1 /*
2 * Copyright (c) 1999, 2011, Oracle and/or its affiliates. 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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "memory/allocation.inline.hpp"
27 #include "opto/addnode.hpp"
28 #include "opto/connode.hpp"
29 #include "opto/divnode.hpp"
30 #include "opto/loopnode.hpp"
31 #include "opto/mulnode.hpp"
32 #include "opto/rootnode.hpp"
33 #include "opto/subnode.hpp"
34
35 //=============================================================================
36 //------------------------------split_thru_phi---------------------------------
37 // Split Node 'n' through merge point if there is enough win.
38 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
39 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
40 // ConvI2L may have type information on it which is unsafe to push up
41 // so disable this for now
42 return NULL;
43 }
44 int wins = 0;
45 assert(!n->is_CFG(), "");
46 assert(region->is_Region(), "");
47
48 const Type* type = n->bottom_type();
49 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr();
50 Node *phi;
51 if (t_oop != NULL && t_oop->is_known_instance_field()) {
52 int iid = t_oop->instance_id();
53 int index = C->get_alias_index(t_oop);
54 int offset = t_oop->offset();
55 phi = new (C,region->req()) PhiNode(region, type, NULL, iid, index, offset);
56 } else {
57 phi = PhiNode::make_blank(region, n);
58 }
59 uint old_unique = C->unique();
60 for (uint i = 1; i < region->req(); i++) {
61 Node *x;
62 Node* the_clone = NULL;
63 if (region->in(i) == C->top()) {
64 x = C->top(); // Dead path? Use a dead data op
65 } else {
66 x = n->clone(); // Else clone up the data op
67 the_clone = x; // Remember for possible deletion.
68 // Alter data node to use pre-phi inputs
69 if (n->in(0) == region)
70 x->set_req( 0, region->in(i) );
71 for (uint j = 1; j < n->req(); j++) {
72 Node *in = n->in(j);
73 if (in->is_Phi() && in->in(0) == region)
74 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
75 }
76 }
77 // Check for a 'win' on some paths
78 const Type *t = x->Value(&_igvn);
79
80 bool singleton = t->singleton();
81
82 // A TOP singleton indicates that there are no possible values incoming
83 // along a particular edge. In most cases, this is OK, and the Phi will
84 // be eliminated later in an Ideal call. However, we can't allow this to
85 // happen if the singleton occurs on loop entry, as the elimination of
86 // the PhiNode may cause the resulting node to migrate back to a previous
87 // loop iteration.
88 if (singleton && t == Type::TOP) {
89 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
90 // irreducible loop may not be indicated by an affirmative is_Loop());
91 // therefore, the only top we can split thru a phi is on a backedge of
92 // a loop.
93 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
94 }
95
96 if (singleton) {
97 wins++;
98 x = ((PhaseGVN&)_igvn).makecon(t);
99 } else {
100 // We now call Identity to try to simplify the cloned node.
101 // Note that some Identity methods call phase->type(this).
102 // Make sure that the type array is big enough for
103 // our new node, even though we may throw the node away.
104 // (Note: This tweaking with igvn only works because x is a new node.)
105 _igvn.set_type(x, t);
106 // If x is a TypeNode, capture any more-precise type permanently into Node
107 // otherwise it will be not updated during igvn->transform since
108 // igvn->type(x) is set to x->Value() already.
109 x->raise_bottom_type(t);
110 Node *y = x->Identity(&_igvn);
111 if (y != x) {
112 wins++;
113 x = y;
114 } else {
115 y = _igvn.hash_find(x);
116 if (y) {
117 wins++;
118 x = y;
119 } else {
120 // Else x is a new node we are keeping
121 // We do not need register_new_node_with_optimizer
122 // because set_type has already been called.
123 _igvn._worklist.push(x);
124 }
125 }
126 }
127 if (x != the_clone && the_clone != NULL)
128 _igvn.remove_dead_node(the_clone);
129 phi->set_req( i, x );
130 }
131 // Too few wins?
132 if (wins <= policy) {
133 _igvn.remove_dead_node(phi);
134 return NULL;
135 }
136
137 // Record Phi
138 register_new_node( phi, region );
139
140 for (uint i2 = 1; i2 < phi->req(); i2++) {
141 Node *x = phi->in(i2);
142 // If we commoned up the cloned 'x' with another existing Node,
143 // the existing Node picks up a new use. We need to make the
144 // existing Node occur higher up so it dominates its uses.
145 Node *old_ctrl;
146 IdealLoopTree *old_loop;
147
148 if (x->is_Con()) {
149 // Constant's control is always root.
150 set_ctrl(x, C->root());
151 continue;
152 }
153 // The occasional new node
154 if (x->_idx >= old_unique) { // Found a new, unplaced node?
155 old_ctrl = NULL;
156 old_loop = NULL; // Not in any prior loop
157 } else {
158 old_ctrl = get_ctrl(x);
159 old_loop = get_loop(old_ctrl); // Get prior loop
160 }
161 // New late point must dominate new use
162 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
163 if (new_ctrl == old_ctrl) // Nothing is changed
164 continue;
165
166 IdealLoopTree *new_loop = get_loop(new_ctrl);
167
168 // Don't move x into a loop if its uses are
169 // outside of loop. Otherwise x will be cloned
170 // for each use outside of this loop.
171 IdealLoopTree *use_loop = get_loop(region);
172 if (!new_loop->is_member(use_loop) &&
173 (old_loop == NULL || !new_loop->is_member(old_loop))) {
174 // Take early control, later control will be recalculated
175 // during next iteration of loop optimizations.
176 new_ctrl = get_early_ctrl(x);
177 new_loop = get_loop(new_ctrl);
178 }
179 // Set new location
180 set_ctrl(x, new_ctrl);
181 // If changing loop bodies, see if we need to collect into new body
182 if (old_loop != new_loop) {
183 if (old_loop && !old_loop->_child)
184 old_loop->_body.yank(x);
185 if (!new_loop->_child)
186 new_loop->_body.push(x); // Collect body info
187 }
188 }
189
190 return phi;
191 }
192
193 //------------------------------dominated_by------------------------------------
194 // Replace the dominated test with an obvious true or false. Place it on the
195 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
196 // live path up to the dominating control.
197 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) {
198 #ifndef PRODUCT
199 if (VerifyLoopOptimizations && PrintOpto) tty->print_cr("dominating test");
200 #endif
201
202
203 // prevdom is the dominating projection of the dominating test.
204 assert( iff->is_If(), "" );
205 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
206 int pop = prevdom->Opcode();
207 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
208 if (flip) {
209 if (pop == Op_IfTrue)
210 pop = Op_IfFalse;
211 else
212 pop = Op_IfTrue;
213 }
214 // 'con' is set to true or false to kill the dominated test.
215 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
216 set_ctrl(con, C->root()); // Constant gets a new use
217 // Hack the dominated test
218 _igvn.hash_delete(iff);
219 iff->set_req(1, con);
220 _igvn._worklist.push(iff);
221
222 // If I dont have a reachable TRUE and FALSE path following the IfNode then
223 // I can assume this path reaches an infinite loop. In this case it's not
224 // important to optimize the data Nodes - either the whole compilation will
225 // be tossed or this path (and all data Nodes) will go dead.
226 if (iff->outcnt() != 2) return;
227
228 // Make control-dependent data Nodes on the live path (path that will remain
229 // once the dominated IF is removed) become control-dependent on the
230 // dominating projection.
231 Node* dp = iff->as_If()->proj_out(pop == Op_IfTrue);
232
233 // Loop predicates may have depending checks which should not
234 // be skipped. For example, range check predicate has two checks
235 // for lower and upper bounds.
236 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp->as_Proj()->_con)->as_Proj();
237 if (exclude_loop_predicate &&
238 is_uncommon_trap_proj(unc_proj, Deoptimization::Reason_predicate))
239 return; // Let IGVN transformation change control dependence.
240
241 IdealLoopTree *old_loop = get_loop(dp);
242
243 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
244 Node* cd = dp->fast_out(i); // Control-dependent node
245 if (cd->depends_only_on_test()) {
246 assert(cd->in(0) == dp, "");
247 _igvn.hash_delete(cd);
248 cd->set_req(0, prevdom);
249 set_early_ctrl(cd);
250 _igvn._worklist.push(cd);
251 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
252 if (old_loop != new_loop) {
253 if (!old_loop->_child) old_loop->_body.yank(cd);
254 if (!new_loop->_child) new_loop->_body.push(cd);
255 }
256 --i;
257 --imax;
258 }
259 }
260 }
261
262 //------------------------------has_local_phi_input----------------------------
263 // Return TRUE if 'n' has Phi inputs from its local block and no other
264 // block-local inputs (all non-local-phi inputs come from earlier blocks)
265 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
266 Node *n_ctrl = get_ctrl(n);
267 // See if some inputs come from a Phi in this block, or from before
268 // this block.
269 uint i;
270 for( i = 1; i < n->req(); i++ ) {
271 Node *phi = n->in(i);
272 if( phi->is_Phi() && phi->in(0) == n_ctrl )
273 break;
274 }
275 if( i >= n->req() )
276 return NULL; // No Phi inputs; nowhere to clone thru
277
278 // Check for inputs created between 'n' and the Phi input. These
279 // must split as well; they have already been given the chance
280 // (courtesy of a post-order visit) and since they did not we must
281 // recover the 'cost' of splitting them by being very profitable
282 // when splitting 'n'. Since this is unlikely we simply give up.
283 for( i = 1; i < n->req(); i++ ) {
284 Node *m = n->in(i);
285 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
286 // We allow the special case of AddP's with no local inputs.
287 // This allows us to split-up address expressions.
288 if (m->is_AddP() &&
289 get_ctrl(m->in(2)) != n_ctrl &&
290 get_ctrl(m->in(3)) != n_ctrl) {
291 // Move the AddP up to dominating point
292 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl)));
293 continue;
294 }
295 return NULL;
296 }
297 }
298
299 return n_ctrl;
300 }
301
302 //------------------------------remix_address_expressions----------------------
303 // Rework addressing expressions to get the most loop-invariant stuff
304 // moved out. We'd like to do all associative operators, but it's especially
305 // important (common) to do address expressions.
306 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
307 if (!has_ctrl(n)) return NULL;
308 Node *n_ctrl = get_ctrl(n);
309 IdealLoopTree *n_loop = get_loop(n_ctrl);
310
311 // See if 'n' mixes loop-varying and loop-invariant inputs and
312 // itself is loop-varying.
313
314 // Only interested in binary ops (and AddP)
315 if( n->req() < 3 || n->req() > 4 ) return NULL;
316
317 Node *n1_ctrl = get_ctrl(n->in( 1));
318 Node *n2_ctrl = get_ctrl(n->in( 2));
319 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
320 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
321 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
322 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
323
324 // Does one of my inputs spin in a tighter loop than self?
325 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
326 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
327 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
328 return NULL; // Leave well enough alone
329
330 // Is at least one of my inputs loop-invariant?
331 if( n1_loop == n_loop &&
332 n2_loop == n_loop &&
333 n3_loop == n_loop )
334 return NULL; // No loop-invariant inputs
335
336
337 int n_op = n->Opcode();
338
339 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
340 if( n_op == Op_LShiftI ) {
341 // Scale is loop invariant
342 Node *scale = n->in(2);
343 Node *scale_ctrl = get_ctrl(scale);
344 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
345 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
346 return NULL;
347 const TypeInt *scale_t = scale->bottom_type()->isa_int();
348 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
349 return NULL; // Dont bother with byte/short masking
350 // Add must vary with loop (else shift would be loop-invariant)
351 Node *add = n->in(1);
352 Node *add_ctrl = get_ctrl(add);
353 IdealLoopTree *add_loop = get_loop(add_ctrl);
354 //assert( n_loop == add_loop, "" );
355 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
356
357 // Convert I-V into I+ (0-V); same for V-I
358 if( add->Opcode() == Op_SubI &&
359 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
360 Node *zero = _igvn.intcon(0);
361 set_ctrl(zero, C->root());
362 Node *neg = new (C, 3) SubINode( _igvn.intcon(0), add->in(2) );
363 register_new_node( neg, get_ctrl(add->in(2) ) );
364 add = new (C, 3) AddINode( add->in(1), neg );
365 register_new_node( add, add_ctrl );
366 }
367 if( add->Opcode() != Op_AddI ) return NULL;
368 // See if one add input is loop invariant
369 Node *add_var = add->in(1);
370 Node *add_var_ctrl = get_ctrl(add_var);
371 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
372 Node *add_invar = add->in(2);
373 Node *add_invar_ctrl = get_ctrl(add_invar);
374 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
375 if( add_var_loop == n_loop ) {
376 } else if( add_invar_loop == n_loop ) {
377 // Swap to find the invariant part
378 add_invar = add_var;
379 add_invar_ctrl = add_var_ctrl;
380 add_invar_loop = add_var_loop;
381 add_var = add->in(2);
382 Node *add_var_ctrl = get_ctrl(add_var);
383 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
384 } else // Else neither input is loop invariant
385 return NULL;
386 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
387 return NULL; // No invariant part of the add?
388
389 // Yes! Reshape address expression!
390 Node *inv_scale = new (C, 3) LShiftINode( add_invar, scale );
391 Node *inv_scale_ctrl =
392 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
393 add_invar_ctrl : scale_ctrl;
394 register_new_node( inv_scale, inv_scale_ctrl );
395 Node *var_scale = new (C, 3) LShiftINode( add_var, scale );
396 register_new_node( var_scale, n_ctrl );
397 Node *var_add = new (C, 3) AddINode( var_scale, inv_scale );
398 register_new_node( var_add, n_ctrl );
399 _igvn.replace_node( n, var_add );
400 return var_add;
401 }
402
403 // Replace (I+V) with (V+I)
404 if( n_op == Op_AddI ||
405 n_op == Op_AddL ||
406 n_op == Op_AddF ||
407 n_op == Op_AddD ||
408 n_op == Op_MulI ||
409 n_op == Op_MulL ||
410 n_op == Op_MulF ||
411 n_op == Op_MulD ) {
412 if( n2_loop == n_loop ) {
413 assert( n1_loop != n_loop, "" );
414 n->swap_edges(1, 2);
415 }
416 }
417
418 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
419 // but not if I2 is a constant.
420 if( n_op == Op_AddP ) {
421 if( n2_loop == n_loop && n3_loop != n_loop ) {
422 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
423 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
424 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
425 IdealLoopTree *n22loop = get_loop( n22_ctrl );
426 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
427 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
428 n23_loop == n_loop ) {
429 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
430 // Stuff new AddP in the loop preheader
431 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
432 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) );
433 register_new_node( add2, n_ctrl );
434 _igvn.replace_node( n, add2 );
435 return add2;
436 }
437 }
438 }
439
440 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
441 if( n2_loop != n_loop && n3_loop == n_loop ) {
442 if( n->in(3)->Opcode() == Op_AddI ) {
443 Node *V = n->in(3)->in(1);
444 Node *I = n->in(3)->in(2);
445 if( is_member(n_loop,get_ctrl(V)) ) {
446 } else {
447 Node *tmp = V; V = I; I = tmp;
448 }
449 if( !is_member(n_loop,get_ctrl(I)) ) {
450 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I );
451 // Stuff new AddP in the loop preheader
452 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
453 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V );
454 register_new_node( add2, n_ctrl );
455 _igvn.replace_node( n, add2 );
456 return add2;
457 }
458 }
459 }
460 }
461
462 return NULL;
463 }
464
465 //------------------------------conditional_move-------------------------------
466 // Attempt to replace a Phi with a conditional move. We have some pretty
467 // strict profitability requirements. All Phis at the merge point must
468 // be converted, so we can remove the control flow. We need to limit the
469 // number of c-moves to a small handful. All code that was in the side-arms
470 // of the CFG diamond is now speculatively executed. This code has to be
471 // "cheap enough". We are pretty much limited to CFG diamonds that merge
472 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
473 Node *PhaseIdealLoop::conditional_move( Node *region ) {
474
475 assert( region->is_Region(), "sanity check" );
476 if( region->req() != 3 ) return NULL;
477
478 // Check for CFG diamond
479 Node *lp = region->in(1);
480 Node *rp = region->in(2);
481 if( !lp || !rp ) return NULL;
482 Node *lp_c = lp->in(0);
483 if( lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If() ) return NULL;
484 IfNode *iff = lp_c->as_If();
485
486 // Check for highly predictable branch. No point in CMOV'ing if
487 // we are going to predict accurately all the time.
488 // %%% This hides patterns produced by utility methods like Math.min.
489 if( iff->_prob < PROB_UNLIKELY_MAG(3) ||
490 iff->_prob > PROB_LIKELY_MAG(3) )
491 return NULL;
492
493 // Check for ops pinned in an arm of the diamond.
494 // Can't remove the control flow in this case
495 if( lp->outcnt() > 1 ) return NULL;
496 if( rp->outcnt() > 1 ) return NULL;
497
498 // Check profitability
499 int cost = 0;
500 int phis = 0;
501 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
502 Node *out = region->fast_out(i);
503 if( !out->is_Phi() ) continue; // Ignore other control edges, etc
504 phis++;
505 PhiNode* phi = out->as_Phi();
506 switch (phi->type()->basic_type()) {
507 case T_LONG:
508 cost++; // Probably encodes as 2 CMOV's
509 case T_INT: // These all CMOV fine
510 case T_FLOAT:
511 case T_DOUBLE:
512 case T_ADDRESS: // (RawPtr)
513 cost++;
514 break;
515 case T_NARROWOOP: // Fall through
516 case T_OBJECT: { // Base oops are OK, but not derived oops
517 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
518 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
519 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
520 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
521 // have a Phi for the base here that we convert to a CMOVE all is well
522 // and good. But if the base is dead, we'll not make a CMOVE. Later
523 // the allocator will have to produce a base by creating a CMOVE of the
524 // relevant bases. This puts the allocator in the business of
525 // manufacturing expensive instructions, generally a bad plan.
526 // Just Say No to Conditionally-Moved Derived Pointers.
527 if( tp && tp->offset() != 0 )
528 return NULL;
529 cost++;
530 break;
531 }
532 default:
533 return NULL; // In particular, can't do memory or I/O
534 }
535 // Add in cost any speculative ops
536 for( uint j = 1; j < region->req(); j++ ) {
537 Node *proj = region->in(j);
538 Node *inp = phi->in(j);
539 if (get_ctrl(inp) == proj) { // Found local op
540 cost++;
541 // Check for a chain of dependent ops; these will all become
542 // speculative in a CMOV.
543 for( uint k = 1; k < inp->req(); k++ )
544 if (get_ctrl(inp->in(k)) == proj)
545 return NULL; // Too much speculative goo
546 }
547 }
548 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
549 // This will likely Split-If, a higher-payoff operation.
550 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
551 Node* use = phi->fast_out(k);
552 if( use->is_Cmp() || use->is_DecodeN() || use->is_EncodeP() )
553 return NULL;
554 }
555 }
556 if( cost >= ConditionalMoveLimit ) return NULL; // Too much goo
557 Node* bol = iff->in(1);
558 assert( bol->Opcode() == Op_Bool, "" );
559 int cmp_op = bol->in(1)->Opcode();
560 // It is expensive to generate flags from a float compare.
561 // Avoid duplicated float compare.
562 if( phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
563
564 // --------------
565 // Now replace all Phis with CMOV's
566 Node *cmov_ctrl = iff->in(0);
567 uint flip = (lp->Opcode() == Op_IfTrue);
568 while( 1 ) {
569 PhiNode* phi = NULL;
570 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
571 Node *out = region->fast_out(i);
572 if (out->is_Phi()) {
573 phi = out->as_Phi();
574 break;
575 }
576 }
577 if (phi == NULL) break;
578 #ifndef PRODUCT
579 if( PrintOpto && VerifyLoopOptimizations ) tty->print_cr("CMOV");
580 #endif
581 // Move speculative ops
582 for( uint j = 1; j < region->req(); j++ ) {
583 Node *proj = region->in(j);
584 Node *inp = phi->in(j);
585 if (get_ctrl(inp) == proj) { // Found local op
586 #ifndef PRODUCT
587 if( PrintOpto && VerifyLoopOptimizations ) {
588 tty->print(" speculate: ");
589 inp->dump();
590 }
591 #endif
592 set_ctrl(inp, cmov_ctrl);
593 }
594 }
595 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) );
596 register_new_node( cmov, cmov_ctrl );
597 _igvn.replace_node( phi, cmov );
598 #ifndef PRODUCT
599 if( VerifyLoopOptimizations ) verify();
600 #endif
601 }
602
603 // The useless CFG diamond will fold up later; see the optimization in
604 // RegionNode::Ideal.
605 _igvn._worklist.push(region);
606
607 return iff->in(1);
608 }
609
610 //------------------------------split_if_with_blocks_pre-----------------------
611 // Do the real work in a non-recursive function. Data nodes want to be
612 // cloned in the pre-order so they can feed each other nicely.
613 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
614 // Cloning these guys is unlikely to win
615 int n_op = n->Opcode();
616 if( n_op == Op_MergeMem ) return n;
617 if( n->is_Proj() ) return n;
618 // Do not clone-up CmpFXXX variations, as these are always
619 // followed by a CmpI
620 if( n->is_Cmp() ) return n;
621 // Attempt to use a conditional move instead of a phi/branch
622 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
623 Node *cmov = conditional_move( n );
624 if( cmov ) return cmov;
625 }
626 if( n->is_CFG() || n->is_LoadStore() )
627 return n;
628 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
629 n_op == Op_Opaque2 ) {
630 if( !C->major_progress() ) // If chance of no more loop opts...
631 _igvn._worklist.push(n); // maybe we'll remove them
632 return n;
633 }
634
635 if( n->is_Con() ) return n; // No cloning for Con nodes
636
637 Node *n_ctrl = get_ctrl(n);
638 if( !n_ctrl ) return n; // Dead node
639
640 // Attempt to remix address expressions for loop invariants
641 Node *m = remix_address_expressions( n );
642 if( m ) return m;
643
644 // Determine if the Node has inputs from some local Phi.
645 // Returns the block to clone thru.
646 Node *n_blk = has_local_phi_input( n );
647 if( !n_blk ) return n;
648 // Do not clone the trip counter through on a CountedLoop
649 // (messes up the canonical shape).
650 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
651
652 // Check for having no control input; not pinned. Allow
653 // dominating control.
654 if( n->in(0) ) {
655 Node *dom = idom(n_blk);
656 if( dom_lca( n->in(0), dom ) != n->in(0) )
657 return n;
658 }
659 // Policy: when is it profitable. You must get more wins than
660 // policy before it is considered profitable. Policy is usually 0,
661 // so 1 win is considered profitable. Big merges will require big
662 // cloning, so get a larger policy.
663 int policy = n_blk->req() >> 2;
664
665 // If the loop is a candidate for range check elimination,
666 // delay splitting through it's phi until a later loop optimization
667 if (n_blk->is_CountedLoop()) {
668 IdealLoopTree *lp = get_loop(n_blk);
669 if (lp && lp->_rce_candidate) {
670 return n;
671 }
672 }
673
674 // Use same limit as split_if_with_blocks_post
675 if( C->unique() > 35000 ) return n; // Method too big
676
677 // Split 'n' through the merge point if it is profitable
678 Node *phi = split_thru_phi( n, n_blk, policy );
679 if( !phi ) return n;
680
681 // Found a Phi to split thru!
682 // Replace 'n' with the new phi
683 _igvn.replace_node( n, phi );
684 // Moved a load around the loop, 'en-registering' something.
685 if( n_blk->Opcode() == Op_Loop && n->is_Load() &&
686 !phi->in(LoopNode::LoopBackControl)->is_Load() )
687 C->set_major_progress();
688
689 return phi;
690 }
691
692 static bool merge_point_too_heavy(Compile* C, Node* region) {
693 // Bail out if the region and its phis have too many users.
694 int weight = 0;
695 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
696 weight += region->fast_out(i)->outcnt();
697 }
698 int nodes_left = MaxNodeLimit - C->unique();
699 if (weight * 8 > nodes_left) {
700 #ifndef PRODUCT
701 if (PrintOpto)
702 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
703 #endif
704 return true;
705 } else {
706 return false;
707 }
708 }
709
710 static bool merge_point_safe(Node* region) {
711 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
712 // having a PhiNode input. This sidesteps the dangerous case where the split
713 // ConvI2LNode may become TOP if the input Value() does not
714 // overlap the ConvI2L range, leaving a node which may not dominate its
715 // uses.
716 // A better fix for this problem can be found in the BugTraq entry, but
717 // expediency for Mantis demands this hack.
718 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop
719 // split_if_with_blocks from splitting a block because we could not move around
720 // the FastLockNode.
721 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
722 Node* n = region->fast_out(i);
723 if (n->is_Phi()) {
724 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
725 Node* m = n->fast_out(j);
726 if (m->is_FastLock())
727 return false;
728 #ifdef _LP64
729 if (m->Opcode() == Op_ConvI2L)
730 return false;
731 #endif
732 }
733 }
734 }
735 return true;
736 }
737
738
739 //------------------------------place_near_use---------------------------------
740 // Place some computation next to use but not inside inner loops.
741 // For inner loop uses move it to the preheader area.
742 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const {
743 IdealLoopTree *u_loop = get_loop( useblock );
744 return (u_loop->_irreducible || u_loop->_child)
745 ? useblock
746 : u_loop->_head->in(LoopNode::EntryControl);
747 }
748
749
750 //------------------------------split_if_with_blocks_post----------------------
751 // Do the real work in a non-recursive function. CFG hackery wants to be
752 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
753 // info.
754 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) {
755
756 // Cloning Cmp through Phi's involves the split-if transform.
757 // FastLock is not used by an If
758 if( n->is_Cmp() && !n->is_FastLock() ) {
759 if( C->unique() > 35000 ) return; // Method too big
760
761 // Do not do 'split-if' if irreducible loops are present.
762 if( _has_irreducible_loops )
763 return;
764
765 Node *n_ctrl = get_ctrl(n);
766 // Determine if the Node has inputs from some local Phi.
767 // Returns the block to clone thru.
768 Node *n_blk = has_local_phi_input( n );
769 if( n_blk != n_ctrl ) return;
770
771 if( merge_point_too_heavy(C, n_ctrl) )
772 return;
773
774 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare?
775 Node *bol = n->unique_out();
776 assert( bol->is_Bool(), "expect a bool here" );
777 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare?
778 Node *iff = bol->unique_out();
779
780 // Check some safety conditions
781 if( iff->is_If() ) { // Classic split-if?
782 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if
783 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
784 if( get_ctrl(iff->in(2)) == n_ctrl ||
785 get_ctrl(iff->in(3)) == n_ctrl )
786 return; // Inputs not yet split-up
787 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) {
788 return; // Loop-invar test gates loop-varying CMOVE
789 }
790 } else {
791 return; // some other kind of node, such as an Allocate
792 }
793
794 // Do not do 'split-if' if some paths are dead. First do dead code
795 // elimination and then see if its still profitable.
796 for( uint i = 1; i < n_ctrl->req(); i++ )
797 if( n_ctrl->in(i) == C->top() )
798 return;
799
800 // When is split-if profitable? Every 'win' on means some control flow
801 // goes dead, so it's almost always a win.
802 int policy = 0;
803 // If trying to do a 'Split-If' at the loop head, it is only
804 // profitable if the cmp folds up on BOTH paths. Otherwise we
805 // risk peeling a loop forever.
806
807 // CNC - Disabled for now. Requires careful handling of loop
808 // body selection for the cloned code. Also, make sure we check
809 // for any input path not being in the same loop as n_ctrl. For
810 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
811 // because the alternative loop entry points won't be converted
812 // into LoopNodes.
813 IdealLoopTree *n_loop = get_loop(n_ctrl);
814 for( uint j = 1; j < n_ctrl->req(); j++ )
815 if( get_loop(n_ctrl->in(j)) != n_loop )
816 return;
817
818 // Check for safety of the merge point.
819 if( !merge_point_safe(n_ctrl) ) {
820 return;
821 }
822
823 // Split compare 'n' through the merge point if it is profitable
824 Node *phi = split_thru_phi( n, n_ctrl, policy );
825 if( !phi ) return;
826
827 // Found a Phi to split thru!
828 // Replace 'n' with the new phi
829 _igvn.replace_node( n, phi );
830
831 // Now split the bool up thru the phi
832 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 );
833 _igvn.replace_node( bol, bolphi );
834 assert( iff->in(1) == bolphi, "" );
835 if( bolphi->Value(&_igvn)->singleton() )
836 return;
837
838 // Conditional-move? Must split up now
839 if( !iff->is_If() ) {
840 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 );
841 _igvn.replace_node( iff, cmovphi );
842 return;
843 }
844
845 // Now split the IF
846 do_split_if( iff );
847 return;
848 }
849
850 // Check for an IF ready to split; one that has its
851 // condition codes input coming from a Phi at the block start.
852 int n_op = n->Opcode();
853
854 // Check for an IF being dominated by another IF same test
855 if( n_op == Op_If ) {
856 Node *bol = n->in(1);
857 uint max = bol->outcnt();
858 // Check for same test used more than once?
859 if( n_op == Op_If && max > 1 && bol->is_Bool() ) {
860 // Search up IDOMs to see if this IF is dominated.
861 Node *cutoff = get_ctrl(bol);
862
863 // Now search up IDOMs till cutoff, looking for a dominating test
864 Node *prevdom = n;
865 Node *dom = idom(prevdom);
866 while( dom != cutoff ) {
867 if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) {
868 // Replace the dominated test with an obvious true or false.
869 // Place it on the IGVN worklist for later cleanup.
870 C->set_major_progress();
871 dominated_by( prevdom, n, false, true );
872 #ifndef PRODUCT
873 if( VerifyLoopOptimizations ) verify();
874 #endif
875 return;
876 }
877 prevdom = dom;
878 dom = idom(prevdom);
879 }
880 }
881 }
882
883 // See if a shared loop-varying computation has no loop-varying uses.
884 // Happens if something is only used for JVM state in uncommon trap exits,
885 // like various versions of induction variable+offset. Clone the
886 // computation per usage to allow it to sink out of the loop.
887 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
888 Node *n_ctrl = get_ctrl(n);
889 IdealLoopTree *n_loop = get_loop(n_ctrl);
890 if( n_loop != _ltree_root ) {
891 DUIterator_Fast imax, i = n->fast_outs(imax);
892 for (; i < imax; i++) {
893 Node* u = n->fast_out(i);
894 if( !has_ctrl(u) ) break; // Found control user
895 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
896 if( u_loop == n_loop ) break; // Found loop-varying use
897 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
898 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
899 }
900 bool did_break = (i < imax); // Did we break out of the previous loop?
901 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
902 Node *late_load_ctrl;
903 if (n->is_Load()) {
904 // If n is a load, get and save the result from get_late_ctrl(),
905 // to be later used in calculating the control for n's clones.
906 clear_dom_lca_tags();
907 late_load_ctrl = get_late_ctrl(n, n_ctrl);
908 }
909 // If n is a load, and the late control is the same as the current
910 // control, then the cloning of n is a pointless exercise, because
911 // GVN will ensure that we end up where we started.
912 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
913 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
914 Node *u = n->last_out(j); // Clone private computation per use
915 _igvn.hash_delete(u);
916 _igvn._worklist.push(u);
917 Node *x = n->clone(); // Clone computation
918 Node *x_ctrl = NULL;
919 if( u->is_Phi() ) {
920 // Replace all uses of normal nodes. Replace Phi uses
921 // individually, so the separate Nodes can sink down
922 // different paths.
923 uint k = 1;
924 while( u->in(k) != n ) k++;
925 u->set_req( k, x );
926 // x goes next to Phi input path
927 x_ctrl = u->in(0)->in(k);
928 --j;
929 } else { // Normal use
930 // Replace all uses
931 for( uint k = 0; k < u->req(); k++ ) {
932 if( u->in(k) == n ) {
933 u->set_req( k, x );
934 --j;
935 }
936 }
937 x_ctrl = get_ctrl(u);
938 }
939
940 // Find control for 'x' next to use but not inside inner loops.
941 // For inner loop uses get the preheader area.
942 x_ctrl = place_near_use(x_ctrl);
943
944 if (n->is_Load()) {
945 // For loads, add a control edge to a CFG node outside of the loop
946 // to force them to not combine and return back inside the loop
947 // during GVN optimization (4641526).
948 //
949 // Because we are setting the actual control input, factor in
950 // the result from get_late_ctrl() so we respect any
951 // anti-dependences. (6233005).
952 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
953
954 // Don't allow the control input to be a CFG splitting node.
955 // Such nodes should only have ProjNodes as outs, e.g. IfNode
956 // should only have IfTrueNode and IfFalseNode (4985384).
957 x_ctrl = find_non_split_ctrl(x_ctrl);
958 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
959
960 x->set_req(0, x_ctrl);
961 }
962 register_new_node(x, x_ctrl);
963
964 // Some institutional knowledge is needed here: 'x' is
965 // yanked because if the optimizer runs GVN on it all the
966 // cloned x's will common up and undo this optimization and
967 // be forced back in the loop. This is annoying because it
968 // makes +VerifyOpto report false-positives on progress. I
969 // tried setting control edges on the x's to force them to
970 // not combine, but the matching gets worried when it tries
971 // to fold a StoreP and an AddP together (as part of an
972 // address expression) and the AddP and StoreP have
973 // different controls.
974 if( !x->is_Load() && !x->is_DecodeN() ) _igvn._worklist.yank(x);
975 }
976 _igvn.remove_dead_node(n);
977 }
978 }
979 }
980 }
981
982 // Check for Opaque2's who's loop has disappeared - who's input is in the
983 // same loop nest as their output. Remove 'em, they are no longer useful.
984 if( n_op == Op_Opaque2 &&
985 n->in(1) != NULL &&
986 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
987 _igvn.replace_node( n, n->in(1) );
988 }
989 }
990
991 //------------------------------split_if_with_blocks---------------------------
992 // Check for aggressive application of 'split-if' optimization,
993 // using basic block level info.
994 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) {
995 Node *n = C->root();
996 visited.set(n->_idx); // first, mark node as visited
997 // Do pre-visit work for root
998 n = split_if_with_blocks_pre( n );
999 uint cnt = n->outcnt();
1000 uint i = 0;
1001 while (true) {
1002 // Visit all children
1003 if (i < cnt) {
1004 Node* use = n->raw_out(i);
1005 ++i;
1006 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1007 // Now do pre-visit work for this use
1008 use = split_if_with_blocks_pre( use );
1009 nstack.push(n, i); // Save parent and next use's index.
1010 n = use; // Process all children of current use.
1011 cnt = use->outcnt();
1012 i = 0;
1013 }
1014 }
1015 else {
1016 // All of n's children have been processed, complete post-processing.
1017 if (cnt != 0 && !n->is_Con()) {
1018 assert(has_node(n), "no dead nodes");
1019 split_if_with_blocks_post( n );
1020 }
1021 if (nstack.is_empty()) {
1022 // Finished all nodes on stack.
1023 break;
1024 }
1025 // Get saved parent node and next use's index. Visit the rest of uses.
1026 n = nstack.node();
1027 cnt = n->outcnt();
1028 i = nstack.index();
1029 nstack.pop();
1030 }
1031 }
1032 }
1033
1034
1035 //=============================================================================
1036 //
1037 // C L O N E A L O O P B O D Y
1038 //
1039
1040 //------------------------------clone_iff--------------------------------------
1041 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1042 // "Nearly" because all Nodes have been cloned from the original in the loop,
1043 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1044 // through the Phi recursively, and return a Bool.
1045 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) {
1046
1047 // Convert this Phi into a Phi merging Bools
1048 uint i;
1049 for( i = 1; i < phi->req(); i++ ) {
1050 Node *b = phi->in(i);
1051 if( b->is_Phi() ) {
1052 _igvn.hash_delete(phi);
1053 _igvn._worklist.push(phi);
1054 phi->set_req(i, clone_iff( b->as_Phi(), loop ));
1055 } else {
1056 assert( b->is_Bool(), "" );
1057 }
1058 }
1059
1060 Node *sample_bool = phi->in(1);
1061 Node *sample_cmp = sample_bool->in(1);
1062
1063 // Make Phis to merge the Cmp's inputs.
1064 int size = phi->in(0)->req();
1065 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1066 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1067 for( i = 1; i < phi->req(); i++ ) {
1068 Node *n1 = phi->in(i)->in(1)->in(1);
1069 Node *n2 = phi->in(i)->in(1)->in(2);
1070 phi1->set_req( i, n1 );
1071 phi2->set_req( i, n2 );
1072 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1073 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1074 }
1075 // See if these Phis have been made before.
1076 // Register with optimizer
1077 Node *hit1 = _igvn.hash_find_insert(phi1);
1078 if( hit1 ) { // Hit, toss just made Phi
1079 _igvn.remove_dead_node(phi1); // Remove new phi
1080 assert( hit1->is_Phi(), "" );
1081 phi1 = (PhiNode*)hit1; // Use existing phi
1082 } else { // Miss
1083 _igvn.register_new_node_with_optimizer(phi1);
1084 }
1085 Node *hit2 = _igvn.hash_find_insert(phi2);
1086 if( hit2 ) { // Hit, toss just made Phi
1087 _igvn.remove_dead_node(phi2); // Remove new phi
1088 assert( hit2->is_Phi(), "" );
1089 phi2 = (PhiNode*)hit2; // Use existing phi
1090 } else { // Miss
1091 _igvn.register_new_node_with_optimizer(phi2);
1092 }
1093 // Register Phis with loop/block info
1094 set_ctrl(phi1, phi->in(0));
1095 set_ctrl(phi2, phi->in(0));
1096 // Make a new Cmp
1097 Node *cmp = sample_cmp->clone();
1098 cmp->set_req( 1, phi1 );
1099 cmp->set_req( 2, phi2 );
1100 _igvn.register_new_node_with_optimizer(cmp);
1101 set_ctrl(cmp, phi->in(0));
1102
1103 // Make a new Bool
1104 Node *b = sample_bool->clone();
1105 b->set_req(1,cmp);
1106 _igvn.register_new_node_with_optimizer(b);
1107 set_ctrl(b, phi->in(0));
1108
1109 assert( b->is_Bool(), "" );
1110 return (BoolNode*)b;
1111 }
1112
1113 //------------------------------clone_bool-------------------------------------
1114 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1115 // "Nearly" because all Nodes have been cloned from the original in the loop,
1116 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1117 // through the Phi recursively, and return a Bool.
1118 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1119 uint i;
1120 // Convert this Phi into a Phi merging Bools
1121 for( i = 1; i < phi->req(); i++ ) {
1122 Node *b = phi->in(i);
1123 if( b->is_Phi() ) {
1124 _igvn.hash_delete(phi);
1125 _igvn._worklist.push(phi);
1126 phi->set_req(i, clone_bool( b->as_Phi(), loop ));
1127 } else {
1128 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1129 }
1130 }
1131
1132 Node *sample_cmp = phi->in(1);
1133
1134 // Make Phis to merge the Cmp's inputs.
1135 int size = phi->in(0)->req();
1136 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1137 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1138 for( uint j = 1; j < phi->req(); j++ ) {
1139 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1140 Node *n1, *n2;
1141 if( cmp_top->is_Cmp() ) {
1142 n1 = cmp_top->in(1);
1143 n2 = cmp_top->in(2);
1144 } else {
1145 n1 = n2 = cmp_top;
1146 }
1147 phi1->set_req( j, n1 );
1148 phi2->set_req( j, n2 );
1149 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1150 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1151 }
1152
1153 // See if these Phis have been made before.
1154 // Register with optimizer
1155 Node *hit1 = _igvn.hash_find_insert(phi1);
1156 if( hit1 ) { // Hit, toss just made Phi
1157 _igvn.remove_dead_node(phi1); // Remove new phi
1158 assert( hit1->is_Phi(), "" );
1159 phi1 = (PhiNode*)hit1; // Use existing phi
1160 } else { // Miss
1161 _igvn.register_new_node_with_optimizer(phi1);
1162 }
1163 Node *hit2 = _igvn.hash_find_insert(phi2);
1164 if( hit2 ) { // Hit, toss just made Phi
1165 _igvn.remove_dead_node(phi2); // Remove new phi
1166 assert( hit2->is_Phi(), "" );
1167 phi2 = (PhiNode*)hit2; // Use existing phi
1168 } else { // Miss
1169 _igvn.register_new_node_with_optimizer(phi2);
1170 }
1171 // Register Phis with loop/block info
1172 set_ctrl(phi1, phi->in(0));
1173 set_ctrl(phi2, phi->in(0));
1174 // Make a new Cmp
1175 Node *cmp = sample_cmp->clone();
1176 cmp->set_req( 1, phi1 );
1177 cmp->set_req( 2, phi2 );
1178 _igvn.register_new_node_with_optimizer(cmp);
1179 set_ctrl(cmp, phi->in(0));
1180
1181 assert( cmp->is_Cmp(), "" );
1182 return (CmpNode*)cmp;
1183 }
1184
1185 //------------------------------sink_use---------------------------------------
1186 // If 'use' was in the loop-exit block, it now needs to be sunk
1187 // below the post-loop merge point.
1188 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1189 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1190 set_ctrl(use, post_loop);
1191 for (DUIterator j = use->outs(); use->has_out(j); j++)
1192 sink_use(use->out(j), post_loop);
1193 }
1194 }
1195
1196 //------------------------------clone_loop-------------------------------------
1197 //
1198 // C L O N E A L O O P B O D Y
1199 //
1200 // This is the basic building block of the loop optimizations. It clones an
1201 // entire loop body. It makes an old_new loop body mapping; with this mapping
1202 // you can find the new-loop equivalent to an old-loop node. All new-loop
1203 // nodes are exactly equal to their old-loop counterparts, all edges are the
1204 // same. All exits from the old-loop now have a RegionNode that merges the
1205 // equivalent new-loop path. This is true even for the normal "loop-exit"
1206 // condition. All uses of loop-invariant old-loop values now come from (one
1207 // or more) Phis that merge their new-loop equivalents.
1208 //
1209 // This operation leaves the graph in an illegal state: there are two valid
1210 // control edges coming from the loop pre-header to both loop bodies. I'll
1211 // definitely have to hack the graph after running this transform.
1212 //
1213 // From this building block I will further edit edges to perform loop peeling
1214 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1215 //
1216 // Parameter side_by_size_idom:
1217 // When side_by_size_idom is NULL, the dominator tree is constructed for
1218 // the clone loop to dominate the original. Used in construction of
1219 // pre-main-post loop sequence.
1220 // When nonnull, the clone and original are side-by-side, both are
1221 // dominated by the side_by_side_idom node. Used in construction of
1222 // unswitched loops.
1223 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1224 Node* side_by_side_idom) {
1225
1226 // Step 1: Clone the loop body. Make the old->new mapping.
1227 uint i;
1228 for( i = 0; i < loop->_body.size(); i++ ) {
1229 Node *old = loop->_body.at(i);
1230 Node *nnn = old->clone();
1231 old_new.map( old->_idx, nnn );
1232 _igvn.register_new_node_with_optimizer(nnn);
1233 }
1234
1235
1236 // Step 2: Fix the edges in the new body. If the old input is outside the
1237 // loop use it. If the old input is INside the loop, use the corresponding
1238 // new node instead.
1239 for( i = 0; i < loop->_body.size(); i++ ) {
1240 Node *old = loop->_body.at(i);
1241 Node *nnn = old_new[old->_idx];
1242 // Fix CFG/Loop controlling the new node
1243 if (has_ctrl(old)) {
1244 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
1245 } else {
1246 set_loop(nnn, loop->_parent);
1247 if (old->outcnt() > 0) {
1248 set_idom( nnn, old_new[idom(old)->_idx], dd );
1249 }
1250 }
1251 // Correct edges to the new node
1252 for( uint j = 0; j < nnn->req(); j++ ) {
1253 Node *n = nnn->in(j);
1254 if( n ) {
1255 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
1256 if( loop->is_member( old_in_loop ) )
1257 nnn->set_req(j, old_new[n->_idx]);
1258 }
1259 }
1260 _igvn.hash_find_insert(nnn);
1261 }
1262 Node *newhead = old_new[loop->_head->_idx];
1263 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1264
1265
1266 // Step 3: Now fix control uses. Loop varying control uses have already
1267 // been fixed up (as part of all input edges in Step 2). Loop invariant
1268 // control uses must be either an IfFalse or an IfTrue. Make a merge
1269 // point to merge the old and new IfFalse/IfTrue nodes; make the use
1270 // refer to this.
1271 ResourceArea *area = Thread::current()->resource_area();
1272 Node_List worklist(area);
1273 uint new_counter = C->unique();
1274 for( i = 0; i < loop->_body.size(); i++ ) {
1275 Node* old = loop->_body.at(i);
1276 if( !old->is_CFG() ) continue;
1277 Node* nnn = old_new[old->_idx];
1278
1279 // Copy uses to a worklist, so I can munge the def-use info
1280 // with impunity.
1281 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1282 worklist.push(old->fast_out(j));
1283
1284 while( worklist.size() ) { // Visit all uses
1285 Node *use = worklist.pop();
1286 if (!has_node(use)) continue; // Ignore dead nodes
1287 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1288 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
1289 // Both OLD and USE are CFG nodes here.
1290 assert( use->is_Proj(), "" );
1291
1292 // Clone the loop exit control projection
1293 Node *newuse = use->clone();
1294 newuse->set_req(0,nnn);
1295 _igvn.register_new_node_with_optimizer(newuse);
1296 set_loop(newuse, use_loop);
1297 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
1298
1299 // We need a Region to merge the exit from the peeled body and the
1300 // exit from the old loop body.
1301 RegionNode *r = new (C, 3) RegionNode(3);
1302 // Map the old use to the new merge point
1303 old_new.map( use->_idx, r );
1304 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
1305 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
1306
1307 // The original user of 'use' uses 'r' instead.
1308 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
1309 Node* useuse = use->last_out(l);
1310 _igvn.hash_delete(useuse);
1311 _igvn._worklist.push(useuse);
1312 uint uses_found = 0;
1313 if( useuse->in(0) == use ) {
1314 useuse->set_req(0, r);
1315 uses_found++;
1316 if( useuse->is_CFG() ) {
1317 assert( dom_depth(useuse) > dd_r, "" );
1318 set_idom(useuse, r, dom_depth(useuse));
1319 }
1320 }
1321 for( uint k = 1; k < useuse->req(); k++ ) {
1322 if( useuse->in(k) == use ) {
1323 useuse->set_req(k, r);
1324 uses_found++;
1325 }
1326 }
1327 l -= uses_found; // we deleted 1 or more copies of this edge
1328 }
1329
1330 // Now finish up 'r'
1331 r->set_req( 1, newuse );
1332 r->set_req( 2, use );
1333 _igvn.register_new_node_with_optimizer(r);
1334 set_loop(r, use_loop);
1335 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
1336 } // End of if a loop-exit test
1337 }
1338 }
1339
1340 // Step 4: If loop-invariant use is not control, it must be dominated by a
1341 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
1342 // there if needed. Make a Phi there merging old and new used values.
1343 Node_List *split_if_set = NULL;
1344 Node_List *split_bool_set = NULL;
1345 Node_List *split_cex_set = NULL;
1346 for( i = 0; i < loop->_body.size(); i++ ) {
1347 Node* old = loop->_body.at(i);
1348 Node* nnn = old_new[old->_idx];
1349 // Copy uses to a worklist, so I can munge the def-use info
1350 // with impunity.
1351 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1352 worklist.push(old->fast_out(j));
1353
1354 while( worklist.size() ) {
1355 Node *use = worklist.pop();
1356 if (!has_node(use)) continue; // Ignore dead nodes
1357 if (use->in(0) == C->top()) continue;
1358 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1359 // Check for data-use outside of loop - at least one of OLD or USE
1360 // must not be a CFG node.
1361 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) {
1362
1363 // If the Data use is an IF, that means we have an IF outside of the
1364 // loop that is switching on a condition that is set inside of the
1365 // loop. Happens if people set a loop-exit flag; then test the flag
1366 // in the loop to break the loop, then test is again outside of the
1367 // loop to determine which way the loop exited.
1368 if( use->is_If() || use->is_CMove() ) {
1369 // Since this code is highly unlikely, we lazily build the worklist
1370 // of such Nodes to go split.
1371 if( !split_if_set )
1372 split_if_set = new Node_List(area);
1373 split_if_set->push(use);
1374 }
1375 if( use->is_Bool() ) {
1376 if( !split_bool_set )
1377 split_bool_set = new Node_List(area);
1378 split_bool_set->push(use);
1379 }
1380 if( use->Opcode() == Op_CreateEx ) {
1381 if( !split_cex_set )
1382 split_cex_set = new Node_List(area);
1383 split_cex_set->push(use);
1384 }
1385
1386
1387 // Get "block" use is in
1388 uint idx = 0;
1389 while( use->in(idx) != old ) idx++;
1390 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1391 assert( !loop->is_member( get_loop( prev ) ), "" );
1392 Node *cfg = prev->_idx >= new_counter
1393 ? prev->in(2)
1394 : idom(prev);
1395 if( use->is_Phi() ) // Phi use is in prior block
1396 cfg = prev->in(idx); // NOT in block of Phi itself
1397 if (cfg->is_top()) { // Use is dead?
1398 _igvn.hash_delete(use);
1399 _igvn._worklist.push(use);
1400 use->set_req(idx, C->top());
1401 continue;
1402 }
1403
1404 while( !loop->is_member( get_loop( cfg ) ) ) {
1405 prev = cfg;
1406 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1407 }
1408 // If the use occurs after merging several exits from the loop, then
1409 // old value must have dominated all those exits. Since the same old
1410 // value was used on all those exits we did not need a Phi at this
1411 // merge point. NOW we do need a Phi here. Each loop exit value
1412 // is now merged with the peeled body exit; each exit gets its own
1413 // private Phi and those Phis need to be merged here.
1414 Node *phi;
1415 if( prev->is_Region() ) {
1416 if( idx == 0 ) { // Updating control edge?
1417 phi = prev; // Just use existing control
1418 } else { // Else need a new Phi
1419 phi = PhiNode::make( prev, old );
1420 // Now recursively fix up the new uses of old!
1421 for( uint i = 1; i < prev->req(); i++ ) {
1422 worklist.push(phi); // Onto worklist once for each 'old' input
1423 }
1424 }
1425 } else {
1426 // Get new RegionNode merging old and new loop exits
1427 prev = old_new[prev->_idx];
1428 assert( prev, "just made this in step 7" );
1429 if( idx == 0 ) { // Updating control edge?
1430 phi = prev; // Just use existing control
1431 } else { // Else need a new Phi
1432 // Make a new Phi merging data values properly
1433 phi = PhiNode::make( prev, old );
1434 phi->set_req( 1, nnn );
1435 }
1436 }
1437 // If inserting a new Phi, check for prior hits
1438 if( idx != 0 ) {
1439 Node *hit = _igvn.hash_find_insert(phi);
1440 if( hit == NULL ) {
1441 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1442 } else { // or
1443 // Remove the new phi from the graph and use the hit
1444 _igvn.remove_dead_node(phi);
1445 phi = hit; // Use existing phi
1446 }
1447 set_ctrl(phi, prev);
1448 }
1449 // Make 'use' use the Phi instead of the old loop body exit value
1450 _igvn.hash_delete(use);
1451 _igvn._worklist.push(use);
1452 use->set_req(idx, phi);
1453 if( use->_idx >= new_counter ) { // If updating new phis
1454 // Not needed for correctness, but prevents a weak assert
1455 // in AddPNode from tripping (when we end up with different
1456 // base & derived Phis that will become the same after
1457 // IGVN does CSE).
1458 Node *hit = _igvn.hash_find_insert(use);
1459 if( hit ) // Go ahead and re-hash for hits.
1460 _igvn.replace_node( use, hit );
1461 }
1462
1463 // If 'use' was in the loop-exit block, it now needs to be sunk
1464 // below the post-loop merge point.
1465 sink_use( use, prev );
1466 }
1467 }
1468 }
1469
1470 // Check for IFs that need splitting/cloning. Happens if an IF outside of
1471 // the loop uses a condition set in the loop. The original IF probably
1472 // takes control from one or more OLD Regions (which in turn get from NEW
1473 // Regions). In any case, there will be a set of Phis for each merge point
1474 // from the IF up to where the original BOOL def exists the loop.
1475 if( split_if_set ) {
1476 while( split_if_set->size() ) {
1477 Node *iff = split_if_set->pop();
1478 if( iff->in(1)->is_Phi() ) {
1479 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop );
1480 _igvn.hash_delete(iff);
1481 _igvn._worklist.push(iff);
1482 iff->set_req(1, b);
1483 }
1484 }
1485 }
1486 if( split_bool_set ) {
1487 while( split_bool_set->size() ) {
1488 Node *b = split_bool_set->pop();
1489 Node *phi = b->in(1);
1490 assert( phi->is_Phi(), "" );
1491 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop );
1492 _igvn.hash_delete(b);
1493 _igvn._worklist.push(b);
1494 b->set_req(1, cmp);
1495 }
1496 }
1497 if( split_cex_set ) {
1498 while( split_cex_set->size() ) {
1499 Node *b = split_cex_set->pop();
1500 assert( b->in(0)->is_Region(), "" );
1501 assert( b->in(1)->is_Phi(), "" );
1502 assert( b->in(0)->in(0) == b->in(1)->in(0), "" );
1503 split_up( b, b->in(0), NULL );
1504 }
1505 }
1506
1507 }
1508
1509
1510 //---------------------- stride_of_possible_iv -------------------------------------
1511 // Looks for an iff/bool/comp with one operand of the compare
1512 // being a cycle involving an add and a phi,
1513 // with an optional truncation (left-shift followed by a right-shift)
1514 // of the add. Returns zero if not an iv.
1515 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
1516 Node* trunc1 = NULL;
1517 Node* trunc2 = NULL;
1518 const TypeInt* ttype = NULL;
1519 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
1520 return 0;
1521 }
1522 BoolNode* bl = iff->in(1)->as_Bool();
1523 Node* cmp = bl->in(1);
1524 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) {
1525 return 0;
1526 }
1527 // Must have an invariant operand
1528 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
1529 return 0;
1530 }
1531 Node* add2 = NULL;
1532 Node* cmp1 = cmp->in(1);
1533 if (cmp1->is_Phi()) {
1534 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
1535 Node* phi = cmp1;
1536 for (uint i = 1; i < phi->req(); i++) {
1537 Node* in = phi->in(i);
1538 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
1539 &trunc1, &trunc2, &ttype);
1540 if (add && add->in(1) == phi) {
1541 add2 = add->in(2);
1542 break;
1543 }
1544 }
1545 } else {
1546 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
1547 Node* addtrunc = cmp1;
1548 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
1549 &trunc1, &trunc2, &ttype);
1550 if (add && add->in(1)->is_Phi()) {
1551 Node* phi = add->in(1);
1552 for (uint i = 1; i < phi->req(); i++) {
1553 if (phi->in(i) == addtrunc) {
1554 add2 = add->in(2);
1555 break;
1556 }
1557 }
1558 }
1559 }
1560 if (add2 != NULL) {
1561 const TypeInt* add2t = _igvn.type(add2)->is_int();
1562 if (add2t->is_con()) {
1563 return add2t->get_con();
1564 }
1565 }
1566 return 0;
1567 }
1568
1569
1570 //---------------------- stay_in_loop -------------------------------------
1571 // Return the (unique) control output node that's in the loop (if it exists.)
1572 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
1573 Node* unique = NULL;
1574 if (!n) return NULL;
1575 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1576 Node* use = n->fast_out(i);
1577 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
1578 if (unique != NULL) {
1579 return NULL;
1580 }
1581 unique = use;
1582 }
1583 }
1584 return unique;
1585 }
1586
1587 //------------------------------ register_node -------------------------------------
1588 // Utility to register node "n" with PhaseIdealLoop
1589 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
1590 _igvn.register_new_node_with_optimizer(n);
1591 loop->_body.push(n);
1592 if (n->is_CFG()) {
1593 set_loop(n, loop);
1594 set_idom(n, pred, ddepth);
1595 } else {
1596 set_ctrl(n, pred);
1597 }
1598 }
1599
1600 //------------------------------ proj_clone -------------------------------------
1601 // Utility to create an if-projection
1602 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
1603 ProjNode* c = p->clone()->as_Proj();
1604 c->set_req(0, iff);
1605 return c;
1606 }
1607
1608 //------------------------------ short_circuit_if -------------------------------------
1609 // Force the iff control output to be the live_proj
1610 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
1611 int proj_con = live_proj->_con;
1612 assert(proj_con == 0 || proj_con == 1, "false or true projection");
1613 Node *con = _igvn.intcon(proj_con);
1614 set_ctrl(con, C->root());
1615 if (iff) {
1616 iff->set_req(1, con);
1617 }
1618 return con;
1619 }
1620
1621 //------------------------------ insert_if_before_proj -------------------------------------
1622 // Insert a new if before an if projection (* - new node)
1623 //
1624 // before
1625 // if(test)
1626 // / \
1627 // v v
1628 // other-proj proj (arg)
1629 //
1630 // after
1631 // if(test)
1632 // / \
1633 // / v
1634 // | * proj-clone
1635 // v |
1636 // other-proj v
1637 // * new_if(relop(cmp[IU](left,right)))
1638 // / \
1639 // v v
1640 // * new-proj proj
1641 // (returned)
1642 //
1643 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
1644 IfNode* iff = proj->in(0)->as_If();
1645 IdealLoopTree *loop = get_loop(proj);
1646 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1647 int ddepth = dom_depth(proj);
1648
1649 _igvn.hash_delete(iff);
1650 _igvn._worklist.push(iff);
1651 _igvn.hash_delete(proj);
1652 _igvn._worklist.push(proj);
1653
1654 proj->set_req(0, NULL); // temporary disconnect
1655 ProjNode* proj2 = proj_clone(proj, iff);
1656 register_node(proj2, loop, iff, ddepth);
1657
1658 Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right);
1659 register_node(cmp, loop, proj2, ddepth);
1660
1661 BoolNode* bol = new (C,2)BoolNode(cmp, relop);
1662 register_node(bol, loop, proj2, ddepth);
1663
1664 IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt);
1665 register_node(new_if, loop, proj2, ddepth);
1666
1667 proj->set_req(0, new_if); // reattach
1668 set_idom(proj, new_if, ddepth);
1669
1670 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
1671 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
1672
1673 return new_exit;
1674 }
1675
1676 //------------------------------ insert_region_before_proj -------------------------------------
1677 // Insert a region before an if projection (* - new node)
1678 //
1679 // before
1680 // if(test)
1681 // / |
1682 // v |
1683 // proj v
1684 // other-proj
1685 //
1686 // after
1687 // if(test)
1688 // / |
1689 // v |
1690 // * proj-clone v
1691 // | other-proj
1692 // v
1693 // * new-region
1694 // |
1695 // v
1696 // * dum_if
1697 // / \
1698 // v \
1699 // * dum-proj v
1700 // proj
1701 //
1702 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
1703 IfNode* iff = proj->in(0)->as_If();
1704 IdealLoopTree *loop = get_loop(proj);
1705 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1706 int ddepth = dom_depth(proj);
1707
1708 _igvn.hash_delete(iff);
1709 _igvn._worklist.push(iff);
1710 _igvn.hash_delete(proj);
1711 _igvn._worklist.push(proj);
1712
1713 proj->set_req(0, NULL); // temporary disconnect
1714 ProjNode* proj2 = proj_clone(proj, iff);
1715 register_node(proj2, loop, iff, ddepth);
1716
1717 RegionNode* reg = new (C,2)RegionNode(2);
1718 reg->set_req(1, proj2);
1719 register_node(reg, loop, iff, ddepth);
1720
1721 IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
1722 register_node(dum_if, loop, reg, ddepth);
1723
1724 proj->set_req(0, dum_if); // reattach
1725 set_idom(proj, dum_if, ddepth);
1726
1727 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
1728 register_node(dum_proj, loop, dum_if, ddepth);
1729
1730 return reg;
1731 }
1732
1733 //------------------------------ insert_cmpi_loop_exit -------------------------------------
1734 // Clone a signed compare loop exit from an unsigned compare and
1735 // insert it before the unsigned cmp on the stay-in-loop path.
1736 // All new nodes inserted in the dominator tree between the original
1737 // if and it's projections. The original if test is replaced with
1738 // a constant to force the stay-in-loop path.
1739 //
1740 // This is done to make sure that the original if and it's projections
1741 // still dominate the same set of control nodes, that the ctrl() relation
1742 // from data nodes to them is preserved, and that their loop nesting is
1743 // preserved.
1744 //
1745 // before
1746 // if(i <u limit) unsigned compare loop exit
1747 // / |
1748 // v v
1749 // exit-proj stay-in-loop-proj
1750 //
1751 // after
1752 // if(stay-in-loop-const) original if
1753 // / |
1754 // / v
1755 // / if(i < limit) new signed test
1756 // / / |
1757 // / / v
1758 // / / if(i <u limit) new cloned unsigned test
1759 // / / / |
1760 // v v v |
1761 // region |
1762 // | |
1763 // dum-if |
1764 // / | |
1765 // ether | |
1766 // v v
1767 // exit-proj stay-in-loop-proj
1768 //
1769 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
1770 const bool Signed = true;
1771 const bool Unsigned = false;
1772
1773 BoolNode* bol = if_cmpu->in(1)->as_Bool();
1774 if (bol->_test._test != BoolTest::lt) return NULL;
1775 CmpNode* cmpu = bol->in(1)->as_Cmp();
1776 if (cmpu->Opcode() != Op_CmpU) return NULL;
1777 int stride = stride_of_possible_iv(if_cmpu);
1778 if (stride == 0) return NULL;
1779
1780 ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj();
1781 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
1782
1783 Node* limit = NULL;
1784 if (stride > 0) {
1785 limit = cmpu->in(2);
1786 } else {
1787 limit = _igvn.makecon(TypeInt::ZERO);
1788 set_ctrl(limit, C->root());
1789 }
1790 // Create a new region on the exit path
1791 RegionNode* reg = insert_region_before_proj(lp_exit);
1792
1793 // Clone the if-cmpu-true-false using a signed compare
1794 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
1795 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
1796 reg->add_req(cmpi_exit);
1797
1798 // Clone the if-cmpu-true-false
1799 BoolTest::mask rel_u = bol->_test._test;
1800 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
1801 reg->add_req(cmpu_exit);
1802
1803 // Force original if to stay in loop.
1804 short_circuit_if(if_cmpu, lp_continue);
1805
1806 return cmpi_exit->in(0)->as_If();
1807 }
1808
1809 //------------------------------ remove_cmpi_loop_exit -------------------------------------
1810 // Remove a previously inserted signed compare loop exit.
1811 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
1812 Node* lp_proj = stay_in_loop(if_cmp, loop);
1813 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
1814 stay_in_loop(lp_proj, loop)->is_If() &&
1815 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
1816 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
1817 set_ctrl(con, C->root());
1818 if_cmp->set_req(1, con);
1819 }
1820
1821 //------------------------------ scheduled_nodelist -------------------------------------
1822 // Create a post order schedule of nodes that are in the
1823 // "member" set. The list is returned in "sched".
1824 // The first node in "sched" is the loop head, followed by
1825 // nodes which have no inputs in the "member" set, and then
1826 // followed by the nodes that have an immediate input dependence
1827 // on a node in "sched".
1828 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
1829
1830 assert(member.test(loop->_head->_idx), "loop head must be in member set");
1831 Arena *a = Thread::current()->resource_area();
1832 VectorSet visited(a);
1833 Node_Stack nstack(a, loop->_body.size());
1834
1835 Node* n = loop->_head; // top of stack is cached in "n"
1836 uint idx = 0;
1837 visited.set(n->_idx);
1838
1839 // Initially push all with no inputs from within member set
1840 for(uint i = 0; i < loop->_body.size(); i++ ) {
1841 Node *elt = loop->_body.at(i);
1842 if (member.test(elt->_idx)) {
1843 bool found = false;
1844 for (uint j = 0; j < elt->req(); j++) {
1845 Node* def = elt->in(j);
1846 if (def && member.test(def->_idx) && def != elt) {
1847 found = true;
1848 break;
1849 }
1850 }
1851 if (!found && elt != loop->_head) {
1852 nstack.push(n, idx);
1853 n = elt;
1854 assert(!visited.test(n->_idx), "not seen yet");
1855 visited.set(n->_idx);
1856 }
1857 }
1858 }
1859
1860 // traverse out's that are in the member set
1861 while (true) {
1862 if (idx < n->outcnt()) {
1863 Node* use = n->raw_out(idx);
1864 idx++;
1865 if (!visited.test_set(use->_idx)) {
1866 if (member.test(use->_idx)) {
1867 nstack.push(n, idx);
1868 n = use;
1869 idx = 0;
1870 }
1871 }
1872 } else {
1873 // All outputs processed
1874 sched.push(n);
1875 if (nstack.is_empty()) break;
1876 n = nstack.node();
1877 idx = nstack.index();
1878 nstack.pop();
1879 }
1880 }
1881 }
1882
1883
1884 //------------------------------ has_use_in_set -------------------------------------
1885 // Has a use in the vector set
1886 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
1887 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1888 Node* use = n->fast_out(j);
1889 if (vset.test(use->_idx)) {
1890 return true;
1891 }
1892 }
1893 return false;
1894 }
1895
1896
1897 //------------------------------ has_use_internal_to_set -------------------------------------
1898 // Has use internal to the vector set (ie. not in a phi at the loop head)
1899 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
1900 Node* head = loop->_head;
1901 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1902 Node* use = n->fast_out(j);
1903 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
1904 return true;
1905 }
1906 }
1907 return false;
1908 }
1909
1910
1911 //------------------------------ clone_for_use_outside_loop -------------------------------------
1912 // clone "n" for uses that are outside of loop
1913 void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
1914
1915 assert(worklist.size() == 0, "should be empty");
1916 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1917 Node* use = n->fast_out(j);
1918 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
1919 worklist.push(use);
1920 }
1921 }
1922 while( worklist.size() ) {
1923 Node *use = worklist.pop();
1924 if (!has_node(use) || use->in(0) == C->top()) continue;
1925 uint j;
1926 for (j = 0; j < use->req(); j++) {
1927 if (use->in(j) == n) break;
1928 }
1929 assert(j < use->req(), "must be there");
1930
1931 // clone "n" and insert it between the inputs of "n" and the use outside the loop
1932 Node* n_clone = n->clone();
1933 _igvn.hash_delete(use);
1934 use->set_req(j, n_clone);
1935 _igvn._worklist.push(use);
1936 Node* use_c;
1937 if (!use->is_Phi()) {
1938 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
1939 } else {
1940 // Use in a phi is considered a use in the associated predecessor block
1941 use_c = use->in(0)->in(j);
1942 }
1943 set_ctrl(n_clone, use_c);
1944 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
1945 get_loop(use_c)->_body.push(n_clone);
1946 _igvn.register_new_node_with_optimizer(n_clone);
1947 #if !defined(PRODUCT)
1948 if (TracePartialPeeling) {
1949 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
1950 }
1951 #endif
1952 }
1953 }
1954
1955
1956 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
1957 // clone "n" for special uses that are in the not_peeled region.
1958 // If these def-uses occur in separate blocks, the code generator
1959 // marks the method as not compilable. For example, if a "BoolNode"
1960 // is in a different basic block than the "IfNode" that uses it, then
1961 // the compilation is aborted in the code generator.
1962 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1963 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
1964 if (n->is_Phi() || n->is_Load()) {
1965 return;
1966 }
1967 assert(worklist.size() == 0, "should be empty");
1968 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1969 Node* use = n->fast_out(j);
1970 if ( not_peel.test(use->_idx) &&
1971 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
1972 use->in(1) == n) {
1973 worklist.push(use);
1974 }
1975 }
1976 if (worklist.size() > 0) {
1977 // clone "n" and insert it between inputs of "n" and the use
1978 Node* n_clone = n->clone();
1979 loop->_body.push(n_clone);
1980 _igvn.register_new_node_with_optimizer(n_clone);
1981 set_ctrl(n_clone, get_ctrl(n));
1982 sink_list.push(n_clone);
1983 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
1984 #if !defined(PRODUCT)
1985 if (TracePartialPeeling) {
1986 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
1987 }
1988 #endif
1989 while( worklist.size() ) {
1990 Node *use = worklist.pop();
1991 _igvn.hash_delete(use);
1992 _igvn._worklist.push(use);
1993 for (uint j = 1; j < use->req(); j++) {
1994 if (use->in(j) == n) {
1995 use->set_req(j, n_clone);
1996 }
1997 }
1998 }
1999 }
2000 }
2001
2002
2003 //------------------------------ insert_phi_for_loop -------------------------------------
2004 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2005 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2006 Node *phi = PhiNode::make(lp, back_edge_val);
2007 phi->set_req(LoopNode::EntryControl, lp_entry_val);
2008 // Use existing phi if it already exists
2009 Node *hit = _igvn.hash_find_insert(phi);
2010 if( hit == NULL ) {
2011 _igvn.register_new_node_with_optimizer(phi);
2012 set_ctrl(phi, lp);
2013 } else {
2014 // Remove the new phi from the graph and use the hit
2015 _igvn.remove_dead_node(phi);
2016 phi = hit;
2017 }
2018 _igvn.hash_delete(use);
2019 _igvn._worklist.push(use);
2020 use->set_req(idx, phi);
2021 }
2022
2023 #ifdef ASSERT
2024 //------------------------------ is_valid_loop_partition -------------------------------------
2025 // Validate the loop partition sets: peel and not_peel
2026 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2027 VectorSet& not_peel ) {
2028 uint i;
2029 // Check that peel_list entries are in the peel set
2030 for (i = 0; i < peel_list.size(); i++) {
2031 if (!peel.test(peel_list.at(i)->_idx)) {
2032 return false;
2033 }
2034 }
2035 // Check at loop members are in one of peel set or not_peel set
2036 for (i = 0; i < loop->_body.size(); i++ ) {
2037 Node *def = loop->_body.at(i);
2038 uint di = def->_idx;
2039 // Check that peel set elements are in peel_list
2040 if (peel.test(di)) {
2041 if (not_peel.test(di)) {
2042 return false;
2043 }
2044 // Must be in peel_list also
2045 bool found = false;
2046 for (uint j = 0; j < peel_list.size(); j++) {
2047 if (peel_list.at(j)->_idx == di) {
2048 found = true;
2049 break;
2050 }
2051 }
2052 if (!found) {
2053 return false;
2054 }
2055 } else if (not_peel.test(di)) {
2056 if (peel.test(di)) {
2057 return false;
2058 }
2059 } else {
2060 return false;
2061 }
2062 }
2063 return true;
2064 }
2065
2066 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2067 // Ensure a use outside of loop is of the right form
2068 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2069 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2070 return (use->is_Phi() &&
2071 use_c->is_Region() && use_c->req() == 3 &&
2072 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2073 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2074 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2075 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2076 }
2077
2078 //------------------------------ is_valid_clone_loop_form -------------------------------------
2079 // Ensure that all uses outside of loop are of the right form
2080 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2081 uint orig_exit_idx, uint clone_exit_idx) {
2082 uint len = peel_list.size();
2083 for (uint i = 0; i < len; i++) {
2084 Node *def = peel_list.at(i);
2085
2086 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2087 Node *use = def->fast_out(j);
2088 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2089 if (!loop->is_member(get_loop(use_c))) {
2090 // use is not in the loop, check for correct structure
2091 if (use->in(0) == def) {
2092 // Okay
2093 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2094 return false;
2095 }
2096 }
2097 }
2098 }
2099 return true;
2100 }
2101 #endif
2102
2103 //------------------------------ partial_peel -------------------------------------
2104 // Partially peel (aka loop rotation) the top portion of a loop (called
2105 // the peel section below) by cloning it and placing one copy just before
2106 // the new loop head and the other copy at the bottom of the new loop.
2107 //
2108 // before after where it came from
2109 //
2110 // stmt1 stmt1
2111 // loop: stmt2 clone
2112 // stmt2 if condA goto exitA clone
2113 // if condA goto exitA new_loop: new
2114 // stmt3 stmt3 clone
2115 // if !condB goto loop if condB goto exitB clone
2116 // exitB: stmt2 orig
2117 // stmt4 if !condA goto new_loop orig
2118 // exitA: goto exitA
2119 // exitB:
2120 // stmt4
2121 // exitA:
2122 //
2123 // Step 1: find the cut point: an exit test on probable
2124 // induction variable.
2125 // Step 2: schedule (with cloning) operations in the peel
2126 // section that can be executed after the cut into
2127 // the section that is not peeled. This may need
2128 // to clone operations into exit blocks. For
2129 // instance, a reference to A[i] in the not-peel
2130 // section and a reference to B[i] in an exit block
2131 // may cause a left-shift of i by 2 to be placed
2132 // in the peel block. This step will clone the left
2133 // shift into the exit block and sink the left shift
2134 // from the peel to the not-peel section.
2135 // Step 3: clone the loop, retarget the control, and insert
2136 // phis for values that are live across the new loop
2137 // head. This is very dependent on the graph structure
2138 // from clone_loop. It creates region nodes for
2139 // exit control and associated phi nodes for values
2140 // flow out of the loop through that exit. The region
2141 // node is dominated by the clone's control projection.
2142 // So the clone's peel section is placed before the
2143 // new loop head, and the clone's not-peel section is
2144 // forms the top part of the new loop. The original
2145 // peel section forms the tail of the new loop.
2146 // Step 4: update the dominator tree and recompute the
2147 // dominator depth.
2148 //
2149 // orig
2150 //
2151 // stmt1
2152 // |
2153 // v
2154 // loop predicate
2155 // |
2156 // v
2157 // loop<----+
2158 // | |
2159 // stmt2 |
2160 // | |
2161 // v |
2162 // ifA |
2163 // / | |
2164 // v v |
2165 // false true ^ <-- last_peel
2166 // / | |
2167 // / ===|==cut |
2168 // / stmt3 | <-- first_not_peel
2169 // / | |
2170 // | v |
2171 // v ifB |
2172 // exitA: / \ |
2173 // / \ |
2174 // v v |
2175 // false true |
2176 // / \ |
2177 // / ----+
2178 // |
2179 // v
2180 // exitB:
2181 // stmt4
2182 //
2183 //
2184 // after clone loop
2185 //
2186 // stmt1
2187 // |
2188 // v
2189 // loop predicate
2190 // / \
2191 // clone / \ orig
2192 // / \
2193 // / \
2194 // v v
2195 // +---->loop loop<----+
2196 // | | | |
2197 // | stmt2 stmt2 |
2198 // | | | |
2199 // | v v |
2200 // | ifA ifA |
2201 // | | \ / | |
2202 // | v v v v |
2203 // ^ true false false true ^ <-- last_peel
2204 // | | ^ \ / | |
2205 // | cut==|== \ \ / ===|==cut |
2206 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2207 // | | dom | | | |
2208 // | v \ 1v v2 v |
2209 // | ifB regionA ifB |
2210 // | / \ | / \ |
2211 // | / \ v / \ |
2212 // | v v exitA: v v |
2213 // | true false false true |
2214 // | / ^ \ / \ |
2215 // +---- \ \ / ----+
2216 // dom \ /
2217 // \ 1v v2
2218 // regionB
2219 // |
2220 // v
2221 // exitB:
2222 // stmt4
2223 //
2224 //
2225 // after partial peel
2226 //
2227 // stmt1
2228 // |
2229 // v
2230 // loop predicate
2231 // /
2232 // clone / orig
2233 // / TOP
2234 // / \
2235 // v v
2236 // TOP->loop loop----+
2237 // | | |
2238 // stmt2 stmt2 |
2239 // | | |
2240 // v v |
2241 // ifA ifA |
2242 // | \ / | |
2243 // v v v v |
2244 // true false false true | <-- last_peel
2245 // | ^ \ / +------|---+
2246 // +->newloop \ \ / === ==cut | |
2247 // | stmt3 \ \ / TOP | |
2248 // | | dom | | stmt3 | | <-- first_not_peel
2249 // | v \ 1v v2 v | |
2250 // | ifB regionA ifB ^ v
2251 // | / \ | / \ | |
2252 // | / \ v / \ | |
2253 // | v v exitA: v v | |
2254 // | true false false true | |
2255 // | / ^ \ / \ | |
2256 // | | \ \ / v | |
2257 // | | dom \ / TOP | |
2258 // | | \ 1v v2 | |
2259 // ^ v regionB | |
2260 // | | | | |
2261 // | | v ^ v
2262 // | | exitB: | |
2263 // | | stmt4 | |
2264 // | +------------>-----------------+ |
2265 // | |
2266 // +-----------------<---------------------+
2267 //
2268 //
2269 // final graph
2270 //
2271 // stmt1
2272 // |
2273 // v
2274 // loop predicate
2275 // |
2276 // v
2277 // stmt2 clone
2278 // |
2279 // v
2280 // ........> ifA clone
2281 // : / |
2282 // dom / |
2283 // : v v
2284 // : false true
2285 // : | |
2286 // : | v
2287 // : | newloop<-----+
2288 // : | | |
2289 // : | stmt3 clone |
2290 // : | | |
2291 // : | v |
2292 // : | ifB |
2293 // : | / \ |
2294 // : | v v |
2295 // : | false true |
2296 // : | | | |
2297 // : | v stmt2 |
2298 // : | exitB: | |
2299 // : | stmt4 v |
2300 // : | ifA orig |
2301 // : | / \ |
2302 // : | / \ |
2303 // : | v v |
2304 // : | false true |
2305 // : | / \ |
2306 // : v v -----+
2307 // RegionA
2308 // |
2309 // v
2310 // exitA
2311 //
2312 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
2313
2314 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
2315 if (!loop->_head->is_Loop()) {
2316 return false; }
2317
2318 LoopNode *head = loop->_head->as_Loop();
2319
2320 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
2321 return false;
2322 }
2323
2324 // Check for complex exit control
2325 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
2326 Node *n = loop->_body.at(ii);
2327 int opc = n->Opcode();
2328 if (n->is_Call() ||
2329 opc == Op_Catch ||
2330 opc == Op_CatchProj ||
2331 opc == Op_Jump ||
2332 opc == Op_JumpProj) {
2333 #if !defined(PRODUCT)
2334 if (TracePartialPeeling) {
2335 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
2336 }
2337 #endif
2338 return false;
2339 }
2340 }
2341
2342 int dd = dom_depth(head);
2343
2344 // Step 1: find cut point
2345
2346 // Walk up dominators to loop head looking for first loop exit
2347 // which is executed on every path thru loop.
2348 IfNode *peel_if = NULL;
2349 IfNode *peel_if_cmpu = NULL;
2350
2351 Node *iff = loop->tail();
2352 while( iff != head ) {
2353 if( iff->is_If() ) {
2354 Node *ctrl = get_ctrl(iff->in(1));
2355 if (ctrl->is_top()) return false; // Dead test on live IF.
2356 // If loop-varying exit-test, check for induction variable
2357 if( loop->is_member(get_loop(ctrl)) &&
2358 loop->is_loop_exit(iff) &&
2359 is_possible_iv_test(iff)) {
2360 Node* cmp = iff->in(1)->in(1);
2361 if (cmp->Opcode() == Op_CmpI) {
2362 peel_if = iff->as_If();
2363 } else {
2364 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
2365 peel_if_cmpu = iff->as_If();
2366 }
2367 }
2368 }
2369 iff = idom(iff);
2370 }
2371 // Prefer signed compare over unsigned compare.
2372 IfNode* new_peel_if = NULL;
2373 if (peel_if == NULL) {
2374 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
2375 return false; // No peel point found
2376 }
2377 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
2378 if (new_peel_if == NULL) {
2379 return false; // No peel point found
2380 }
2381 peel_if = new_peel_if;
2382 }
2383 Node* last_peel = stay_in_loop(peel_if, loop);
2384 Node* first_not_peeled = stay_in_loop(last_peel, loop);
2385 if (first_not_peeled == NULL || first_not_peeled == head) {
2386 return false;
2387 }
2388
2389 #if !defined(PRODUCT)
2390 if (TraceLoopOpts) {
2391 tty->print("PartialPeel ");
2392 loop->dump_head();
2393 }
2394
2395 if (TracePartialPeeling) {
2396 tty->print_cr("before partial peel one iteration");
2397 Node_List wl;
2398 Node* t = head->in(2);
2399 while (true) {
2400 wl.push(t);
2401 if (t == head) break;
2402 t = idom(t);
2403 }
2404 while (wl.size() > 0) {
2405 Node* tt = wl.pop();
2406 tt->dump();
2407 if (tt == last_peel) tty->print_cr("-- cut --");
2408 }
2409 }
2410 #endif
2411 ResourceArea *area = Thread::current()->resource_area();
2412 VectorSet peel(area);
2413 VectorSet not_peel(area);
2414 Node_List peel_list(area);
2415 Node_List worklist(area);
2416 Node_List sink_list(area);
2417
2418 // Set of cfg nodes to peel are those that are executable from
2419 // the head through last_peel.
2420 assert(worklist.size() == 0, "should be empty");
2421 worklist.push(head);
2422 peel.set(head->_idx);
2423 while (worklist.size() > 0) {
2424 Node *n = worklist.pop();
2425 if (n != last_peel) {
2426 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2427 Node* use = n->fast_out(j);
2428 if (use->is_CFG() &&
2429 loop->is_member(get_loop(use)) &&
2430 !peel.test_set(use->_idx)) {
2431 worklist.push(use);
2432 }
2433 }
2434 }
2435 }
2436
2437 // Set of non-cfg nodes to peel are those that are control
2438 // dependent on the cfg nodes.
2439 uint i;
2440 for(i = 0; i < loop->_body.size(); i++ ) {
2441 Node *n = loop->_body.at(i);
2442 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
2443 if (peel.test(n_c->_idx)) {
2444 peel.set(n->_idx);
2445 } else {
2446 not_peel.set(n->_idx);
2447 }
2448 }
2449
2450 // Step 2: move operations from the peeled section down into the
2451 // not-peeled section
2452
2453 // Get a post order schedule of nodes in the peel region
2454 // Result in right-most operand.
2455 scheduled_nodelist(loop, peel, peel_list );
2456
2457 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2458
2459 // For future check for too many new phis
2460 uint old_phi_cnt = 0;
2461 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2462 Node* use = head->fast_out(j);
2463 if (use->is_Phi()) old_phi_cnt++;
2464 }
2465
2466 #if !defined(PRODUCT)
2467 if (TracePartialPeeling) {
2468 tty->print_cr("\npeeled list");
2469 }
2470 #endif
2471
2472 // Evacuate nodes in peel region into the not_peeled region if possible
2473 uint new_phi_cnt = 0;
2474 for (i = 0; i < peel_list.size();) {
2475 Node* n = peel_list.at(i);
2476 #if !defined(PRODUCT)
2477 if (TracePartialPeeling) n->dump();
2478 #endif
2479 bool incr = true;
2480 if ( !n->is_CFG() ) {
2481
2482 if ( has_use_in_set(n, not_peel) ) {
2483
2484 // If not used internal to the peeled region,
2485 // move "n" from peeled to not_peeled region.
2486
2487 if ( !has_use_internal_to_set(n, peel, loop) ) {
2488
2489 // if not pinned and not a load (which maybe anti-dependent on a store)
2490 // and not a CMove (Matcher expects only bool->cmove).
2491 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
2492 clone_for_use_outside_loop( loop, n, worklist );
2493
2494 sink_list.push(n);
2495 peel >>= n->_idx; // delete n from peel set.
2496 not_peel <<= n->_idx; // add n to not_peel set.
2497 peel_list.remove(i);
2498 incr = false;
2499 #if !defined(PRODUCT)
2500 if (TracePartialPeeling) {
2501 tty->print_cr("sink to not_peeled region: %d newbb: %d",
2502 n->_idx, get_ctrl(n)->_idx);
2503 }
2504 #endif
2505 }
2506 } else {
2507 // Otherwise check for special def-use cases that span
2508 // the peel/not_peel boundary such as bool->if
2509 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
2510 new_phi_cnt++;
2511 }
2512 }
2513 }
2514 if (incr) i++;
2515 }
2516
2517 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
2518 #if !defined(PRODUCT)
2519 if (TracePartialPeeling) {
2520 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
2521 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
2522 }
2523 #endif
2524 if (new_peel_if != NULL) {
2525 remove_cmpi_loop_exit(new_peel_if, loop);
2526 }
2527 // Inhibit more partial peeling on this loop
2528 assert(!head->is_partial_peel_loop(), "not partial peeled");
2529 head->mark_partial_peel_failed();
2530 return false;
2531 }
2532
2533 // Step 3: clone loop, retarget control, and insert new phis
2534
2535 // Create new loop head for new phis and to hang
2536 // the nodes being moved (sinked) from the peel region.
2537 LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel);
2538 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
2539 _igvn.register_new_node_with_optimizer(new_head);
2540 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
2541 first_not_peeled->set_req(0, new_head);
2542 set_loop(new_head, loop);
2543 loop->_body.push(new_head);
2544 not_peel.set(new_head->_idx);
2545 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
2546 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
2547
2548 while (sink_list.size() > 0) {
2549 Node* n = sink_list.pop();
2550 set_ctrl(n, new_head);
2551 }
2552
2553 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2554
2555 clone_loop( loop, old_new, dd );
2556
2557 const uint clone_exit_idx = 1;
2558 const uint orig_exit_idx = 2;
2559 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
2560
2561 Node* head_clone = old_new[head->_idx];
2562 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
2563 Node* orig_tail_clone = head_clone->in(2);
2564
2565 // Add phi if "def" node is in peel set and "use" is not
2566
2567 for(i = 0; i < peel_list.size(); i++ ) {
2568 Node *def = peel_list.at(i);
2569 if (!def->is_CFG()) {
2570 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2571 Node *use = def->fast_out(j);
2572 if (has_node(use) && use->in(0) != C->top() &&
2573 (!peel.test(use->_idx) ||
2574 (use->is_Phi() && use->in(0) == head)) ) {
2575 worklist.push(use);
2576 }
2577 }
2578 while( worklist.size() ) {
2579 Node *use = worklist.pop();
2580 for (uint j = 1; j < use->req(); j++) {
2581 Node* n = use->in(j);
2582 if (n == def) {
2583
2584 // "def" is in peel set, "use" is not in peel set
2585 // or "use" is in the entry boundary (a phi) of the peel set
2586
2587 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
2588
2589 if ( loop->is_member(get_loop( use_c )) ) {
2590 // use is in loop
2591 if (old_new[use->_idx] != NULL) { // null for dead code
2592 Node* use_clone = old_new[use->_idx];
2593 _igvn.hash_delete(use);
2594 use->set_req(j, C->top());
2595 _igvn._worklist.push(use);
2596 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
2597 }
2598 } else {
2599 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
2600 // use is not in the loop, check if the live range includes the cut
2601 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
2602 if (not_peel.test(lp_if->_idx)) {
2603 assert(j == orig_exit_idx, "use from original loop");
2604 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
2605 }
2606 }
2607 }
2608 }
2609 }
2610 }
2611 }
2612
2613 // Step 3b: retarget control
2614
2615 // Redirect control to the new loop head if a cloned node in
2616 // the not_peeled region has control that points into the peeled region.
2617 // This necessary because the cloned peeled region will be outside
2618 // the loop.
2619 // from to
2620 // cloned-peeled <---+
2621 // new_head_clone: | <--+
2622 // cloned-not_peeled in(0) in(0)
2623 // orig-peeled
2624
2625 for(i = 0; i < loop->_body.size(); i++ ) {
2626 Node *n = loop->_body.at(i);
2627 if (!n->is_CFG() && n->in(0) != NULL &&
2628 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
2629 Node* n_clone = old_new[n->_idx];
2630 _igvn.hash_delete(n_clone);
2631 n_clone->set_req(0, new_head_clone);
2632 _igvn._worklist.push(n_clone);
2633 }
2634 }
2635
2636 // Backedge of the surviving new_head (the clone) is original last_peel
2637 _igvn.hash_delete(new_head_clone);
2638 new_head_clone->set_req(LoopNode::LoopBackControl, last_peel);
2639 _igvn._worklist.push(new_head_clone);
2640
2641 // Cut first node in original not_peel set
2642 _igvn.hash_delete(new_head);
2643 new_head->set_req(LoopNode::EntryControl, C->top());
2644 new_head->set_req(LoopNode::LoopBackControl, C->top());
2645 _igvn._worklist.push(new_head);
2646
2647 // Copy head_clone back-branch info to original head
2648 // and remove original head's loop entry and
2649 // clone head's back-branch
2650 _igvn.hash_delete(head);
2651 _igvn.hash_delete(head_clone);
2652 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
2653 head->set_req(LoopNode::LoopBackControl, C->top());
2654 head_clone->set_req(LoopNode::LoopBackControl, C->top());
2655 _igvn._worklist.push(head);
2656 _igvn._worklist.push(head_clone);
2657
2658 // Similarly modify the phis
2659 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
2660 Node* use = head->fast_out(k);
2661 if (use->is_Phi() && use->outcnt() > 0) {
2662 Node* use_clone = old_new[use->_idx];
2663 _igvn.hash_delete(use);
2664 _igvn.hash_delete(use_clone);
2665 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
2666 use->set_req(LoopNode::LoopBackControl, C->top());
2667 use_clone->set_req(LoopNode::LoopBackControl, C->top());
2668 _igvn._worklist.push(use);
2669 _igvn._worklist.push(use_clone);
2670 }
2671 }
2672
2673 // Step 4: update dominator tree and dominator depth
2674
2675 set_idom(head, orig_tail_clone, dd);
2676 recompute_dom_depth();
2677
2678 // Inhibit more partial peeling on this loop
2679 new_head_clone->set_partial_peel_loop();
2680 C->set_major_progress();
2681
2682 #if !defined(PRODUCT)
2683 if (TracePartialPeeling) {
2684 tty->print_cr("\nafter partial peel one iteration");
2685 Node_List wl(area);
2686 Node* t = last_peel;
2687 while (true) {
2688 wl.push(t);
2689 if (t == head_clone) break;
2690 t = idom(t);
2691 }
2692 while (wl.size() > 0) {
2693 Node* tt = wl.pop();
2694 if (tt == head) tty->print_cr("orig head");
2695 else if (tt == new_head_clone) tty->print_cr("new head");
2696 else if (tt == head_clone) tty->print_cr("clone head");
2697 tt->dump();
2698 }
2699 }
2700 #endif
2701 return true;
2702 }
2703
2704 //------------------------------reorg_offsets----------------------------------
2705 // Reorganize offset computations to lower register pressure. Mostly
2706 // prevent loop-fallout uses of the pre-incremented trip counter (which are
2707 // then alive with the post-incremented trip counter forcing an extra
2708 // register move)
2709 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
2710 // Perform it only for canonical counted loops.
2711 // Loop's shape could be messed up by iteration_split_impl.
2712 if (!loop->_head->is_CountedLoop())
2713 return;
2714 if (!loop->_head->as_Loop()->is_valid_counted_loop())
2715 return;
2716
2717 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2718 CountedLoopEndNode *cle = cl->loopexit();
2719 Node *exit = cle->proj_out(false);
2720 Node *phi = cl->phi();
2721
2722 // Check for the special case of folks using the pre-incremented
2723 // trip-counter on the fall-out path (forces the pre-incremented
2724 // and post-incremented trip counter to be live at the same time).
2725 // Fix this by adjusting to use the post-increment trip counter.
2726
2727 bool progress = true;
2728 while (progress) {
2729 progress = false;
2730 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
2731 Node* use = phi->fast_out(i); // User of trip-counter
2732 if (!has_ctrl(use)) continue;
2733 Node *u_ctrl = get_ctrl(use);
2734 if (use->is_Phi()) {
2735 u_ctrl = NULL;
2736 for (uint j = 1; j < use->req(); j++)
2737 if (use->in(j) == phi)
2738 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
2739 }
2740 IdealLoopTree *u_loop = get_loop(u_ctrl);
2741 // Look for loop-invariant use
2742 if (u_loop == loop) continue;
2743 if (loop->is_member(u_loop)) continue;
2744 // Check that use is live out the bottom. Assuming the trip-counter
2745 // update is right at the bottom, uses of of the loop middle are ok.
2746 if (dom_lca(exit, u_ctrl) != exit) continue;
2747 // Hit! Refactor use to use the post-incremented tripcounter.
2748 // Compute a post-increment tripcounter.
2749 Node *opaq = new (C, 2) Opaque2Node( C, cle->incr() );
2750 register_new_node( opaq, u_ctrl );
2751 Node *neg_stride = _igvn.intcon(-cle->stride_con());
2752 set_ctrl(neg_stride, C->root());
2753 Node *post = new (C, 3) AddINode( opaq, neg_stride);
2754 register_new_node( post, u_ctrl );
2755 _igvn.hash_delete(use);
2756 _igvn._worklist.push(use);
2757 for (uint j = 1; j < use->req(); j++) {
2758 if (use->in(j) == phi)
2759 use->set_req(j, post);
2760 }
2761 // Since DU info changed, rerun loop
2762 progress = true;
2763 break;
2764 }
2765 }
2766
2767 }