn.hash_delete(use);
1985 _igvn._worklist.push(use);
1986 use->set_req(idx, phi);
1987 }
1988
1989 #ifdef ASSERT
1990 //------------------------------ is_valid_loop_partition -------------------------------------
1991 // Validate the loop partition sets: peel and not_peel
1992 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
1993 VectorSet& not_peel ) {
1994 uint i;
1995 // Check that peel_list entries are in the peel set
1996 for (i = 0; i < peel_list.size(); i++) {
1997 if (!peel.test(peel_list.at(i)->_idx)) {
1998 return false;
1999 }
2000 }
2001 // Check at loop members are in one of peel set or not_peel set
2002 for (i = 0; i < loop->_body.size(); i++ ) {
2003 Node *def = loop->_body.at(i);
2004 uint di = def->_idx;
2005 // Check that peel set elements are in peel_list
2006 if (peel.test(di)) {
2007 if (not_peel.test(di)) {
2008 return false;
2009 }
2010 // Must be in peel_list also
2011 bool found = false;
2012 for (uint j = 0; j < peel_list.size(); j++) {
2013 if (peel_list.at(j)->_idx == di) {
2014 found = true;
2015 break;
2016 }
2017 }
2018 if (!found) {
2019 return false;
2020 }
2021 } else if (not_peel.test(di)) {
2022 if (peel.test(di)) {
2023 return false;
2024 }
2025 } else {
2026 return false;
2027 }
2028 }
2029 return true;
2030 }
2031
2032 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2033 // Ensure a use outside of loop is of the right form
2034 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2035 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2036 return (use->is_Phi() &&
2037 use_c->is_Region() && use_c->req() == 3 &&
2038 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2039 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2040 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2041 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2042 }
2043
2044 //------------------------------ is_valid_clone_loop_form -------------------------------------
2045 // Ensure that all uses outside of loop are of the right form
2046 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2047 uint orig_exit_idx, uint clone_exit_idx) {
2048 uint len = peel_list.size();
2049 for (uint i = 0; i < len; i++) {
2050 Node *def = peel_list.at(i);
2051
2052 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2053 Node *use = def->fast_out(j);
2054 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2055 if (!loop->is_member(get_loop(use_c))) {
2056 // use is not in the loop, check for correct structure
2057 if (use->in(0) == def) {
2058 // Okay
2059 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2060 return false;
2061 }
2062 }
2063 }
2064 }
2065 return true;
2066 }
2067 #endif
2068
2069 //------------------------------ partial_peel -------------------------------------
2070 // Partially peel (aka loop rotation) the top portion of a loop (called
2071 // the peel section below) by cloning it and placing one copy just before
2072 // the new loop head and the other copy at the bottom of the new loop.
2073 //
2074 // before after where it came from
2075 //
2076 // stmt1 stmt1
2077 // loop: stmt2 clone
2078 // stmt2 if condA goto exitA clone
2079 // if condA goto exitA new_loop: new
2080 // stmt3 stmt3 clone
2081 // if !condB goto loop if condB goto exitB clone
2082 // exitB: stmt2 orig
2083 // stmt4 if !condA goto new_loop orig
2084 // exitA: goto exitA
2085 // exitB:
2086 // stmt4
2087 // exitA:
2088 //
2089 // Step 1: find the cut point: an exit test on probable
2090 // induction variable.
2091 // Step 2: schedule (with cloning) operations in the peel
2092 // section that can be executed after the cut into
2093 // the section that is not peeled. This may need
2094 // to clone operations into exit blocks. For
2095 // instance, a reference to A[i] in the not-peel
2096 // section and a reference to B[i] in an exit block
2097 // may cause a left-shift of i by 2 to be placed
2098 // in the peel block. This step will clone the left
2099 // shift into the exit block and sink the left shift
2100 // from the peel to the not-peel section.
2101 // Step 3: clone the loop, retarget the control, and insert
2102 // phis for values that are live across the new loop
2103 // head. This is very dependent on the graph structure
2104 // from clone_loop. It creates region nodes for
2105 // exit control and associated phi nodes for values
2106 // flow out of the loop through that exit. The region
2107 // node is dominated by the clone's control projection.
2108 // So the clone's peel section is placed before the
2109 // new loop head, and the clone's not-peel section is
2110 // forms the top part of the new loop. The original
2111 // peel section forms the tail of the new loop.
2112 // Step 4: update the dominator tree and recompute the
2113 // dominator depth.
2114 //
2115 // orig
2116 //
2117 // stmt1
2118 // |
2119 // v
2120 // loop<----+
2121 // | |
2122 // stmt2 |
2123 // | |
2124 // v |
2125 // ifA |
2126 // / | |
2127 // v v |
2128 // false true ^ <-- last_peel
2129 // / | |
2130 // / ===|==cut |
2131 // / stmt3 | <-- first_not_peel
2132 // / | |
2133 // | v |
2134 // v ifB |
2135 // exitA: / \ |
2136 // / \ |
2137 // v v |
2138 // false true |
2139 // / \ |
2140 // / ----+
2141 // |
2142 // v
2143 // exitB:
2144 // stmt4
2145 //
2146 //
2147 // after clone loop
2148 //
2149 // stmt1
2150 // / \
2151 // clone / \ orig
2152 // / \
2153 // / \
2154 // v v
2155 // +---->loop loop<----+
2156 // | | | |
2157 // | stmt2 stmt2 |
2158 // | | | |
2159 // | v v |
2160 // | ifA ifA |
2161 // | | \ / | |
2162 // | v v v v |
2163 // ^ true false false true ^ <-- last_peel
2164 // | | ^ \ / | |
2165 // | cut==|== \ \ / ===|==cut |
2166 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2167 // | | dom | | | |
2168 // | v \ 1v v2 v |
2169 // | ifB regionA ifB |
2170 // | / \ | / \ |
2171 // | / \ v / \ |
2172 // | v v exitA: v v |
2173 // | true false false true |
2174 // | / ^ \ / \ |
2175 // +---- \ \ / ----+
2176 // dom \ /
2177 // \ 1v v2
2178 // regionB
2179 // |
2180 // v
2181 // exitB:
2182 // stmt4
2183 //
2184 //
2185 // after partial peel
2186 //
2187 // stmt1
2188 // /
2189 // clone / orig
2190 // / TOP
2191 // / \
2192 // v v
2193 // TOP->region region----+
2194 // | | |
2195 // stmt2 stmt2 |
2196 // | | |
2197 // v v |
2198 // ifA ifA |
2199 // | \ / | |
2200 // v v v v |
2201 // true false false true | <-- last_peel
2202 // | ^ \ / +------|---+
2203 // +->newloop \ \ / === ==cut | |
2204 // | stmt3 \ \ / TOP | |
2205 // | | dom | | stmt3 | | <-- first_not_peel
2206 // | v \ 1v v2 v | |
2207 // | ifB regionA ifB ^ v
2208 // | / \ | / \ | |
2209 // | / \ v / \ | |
2210 // | v v exitA: v v | |
2211 // | true false false true | |
2212 // | / ^ \ / \ | |
2213 // | | \ \ / v | |
2214 // | | dom \ / TOP | |
2215 // | | \ 1v v2 | |
2216 // ^ v regionB | |
2217 // | | | | |
2218 // | | v ^ v
2219 // | | exitB: | |
2220 // | | stmt4 | |
2221 // | +------------>-----------------+ |
2222 // | |
2223 // +-----------------<---------------------+
2224 //
2225 //
2226 // final graph
2227 //
2228 // stmt1
2229 // |
2230 // v
2231 // ........> ifA clone
2232 // : / |
2233 // dom / |
2234 // : v v
2235 // : false true
2236 // : | |
2237 // : | stmt2 clone
2238 // : | |
2239 // : | v
2240 // : | newloop<-----+
2241 // : | | |
2242 // : | stmt3 clone |
2243 // : | | |
2244 // : | v |
2245 // : | ifB |
2246 // : | / \ |
2247 // : | v v |
2248 // : | false true |
2249 // : | | | |
2250 // : | v stmt2 |
2251 // : | exitB: | |
2252 // : | stmt4 v |
2253 // : | ifA orig |
2254 // : | / \ |
2255 // : | / \ |
2256 // : | v v |
2257 // : | false true |
2258 // : | / \ |
2259 // : v v -----+
2260 // RegionA
2261 // |
2262 // v
2263 // exitA
2264 //
2265 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
2266
2267 if (!loop->_head->is_Loop()) {
2268 return false; }
2269
2270 LoopNode *head = loop->_head->as_Loop();
2271
2272 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
2273 return false;
2274 }
2275
2276 // Check for complex exit control
2277 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
2278 Node *n = loop->_body.at(ii);
2279 int opc = n->Opcode();
2280 if (n->is_Call() ||
2281 opc == Op_Catch ||
2282 opc == Op_CatchProj ||
2283 opc == Op_Jump ||
2284 opc == Op_JumpProj) {
2285 #if !defined(PRODUCT)
2286 if (TracePartialPeeling) {
2287 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
2288 }
2289 #endif
2290 return false;
2291 }
2292 }
2293
2294 int dd = dom_depth(head);
2295
2296 // Step 1: find cut point
2297
2298 // Walk up dominators to loop head looking for first loop exit
2299 // which is executed on every path thru loop.
2300 IfNode *peel_if = NULL;
2301 IfNode *peel_if_cmpu = NULL;
2302
2303 Node *iff = loop->tail();
2304 while( iff != head ) {
2305 if( iff->is_If() ) {
2306 Node *ctrl = get_ctrl(iff->in(1));
2307 if (ctrl->is_top()) return false; // Dead test on live IF.
2308 // If loop-varying exit-test, check for induction variable
2309 if( loop->is_member(get_loop(ctrl)) &&
2310 loop->is_loop_exit(iff) &&
2311 is_possible_iv_test(iff)) {
2312 Node* cmp = iff->in(1)->in(1);
2313 if (cmp->Opcode() == Op_CmpI) {
2314 peel_if = iff->as_If();
2315 } else {
2316 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
2317 peel_if_cmpu = iff->as_If();
2318 }
2319 }
2320 }
2321 iff = idom(iff);
2322 }
2323 // Prefer signed compare over unsigned compare.
2324 IfNode* new_peel_if = NULL;
2325 if (peel_if == NULL) {
2326 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
2327 return false; // No peel point found
2328 }
2329 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
2330 if (new_peel_if == NULL) {
2331 return false; // No peel point found
2332 }
2333 peel_if = new_peel_if;
2334 }
2335 Node* last_peel = stay_in_loop(peel_if, loop);
2336 Node* first_not_peeled = stay_in_loop(last_peel, loop);
2337 if (first_not_peeled == NULL || first_not_peeled == head) {
2338 return false;
2339 }
2340
2341 #if !defined(PRODUCT)
2342 if (TracePartialPeeling) {
2343 tty->print_cr("before partial peel one iteration");
2344 Node_List wl;
2345 Node* t = head->in(2);
2346 while (true) {
2347 wl.push(t);
2348 if (t == head) break;
2349 t = idom(t);
2350 }
2351 while (wl.size() > 0) {
2352 Node* tt = wl.pop();
2353 tt->dump();
2354 if (tt == last_peel) tty->print_cr("-- cut --");
2355 }
2356 }
2357 #endif
2358 ResourceArea *area = Thread::current()->resource_area();
2359 VectorSet peel(area);
2360 VectorSet not_peel(area);
2361 Node_List peel_list(area);
2362 Node_List worklist(area);
2363 Node_List sink_list(area);
2364
2365 // Set of cfg nodes to peel are those that are executable from
2366 // the head through last_peel.
2367 assert(worklist.size() == 0, "should be empty");
2368 worklist.push(head);
2369 peel.set(head->_idx);
2370 while (worklist.size() > 0) {
2371 Node *n = worklist.pop();
2372 if (n != last_peel) {
2373 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2374 Node* use = n->fast_out(j);
2375 if (use->is_CFG() &&
2376 loop->is_member(get_loop(use)) &&
2377 !peel.test_set(use->_idx)) {
2378 worklist.push(use);
2379 }
2380 }
2381 }
2382 }
2383
2384 // Set of non-cfg nodes to peel are those that are control
2385 // dependent on the cfg nodes.
2386 uint i;
2387 for(i = 0; i < loop->_body.size(); i++ ) {
2388 Node *n = loop->_body.at(i);
2389 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
2390 if (peel.test(n_c->_idx)) {
2391 peel.set(n->_idx);
2392 } else {
2393 not_peel.set(n->_idx);
2394 }
2395 }
2396
2397 // Step 2: move operations from the peeled section down into the
2398 // not-peeled section
2399
2400 // Get a post order schedule of nodes in the peel region
2401 // Result in right-most operand.
2402 scheduled_nodelist(loop, peel, peel_list );
2403
2404 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2405
2406 // For future check for too many new phis
2407 uint old_phi_cnt = 0;
2408 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2409 Node* use = head->fast_out(j);
2410 if (use->is_Phi()) old_phi_cnt++;
2411 }
2412
2413 #if !defined(PRODUCT)
2414 if (TracePartialPeeling) {
2415 tty->print_cr("\npeeled list");
2416 }
2417 #endif
2418
2419 // Evacuate nodes in peel region into the not_peeled region if possible
2420 uint new_phi_cnt = 0;
2421 for (i = 0; i < peel_list.size();) {
2422 Node* n = peel_list.at(i);
2423 #if !defined(PRODUCT)
2424 if (TracePartialPeeling) n->dump();
2425 #endif
2426 bool incr = true;
2427 if ( !n->is_CFG() ) {
2428
2429 if ( has_use_in_set(n, not_peel) ) {
2430
2431 // If not used internal to the peeled region,
2432 // move "n" from peeled to not_peeled region.
2433
2434 if ( !has_use_internal_to_set(n, peel, loop) ) {
2435
2436 // if not pinned and not a load (which maybe anti-dependent on a store)
2437 // and not a CMove (Matcher expects only bool->cmove).
2438 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
2439 clone_for_use_outside_loop( loop, n, worklist );
2440
2441 sink_list.push(n);
2442 peel >>= n->_idx; // delete n from peel set.
2443 not_peel <<= n->_idx; // add n to not_peel set.
2444 peel_list.remove(i);
2445 incr = false;
2446 #if !defined(PRODUCT)
2447 if (TracePartialPeeling) {
2448 tty->print_cr("sink to not_peeled region: %d newbb: %d",
2449 n->_idx, get_ctrl(n)->_idx);
2450 }
2451 #endif
2452 }
2453 } else {
2454 // Otherwise check for special def-use cases that span
2455 // the peel/not_peel boundary such as bool->if
2456 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
2457 new_phi_cnt++;
2458 }
2459 }
2460 }
2461 if (incr) i++;
2462 }
2463
2464 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
2465 #if !defined(PRODUCT)
2466 if (TracePartialPeeling) {
2467 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
2468 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
2469 }
2470 #endif
2471 if (new_peel_if != NULL) {
2472 remove_cmpi_loop_exit(new_peel_if, loop);
2473 }
2474 // Inhibit more partial peeling on this loop
2475 assert(!head->is_partial_peel_loop(), "not partial peeled");
2476 head->mark_partial_peel_failed();
2477 return false;
2478 }
2479
2480 // Step 3: clone loop, retarget control, and insert new phis
2481
2482 // Create new loop head for new phis and to hang
2483 // the nodes being moved (sinked) from the peel region.
2484 LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel);
2485 _igvn.register_new_node_with_optimizer(new_head);
2486 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
2487 first_not_peeled->set_req(0, new_head);
2488 set_loop(new_head, loop);
2489 loop->_body.push(new_head);
2490 not_peel.set(new_head->_idx);
2491 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
2492 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
2493
2494 while (sink_list.size() > 0) {
2495 Node* n = sink_list.pop();
2496 set_ctrl(n, new_head);
2497 }
2498
2499 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2500
2501 clone_loop( loop, old_new, dd );
2502
2503 const uint clone_exit_idx = 1;
2504 const uint orig_exit_idx = 2;
2505 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
2506
2507 Node* head_clone = old_new[head->_idx];
2508 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
2509 Node* orig_tail_clone = head_clone->in(2);
2510
2511 // Add phi if "def" node is in peel set and "use" is not
2512
2513 for(i = 0; i < peel_list.size(); i++ ) {
2514 Node *def = peel_list.at(i);
2515 if (!def->is_CFG()) {
2516 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2517 Node *use = def->fast_out(j);
2518 if (has_node(use) && use->in(0) != C->top() &&
2519 (!peel.test(use->_idx) ||
2520 (use->is_Phi() && use->in(0) == head)) ) {
2521 worklist.push(use);
2522 }
2523 }
2524 while( worklist.size() ) {
2525 Node *use = worklist.pop();
2526 for (uint j = 1; j < use->req(); j++) {
2527 Node* n = use->in(j);
2528 if (n == def) {
2529
2530 // "def" is in peel set, "use" is not in peel set
2531 // or "use" is in the entry boundary (a phi) of the peel set
2532
2533 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
2534
2535 if ( loop->is_member(get_loop( use_c )) ) {
2536 // use is in loop
2537 if (old_new[use->_idx] != NULL) { // null for dead code
2538 Node* use_clone = old_new[use->_idx];
2539 _igvn.hash_delete(use);
2540 use->set_req(j, C->top());
2541 _igvn._worklist.push(use);
2542 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
2543 }
2544 } else {
2545 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
2546 // use is not in the loop, check if the live range includes the cut
2547 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
2548 if (not_peel.test(lp_if->_idx)) {
2549 assert(j == orig_exit_idx, "use from original loop");
2550 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
2551 }
2552 }
2553 }
2554 }
2555 }
2556 }
2557 }
2558
2559 // Step 3b: retarget control
2560
2561 // Redirect control to the new loop head if a cloned node in
2562 // the not_peeled region has control that points into the peeled region.
2563 // This necessary because the cloned peeled region will be outside
2564 // the loop.
2565 // from to
2566 // cloned-peeled <---+
2567 // new_head_clone: | <--+
2568 // cloned-not_peeled in(0) in(0)
2569 // orig-peeled
2570
2571 for(i = 0; i < loop->_body.size(); i++ ) {
2572 Node *n = loop->_body.at(i);
2573 if (!n->is_CFG() && n->in(0) != NULL &&
2574 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
2575 Node* n_clone = old_new[n->_idx];
2576 _igvn.hash_delete(n_clone);
2577 n_clone->set_req(0, new_head_clone);
2578 _igvn._worklist.push(n_clone);
2579 }
2580 }
2581
2582 // Backedge of the surviving new_head (the clone) is original last_peel
2583 _igvn.hash_delete(new_head_clone);
2584 new_head_clone->set_req(LoopNode::LoopBackControl, last_peel);
2585 _igvn._worklist.push(new_head_clone);
2586
2587 // Cut first node in original not_peel set
2588 _igvn.hash_delete(new_head);
2589 new_head->set_req(LoopNode::EntryControl, C->top());
2590 new_head->set_req(LoopNode::LoopBackControl, C->top());
2591 _igvn._worklist.push(new_head);
2592
2593 // Copy head_clone back-branch info to original head
2594 // and remove original head's loop entry and
2595 // clone head's back-branch
2596 _igvn.hash_delete(head);
2597 _igvn.hash_delete(head_clone);
2598 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
2599 head->set_req(LoopNode::LoopBackControl, C->top());
2600 head_clone->set_req(LoopNode::LoopBackControl, C->top());
2601 _igvn._worklist.push(head);
2602 _igvn._worklist.push(head_clone);
2603
2604 // Similarly modify the phis
2605 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
2606 Node* use = head->fast_out(k);
2607 if (use->is_Phi() && use->outcnt() > 0) {
2608 Node* use_clone = old_new[use->_idx];
2609 _igvn.hash_delete(use);
2610 _igvn.hash_delete(use_clone);
2611 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
2612 use->set_req(LoopNode::LoopBackControl, C->top());
2613 use_clone->set_req(LoopNode::LoopBackControl, C->top());
2614 _igvn._worklist.push(use);
2615 _igvn._worklist.push(use_clone);
2616 }
2617 }
2618
2619 // Step 4: update dominator tree and dominator depth
2620
2621 set_idom(head, orig_tail_clone, dd);
2622 recompute_dom_depth();
2623
2624 // Inhibit more partial peeling on this loop
2625 new_head_clone->set_partial_peel_loop();
2626 C->set_major_progress();
2627
2628 #if !defined(PRODUCT)
2629 if (TracePartialPeeling) {
2630 tty->print_cr("\nafter partial peel one iteration");
2631 Node_List wl(area);
2632 Node* t = last_peel;
2633 while (true) {
2634 wl.push(t);
2635 if (t == head_clone) break;
2636 t = idom(t);
2637 }
2638 while (wl.size() > 0) {
2639 Node* tt = wl.pop();
2640 if (tt == head) tty->print_cr("orig head");
2641 else if (tt == new_head_clone) tty->print_cr("new head");
2642 else if (tt == head_clone) tty->print_cr("clone head");
2643 tt->dump();
2644 }
2645 }
2646 #endif
2647 return true;
2648 }
2649
2650 //------------------------------reorg_offsets----------------------------------
2651 // Reorganize offset computations to lower register pressure. Mostly
2652 // prevent loop-fallout uses of the pre-incremented trip counter (which are
2653 // then alive with the post-incremented trip counter forcing an extra
2654 // register move)
2655 void PhaseIdealLoop::reorg_offsets( IdealLoopTree *loop ) {
2656
2657 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2658 CountedLoopEndNode *cle = cl->loopexit();
2659 if( !cle ) return; // The occasional dead loop
2660 // Find loop exit control
2661 Node *exit = cle->proj_out(false);
2662 assert( exit->Opcode() == Op_IfFalse, "" );
2663
2664 // Check for the special case of folks using the pre-incremented
2665 // trip-counter on the fall-out path (forces the pre-incremented
2666 // and post-incremented trip counter to be live at the same time).
2667 // Fix this by adjusting to use the post-increment trip counter.
2668 Node *phi = cl->phi();
2669 if( !phi ) return; // Dead infinite loop
2670 bool progress = true;
2671 while (progress) {
2672 progress = false;
2673 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
2674 Node* use = phi->fast_out(i); // User of trip-counter
2675 if (!has_ctrl(use)) continue;
2676 Node *u_ctrl = get_ctrl(use);
2677 if( use->is_Phi() ) {
2678 u_ctrl = NULL;
2679 for( uint j = 1; j < use->req(); j++ )
2680 if( use->in(j) == phi )
2681 u_ctrl = dom_lca( u_ctrl, use->in(0)->in(j) );
2682 }
2683 IdealLoopTree *u_loop = get_loop(u_ctrl);
2684 // Look for loop-invariant use
2685 if( u_loop == loop ) continue;
2686 if( loop->is_member( u_loop ) ) continue;
2687 // Check that use is live out the bottom. Assuming the trip-counter
2688 // update is right at the bottom, uses of of the loop middle are ok.
2689 if( dom_lca( exit, u_ctrl ) != exit ) continue;
2690 // protect against stride not being a constant
2691 if( !cle->stride_is_con() ) continue;
2692 // Hit! Refactor use to use the post-incremented tripcounter.
2693 // Compute a post-increment tripcounter.
2694 Node *opaq = new (C, 2) Opaque2Node( C, cle->incr() );
2695 register_new_node( opaq, u_ctrl );
2696 Node *neg_stride = _igvn.intcon(-cle->stride_con());
2697 set_ctrl(neg_stride, C->root());
2698 Node *post = new (C, 3) AddINode( opaq, neg_stride);
2699 register_new_node( post, u_ctrl );
2700 _igvn.hash_delete(use);
2701 _igvn._worklist.push(use);
2702 for( uint j = 1; j < use->req(); j++ )
2703 if( use->in(j) == phi )
2704 use->set_req(j, post);
2705 // Since DU info changed, rerun loop
2706 progress = true;
2707 break;
2708 }
2709 }
2710
2711 }