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