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