1 /*
2 * Copyright (c) 1999, 2015, 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/matcher.hpp"
32 #include "opto/mulnode.hpp"
33 #include "opto/movenode.hpp"
34 #include "opto/opaquenode.hpp"
35 #include "opto/rootnode.hpp"
36 #include "opto/shenandoahSupport.hpp"
37 #include "opto/subnode.hpp"
38
39 //=============================================================================
40 //------------------------------split_thru_phi---------------------------------
41 // Split Node 'n' through merge point if there is enough win.
42 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
43 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
44 // ConvI2L may have type information on it which is unsafe to push up
45 // so disable this for now
46 return NULL;
47 }
48
49 int wins = 0;
50 assert(!n->is_CFG(), "");
51 assert(region->is_Region(), "");
52
53 const Type* type = n->bottom_type();
54 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr();
55 Node *phi;
56 if (t_oop != NULL && t_oop->is_known_instance_field()) {
57 int iid = t_oop->instance_id();
58 int index = C->get_alias_index(t_oop);
59 int offset = t_oop->offset();
60 phi = new PhiNode(region, type, NULL, iid, index, offset);
61 } else {
62 phi = PhiNode::make_blank(region, n);
63 }
64 uint old_unique = C->unique();
65 for (uint i = 1; i < region->req(); i++) {
66 Node *x;
67 Node* the_clone = NULL;
68 if (region->in(i) == C->top()) {
69 x = C->top(); // Dead path? Use a dead data op
70 } else {
71 x = n->clone(); // Else clone up the data op
72 the_clone = x; // Remember for possible deletion.
73 // Alter data node to use pre-phi inputs
74 if (n->in(0) == region)
75 x->set_req( 0, region->in(i) );
76 for (uint j = 1; j < n->req(); j++) {
77 Node *in = n->in(j);
78 if (in->is_Phi() && in->in(0) == region)
79 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
80 }
81 }
82 // Check for a 'win' on some paths
83 const Type *t = x->Value(&_igvn);
84
85 bool singleton = t->singleton();
86
87 // A TOP singleton indicates that there are no possible values incoming
88 // along a particular edge. In most cases, this is OK, and the Phi will
89 // be eliminated later in an Ideal call. However, we can't allow this to
90 // happen if the singleton occurs on loop entry, as the elimination of
91 // the PhiNode may cause the resulting node to migrate back to a previous
92 // loop iteration.
93 if (singleton && t == Type::TOP) {
94 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
95 // irreducible loop may not be indicated by an affirmative is_Loop());
96 // therefore, the only top we can split thru a phi is on a backedge of
97 // a loop.
98 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
99 }
100
101 if (singleton) {
102 wins++;
103 x = ((PhaseGVN&)_igvn).makecon(t);
104 } else {
105 // We now call Identity to try to simplify the cloned node.
106 // Note that some Identity methods call phase->type(this).
107 // Make sure that the type array is big enough for
108 // our new node, even though we may throw the node away.
109 // (Note: This tweaking with igvn only works because x is a new node.)
110 _igvn.set_type(x, t);
111 // If x is a TypeNode, capture any more-precise type permanently into Node
112 // otherwise it will be not updated during igvn->transform since
113 // igvn->type(x) is set to x->Value() already.
114 x->raise_bottom_type(t);
115 if (x->Opcode() != Op_ShenandoahWriteBarrier) {
116 Node *y = x->Identity(&_igvn);
117 if (y != x) {
118 wins++;
119 x = y;
120 } else {
121 y = _igvn.hash_find(x);
122 if (y) {
123 wins++;
124 x = y;
125 } else {
126 // Else x is a new node we are keeping
127 // We do not need register_new_node_with_optimizer
128 // because set_type has already been called.
129 _igvn._worklist.push(x);
130 }
131 }
132 } else {
133 _igvn._worklist.push(x);
134 }
135 }
136 if (x != the_clone && the_clone != NULL)
137 _igvn.remove_dead_node(the_clone);
138 phi->set_req( i, x );
139 }
140 // Too few wins?
141 if (wins <= policy) {
142 _igvn.remove_dead_node(phi);
143 return NULL;
144 }
145
146 // Record Phi
147 register_new_node( phi, region );
148
149 for (uint i2 = 1; i2 < phi->req(); i2++) {
150 Node *x = phi->in(i2);
151 // If we commoned up the cloned 'x' with another existing Node,
152 // the existing Node picks up a new use. We need to make the
153 // existing Node occur higher up so it dominates its uses.
154 Node *old_ctrl;
155 IdealLoopTree *old_loop;
156
157 if (x->is_Con()) {
158 // Constant's control is always root.
159 set_ctrl(x, C->root());
160 continue;
161 }
162 // The occasional new node
163 if (x->_idx >= old_unique) { // Found a new, unplaced node?
164 old_ctrl = NULL;
165 old_loop = NULL; // Not in any prior loop
166 } else {
167 old_ctrl = get_ctrl(x);
168 old_loop = get_loop(old_ctrl); // Get prior loop
169 }
170 // New late point must dominate new use
171 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
172 if (new_ctrl == old_ctrl) // Nothing is changed
173 continue;
174
175 IdealLoopTree *new_loop = get_loop(new_ctrl);
176
177 // Don't move x into a loop if its uses are
178 // outside of loop. Otherwise x will be cloned
179 // for each use outside of this loop.
180 IdealLoopTree *use_loop = get_loop(region);
181 if (!new_loop->is_member(use_loop) &&
182 (old_loop == NULL || !new_loop->is_member(old_loop))) {
183 // Take early control, later control will be recalculated
184 // during next iteration of loop optimizations.
185 new_ctrl = get_early_ctrl(x);
186 new_loop = get_loop(new_ctrl);
187 }
188 // Set new location
189 set_ctrl(x, new_ctrl);
190 // If changing loop bodies, see if we need to collect into new body
191 if (old_loop != new_loop) {
192 if (old_loop && !old_loop->_child)
193 old_loop->_body.yank(x);
194 if (!new_loop->_child)
195 new_loop->_body.push(x); // Collect body info
196 }
197 }
198
199 return phi;
200 }
201
202 /**
203 * When splitting a Shenandoah write barrier through a phi, we
204 * can not replace the write-barrier input of the ShenandoahWBMemProj
205 * with the phi. We must also split the ShenandoahWBMemProj through the
206 * phi and generate a new memory phi for it.
207 */
208 void PhaseIdealLoop::split_mem_thru_phi(Node* n, Node* r, Node* phi) {
209 if (n->Opcode() == Op_ShenandoahWriteBarrier) {
210 if (n->has_out_with(Op_ShenandoahWBMemProj)) {
211 Node* old_mem_phi = n->in(ShenandoahBarrierNode::Memory);
212 assert(r->is_Region(), "need region to control phi");
213 assert(phi->is_Phi(), "expect phi");
214 Node* memphi = PhiNode::make(r, old_mem_phi, Type::MEMORY, C->alias_type(n->adr_type())->adr_type());
215 for (uint i = 1; i < r->req(); i++) {
216 Node* wb = phi->in(i);
217 if (wb->Opcode() == Op_ShenandoahWriteBarrier) {
218 // assert(! wb->has_out_with(Op_ShenandoahWBMemProj), "new clone does not have mem proj");
219 Node* new_proj = new ShenandoahWBMemProjNode(wb);
220 register_new_node(new_proj, r->in(i));
221 memphi->set_req(i, new_proj);
222 } else {
223 if (old_mem_phi->is_Phi() && old_mem_phi->in(0) == r) {
224 memphi->set_req(i, old_mem_phi->in(i));
225 }
226 }
227 }
228 register_new_node(memphi, r);
229 Node* old_mem_out = n->find_out_with(Op_ShenandoahWBMemProj);
230 while (old_mem_out != NULL) {
231 assert(old_mem_out != NULL, "expect memory projection");
232 _igvn.replace_node(old_mem_out, memphi);
233 old_mem_out = n->find_out_with(Op_ShenandoahWBMemProj);
234 }
235 }
236 assert(! n->has_out_with(Op_ShenandoahWBMemProj), "no more memory outs");
237 }
238 }
239
240 //------------------------------dominated_by------------------------------------
241 // Replace the dominated test with an obvious true or false. Place it on the
242 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
243 // live path up to the dominating control.
244 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) {
245 #ifndef PRODUCT
246 if (VerifyLoopOptimizations && PrintOpto) tty->print_cr("dominating test");
247 #endif
248
249
250 // prevdom is the dominating projection of the dominating test.
251 assert( iff->is_If(), "" );
252 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
253 int pop = prevdom->Opcode();
254 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
255 if (flip) {
256 if (pop == Op_IfTrue)
257 pop = Op_IfFalse;
258 else
259 pop = Op_IfTrue;
260 }
261 // 'con' is set to true or false to kill the dominated test.
262 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
263 set_ctrl(con, C->root()); // Constant gets a new use
264 // Hack the dominated test
265 _igvn.replace_input_of(iff, 1, con);
266
267 // If I dont have a reachable TRUE and FALSE path following the IfNode then
268 // I can assume this path reaches an infinite loop. In this case it's not
269 // important to optimize the data Nodes - either the whole compilation will
270 // be tossed or this path (and all data Nodes) will go dead.
271 if (iff->outcnt() != 2) return;
272
273 // Make control-dependent data Nodes on the live path (path that will remain
274 // once the dominated IF is removed) become control-dependent on the
275 // dominating projection.
276 Node* dp = iff->as_If()->proj_out(pop == Op_IfTrue);
277
278 // Loop predicates may have depending checks which should not
279 // be skipped. For example, range check predicate has two checks
280 // for lower and upper bounds.
281 if (dp == NULL)
282 return;
283
284 ProjNode* dp_proj = dp->as_Proj();
285 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj();
286 if (exclude_loop_predicate &&
287 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL ||
288 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) {
289 // If this is a range check (IfNode::is_range_check), do not
290 // reorder because Compile::allow_range_check_smearing might have
291 // changed the check.
292 return; // Let IGVN transformation change control dependence.
293 }
294
295 IdealLoopTree *old_loop = get_loop(dp);
296
297 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
298 Node* cd = dp->fast_out(i); // Control-dependent node
299 if (cd->depends_only_on_test()) {
300 assert(cd->in(0) == dp, "");
301 _igvn.replace_input_of(cd, 0, prevdom);
302 set_early_ctrl(cd);
303 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
304 if (old_loop != new_loop) {
305 if (!old_loop->_child) old_loop->_body.yank(cd);
306 if (!new_loop->_child) new_loop->_body.push(cd);
307 }
308 --i;
309 --imax;
310 }
311 }
312 }
313
314 //------------------------------has_local_phi_input----------------------------
315 // Return TRUE if 'n' has Phi inputs from its local block and no other
316 // block-local inputs (all non-local-phi inputs come from earlier blocks)
317 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
318 Node *n_ctrl = get_ctrl(n);
319 // See if some inputs come from a Phi in this block, or from before
320 // this block.
321 uint i;
322 for( i = 1; i < n->req(); i++ ) {
323 Node *phi = n->in(i);
324 if( phi->is_Phi() && phi->in(0) == n_ctrl )
325 break;
326 }
327 if( i >= n->req() )
328 return NULL; // No Phi inputs; nowhere to clone thru
329
330 // Check for inputs created between 'n' and the Phi input. These
331 // must split as well; they have already been given the chance
332 // (courtesy of a post-order visit) and since they did not we must
333 // recover the 'cost' of splitting them by being very profitable
334 // when splitting 'n'. Since this is unlikely we simply give up.
335 for( i = 1; i < n->req(); i++ ) {
336 Node *m = n->in(i);
337 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
338 // We allow the special case of AddP's with no local inputs.
339 // This allows us to split-up address expressions.
340 if (m->is_AddP() &&
341 get_ctrl(m->in(2)) != n_ctrl &&
342 get_ctrl(m->in(3)) != n_ctrl) {
343 // Move the AddP up to dominating point
344 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl)));
345 continue;
346 }
347 return NULL;
348 }
349 }
350
351 return n_ctrl;
352 }
353
354 //------------------------------remix_address_expressions----------------------
355 // Rework addressing expressions to get the most loop-invariant stuff
356 // moved out. We'd like to do all associative operators, but it's especially
357 // important (common) to do address expressions.
358 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
359 if (!has_ctrl(n)) return NULL;
360 Node *n_ctrl = get_ctrl(n);
361 IdealLoopTree *n_loop = get_loop(n_ctrl);
362
363 // See if 'n' mixes loop-varying and loop-invariant inputs and
364 // itself is loop-varying.
365
366 // Only interested in binary ops (and AddP)
367 if( n->req() < 3 || n->req() > 4 ) return NULL;
368
369 Node *n1_ctrl = get_ctrl(n->in( 1));
370 Node *n2_ctrl = get_ctrl(n->in( 2));
371 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
372 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
373 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
374 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
375
376 // Does one of my inputs spin in a tighter loop than self?
377 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
378 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
379 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
380 return NULL; // Leave well enough alone
381
382 // Is at least one of my inputs loop-invariant?
383 if( n1_loop == n_loop &&
384 n2_loop == n_loop &&
385 n3_loop == n_loop )
386 return NULL; // No loop-invariant inputs
387
388
389 int n_op = n->Opcode();
390
391 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
392 if( n_op == Op_LShiftI ) {
393 // Scale is loop invariant
394 Node *scale = n->in(2);
395 Node *scale_ctrl = get_ctrl(scale);
396 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
397 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
398 return NULL;
399 const TypeInt *scale_t = scale->bottom_type()->isa_int();
400 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
401 return NULL; // Dont bother with byte/short masking
402 // Add must vary with loop (else shift would be loop-invariant)
403 Node *add = n->in(1);
404 Node *add_ctrl = get_ctrl(add);
405 IdealLoopTree *add_loop = get_loop(add_ctrl);
406 //assert( n_loop == add_loop, "" );
407 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
408
409 // Convert I-V into I+ (0-V); same for V-I
410 if( add->Opcode() == Op_SubI &&
411 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
412 Node *zero = _igvn.intcon(0);
413 set_ctrl(zero, C->root());
414 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) );
415 register_new_node( neg, get_ctrl(add->in(2) ) );
416 add = new AddINode( add->in(1), neg );
417 register_new_node( add, add_ctrl );
418 }
419 if( add->Opcode() != Op_AddI ) return NULL;
420 // See if one add input is loop invariant
421 Node *add_var = add->in(1);
422 Node *add_var_ctrl = get_ctrl(add_var);
423 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
424 Node *add_invar = add->in(2);
425 Node *add_invar_ctrl = get_ctrl(add_invar);
426 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
427 if( add_var_loop == n_loop ) {
428 } else if( add_invar_loop == n_loop ) {
429 // Swap to find the invariant part
430 add_invar = add_var;
431 add_invar_ctrl = add_var_ctrl;
432 add_invar_loop = add_var_loop;
433 add_var = add->in(2);
434 Node *add_var_ctrl = get_ctrl(add_var);
435 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
436 } else // Else neither input is loop invariant
437 return NULL;
438 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
439 return NULL; // No invariant part of the add?
440
441 // Yes! Reshape address expression!
442 Node *inv_scale = new LShiftINode( add_invar, scale );
443 Node *inv_scale_ctrl =
444 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
445 add_invar_ctrl : scale_ctrl;
446 register_new_node( inv_scale, inv_scale_ctrl );
447 Node *var_scale = new LShiftINode( add_var, scale );
448 register_new_node( var_scale, n_ctrl );
449 Node *var_add = new AddINode( var_scale, inv_scale );
450 register_new_node( var_add, n_ctrl );
451 _igvn.replace_node( n, var_add );
452 return var_add;
453 }
454
455 // Replace (I+V) with (V+I)
456 if( n_op == Op_AddI ||
457 n_op == Op_AddL ||
458 n_op == Op_AddF ||
459 n_op == Op_AddD ||
460 n_op == Op_MulI ||
461 n_op == Op_MulL ||
462 n_op == Op_MulF ||
463 n_op == Op_MulD ) {
464 if( n2_loop == n_loop ) {
465 assert( n1_loop != n_loop, "" );
466 n->swap_edges(1, 2);
467 }
468 }
469
470 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
471 // but not if I2 is a constant.
472 if( n_op == Op_AddP ) {
473 if( n2_loop == n_loop && n3_loop != n_loop ) {
474 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
475 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
476 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
477 IdealLoopTree *n22loop = get_loop( n22_ctrl );
478 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
479 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
480 n23_loop == n_loop ) {
481 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
482 // Stuff new AddP in the loop preheader
483 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
484 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) );
485 register_new_node( add2, n_ctrl );
486 _igvn.replace_node( n, add2 );
487 return add2;
488 }
489 }
490 }
491
492 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
493 if( n2_loop != n_loop && n3_loop == n_loop ) {
494 if( n->in(3)->Opcode() == Op_AddI ) {
495 Node *V = n->in(3)->in(1);
496 Node *I = n->in(3)->in(2);
497 if( is_member(n_loop,get_ctrl(V)) ) {
498 } else {
499 Node *tmp = V; V = I; I = tmp;
500 }
501 if( !is_member(n_loop,get_ctrl(I)) ) {
502 Node *add1 = new AddPNode( n->in(1), n->in(2), I );
503 // Stuff new AddP in the loop preheader
504 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
505 Node *add2 = new AddPNode( n->in(1), add1, V );
506 register_new_node( add2, n_ctrl );
507 _igvn.replace_node( n, add2 );
508 return add2;
509 }
510 }
511 }
512 }
513
514 return NULL;
515 }
516
517 //------------------------------conditional_move-------------------------------
518 // Attempt to replace a Phi with a conditional move. We have some pretty
519 // strict profitability requirements. All Phis at the merge point must
520 // be converted, so we can remove the control flow. We need to limit the
521 // number of c-moves to a small handful. All code that was in the side-arms
522 // of the CFG diamond is now speculatively executed. This code has to be
523 // "cheap enough". We are pretty much limited to CFG diamonds that merge
524 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
525 Node *PhaseIdealLoop::conditional_move( Node *region ) {
526
527 assert(region->is_Region(), "sanity check");
528 if (region->req() != 3) return NULL;
529
530 // Check for CFG diamond
531 Node *lp = region->in(1);
532 Node *rp = region->in(2);
533 if (!lp || !rp) return NULL;
534 Node *lp_c = lp->in(0);
535 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
536 IfNode *iff = lp_c->as_If();
537
538 // Check for ops pinned in an arm of the diamond.
539 // Can't remove the control flow in this case
540 if (lp->outcnt() > 1) return NULL;
541 if (rp->outcnt() > 1) return NULL;
542
543 IdealLoopTree* r_loop = get_loop(region);
544 assert(r_loop == get_loop(iff), "sanity");
545 // Always convert to CMOVE if all results are used only outside this loop.
546 bool used_inside_loop = (r_loop == _ltree_root);
547
548 // Check profitability
549 int cost = 0;
550 int phis = 0;
551 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
552 Node *out = region->fast_out(i);
553 if (!out->is_Phi()) continue; // Ignore other control edges, etc
554 phis++;
555 PhiNode* phi = out->as_Phi();
556 BasicType bt = phi->type()->basic_type();
557 switch (bt) {
558 case T_FLOAT:
559 case T_DOUBLE: {
560 cost += Matcher::float_cmove_cost(); // Could be very expensive
561 break;
562 }
563 case T_LONG: {
564 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
565 }
566 case T_INT: // These all CMOV fine
567 case T_ADDRESS: { // (RawPtr)
568 cost++;
569 break;
570 }
571 case T_NARROWOOP: // Fall through
572 case T_OBJECT: { // Base oops are OK, but not derived oops
573 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
574 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
575 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
576 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
577 // have a Phi for the base here that we convert to a CMOVE all is well
578 // and good. But if the base is dead, we'll not make a CMOVE. Later
579 // the allocator will have to produce a base by creating a CMOVE of the
580 // relevant bases. This puts the allocator in the business of
581 // manufacturing expensive instructions, generally a bad plan.
582 // Just Say No to Conditionally-Moved Derived Pointers.
583 if (tp && tp->offset() != 0)
584 return NULL;
585 cost++;
586 break;
587 }
588 default:
589 return NULL; // In particular, can't do memory or I/O
590 }
591 // Add in cost any speculative ops
592 for (uint j = 1; j < region->req(); j++) {
593 Node *proj = region->in(j);
594 Node *inp = phi->in(j);
595 if (get_ctrl(inp) == proj) { // Found local op
596 cost++;
597 // Check for a chain of dependent ops; these will all become
598 // speculative in a CMOV.
599 for (uint k = 1; k < inp->req(); k++)
600 if (get_ctrl(inp->in(k)) == proj)
601 cost += ConditionalMoveLimit; // Too much speculative goo
602 }
603 }
604 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
605 // This will likely Split-If, a higher-payoff operation.
606 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
607 Node* use = phi->fast_out(k);
608 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
609 cost += ConditionalMoveLimit;
610 // Is there a use inside the loop?
611 // Note: check only basic types since CMoveP is pinned.
612 if (!used_inside_loop && is_java_primitive(bt)) {
613 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
614 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
615 used_inside_loop = true;
616 }
617 }
618 }
619 }
620 Node* bol = iff->in(1);
621 assert(bol->Opcode() == Op_Bool, "");
622 int cmp_op = bol->in(1)->Opcode();
623 // It is expensive to generate flags from a float compare.
624 // Avoid duplicated float compare.
625 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
626
627 float infrequent_prob = PROB_UNLIKELY_MAG(3);
628 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
629 if (used_inside_loop) {
630 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
631
632 // BlockLayoutByFrequency optimization moves infrequent branch
633 // from hot path. No point in CMOV'ing in such case (110 is used
634 // instead of 100 to take into account not exactness of float value).
635 if (BlockLayoutByFrequency) {
636 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
637 }
638 }
639 // Check for highly predictable branch. No point in CMOV'ing if
640 // we are going to predict accurately all the time.
641 if (iff->_prob < infrequent_prob ||
642 iff->_prob > (1.0f - infrequent_prob))
643 return NULL;
644
645 // --------------
646 // Now replace all Phis with CMOV's
647 Node *cmov_ctrl = iff->in(0);
648 uint flip = (lp->Opcode() == Op_IfTrue);
649 while (1) {
650 PhiNode* phi = NULL;
651 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
652 Node *out = region->fast_out(i);
653 if (out->is_Phi()) {
654 phi = out->as_Phi();
655 break;
656 }
657 }
658 if (phi == NULL) break;
659 #ifndef PRODUCT
660 if (PrintOpto && VerifyLoopOptimizations) tty->print_cr("CMOV");
661 #endif
662 // Move speculative ops
663 for (uint j = 1; j < region->req(); j++) {
664 Node *proj = region->in(j);
665 Node *inp = phi->in(j);
666 if (get_ctrl(inp) == proj) { // Found local op
667 #ifndef PRODUCT
668 if (PrintOpto && VerifyLoopOptimizations) {
669 tty->print(" speculate: ");
670 inp->dump();
671 }
672 #endif
673 set_ctrl(inp, cmov_ctrl);
674 }
675 }
676 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
677 register_new_node( cmov, cmov_ctrl );
678 _igvn.replace_node( phi, cmov );
679 #ifndef PRODUCT
680 if (TraceLoopOpts) {
681 tty->print("CMOV ");
682 r_loop->dump_head();
683 if (Verbose) {
684 bol->in(1)->dump(1);
685 cmov->dump(1);
686 }
687 }
688 if (VerifyLoopOptimizations) verify();
689 #endif
690 }
691
692 // The useless CFG diamond will fold up later; see the optimization in
693 // RegionNode::Ideal.
694 _igvn._worklist.push(region);
695
696 return iff->in(1);
697 }
698
699 #ifdef ASSERT
700 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
701 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
702 Node* u = m->fast_out(i);
703 if (u->is_CFG()) {
704 if (u->Opcode() == Op_NeverBranch) {
705 u = ((NeverBranchNode*)u)->proj_out(0);
706 enqueue_cfg_uses(u, wq);
707 } else {
708 wq.push(u);
709 }
710 }
711 }
712 }
713 #endif
714
715 // Try moving a store out of a loop, right before the loop
716 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
717 // Store has to be first in the loop body
718 IdealLoopTree *n_loop = get_loop(n_ctrl);
719 if (n->is_Store() && n_loop != _ltree_root && n_loop->is_loop() && n->in(0) != NULL) {
720 Node* address = n->in(MemNode::Address);
721 Node* value = n->in(MemNode::ValueIn);
722 Node* mem = n->in(MemNode::Memory);
723 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
724 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
725
726 // - address and value must be loop invariant
727 // - memory must be a memory Phi for the loop
728 // - Store must be the only store on this memory slice in the
729 // loop: if there's another store following this one then value
730 // written at iteration i by the second store could be overwritten
731 // at iteration i+n by the first store: it's not safe to move the
732 // first store out of the loop
733 // - nothing must observe the Phi memory: it guarantees no read
734 // before the store and no early exit out of the loop
735 // With those conditions, we are also guaranteed the store post
736 // dominates the loop head. Otherwise there would be extra Phi
737 // involved between the loop's Phi and the store.
738
739 if (!n_loop->is_member(address_loop) &&
740 !n_loop->is_member(value_loop) &&
741 mem->is_Phi() && mem->in(0) == n_loop->_head &&
742 mem->outcnt() == 1 &&
743 mem->in(LoopNode::LoopBackControl) == n) {
744
745 #ifdef ASSERT
746 // Verify that store's control does post dominate loop entry and
747 // that there's no early exit of the loop before the store.
748 bool ctrl_ok = false;
749 {
750 // Follow control from loop head until n, we exit the loop or
751 // we reach the tail
752 ResourceMark rm;
753 Unique_Node_List wq;
754 wq.push(n_loop->_head);
755 assert(n_loop->_tail != NULL, "need a tail");
756 for (uint next = 0; next < wq.size(); ++next) {
757 Node *m = wq.at(next);
758 if (m == n->in(0)) {
759 ctrl_ok = true;
760 continue;
761 }
762 assert(!has_ctrl(m), "should be CFG");
763 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
764 ctrl_ok = false;
765 break;
766 }
767 enqueue_cfg_uses(m, wq);
768 }
769 }
770 assert(ctrl_ok, "bad control");
771 #endif
772
773 // move the Store
774 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
775 _igvn.replace_input_of(n, 0, n_loop->_head->in(LoopNode::EntryControl));
776 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
777 // Disconnect the phi now. An empty phi can confuse other
778 // optimizations in this pass of loop opts.
779 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
780 n_loop->_body.yank(mem);
781
782 IdealLoopTree* new_loop = get_loop(n->in(0));
783 set_ctrl_and_loop(n, n->in(0));
784
785 return n;
786 }
787 }
788 return NULL;
789 }
790
791 // Try moving a store out of a loop, right after the loop
792 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
793 if (n->is_Store() && n->in(0) != NULL) {
794 Node *n_ctrl = get_ctrl(n);
795 IdealLoopTree *n_loop = get_loop(n_ctrl);
796 // Store must be in a loop
797 if (n_loop != _ltree_root && !n_loop->_irreducible) {
798 Node* address = n->in(MemNode::Address);
799 Node* value = n->in(MemNode::ValueIn);
800 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
801 // address must be loop invariant
802 if (!n_loop->is_member(address_loop)) {
803 // Store must be last on this memory slice in the loop and
804 // nothing in the loop must observe it
805 Node* phi = NULL;
806 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
807 Node* u = n->fast_out(i);
808 if (has_ctrl(u)) { // control use?
809 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
810 if (!n_loop->is_member(u_loop)) {
811 continue;
812 }
813 if (u->is_Phi() && u->in(0) == n_loop->_head) {
814 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
815 assert(phi == NULL, "already found");
816 phi = u;
817 continue;
818 }
819 }
820 phi = NULL;
821 break;
822 }
823 if (phi != NULL) {
824 // Nothing in the loop before the store (next iteration)
825 // must observe the stored value
826 bool mem_ok = true;
827 {
828 ResourceMark rm;
829 Unique_Node_List wq;
830 wq.push(phi);
831 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
832 Node *m = wq.at(next);
833 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
834 Node* u = m->fast_out(i);
835 if (u->is_Store() || u->is_Phi()) {
836 if (u != n) {
837 wq.push(u);
838 mem_ok = (wq.size() <= 10);
839 }
840 } else {
841 mem_ok = false;
842 break;
843 }
844 }
845 }
846 }
847 if (mem_ok) {
848 // Move the Store out of the loop creating clones along
849 // all paths out of the loop that observe the stored value
850 _igvn.rehash_node_delayed(phi);
851 int count = phi->replace_edge(n, n->in(MemNode::Memory));
852 assert(count > 0, "inconsistent phi");
853 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
854 Node* u = n->fast_out(i);
855 Node* c = get_ctrl(u);
856
857 if (u->is_Phi()) {
858 c = u->in(0)->in(u->find_edge(n));
859 }
860 IdealLoopTree *u_loop = get_loop(c);
861 assert (!n_loop->is_member(u_loop), "only the phi should have been a use in the loop");
862 while(true) {
863 Node* next_c = find_non_split_ctrl(idom(c));
864 if (n_loop->is_member(get_loop(next_c))) {
865 break;
866 }
867 c = next_c;
868 }
869
870 Node* st = n->clone();
871 st->set_req(0, c);
872 _igvn.register_new_node_with_optimizer(st);
873
874 set_ctrl(st, c);
875 IdealLoopTree* new_loop = get_loop(c);
876 assert(new_loop != n_loop, "should be moved out of loop");
877 if (new_loop->_child == NULL) new_loop->_body.push(st);
878
879 _igvn.replace_input_of(u, u->find_edge(n), st);
880 --imax;
881 --i;
882 }
883
884
885 assert(n->outcnt() == 0, "all uses should be gone");
886 _igvn.replace_input_of(n, MemNode::Memory, C->top());
887 // Disconnect the phi now. An empty phi can confuse other
888 // optimizations in this pass of loop opts..
889 if (phi->in(LoopNode::LoopBackControl) == phi) {
890 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
891 n_loop->_body.yank(phi);
892 }
893 }
894 }
895 }
896 }
897 }
898 }
899
900 //------------------------------split_if_with_blocks_pre-----------------------
901 // Do the real work in a non-recursive function. Data nodes want to be
902 // cloned in the pre-order so they can feed each other nicely.
903 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
904 // Cloning these guys is unlikely to win
905 int n_op = n->Opcode();
906 if( n_op == Op_MergeMem ) return n;
907 if( n->is_Proj() ) return n;
908 // Do not clone-up CmpFXXX variations, as these are always
909 // followed by a CmpI
910 if( n->is_Cmp() ) return n;
911 // Attempt to use a conditional move instead of a phi/branch
912 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
913 Node *cmov = conditional_move( n );
914 if( cmov ) return cmov;
915 }
916 if( n->is_CFG() || n->is_LoadStore() )
917 return n;
918 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
919 n_op == Op_Opaque2 ) {
920 if( !C->major_progress() ) // If chance of no more loop opts...
921 _igvn._worklist.push(n); // maybe we'll remove them
922 return n;
923 }
924
925 if( n->is_Con() ) return n; // No cloning for Con nodes
926
927 Node *n_ctrl = get_ctrl(n);
928 if( !n_ctrl ) return n; // Dead node
929
930 Node* res = try_move_store_before_loop(n, n_ctrl);
931 if (res != NULL) {
932 return n;
933 }
934
935 // Attempt to remix address expressions for loop invariants
936 Node *m = remix_address_expressions( n );
937 if( m ) return m;
938
939 // Determine if the Node has inputs from some local Phi.
940 // Returns the block to clone thru.
941 Node *n_blk = has_local_phi_input( n );
942 if( !n_blk ) return n;
943
944 // Do not clone the trip counter through on a CountedLoop
945 // (messes up the canonical shape).
946 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
947
948 // Check for having no control input; not pinned. Allow
949 // dominating control.
950 if (n->in(0)) {
951 Node *dom = idom(n_blk);
952 if (dom_lca(n->in(0), dom) != n->in(0)) {
953 return n;
954 }
955 }
956 // Policy: when is it profitable. You must get more wins than
957 // policy before it is considered profitable. Policy is usually 0,
958 // so 1 win is considered profitable. Big merges will require big
959 // cloning, so get a larger policy.
960 int policy = n_blk->req() >> 2;
961
962 // If the loop is a candidate for range check elimination,
963 // delay splitting through it's phi until a later loop optimization
964 if (n_blk->is_CountedLoop()) {
965 IdealLoopTree *lp = get_loop(n_blk);
966 if (lp && lp->_rce_candidate) {
967 return n;
968 }
969 }
970
971 // Use same limit as split_if_with_blocks_post
972 if( C->live_nodes() > 35000 ) return n; // Method too big
973
974 // Split 'n' through the merge point if it is profitable
975 Node *phi = split_thru_phi( n, n_blk, policy );
976 if (!phi) return n;
977
978 // Found a Phi to split thru!
979 // Replace 'n' with the new phi
980 split_mem_thru_phi(n, n_blk, phi);
981 _igvn.replace_node( n, phi );
982 // Moved a load around the loop, 'en-registering' something.
983 if (n_blk->is_Loop() && n->is_Load() &&
984 !phi->in(LoopNode::LoopBackControl)->is_Load())
985 C->set_major_progress();
986
987 // Moved a barrier around the loop, 'en-registering' something.
988 if (n_blk->is_Loop() && n->is_ShenandoahBarrier() &&
989 !phi->in(LoopNode::LoopBackControl)->is_ShenandoahBarrier())
990 C->set_major_progress();
991
992 return phi;
993 }
994
995 static bool merge_point_too_heavy(Compile* C, Node* region) {
996 // Bail out if the region and its phis have too many users.
997 int weight = 0;
998 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
999 weight += region->fast_out(i)->outcnt();
1000 }
1001 int nodes_left = C->max_node_limit() - C->live_nodes();
1002 if (weight * 8 > nodes_left) {
1003 #ifndef PRODUCT
1004 if (PrintOpto)
1005 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1006 #endif
1007 return true;
1008 } else {
1009 return false;
1010 }
1011 }
1012
1013 static bool merge_point_safe(Node* region) {
1014 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1015 // having a PhiNode input. This sidesteps the dangerous case where the split
1016 // ConvI2LNode may become TOP if the input Value() does not
1017 // overlap the ConvI2L range, leaving a node which may not dominate its
1018 // uses.
1019 // A better fix for this problem can be found in the BugTraq entry, but
1020 // expediency for Mantis demands this hack.
1021 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop
1022 // split_if_with_blocks from splitting a block because we could not move around
1023 // the FastLockNode.
1024 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1025 Node* n = region->fast_out(i);
1026 if (n->is_Phi()) {
1027 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1028 Node* m = n->fast_out(j);
1029 if (m->is_FastLock())
1030 return false;
1031 #ifdef _LP64
1032 if (m->Opcode() == Op_ConvI2L)
1033 return false;
1034 #endif
1035 }
1036 }
1037 }
1038 return true;
1039 }
1040
1041
1042 //------------------------------place_near_use---------------------------------
1043 // Place some computation next to use but not inside inner loops.
1044 // For inner loop uses move it to the preheader area.
1045 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const {
1046 IdealLoopTree *u_loop = get_loop( useblock );
1047 return (u_loop->_irreducible || u_loop->_child)
1048 ? useblock
1049 : u_loop->_head->in(LoopNode::EntryControl);
1050 }
1051
1052
1053 //------------------------------split_if_with_blocks_post----------------------
1054 // Do the real work in a non-recursive function. CFG hackery wants to be
1055 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1056 // info.
1057 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) {
1058
1059 // Cloning Cmp through Phi's involves the split-if transform.
1060 // FastLock is not used by an If
1061 if( n->is_Cmp() && !n->is_FastLock() ) {
1062 if( C->live_nodes() > 35000 ) return; // Method too big
1063
1064 // Do not do 'split-if' if irreducible loops are present.
1065 if( _has_irreducible_loops )
1066 return;
1067
1068 Node *n_ctrl = get_ctrl(n);
1069 // Determine if the Node has inputs from some local Phi.
1070 // Returns the block to clone thru.
1071 Node *n_blk = has_local_phi_input( n );
1072 if( n_blk != n_ctrl ) return;
1073
1074 if( merge_point_too_heavy(C, n_ctrl) )
1075 return;
1076
1077 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare?
1078 Node *bol = n->unique_out();
1079 assert( bol->is_Bool(), "expect a bool here" );
1080 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare?
1081 Node *iff = bol->unique_out();
1082
1083 // Check some safety conditions
1084 if( iff->is_If() ) { // Classic split-if?
1085 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if
1086 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1087 // Can't split CMove with different control edge.
1088 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) return;
1089 if( get_ctrl(iff->in(2)) == n_ctrl ||
1090 get_ctrl(iff->in(3)) == n_ctrl )
1091 return; // Inputs not yet split-up
1092 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) {
1093 return; // Loop-invar test gates loop-varying CMOVE
1094 }
1095 } else {
1096 return; // some other kind of node, such as an Allocate
1097 }
1098
1099 // Do not do 'split-if' if some paths are dead. First do dead code
1100 // elimination and then see if its still profitable.
1101 for( uint i = 1; i < n_ctrl->req(); i++ )
1102 if( n_ctrl->in(i) == C->top() )
1103 return;
1104
1105 // When is split-if profitable? Every 'win' on means some control flow
1106 // goes dead, so it's almost always a win.
1107 int policy = 0;
1108 // If trying to do a 'Split-If' at the loop head, it is only
1109 // profitable if the cmp folds up on BOTH paths. Otherwise we
1110 // risk peeling a loop forever.
1111
1112 // CNC - Disabled for now. Requires careful handling of loop
1113 // body selection for the cloned code. Also, make sure we check
1114 // for any input path not being in the same loop as n_ctrl. For
1115 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1116 // because the alternative loop entry points won't be converted
1117 // into LoopNodes.
1118 IdealLoopTree *n_loop = get_loop(n_ctrl);
1119 for( uint j = 1; j < n_ctrl->req(); j++ )
1120 if( get_loop(n_ctrl->in(j)) != n_loop )
1121 return;
1122
1123 // Check for safety of the merge point.
1124 if( !merge_point_safe(n_ctrl) ) {
1125 return;
1126 }
1127
1128 // Split compare 'n' through the merge point if it is profitable
1129 Node *phi = split_thru_phi( n, n_ctrl, policy );
1130 if( !phi ) return;
1131
1132 // Found a Phi to split thru!
1133 // Replace 'n' with the new phi
1134 _igvn.replace_node( n, phi );
1135
1136 // Now split the bool up thru the phi
1137 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 );
1138 guarantee(bolphi != NULL, "null boolean phi node");
1139
1140 _igvn.replace_node( bol, bolphi );
1141 assert( iff->in(1) == bolphi, "" );
1142
1143 if( bolphi->Value(&_igvn)->singleton() )
1144 return;
1145
1146 // Conditional-move? Must split up now
1147 if( !iff->is_If() ) {
1148 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 );
1149 _igvn.replace_node( iff, cmovphi );
1150 return;
1151 }
1152
1153 // Now split the IF
1154 do_split_if( iff );
1155 return;
1156 }
1157
1158 // Check for an IF ready to split; one that has its
1159 // condition codes input coming from a Phi at the block start.
1160 int n_op = n->Opcode();
1161
1162 // Check for an IF being dominated by another IF same test
1163 if (n_op == Op_If) {
1164 Node *bol = n->in(1);
1165 uint max = bol->outcnt();
1166 // Check for same test used more than once?
1167 if (max > 1 && bol->is_Bool()) {
1168 // Search up IDOMs to see if this IF is dominated.
1169 Node *cutoff = get_ctrl(bol);
1170
1171 // Now search up IDOMs till cutoff, looking for a dominating test
1172 Node *prevdom = n;
1173 Node *dom = idom(prevdom);
1174 while (dom != cutoff) {
1175 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) {
1176 // Replace the dominated test with an obvious true or false.
1177 // Place it on the IGVN worklist for later cleanup.
1178 C->set_major_progress();
1179 dominated_by(prevdom, n, false, true);
1180 #ifndef PRODUCT
1181 if( VerifyLoopOptimizations ) verify();
1182 #endif
1183 return;
1184 }
1185 prevdom = dom;
1186 dom = idom(prevdom);
1187 }
1188 }
1189 }
1190
1191 // See if a shared loop-varying computation has no loop-varying uses.
1192 // Happens if something is only used for JVM state in uncommon trap exits,
1193 // like various versions of induction variable+offset. Clone the
1194 // computation per usage to allow it to sink out of the loop.
1195 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
1196 Node *n_ctrl = get_ctrl(n);
1197 IdealLoopTree *n_loop = get_loop(n_ctrl);
1198 if( n_loop != _ltree_root ) {
1199 DUIterator_Fast imax, i = n->fast_outs(imax);
1200 for (; i < imax; i++) {
1201 Node* u = n->fast_out(i);
1202 if( !has_ctrl(u) ) break; // Found control user
1203 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1204 if( u_loop == n_loop ) break; // Found loop-varying use
1205 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
1206 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
1207 }
1208 bool did_break = (i < imax); // Did we break out of the previous loop?
1209 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
1210 Node *late_load_ctrl = NULL;
1211 if (n->is_Load()) {
1212 // If n is a load, get and save the result from get_late_ctrl(),
1213 // to be later used in calculating the control for n's clones.
1214 clear_dom_lca_tags();
1215 late_load_ctrl = get_late_ctrl(n, n_ctrl);
1216 }
1217 // If n is a load, and the late control is the same as the current
1218 // control, then the cloning of n is a pointless exercise, because
1219 // GVN will ensure that we end up where we started.
1220 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
1221 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
1222 Node *u = n->last_out(j); // Clone private computation per use
1223 _igvn.rehash_node_delayed(u);
1224 Node *x = n->clone(); // Clone computation
1225 Node *x_ctrl = NULL;
1226 if( u->is_Phi() ) {
1227 // Replace all uses of normal nodes. Replace Phi uses
1228 // individually, so the separate Nodes can sink down
1229 // different paths.
1230 uint k = 1;
1231 while( u->in(k) != n ) k++;
1232 u->set_req( k, x );
1233 // x goes next to Phi input path
1234 x_ctrl = u->in(0)->in(k);
1235 --j;
1236 } else { // Normal use
1237 // Replace all uses
1238 for( uint k = 0; k < u->req(); k++ ) {
1239 if( u->in(k) == n ) {
1240 u->set_req( k, x );
1241 --j;
1242 }
1243 }
1244 x_ctrl = get_ctrl(u);
1245 }
1246
1247 // Find control for 'x' next to use but not inside inner loops.
1248 // For inner loop uses get the preheader area.
1249 x_ctrl = place_near_use(x_ctrl);
1250
1251 if (n->is_Load()) {
1252 // For loads, add a control edge to a CFG node outside of the loop
1253 // to force them to not combine and return back inside the loop
1254 // during GVN optimization (4641526).
1255 //
1256 // Because we are setting the actual control input, factor in
1257 // the result from get_late_ctrl() so we respect any
1258 // anti-dependences. (6233005).
1259 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
1260
1261 // Don't allow the control input to be a CFG splitting node.
1262 // Such nodes should only have ProjNodes as outs, e.g. IfNode
1263 // should only have IfTrueNode and IfFalseNode (4985384).
1264 x_ctrl = find_non_split_ctrl(x_ctrl);
1265 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1266
1267 x->set_req(0, x_ctrl);
1268 }
1269 register_new_node(x, x_ctrl);
1270
1271 // Some institutional knowledge is needed here: 'x' is
1272 // yanked because if the optimizer runs GVN on it all the
1273 // cloned x's will common up and undo this optimization and
1274 // be forced back in the loop. This is annoying because it
1275 // makes +VerifyOpto report false-positives on progress. I
1276 // tried setting control edges on the x's to force them to
1277 // not combine, but the matching gets worried when it tries
1278 // to fold a StoreP and an AddP together (as part of an
1279 // address expression) and the AddP and StoreP have
1280 // different controls.
1281 if (!x->is_Load() && !x->is_DecodeNarrowPtr() && !x->is_ShenandoahBarrier()) _igvn._worklist.yank(x);
1282 }
1283 _igvn.remove_dead_node(n);
1284 }
1285 }
1286 }
1287 }
1288
1289 try_move_store_after_loop(n);
1290
1291 // Check for Opaque2's who's loop has disappeared - who's input is in the
1292 // same loop nest as their output. Remove 'em, they are no longer useful.
1293 if( n_op == Op_Opaque2 &&
1294 n->in(1) != NULL &&
1295 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1296 _igvn.replace_node( n, n->in(1) );
1297 }
1298 }
1299
1300 //------------------------------split_if_with_blocks---------------------------
1301 // Check for aggressive application of 'split-if' optimization,
1302 // using basic block level info.
1303 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) {
1304 Node *n = C->root();
1305 visited.set(n->_idx); // first, mark node as visited
1306 // Do pre-visit work for root
1307 n = split_if_with_blocks_pre( n );
1308 uint cnt = n->outcnt();
1309 uint i = 0;
1310 while (true) {
1311 // Visit all children
1312 if (i < cnt) {
1313 Node* use = n->raw_out(i);
1314 ++i;
1315 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1316 // Now do pre-visit work for this use
1317 use = split_if_with_blocks_pre( use );
1318 nstack.push(n, i); // Save parent and next use's index.
1319 n = use; // Process all children of current use.
1320 cnt = use->outcnt();
1321 i = 0;
1322 }
1323 }
1324 else {
1325 // All of n's children have been processed, complete post-processing.
1326 if (cnt != 0 && !n->is_Con()) {
1327 assert(has_node(n), "no dead nodes");
1328 split_if_with_blocks_post( n );
1329 }
1330 if (nstack.is_empty()) {
1331 // Finished all nodes on stack.
1332 break;
1333 }
1334 // Get saved parent node and next use's index. Visit the rest of uses.
1335 n = nstack.node();
1336 cnt = n->outcnt();
1337 i = nstack.index();
1338 nstack.pop();
1339 }
1340 }
1341 }
1342
1343
1344 //=============================================================================
1345 //
1346 // C L O N E A L O O P B O D Y
1347 //
1348
1349 //------------------------------clone_iff--------------------------------------
1350 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1351 // "Nearly" because all Nodes have been cloned from the original in the loop,
1352 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1353 // through the Phi recursively, and return a Bool.
1354 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) {
1355
1356 // Convert this Phi into a Phi merging Bools
1357 uint i;
1358 for( i = 1; i < phi->req(); i++ ) {
1359 Node *b = phi->in(i);
1360 if( b->is_Phi() ) {
1361 _igvn.replace_input_of(phi, i, clone_iff( b->as_Phi(), loop ));
1362 } else {
1363 assert( b->is_Bool(), "" );
1364 }
1365 }
1366
1367 Node *sample_bool = phi->in(1);
1368 Node *sample_cmp = sample_bool->in(1);
1369
1370 // Make Phis to merge the Cmp's inputs.
1371 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1372 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1373 for( i = 1; i < phi->req(); i++ ) {
1374 Node *n1 = phi->in(i)->in(1)->in(1);
1375 Node *n2 = phi->in(i)->in(1)->in(2);
1376 phi1->set_req( i, n1 );
1377 phi2->set_req( i, n2 );
1378 phi1->set_type( phi1->type()->meet_speculative(n1->bottom_type()));
1379 phi2->set_type( phi2->type()->meet_speculative(n2->bottom_type()));
1380 }
1381 // See if these Phis have been made before.
1382 // Register with optimizer
1383 Node *hit1 = _igvn.hash_find_insert(phi1);
1384 if( hit1 ) { // Hit, toss just made Phi
1385 _igvn.remove_dead_node(phi1); // Remove new phi
1386 assert( hit1->is_Phi(), "" );
1387 phi1 = (PhiNode*)hit1; // Use existing phi
1388 } else { // Miss
1389 _igvn.register_new_node_with_optimizer(phi1);
1390 }
1391 Node *hit2 = _igvn.hash_find_insert(phi2);
1392 if( hit2 ) { // Hit, toss just made Phi
1393 _igvn.remove_dead_node(phi2); // Remove new phi
1394 assert( hit2->is_Phi(), "" );
1395 phi2 = (PhiNode*)hit2; // Use existing phi
1396 } else { // Miss
1397 _igvn.register_new_node_with_optimizer(phi2);
1398 }
1399 // Register Phis with loop/block info
1400 set_ctrl(phi1, phi->in(0));
1401 set_ctrl(phi2, phi->in(0));
1402 // Make a new Cmp
1403 Node *cmp = sample_cmp->clone();
1404 cmp->set_req( 1, phi1 );
1405 cmp->set_req( 2, phi2 );
1406 _igvn.register_new_node_with_optimizer(cmp);
1407 set_ctrl(cmp, phi->in(0));
1408
1409 // Make a new Bool
1410 Node *b = sample_bool->clone();
1411 b->set_req(1,cmp);
1412 _igvn.register_new_node_with_optimizer(b);
1413 set_ctrl(b, phi->in(0));
1414
1415 assert( b->is_Bool(), "" );
1416 return (BoolNode*)b;
1417 }
1418
1419 //------------------------------clone_bool-------------------------------------
1420 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1421 // "Nearly" because all Nodes have been cloned from the original in the loop,
1422 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1423 // through the Phi recursively, and return a Bool.
1424 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1425 uint i;
1426 // Convert this Phi into a Phi merging Bools
1427 for( i = 1; i < phi->req(); i++ ) {
1428 Node *b = phi->in(i);
1429 if( b->is_Phi() ) {
1430 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1431 } else {
1432 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1433 }
1434 }
1435
1436 Node *sample_cmp = phi->in(1);
1437
1438 // Make Phis to merge the Cmp's inputs.
1439 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1440 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1441 for( uint j = 1; j < phi->req(); j++ ) {
1442 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1443 Node *n1, *n2;
1444 if( cmp_top->is_Cmp() ) {
1445 n1 = cmp_top->in(1);
1446 n2 = cmp_top->in(2);
1447 } else {
1448 n1 = n2 = cmp_top;
1449 }
1450 phi1->set_req( j, n1 );
1451 phi2->set_req( j, n2 );
1452 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1453 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1454 }
1455
1456 // See if these Phis have been made before.
1457 // Register with optimizer
1458 Node *hit1 = _igvn.hash_find_insert(phi1);
1459 if( hit1 ) { // Hit, toss just made Phi
1460 _igvn.remove_dead_node(phi1); // Remove new phi
1461 assert( hit1->is_Phi(), "" );
1462 phi1 = (PhiNode*)hit1; // Use existing phi
1463 } else { // Miss
1464 _igvn.register_new_node_with_optimizer(phi1);
1465 }
1466 Node *hit2 = _igvn.hash_find_insert(phi2);
1467 if( hit2 ) { // Hit, toss just made Phi
1468 _igvn.remove_dead_node(phi2); // Remove new phi
1469 assert( hit2->is_Phi(), "" );
1470 phi2 = (PhiNode*)hit2; // Use existing phi
1471 } else { // Miss
1472 _igvn.register_new_node_with_optimizer(phi2);
1473 }
1474 // Register Phis with loop/block info
1475 set_ctrl(phi1, phi->in(0));
1476 set_ctrl(phi2, phi->in(0));
1477 // Make a new Cmp
1478 Node *cmp = sample_cmp->clone();
1479 cmp->set_req( 1, phi1 );
1480 cmp->set_req( 2, phi2 );
1481 _igvn.register_new_node_with_optimizer(cmp);
1482 set_ctrl(cmp, phi->in(0));
1483
1484 assert( cmp->is_Cmp(), "" );
1485 return (CmpNode*)cmp;
1486 }
1487
1488 //------------------------------sink_use---------------------------------------
1489 // If 'use' was in the loop-exit block, it now needs to be sunk
1490 // below the post-loop merge point.
1491 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1492 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1493 set_ctrl(use, post_loop);
1494 for (DUIterator j = use->outs(); use->has_out(j); j++)
1495 sink_use(use->out(j), post_loop);
1496 }
1497 }
1498
1499 //------------------------------clone_loop-------------------------------------
1500 //
1501 // C L O N E A L O O P B O D Y
1502 //
1503 // This is the basic building block of the loop optimizations. It clones an
1504 // entire loop body. It makes an old_new loop body mapping; with this mapping
1505 // you can find the new-loop equivalent to an old-loop node. All new-loop
1506 // nodes are exactly equal to their old-loop counterparts, all edges are the
1507 // same. All exits from the old-loop now have a RegionNode that merges the
1508 // equivalent new-loop path. This is true even for the normal "loop-exit"
1509 // condition. All uses of loop-invariant old-loop values now come from (one
1510 // or more) Phis that merge their new-loop equivalents.
1511 //
1512 // This operation leaves the graph in an illegal state: there are two valid
1513 // control edges coming from the loop pre-header to both loop bodies. I'll
1514 // definitely have to hack the graph after running this transform.
1515 //
1516 // From this building block I will further edit edges to perform loop peeling
1517 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1518 //
1519 // Parameter side_by_size_idom:
1520 // When side_by_size_idom is NULL, the dominator tree is constructed for
1521 // the clone loop to dominate the original. Used in construction of
1522 // pre-main-post loop sequence.
1523 // When nonnull, the clone and original are side-by-side, both are
1524 // dominated by the side_by_side_idom node. Used in construction of
1525 // unswitched loops.
1526 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1527 Node* side_by_side_idom) {
1528
1529 #ifndef PRODUCT
1530 if (C->do_vector_loop() && PrintOpto) {
1531 const char* mname = C->method()->name()->as_quoted_ascii();
1532 if (mname != NULL) {
1533 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
1534 }
1535 }
1536 #endif
1537
1538 CloneMap& cm = C->clone_map();
1539 Dict* dict = cm.dict();
1540 if (C->do_vector_loop()) {
1541 cm.set_clone_idx(cm.max_gen()+1);
1542 #ifndef PRODUCT
1543 if (PrintOpto) {
1544 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
1545 loop->dump_head();
1546 }
1547 #endif
1548 }
1549
1550 // Step 1: Clone the loop body. Make the old->new mapping.
1551 uint i;
1552 for( i = 0; i < loop->_body.size(); i++ ) {
1553 Node *old = loop->_body.at(i);
1554 Node *nnn = old->clone();
1555 old_new.map( old->_idx, nnn );
1556 if (C->do_vector_loop()) {
1557 cm.verify_insert_and_clone(old, nnn, cm.clone_idx());
1558 }
1559 _igvn.register_new_node_with_optimizer(nnn);
1560 }
1561
1562
1563 // Step 2: Fix the edges in the new body. If the old input is outside the
1564 // loop use it. If the old input is INside the loop, use the corresponding
1565 // new node instead.
1566 for( i = 0; i < loop->_body.size(); i++ ) {
1567 Node *old = loop->_body.at(i);
1568 Node *nnn = old_new[old->_idx];
1569 // Fix CFG/Loop controlling the new node
1570 if (has_ctrl(old)) {
1571 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
1572 } else {
1573 set_loop(nnn, loop->_parent);
1574 if (old->outcnt() > 0) {
1575 set_idom( nnn, old_new[idom(old)->_idx], dd );
1576 }
1577 }
1578 // Correct edges to the new node
1579 for( uint j = 0; j < nnn->req(); j++ ) {
1580 Node *n = nnn->in(j);
1581 if( n ) {
1582 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
1583 if( loop->is_member( old_in_loop ) )
1584 nnn->set_req(j, old_new[n->_idx]);
1585 }
1586 }
1587 _igvn.hash_find_insert(nnn);
1588 }
1589 Node *newhead = old_new[loop->_head->_idx];
1590 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1591
1592
1593 // Step 3: Now fix control uses. Loop varying control uses have already
1594 // been fixed up (as part of all input edges in Step 2). Loop invariant
1595 // control uses must be either an IfFalse or an IfTrue. Make a merge
1596 // point to merge the old and new IfFalse/IfTrue nodes; make the use
1597 // refer to this.
1598 ResourceArea *area = Thread::current()->resource_area();
1599 Node_List worklist(area);
1600 uint new_counter = C->unique();
1601 for( i = 0; i < loop->_body.size(); i++ ) {
1602 Node* old = loop->_body.at(i);
1603 if( !old->is_CFG() ) continue;
1604 Node* nnn = old_new[old->_idx];
1605
1606 // Copy uses to a worklist, so I can munge the def-use info
1607 // with impunity.
1608 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1609 worklist.push(old->fast_out(j));
1610
1611 while( worklist.size() ) { // Visit all uses
1612 Node *use = worklist.pop();
1613 if (!has_node(use)) continue; // Ignore dead nodes
1614 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1615 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
1616 // Both OLD and USE are CFG nodes here.
1617 assert( use->is_Proj(), "" );
1618
1619 // Clone the loop exit control projection
1620 Node *newuse = use->clone();
1621 if (C->do_vector_loop()) {
1622 cm.verify_insert_and_clone(use, newuse, cm.clone_idx());
1623 }
1624 newuse->set_req(0,nnn);
1625 _igvn.register_new_node_with_optimizer(newuse);
1626 set_loop(newuse, use_loop);
1627 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
1628
1629 // We need a Region to merge the exit from the peeled body and the
1630 // exit from the old loop body.
1631 RegionNode *r = new RegionNode(3);
1632 // Map the old use to the new merge point
1633 old_new.map( use->_idx, r );
1634 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
1635 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
1636
1637 // The original user of 'use' uses 'r' instead.
1638 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
1639 Node* useuse = use->last_out(l);
1640 _igvn.rehash_node_delayed(useuse);
1641 uint uses_found = 0;
1642 if( useuse->in(0) == use ) {
1643 useuse->set_req(0, r);
1644 uses_found++;
1645 if( useuse->is_CFG() ) {
1646 assert( dom_depth(useuse) > dd_r, "" );
1647 set_idom(useuse, r, dom_depth(useuse));
1648 }
1649 }
1650 for( uint k = 1; k < useuse->req(); k++ ) {
1651 if( useuse->in(k) == use ) {
1652 useuse->set_req(k, r);
1653 uses_found++;
1654 }
1655 }
1656 l -= uses_found; // we deleted 1 or more copies of this edge
1657 }
1658
1659 // Now finish up 'r'
1660 r->set_req( 1, newuse );
1661 r->set_req( 2, use );
1662 _igvn.register_new_node_with_optimizer(r);
1663 set_loop(r, use_loop);
1664 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
1665 } // End of if a loop-exit test
1666 }
1667 }
1668
1669 // Step 4: If loop-invariant use is not control, it must be dominated by a
1670 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
1671 // there if needed. Make a Phi there merging old and new used values.
1672 Node_List *split_if_set = NULL;
1673 Node_List *split_bool_set = NULL;
1674 Node_List *split_cex_set = NULL;
1675 for( i = 0; i < loop->_body.size(); i++ ) {
1676 Node* old = loop->_body.at(i);
1677 Node* nnn = old_new[old->_idx];
1678 // Copy uses to a worklist, so I can munge the def-use info
1679 // with impunity.
1680 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1681 worklist.push(old->fast_out(j));
1682
1683 while( worklist.size() ) {
1684 Node *use = worklist.pop();
1685 if (!has_node(use)) continue; // Ignore dead nodes
1686 if (use->in(0) == C->top()) continue;
1687 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1688 // Check for data-use outside of loop - at least one of OLD or USE
1689 // must not be a CFG node.
1690 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) {
1691
1692 // If the Data use is an IF, that means we have an IF outside of the
1693 // loop that is switching on a condition that is set inside of the
1694 // loop. Happens if people set a loop-exit flag; then test the flag
1695 // in the loop to break the loop, then test is again outside of the
1696 // loop to determine which way the loop exited.
1697 // Loop predicate If node connects to Bool node through Opaque1 node.
1698 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use)) {
1699 // Since this code is highly unlikely, we lazily build the worklist
1700 // of such Nodes to go split.
1701 if( !split_if_set )
1702 split_if_set = new Node_List(area);
1703 split_if_set->push(use);
1704 }
1705 if( use->is_Bool() ) {
1706 if( !split_bool_set )
1707 split_bool_set = new Node_List(area);
1708 split_bool_set->push(use);
1709 }
1710 if( use->Opcode() == Op_CreateEx ) {
1711 if( !split_cex_set )
1712 split_cex_set = new Node_List(area);
1713 split_cex_set->push(use);
1714 }
1715
1716
1717 // Get "block" use is in
1718 uint idx = 0;
1719 while( use->in(idx) != old ) idx++;
1720 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1721 assert( !loop->is_member( get_loop( prev ) ), "" );
1722 Node *cfg = prev->_idx >= new_counter
1723 ? prev->in(2)
1724 : idom(prev);
1725 if( use->is_Phi() ) // Phi use is in prior block
1726 cfg = prev->in(idx); // NOT in block of Phi itself
1727 if (cfg->is_top()) { // Use is dead?
1728 _igvn.replace_input_of(use, idx, C->top());
1729 continue;
1730 }
1731
1732 while( !loop->is_member( get_loop( cfg ) ) ) {
1733 prev = cfg;
1734 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1735 }
1736 // If the use occurs after merging several exits from the loop, then
1737 // old value must have dominated all those exits. Since the same old
1738 // value was used on all those exits we did not need a Phi at this
1739 // merge point. NOW we do need a Phi here. Each loop exit value
1740 // is now merged with the peeled body exit; each exit gets its own
1741 // private Phi and those Phis need to be merged here.
1742 Node *phi;
1743 if( prev->is_Region() ) {
1744 if( idx == 0 && use->Opcode() != Op_ShenandoahWBMemProj) { // Updating control edge?
1745 phi = prev; // Just use existing control
1746 } else { // Else need a new Phi
1747 phi = PhiNode::make( prev, old );
1748 // Now recursively fix up the new uses of old!
1749 uint first = use->Opcode() != Op_ShenandoahWBMemProj ? 1 : 0;
1750 for( uint i = first; i < prev->req(); i++ ) {
1751 worklist.push(phi); // Onto worklist once for each 'old' input
1752 }
1753 }
1754 } else {
1755 // Get new RegionNode merging old and new loop exits
1756 prev = old_new[prev->_idx];
1757 assert( prev, "just made this in step 7" );
1758 if( idx == 0 && use->Opcode() != Op_ShenandoahWBMemProj) { // Updating control edge?
1759 phi = prev; // Just use existing control
1760 } else { // Else need a new Phi
1761 // Make a new Phi merging data values properly
1762 phi = PhiNode::make( prev, old );
1763 phi->set_req( 1, nnn );
1764 }
1765 }
1766 // If inserting a new Phi, check for prior hits
1767 if( idx != 0 ) {
1768 Node *hit = _igvn.hash_find_insert(phi);
1769 if( hit == NULL ) {
1770 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1771 } else { // or
1772 // Remove the new phi from the graph and use the hit
1773 _igvn.remove_dead_node(phi);
1774 phi = hit; // Use existing phi
1775 }
1776 set_ctrl(phi, prev);
1777 }
1778 // Make 'use' use the Phi instead of the old loop body exit value
1779 _igvn.replace_input_of(use, idx, phi);
1780 if( use->_idx >= new_counter ) { // If updating new phis
1781 // Not needed for correctness, but prevents a weak assert
1782 // in AddPNode from tripping (when we end up with different
1783 // base & derived Phis that will become the same after
1784 // IGVN does CSE).
1785 Node *hit = _igvn.hash_find_insert(use);
1786 if( hit ) // Go ahead and re-hash for hits.
1787 _igvn.replace_node( use, hit );
1788 }
1789
1790 // If 'use' was in the loop-exit block, it now needs to be sunk
1791 // below the post-loop merge point.
1792 sink_use( use, prev );
1793 }
1794 }
1795 }
1796
1797 // Check for IFs that need splitting/cloning. Happens if an IF outside of
1798 // the loop uses a condition set in the loop. The original IF probably
1799 // takes control from one or more OLD Regions (which in turn get from NEW
1800 // Regions). In any case, there will be a set of Phis for each merge point
1801 // from the IF up to where the original BOOL def exists the loop.
1802 if( split_if_set ) {
1803 while( split_if_set->size() ) {
1804 Node *iff = split_if_set->pop();
1805 if( iff->in(1)->is_Phi() ) {
1806 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop );
1807 _igvn.replace_input_of(iff, 1, b);
1808 }
1809 }
1810 }
1811 if( split_bool_set ) {
1812 while( split_bool_set->size() ) {
1813 Node *b = split_bool_set->pop();
1814 Node *phi = b->in(1);
1815 assert( phi->is_Phi(), "" );
1816 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop );
1817 _igvn.replace_input_of(b, 1, cmp);
1818 }
1819 }
1820 if( split_cex_set ) {
1821 while( split_cex_set->size() ) {
1822 Node *b = split_cex_set->pop();
1823 assert( b->in(0)->is_Region(), "" );
1824 assert( b->in(1)->is_Phi(), "" );
1825 assert( b->in(0)->in(0) == b->in(1)->in(0), "" );
1826 split_up( b, b->in(0), NULL );
1827 }
1828 }
1829
1830 }
1831
1832
1833 //---------------------- stride_of_possible_iv -------------------------------------
1834 // Looks for an iff/bool/comp with one operand of the compare
1835 // being a cycle involving an add and a phi,
1836 // with an optional truncation (left-shift followed by a right-shift)
1837 // of the add. Returns zero if not an iv.
1838 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
1839 Node* trunc1 = NULL;
1840 Node* trunc2 = NULL;
1841 const TypeInt* ttype = NULL;
1842 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
1843 return 0;
1844 }
1845 BoolNode* bl = iff->in(1)->as_Bool();
1846 Node* cmp = bl->in(1);
1847 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) {
1848 return 0;
1849 }
1850 // Must have an invariant operand
1851 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
1852 return 0;
1853 }
1854 Node* add2 = NULL;
1855 Node* cmp1 = cmp->in(1);
1856 if (cmp1->is_Phi()) {
1857 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
1858 Node* phi = cmp1;
1859 for (uint i = 1; i < phi->req(); i++) {
1860 Node* in = phi->in(i);
1861 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
1862 &trunc1, &trunc2, &ttype);
1863 if (add && add->in(1) == phi) {
1864 add2 = add->in(2);
1865 break;
1866 }
1867 }
1868 } else {
1869 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
1870 Node* addtrunc = cmp1;
1871 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
1872 &trunc1, &trunc2, &ttype);
1873 if (add && add->in(1)->is_Phi()) {
1874 Node* phi = add->in(1);
1875 for (uint i = 1; i < phi->req(); i++) {
1876 if (phi->in(i) == addtrunc) {
1877 add2 = add->in(2);
1878 break;
1879 }
1880 }
1881 }
1882 }
1883 if (add2 != NULL) {
1884 const TypeInt* add2t = _igvn.type(add2)->is_int();
1885 if (add2t->is_con()) {
1886 return add2t->get_con();
1887 }
1888 }
1889 return 0;
1890 }
1891
1892
1893 //---------------------- stay_in_loop -------------------------------------
1894 // Return the (unique) control output node that's in the loop (if it exists.)
1895 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
1896 Node* unique = NULL;
1897 if (!n) return NULL;
1898 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1899 Node* use = n->fast_out(i);
1900 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
1901 if (unique != NULL) {
1902 return NULL;
1903 }
1904 unique = use;
1905 }
1906 }
1907 return unique;
1908 }
1909
1910 //------------------------------ register_node -------------------------------------
1911 // Utility to register node "n" with PhaseIdealLoop
1912 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
1913 _igvn.register_new_node_with_optimizer(n);
1914 loop->_body.push(n);
1915 if (n->is_CFG()) {
1916 set_loop(n, loop);
1917 set_idom(n, pred, ddepth);
1918 } else {
1919 set_ctrl(n, pred);
1920 }
1921 }
1922
1923 //------------------------------ proj_clone -------------------------------------
1924 // Utility to create an if-projection
1925 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
1926 ProjNode* c = p->clone()->as_Proj();
1927 c->set_req(0, iff);
1928 return c;
1929 }
1930
1931 //------------------------------ short_circuit_if -------------------------------------
1932 // Force the iff control output to be the live_proj
1933 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
1934 guarantee(live_proj != NULL, "null projection");
1935 int proj_con = live_proj->_con;
1936 assert(proj_con == 0 || proj_con == 1, "false or true projection");
1937 Node *con = _igvn.intcon(proj_con);
1938 set_ctrl(con, C->root());
1939 if (iff) {
1940 iff->set_req(1, con);
1941 }
1942 return con;
1943 }
1944
1945 //------------------------------ insert_if_before_proj -------------------------------------
1946 // Insert a new if before an if projection (* - new node)
1947 //
1948 // before
1949 // if(test)
1950 // / \
1951 // v v
1952 // other-proj proj (arg)
1953 //
1954 // after
1955 // if(test)
1956 // / \
1957 // / v
1958 // | * proj-clone
1959 // v |
1960 // other-proj v
1961 // * new_if(relop(cmp[IU](left,right)))
1962 // / \
1963 // v v
1964 // * new-proj proj
1965 // (returned)
1966 //
1967 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
1968 IfNode* iff = proj->in(0)->as_If();
1969 IdealLoopTree *loop = get_loop(proj);
1970 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1971 int ddepth = dom_depth(proj);
1972
1973 _igvn.rehash_node_delayed(iff);
1974 _igvn.rehash_node_delayed(proj);
1975
1976 proj->set_req(0, NULL); // temporary disconnect
1977 ProjNode* proj2 = proj_clone(proj, iff);
1978 register_node(proj2, loop, iff, ddepth);
1979
1980 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
1981 register_node(cmp, loop, proj2, ddepth);
1982
1983 BoolNode* bol = new BoolNode(cmp, relop);
1984 register_node(bol, loop, proj2, ddepth);
1985
1986 IfNode* new_if = new IfNode(proj2, bol, iff->_prob, iff->_fcnt);
1987 register_node(new_if, loop, proj2, ddepth);
1988
1989 proj->set_req(0, new_if); // reattach
1990 set_idom(proj, new_if, ddepth);
1991
1992 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
1993 guarantee(new_exit != NULL, "null exit node");
1994 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
1995
1996 return new_exit;
1997 }
1998
1999 //------------------------------ insert_region_before_proj -------------------------------------
2000 // Insert a region before an if projection (* - new node)
2001 //
2002 // before
2003 // if(test)
2004 // / |
2005 // v |
2006 // proj v
2007 // other-proj
2008 //
2009 // after
2010 // if(test)
2011 // / |
2012 // v |
2013 // * proj-clone v
2014 // | other-proj
2015 // v
2016 // * new-region
2017 // |
2018 // v
2019 // * dum_if
2020 // / \
2021 // v \
2022 // * dum-proj v
2023 // proj
2024 //
2025 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2026 IfNode* iff = proj->in(0)->as_If();
2027 IdealLoopTree *loop = get_loop(proj);
2028 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2029 int ddepth = dom_depth(proj);
2030
2031 _igvn.rehash_node_delayed(iff);
2032 _igvn.rehash_node_delayed(proj);
2033
2034 proj->set_req(0, NULL); // temporary disconnect
2035 ProjNode* proj2 = proj_clone(proj, iff);
2036 register_node(proj2, loop, iff, ddepth);
2037
2038 RegionNode* reg = new RegionNode(2);
2039 reg->set_req(1, proj2);
2040 register_node(reg, loop, iff, ddepth);
2041
2042 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
2043 register_node(dum_if, loop, reg, ddepth);
2044
2045 proj->set_req(0, dum_if); // reattach
2046 set_idom(proj, dum_if, ddepth);
2047
2048 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
2049 register_node(dum_proj, loop, dum_if, ddepth);
2050
2051 return reg;
2052 }
2053
2054 //------------------------------ insert_cmpi_loop_exit -------------------------------------
2055 // Clone a signed compare loop exit from an unsigned compare and
2056 // insert it before the unsigned cmp on the stay-in-loop path.
2057 // All new nodes inserted in the dominator tree between the original
2058 // if and it's projections. The original if test is replaced with
2059 // a constant to force the stay-in-loop path.
2060 //
2061 // This is done to make sure that the original if and it's projections
2062 // still dominate the same set of control nodes, that the ctrl() relation
2063 // from data nodes to them is preserved, and that their loop nesting is
2064 // preserved.
2065 //
2066 // before
2067 // if(i <u limit) unsigned compare loop exit
2068 // / |
2069 // v v
2070 // exit-proj stay-in-loop-proj
2071 //
2072 // after
2073 // if(stay-in-loop-const) original if
2074 // / |
2075 // / v
2076 // / if(i < limit) new signed test
2077 // / / |
2078 // / / v
2079 // / / if(i <u limit) new cloned unsigned test
2080 // / / / |
2081 // v v v |
2082 // region |
2083 // | |
2084 // dum-if |
2085 // / | |
2086 // ether | |
2087 // v v
2088 // exit-proj stay-in-loop-proj
2089 //
2090 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
2091 const bool Signed = true;
2092 const bool Unsigned = false;
2093
2094 BoolNode* bol = if_cmpu->in(1)->as_Bool();
2095 if (bol->_test._test != BoolTest::lt) return NULL;
2096 CmpNode* cmpu = bol->in(1)->as_Cmp();
2097 if (cmpu->Opcode() != Op_CmpU) return NULL;
2098 int stride = stride_of_possible_iv(if_cmpu);
2099 if (stride == 0) return NULL;
2100
2101 Node* lp_proj = stay_in_loop(if_cmpu, loop);
2102 guarantee(lp_proj != NULL, "null loop node");
2103
2104 ProjNode* lp_continue = lp_proj->as_Proj();
2105 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
2106
2107 Node* limit = NULL;
2108 if (stride > 0) {
2109 limit = cmpu->in(2);
2110 } else {
2111 limit = _igvn.makecon(TypeInt::ZERO);
2112 set_ctrl(limit, C->root());
2113 }
2114 // Create a new region on the exit path
2115 RegionNode* reg = insert_region_before_proj(lp_exit);
2116 guarantee(reg != NULL, "null region node");
2117
2118 // Clone the if-cmpu-true-false using a signed compare
2119 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
2120 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
2121 reg->add_req(cmpi_exit);
2122
2123 // Clone the if-cmpu-true-false
2124 BoolTest::mask rel_u = bol->_test._test;
2125 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
2126 reg->add_req(cmpu_exit);
2127
2128 // Force original if to stay in loop.
2129 short_circuit_if(if_cmpu, lp_continue);
2130
2131 return cmpi_exit->in(0)->as_If();
2132 }
2133
2134 //------------------------------ remove_cmpi_loop_exit -------------------------------------
2135 // Remove a previously inserted signed compare loop exit.
2136 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
2137 Node* lp_proj = stay_in_loop(if_cmp, loop);
2138 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
2139 stay_in_loop(lp_proj, loop)->is_If() &&
2140 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
2141 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
2142 set_ctrl(con, C->root());
2143 if_cmp->set_req(1, con);
2144 }
2145
2146 //------------------------------ scheduled_nodelist -------------------------------------
2147 // Create a post order schedule of nodes that are in the
2148 // "member" set. The list is returned in "sched".
2149 // The first node in "sched" is the loop head, followed by
2150 // nodes which have no inputs in the "member" set, and then
2151 // followed by the nodes that have an immediate input dependence
2152 // on a node in "sched".
2153 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
2154
2155 assert(member.test(loop->_head->_idx), "loop head must be in member set");
2156 Arena *a = Thread::current()->resource_area();
2157 VectorSet visited(a);
2158 Node_Stack nstack(a, loop->_body.size());
2159
2160 Node* n = loop->_head; // top of stack is cached in "n"
2161 uint idx = 0;
2162 visited.set(n->_idx);
2163
2164 // Initially push all with no inputs from within member set
2165 for(uint i = 0; i < loop->_body.size(); i++ ) {
2166 Node *elt = loop->_body.at(i);
2167 if (member.test(elt->_idx)) {
2168 bool found = false;
2169 for (uint j = 0; j < elt->req(); j++) {
2170 Node* def = elt->in(j);
2171 if (def && member.test(def->_idx) && def != elt) {
2172 found = true;
2173 break;
2174 }
2175 }
2176 if (!found && elt != loop->_head) {
2177 nstack.push(n, idx);
2178 n = elt;
2179 assert(!visited.test(n->_idx), "not seen yet");
2180 visited.set(n->_idx);
2181 }
2182 }
2183 }
2184
2185 // traverse out's that are in the member set
2186 while (true) {
2187 if (idx < n->outcnt()) {
2188 Node* use = n->raw_out(idx);
2189 idx++;
2190 if (!visited.test_set(use->_idx)) {
2191 if (member.test(use->_idx)) {
2192 nstack.push(n, idx);
2193 n = use;
2194 idx = 0;
2195 }
2196 }
2197 } else {
2198 // All outputs processed
2199 sched.push(n);
2200 if (nstack.is_empty()) break;
2201 n = nstack.node();
2202 idx = nstack.index();
2203 nstack.pop();
2204 }
2205 }
2206 }
2207
2208
2209 //------------------------------ has_use_in_set -------------------------------------
2210 // Has a use in the vector set
2211 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
2212 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2213 Node* use = n->fast_out(j);
2214 if (vset.test(use->_idx)) {
2215 return true;
2216 }
2217 }
2218 return false;
2219 }
2220
2221
2222 //------------------------------ has_use_internal_to_set -------------------------------------
2223 // Has use internal to the vector set (ie. not in a phi at the loop head)
2224 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
2225 Node* head = loop->_head;
2226 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2227 Node* use = n->fast_out(j);
2228 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
2229 return true;
2230 }
2231 }
2232 return false;
2233 }
2234
2235
2236 //------------------------------ clone_for_use_outside_loop -------------------------------------
2237 // clone "n" for uses that are outside of loop
2238 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
2239 int cloned = 0;
2240 assert(worklist.size() == 0, "should be empty");
2241 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2242 Node* use = n->fast_out(j);
2243 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
2244 worklist.push(use);
2245 }
2246 }
2247 while( worklist.size() ) {
2248 Node *use = worklist.pop();
2249 if (!has_node(use) || use->in(0) == C->top()) continue;
2250 uint j;
2251 for (j = 0; j < use->req(); j++) {
2252 if (use->in(j) == n) break;
2253 }
2254 assert(j < use->req(), "must be there");
2255
2256 // clone "n" and insert it between the inputs of "n" and the use outside the loop
2257 Node* n_clone = n->clone();
2258 _igvn.replace_input_of(use, j, n_clone);
2259 cloned++;
2260 Node* use_c;
2261 if (!use->is_Phi()) {
2262 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
2263 } else {
2264 // Use in a phi is considered a use in the associated predecessor block
2265 use_c = use->in(0)->in(j);
2266 }
2267 set_ctrl(n_clone, use_c);
2268 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
2269 get_loop(use_c)->_body.push(n_clone);
2270 _igvn.register_new_node_with_optimizer(n_clone);
2271 #if !defined(PRODUCT)
2272 if (TracePartialPeeling) {
2273 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
2274 }
2275 #endif
2276 }
2277 return cloned;
2278 }
2279
2280
2281 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
2282 // clone "n" for special uses that are in the not_peeled region.
2283 // If these def-uses occur in separate blocks, the code generator
2284 // marks the method as not compilable. For example, if a "BoolNode"
2285 // is in a different basic block than the "IfNode" that uses it, then
2286 // the compilation is aborted in the code generator.
2287 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
2288 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
2289 if (n->is_Phi() || n->is_Load()) {
2290 return;
2291 }
2292 assert(worklist.size() == 0, "should be empty");
2293 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2294 Node* use = n->fast_out(j);
2295 if ( not_peel.test(use->_idx) &&
2296 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
2297 use->in(1) == n) {
2298 worklist.push(use);
2299 }
2300 }
2301 if (worklist.size() > 0) {
2302 // clone "n" and insert it between inputs of "n" and the use
2303 Node* n_clone = n->clone();
2304 loop->_body.push(n_clone);
2305 _igvn.register_new_node_with_optimizer(n_clone);
2306 set_ctrl(n_clone, get_ctrl(n));
2307 sink_list.push(n_clone);
2308 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
2309 #if !defined(PRODUCT)
2310 if (TracePartialPeeling) {
2311 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2312 }
2313 #endif
2314 while( worklist.size() ) {
2315 Node *use = worklist.pop();
2316 _igvn.rehash_node_delayed(use);
2317 for (uint j = 1; j < use->req(); j++) {
2318 if (use->in(j) == n) {
2319 use->set_req(j, n_clone);
2320 }
2321 }
2322 }
2323 }
2324 }
2325
2326
2327 //------------------------------ insert_phi_for_loop -------------------------------------
2328 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2329 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2330 Node *phi = PhiNode::make(lp, back_edge_val);
2331 phi->set_req(LoopNode::EntryControl, lp_entry_val);
2332 // Use existing phi if it already exists
2333 Node *hit = _igvn.hash_find_insert(phi);
2334 if( hit == NULL ) {
2335 _igvn.register_new_node_with_optimizer(phi);
2336 set_ctrl(phi, lp);
2337 } else {
2338 // Remove the new phi from the graph and use the hit
2339 _igvn.remove_dead_node(phi);
2340 phi = hit;
2341 }
2342 _igvn.replace_input_of(use, idx, phi);
2343 }
2344
2345 #ifdef ASSERT
2346 //------------------------------ is_valid_loop_partition -------------------------------------
2347 // Validate the loop partition sets: peel and not_peel
2348 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2349 VectorSet& not_peel ) {
2350 uint i;
2351 // Check that peel_list entries are in the peel set
2352 for (i = 0; i < peel_list.size(); i++) {
2353 if (!peel.test(peel_list.at(i)->_idx)) {
2354 return false;
2355 }
2356 }
2357 // Check at loop members are in one of peel set or not_peel set
2358 for (i = 0; i < loop->_body.size(); i++ ) {
2359 Node *def = loop->_body.at(i);
2360 uint di = def->_idx;
2361 // Check that peel set elements are in peel_list
2362 if (peel.test(di)) {
2363 if (not_peel.test(di)) {
2364 return false;
2365 }
2366 // Must be in peel_list also
2367 bool found = false;
2368 for (uint j = 0; j < peel_list.size(); j++) {
2369 if (peel_list.at(j)->_idx == di) {
2370 found = true;
2371 break;
2372 }
2373 }
2374 if (!found) {
2375 return false;
2376 }
2377 } else if (not_peel.test(di)) {
2378 if (peel.test(di)) {
2379 return false;
2380 }
2381 } else {
2382 return false;
2383 }
2384 }
2385 return true;
2386 }
2387
2388 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2389 // Ensure a use outside of loop is of the right form
2390 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2391 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2392 return (use->is_Phi() &&
2393 use_c->is_Region() && use_c->req() == 3 &&
2394 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2395 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2396 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2397 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2398 }
2399
2400 //------------------------------ is_valid_clone_loop_form -------------------------------------
2401 // Ensure that all uses outside of loop are of the right form
2402 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2403 uint orig_exit_idx, uint clone_exit_idx) {
2404 uint len = peel_list.size();
2405 for (uint i = 0; i < len; i++) {
2406 Node *def = peel_list.at(i);
2407
2408 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2409 Node *use = def->fast_out(j);
2410 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2411 if (!loop->is_member(get_loop(use_c))) {
2412 // use is not in the loop, check for correct structure
2413 if (use->in(0) == def) {
2414 // Okay
2415 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2416 return false;
2417 }
2418 }
2419 }
2420 }
2421 return true;
2422 }
2423 #endif
2424
2425 //------------------------------ partial_peel -------------------------------------
2426 // Partially peel (aka loop rotation) the top portion of a loop (called
2427 // the peel section below) by cloning it and placing one copy just before
2428 // the new loop head and the other copy at the bottom of the new loop.
2429 //
2430 // before after where it came from
2431 //
2432 // stmt1 stmt1
2433 // loop: stmt2 clone
2434 // stmt2 if condA goto exitA clone
2435 // if condA goto exitA new_loop: new
2436 // stmt3 stmt3 clone
2437 // if !condB goto loop if condB goto exitB clone
2438 // exitB: stmt2 orig
2439 // stmt4 if !condA goto new_loop orig
2440 // exitA: goto exitA
2441 // exitB:
2442 // stmt4
2443 // exitA:
2444 //
2445 // Step 1: find the cut point: an exit test on probable
2446 // induction variable.
2447 // Step 2: schedule (with cloning) operations in the peel
2448 // section that can be executed after the cut into
2449 // the section that is not peeled. This may need
2450 // to clone operations into exit blocks. For
2451 // instance, a reference to A[i] in the not-peel
2452 // section and a reference to B[i] in an exit block
2453 // may cause a left-shift of i by 2 to be placed
2454 // in the peel block. This step will clone the left
2455 // shift into the exit block and sink the left shift
2456 // from the peel to the not-peel section.
2457 // Step 3: clone the loop, retarget the control, and insert
2458 // phis for values that are live across the new loop
2459 // head. This is very dependent on the graph structure
2460 // from clone_loop. It creates region nodes for
2461 // exit control and associated phi nodes for values
2462 // flow out of the loop through that exit. The region
2463 // node is dominated by the clone's control projection.
2464 // So the clone's peel section is placed before the
2465 // new loop head, and the clone's not-peel section is
2466 // forms the top part of the new loop. The original
2467 // peel section forms the tail of the new loop.
2468 // Step 4: update the dominator tree and recompute the
2469 // dominator depth.
2470 //
2471 // orig
2472 //
2473 // stmt1
2474 // |
2475 // v
2476 // loop predicate
2477 // |
2478 // v
2479 // loop<----+
2480 // | |
2481 // stmt2 |
2482 // | |
2483 // v |
2484 // ifA |
2485 // / | |
2486 // v v |
2487 // false true ^ <-- last_peel
2488 // / | |
2489 // / ===|==cut |
2490 // / stmt3 | <-- first_not_peel
2491 // / | |
2492 // | v |
2493 // v ifB |
2494 // exitA: / \ |
2495 // / \ |
2496 // v v |
2497 // false true |
2498 // / \ |
2499 // / ----+
2500 // |
2501 // v
2502 // exitB:
2503 // stmt4
2504 //
2505 //
2506 // after clone loop
2507 //
2508 // stmt1
2509 // |
2510 // v
2511 // loop predicate
2512 // / \
2513 // clone / \ orig
2514 // / \
2515 // / \
2516 // v v
2517 // +---->loop loop<----+
2518 // | | | |
2519 // | stmt2 stmt2 |
2520 // | | | |
2521 // | v v |
2522 // | ifA ifA |
2523 // | | \ / | |
2524 // | v v v v |
2525 // ^ true false false true ^ <-- last_peel
2526 // | | ^ \ / | |
2527 // | cut==|== \ \ / ===|==cut |
2528 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2529 // | | dom | | | |
2530 // | v \ 1v v2 v |
2531 // | ifB regionA ifB |
2532 // | / \ | / \ |
2533 // | / \ v / \ |
2534 // | v v exitA: v v |
2535 // | true false false true |
2536 // | / ^ \ / \ |
2537 // +---- \ \ / ----+
2538 // dom \ /
2539 // \ 1v v2
2540 // regionB
2541 // |
2542 // v
2543 // exitB:
2544 // stmt4
2545 //
2546 //
2547 // after partial peel
2548 //
2549 // stmt1
2550 // |
2551 // v
2552 // loop predicate
2553 // /
2554 // clone / orig
2555 // / TOP
2556 // / \
2557 // v v
2558 // TOP->loop loop----+
2559 // | | |
2560 // stmt2 stmt2 |
2561 // | | |
2562 // v v |
2563 // ifA ifA |
2564 // | \ / | |
2565 // v v v v |
2566 // true false false true | <-- last_peel
2567 // | ^ \ / +------|---+
2568 // +->newloop \ \ / === ==cut | |
2569 // | stmt3 \ \ / TOP | |
2570 // | | dom | | stmt3 | | <-- first_not_peel
2571 // | v \ 1v v2 v | |
2572 // | ifB regionA ifB ^ v
2573 // | / \ | / \ | |
2574 // | / \ v / \ | |
2575 // | v v exitA: v v | |
2576 // | true false false true | |
2577 // | / ^ \ / \ | |
2578 // | | \ \ / v | |
2579 // | | dom \ / TOP | |
2580 // | | \ 1v v2 | |
2581 // ^ v regionB | |
2582 // | | | | |
2583 // | | v ^ v
2584 // | | exitB: | |
2585 // | | stmt4 | |
2586 // | +------------>-----------------+ |
2587 // | |
2588 // +-----------------<---------------------+
2589 //
2590 //
2591 // final graph
2592 //
2593 // stmt1
2594 // |
2595 // v
2596 // loop predicate
2597 // |
2598 // v
2599 // stmt2 clone
2600 // |
2601 // v
2602 // ........> ifA clone
2603 // : / |
2604 // dom / |
2605 // : v v
2606 // : false true
2607 // : | |
2608 // : | v
2609 // : | newloop<-----+
2610 // : | | |
2611 // : | stmt3 clone |
2612 // : | | |
2613 // : | v |
2614 // : | ifB |
2615 // : | / \ |
2616 // : | v v |
2617 // : | false true |
2618 // : | | | |
2619 // : | v stmt2 |
2620 // : | exitB: | |
2621 // : | stmt4 v |
2622 // : | ifA orig |
2623 // : | / \ |
2624 // : | / \ |
2625 // : | v v |
2626 // : | false true |
2627 // : | / \ |
2628 // : v v -----+
2629 // RegionA
2630 // |
2631 // v
2632 // exitA
2633 //
2634 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
2635
2636 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
2637 if (!loop->_head->is_Loop()) {
2638 return false; }
2639
2640 LoopNode *head = loop->_head->as_Loop();
2641
2642 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
2643 return false;
2644 }
2645
2646 // Check for complex exit control
2647 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
2648 Node *n = loop->_body.at(ii);
2649 int opc = n->Opcode();
2650 if (n->is_Call() ||
2651 opc == Op_Catch ||
2652 opc == Op_CatchProj ||
2653 opc == Op_Jump ||
2654 opc == Op_JumpProj) {
2655 #if !defined(PRODUCT)
2656 if (TracePartialPeeling) {
2657 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
2658 }
2659 #endif
2660 return false;
2661 }
2662 }
2663
2664 int dd = dom_depth(head);
2665
2666 // Step 1: find cut point
2667
2668 // Walk up dominators to loop head looking for first loop exit
2669 // which is executed on every path thru loop.
2670 IfNode *peel_if = NULL;
2671 IfNode *peel_if_cmpu = NULL;
2672
2673 Node *iff = loop->tail();
2674 while( iff != head ) {
2675 if( iff->is_If() ) {
2676 Node *ctrl = get_ctrl(iff->in(1));
2677 if (ctrl->is_top()) return false; // Dead test on live IF.
2678 // If loop-varying exit-test, check for induction variable
2679 if( loop->is_member(get_loop(ctrl)) &&
2680 loop->is_loop_exit(iff) &&
2681 is_possible_iv_test(iff)) {
2682 Node* cmp = iff->in(1)->in(1);
2683 if (cmp->Opcode() == Op_CmpI) {
2684 peel_if = iff->as_If();
2685 } else {
2686 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
2687 peel_if_cmpu = iff->as_If();
2688 }
2689 }
2690 }
2691 iff = idom(iff);
2692 }
2693 // Prefer signed compare over unsigned compare.
2694 IfNode* new_peel_if = NULL;
2695 if (peel_if == NULL) {
2696 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
2697 return false; // No peel point found
2698 }
2699 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
2700 if (new_peel_if == NULL) {
2701 return false; // No peel point found
2702 }
2703 peel_if = new_peel_if;
2704 }
2705 Node* last_peel = stay_in_loop(peel_if, loop);
2706 Node* first_not_peeled = stay_in_loop(last_peel, loop);
2707 if (first_not_peeled == NULL || first_not_peeled == head) {
2708 return false;
2709 }
2710
2711 #if !defined(PRODUCT)
2712 if (TraceLoopOpts) {
2713 tty->print("PartialPeel ");
2714 loop->dump_head();
2715 }
2716
2717 if (TracePartialPeeling) {
2718 tty->print_cr("before partial peel one iteration");
2719 Node_List wl;
2720 Node* t = head->in(2);
2721 while (true) {
2722 wl.push(t);
2723 if (t == head) break;
2724 t = idom(t);
2725 }
2726 while (wl.size() > 0) {
2727 Node* tt = wl.pop();
2728 tt->dump();
2729 if (tt == last_peel) tty->print_cr("-- cut --");
2730 }
2731 }
2732 #endif
2733 ResourceArea *area = Thread::current()->resource_area();
2734 VectorSet peel(area);
2735 VectorSet not_peel(area);
2736 Node_List peel_list(area);
2737 Node_List worklist(area);
2738 Node_List sink_list(area);
2739
2740 // Set of cfg nodes to peel are those that are executable from
2741 // the head through last_peel.
2742 assert(worklist.size() == 0, "should be empty");
2743 worklist.push(head);
2744 peel.set(head->_idx);
2745 while (worklist.size() > 0) {
2746 Node *n = worklist.pop();
2747 if (n != last_peel) {
2748 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2749 Node* use = n->fast_out(j);
2750 if (use->is_CFG() &&
2751 loop->is_member(get_loop(use)) &&
2752 !peel.test_set(use->_idx)) {
2753 worklist.push(use);
2754 }
2755 }
2756 }
2757 }
2758
2759 // Set of non-cfg nodes to peel are those that are control
2760 // dependent on the cfg nodes.
2761 uint i;
2762 for(i = 0; i < loop->_body.size(); i++ ) {
2763 Node *n = loop->_body.at(i);
2764 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
2765 if (peel.test(n_c->_idx)) {
2766 peel.set(n->_idx);
2767 } else {
2768 not_peel.set(n->_idx);
2769 }
2770 }
2771
2772 // Step 2: move operations from the peeled section down into the
2773 // not-peeled section
2774
2775 // Get a post order schedule of nodes in the peel region
2776 // Result in right-most operand.
2777 scheduled_nodelist(loop, peel, peel_list );
2778
2779 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2780
2781 // For future check for too many new phis
2782 uint old_phi_cnt = 0;
2783 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2784 Node* use = head->fast_out(j);
2785 if (use->is_Phi()) old_phi_cnt++;
2786 }
2787
2788 #if !defined(PRODUCT)
2789 if (TracePartialPeeling) {
2790 tty->print_cr("\npeeled list");
2791 }
2792 #endif
2793
2794 // Evacuate nodes in peel region into the not_peeled region if possible
2795 uint new_phi_cnt = 0;
2796 uint cloned_for_outside_use = 0;
2797 for (i = 0; i < peel_list.size();) {
2798 Node* n = peel_list.at(i);
2799 #if !defined(PRODUCT)
2800 if (TracePartialPeeling) n->dump();
2801 #endif
2802 bool incr = true;
2803 if ( !n->is_CFG() ) {
2804
2805 if ( has_use_in_set(n, not_peel) ) {
2806
2807 // If not used internal to the peeled region,
2808 // move "n" from peeled to not_peeled region.
2809
2810 if ( !has_use_internal_to_set(n, peel, loop) ) {
2811
2812 // if not pinned and not a load (which maybe anti-dependent on a store)
2813 // and not a CMove (Matcher expects only bool->cmove).
2814 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
2815 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist );
2816 sink_list.push(n);
2817 peel >>= n->_idx; // delete n from peel set.
2818 not_peel <<= n->_idx; // add n to not_peel set.
2819 peel_list.remove(i);
2820 incr = false;
2821 #if !defined(PRODUCT)
2822 if (TracePartialPeeling) {
2823 tty->print_cr("sink to not_peeled region: %d newbb: %d",
2824 n->_idx, get_ctrl(n)->_idx);
2825 }
2826 #endif
2827 }
2828 } else {
2829 // Otherwise check for special def-use cases that span
2830 // the peel/not_peel boundary such as bool->if
2831 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
2832 new_phi_cnt++;
2833 }
2834 }
2835 }
2836 if (incr) i++;
2837 }
2838
2839 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
2840 #if !defined(PRODUCT)
2841 if (TracePartialPeeling) {
2842 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
2843 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
2844 }
2845 #endif
2846 if (new_peel_if != NULL) {
2847 remove_cmpi_loop_exit(new_peel_if, loop);
2848 }
2849 // Inhibit more partial peeling on this loop
2850 assert(!head->is_partial_peel_loop(), "not partial peeled");
2851 head->mark_partial_peel_failed();
2852 if (cloned_for_outside_use > 0) {
2853 // Terminate this round of loop opts because
2854 // the graph outside this loop was changed.
2855 C->set_major_progress();
2856 return true;
2857 }
2858 return false;
2859 }
2860
2861 // Step 3: clone loop, retarget control, and insert new phis
2862
2863 // Create new loop head for new phis and to hang
2864 // the nodes being moved (sinked) from the peel region.
2865 LoopNode* new_head = new LoopNode(last_peel, last_peel);
2866 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
2867 _igvn.register_new_node_with_optimizer(new_head);
2868 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
2869 _igvn.replace_input_of(first_not_peeled, 0, new_head);
2870 set_loop(new_head, loop);
2871 loop->_body.push(new_head);
2872 not_peel.set(new_head->_idx);
2873 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
2874 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
2875
2876 while (sink_list.size() > 0) {
2877 Node* n = sink_list.pop();
2878 set_ctrl(n, new_head);
2879 }
2880
2881 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2882
2883 clone_loop( loop, old_new, dd );
2884
2885 const uint clone_exit_idx = 1;
2886 const uint orig_exit_idx = 2;
2887 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
2888
2889 Node* head_clone = old_new[head->_idx];
2890 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
2891 Node* orig_tail_clone = head_clone->in(2);
2892
2893 // Add phi if "def" node is in peel set and "use" is not
2894
2895 for(i = 0; i < peel_list.size(); i++ ) {
2896 Node *def = peel_list.at(i);
2897 if (!def->is_CFG()) {
2898 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2899 Node *use = def->fast_out(j);
2900 if (has_node(use) && use->in(0) != C->top() &&
2901 (!peel.test(use->_idx) ||
2902 (use->is_Phi() && use->in(0) == head)) ) {
2903 worklist.push(use);
2904 }
2905 }
2906 while( worklist.size() ) {
2907 Node *use = worklist.pop();
2908 for (uint j = 1; j < use->req(); j++) {
2909 Node* n = use->in(j);
2910 if (n == def) {
2911
2912 // "def" is in peel set, "use" is not in peel set
2913 // or "use" is in the entry boundary (a phi) of the peel set
2914
2915 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
2916
2917 if ( loop->is_member(get_loop( use_c )) ) {
2918 // use is in loop
2919 if (old_new[use->_idx] != NULL) { // null for dead code
2920 Node* use_clone = old_new[use->_idx];
2921 _igvn.replace_input_of(use, j, C->top());
2922 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
2923 }
2924 } else {
2925 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
2926 // use is not in the loop, check if the live range includes the cut
2927 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
2928 if (not_peel.test(lp_if->_idx)) {
2929 assert(j == orig_exit_idx, "use from original loop");
2930 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
2931 }
2932 }
2933 }
2934 }
2935 }
2936 }
2937 }
2938
2939 // Step 3b: retarget control
2940
2941 // Redirect control to the new loop head if a cloned node in
2942 // the not_peeled region has control that points into the peeled region.
2943 // This necessary because the cloned peeled region will be outside
2944 // the loop.
2945 // from to
2946 // cloned-peeled <---+
2947 // new_head_clone: | <--+
2948 // cloned-not_peeled in(0) in(0)
2949 // orig-peeled
2950
2951 for(i = 0; i < loop->_body.size(); i++ ) {
2952 Node *n = loop->_body.at(i);
2953 if (!n->is_CFG() && n->in(0) != NULL &&
2954 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
2955 Node* n_clone = old_new[n->_idx];
2956 _igvn.replace_input_of(n_clone, 0, new_head_clone);
2957 }
2958 }
2959
2960 // Backedge of the surviving new_head (the clone) is original last_peel
2961 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
2962
2963 // Cut first node in original not_peel set
2964 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
2965 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
2966 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
2967
2968 // Copy head_clone back-branch info to original head
2969 // and remove original head's loop entry and
2970 // clone head's back-branch
2971 _igvn.rehash_node_delayed(head); // Multiple edge updates
2972 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
2973 head->set_req(LoopNode::LoopBackControl, C->top());
2974 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
2975
2976 // Similarly modify the phis
2977 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
2978 Node* use = head->fast_out(k);
2979 if (use->is_Phi() && use->outcnt() > 0) {
2980 Node* use_clone = old_new[use->_idx];
2981 _igvn.rehash_node_delayed(use); // Multiple edge updates
2982 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
2983 use->set_req(LoopNode::LoopBackControl, C->top());
2984 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
2985 }
2986 }
2987
2988 // Step 4: update dominator tree and dominator depth
2989
2990 set_idom(head, orig_tail_clone, dd);
2991 recompute_dom_depth();
2992
2993 // Inhibit more partial peeling on this loop
2994 new_head_clone->set_partial_peel_loop();
2995 C->set_major_progress();
2996 loop->record_for_igvn();
2997
2998 #if !defined(PRODUCT)
2999 if (TracePartialPeeling) {
3000 tty->print_cr("\nafter partial peel one iteration");
3001 Node_List wl(area);
3002 Node* t = last_peel;
3003 while (true) {
3004 wl.push(t);
3005 if (t == head_clone) break;
3006 t = idom(t);
3007 }
3008 while (wl.size() > 0) {
3009 Node* tt = wl.pop();
3010 if (tt == head) tty->print_cr("orig head");
3011 else if (tt == new_head_clone) tty->print_cr("new head");
3012 else if (tt == head_clone) tty->print_cr("clone head");
3013 tt->dump();
3014 }
3015 }
3016 #endif
3017 return true;
3018 }
3019
3020 //------------------------------reorg_offsets----------------------------------
3021 // Reorganize offset computations to lower register pressure. Mostly
3022 // prevent loop-fallout uses of the pre-incremented trip counter (which are
3023 // then alive with the post-incremented trip counter forcing an extra
3024 // register move)
3025 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
3026 // Perform it only for canonical counted loops.
3027 // Loop's shape could be messed up by iteration_split_impl.
3028 if (!loop->_head->is_CountedLoop())
3029 return;
3030 if (!loop->_head->as_Loop()->is_valid_counted_loop())
3031 return;
3032
3033 CountedLoopNode *cl = loop->_head->as_CountedLoop();
3034 CountedLoopEndNode *cle = cl->loopexit();
3035 Node *exit = cle->proj_out(false);
3036 Node *phi = cl->phi();
3037
3038 // Check for the special case of folks using the pre-incremented
3039 // trip-counter on the fall-out path (forces the pre-incremented
3040 // and post-incremented trip counter to be live at the same time).
3041 // Fix this by adjusting to use the post-increment trip counter.
3042
3043 bool progress = true;
3044 while (progress) {
3045 progress = false;
3046 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3047 Node* use = phi->fast_out(i); // User of trip-counter
3048 if (!has_ctrl(use)) continue;
3049 Node *u_ctrl = get_ctrl(use);
3050 if (use->is_Phi()) {
3051 u_ctrl = NULL;
3052 for (uint j = 1; j < use->req(); j++)
3053 if (use->in(j) == phi)
3054 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
3055 }
3056 IdealLoopTree *u_loop = get_loop(u_ctrl);
3057 // Look for loop-invariant use
3058 if (u_loop == loop) continue;
3059 if (loop->is_member(u_loop)) continue;
3060 // Check that use is live out the bottom. Assuming the trip-counter
3061 // update is right at the bottom, uses of of the loop middle are ok.
3062 if (dom_lca(exit, u_ctrl) != exit) continue;
3063 // Hit! Refactor use to use the post-incremented tripcounter.
3064 // Compute a post-increment tripcounter.
3065 Node *opaq = new Opaque2Node( C, cle->incr() );
3066 register_new_node(opaq, exit);
3067 Node *neg_stride = _igvn.intcon(-cle->stride_con());
3068 set_ctrl(neg_stride, C->root());
3069 Node *post = new AddINode( opaq, neg_stride);
3070 register_new_node(post, exit);
3071 _igvn.rehash_node_delayed(use);
3072 for (uint j = 1; j < use->req(); j++) {
3073 if (use->in(j) == phi)
3074 use->set_req(j, post);
3075 }
3076 // Since DU info changed, rerun loop
3077 progress = true;
3078 break;
3079 }
3080 }
3081
3082 }