1 /*
2 * Copyright (c) 1999, 2018, 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 "memory/resourceArea.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/castnode.hpp"
31 #include "opto/connode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/loopnode.hpp"
35 #include "opto/matcher.hpp"
36 #include "opto/mulnode.hpp"
37 #include "opto/movenode.hpp"
38 #include "opto/opaquenode.hpp"
39 #include "opto/rootnode.hpp"
40 #include "opto/subnode.hpp"
41
42 //=============================================================================
43 //------------------------------split_thru_phi---------------------------------
44 // Split Node 'n' through merge point if there is enough win.
45 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
46 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
47 // ConvI2L may have type information on it which is unsafe to push up
48 // so disable this for now
49 return NULL;
50 }
51
52 // Splitting range check CastIIs through a loop induction Phi can
53 // cause new Phis to be created that are left unrelated to the loop
54 // induction Phi and prevent optimizations (vectorization)
55 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() &&
56 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) {
57 return NULL;
58 }
59
60 int wins = 0;
61 assert(!n->is_CFG(), "");
62 assert(region->is_Region(), "");
63
64 const Type* type = n->bottom_type();
65 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr();
66 Node *phi;
67 if (t_oop != NULL && t_oop->is_known_instance_field()) {
68 int iid = t_oop->instance_id();
69 int index = C->get_alias_index(t_oop);
70 int offset = t_oop->offset();
71 phi = new PhiNode(region, type, NULL, iid, index, offset);
72 } else {
73 phi = PhiNode::make_blank(region, n);
74 }
75 uint old_unique = C->unique();
76 for (uint i = 1; i < region->req(); i++) {
77 Node *x;
78 Node* the_clone = NULL;
79 if (region->in(i) == C->top()) {
80 x = C->top(); // Dead path? Use a dead data op
81 } else {
82 x = n->clone(); // Else clone up the data op
83 the_clone = x; // Remember for possible deletion.
84 // Alter data node to use pre-phi inputs
85 if (n->in(0) == region)
86 x->set_req( 0, region->in(i) );
87 for (uint j = 1; j < n->req(); j++) {
88 Node *in = n->in(j);
89 if (in->is_Phi() && in->in(0) == region)
90 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
91 }
92 }
93 // Check for a 'win' on some paths
94 const Type *t = x->Value(&_igvn);
95
96 bool singleton = t->singleton();
97
98 // A TOP singleton indicates that there are no possible values incoming
99 // along a particular edge. In most cases, this is OK, and the Phi will
100 // be eliminated later in an Ideal call. However, we can't allow this to
101 // happen if the singleton occurs on loop entry, as the elimination of
102 // the PhiNode may cause the resulting node to migrate back to a previous
103 // loop iteration.
104 if (singleton && t == Type::TOP) {
105 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
106 // irreducible loop may not be indicated by an affirmative is_Loop());
107 // therefore, the only top we can split thru a phi is on a backedge of
108 // a loop.
109 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
110 }
111
112 if (singleton) {
113 wins++;
114 x = ((PhaseGVN&)_igvn).makecon(t);
115 } else {
116 // We now call Identity to try to simplify the cloned node.
117 // Note that some Identity methods call phase->type(this).
118 // Make sure that the type array is big enough for
119 // our new node, even though we may throw the node away.
120 // (Note: This tweaking with igvn only works because x is a new node.)
121 _igvn.set_type(x, t);
122 // If x is a TypeNode, capture any more-precise type permanently into Node
123 // otherwise it will be not updated during igvn->transform since
124 // igvn->type(x) is set to x->Value() already.
125 x->raise_bottom_type(t);
126 Node *y = x->Identity(&_igvn);
127 if (y != x) {
128 wins++;
129 x = y;
130 } else {
131 y = _igvn.hash_find(x);
132 if (y) {
133 wins++;
134 x = y;
135 } else {
136 // Else x is a new node we are keeping
137 // We do not need register_new_node_with_optimizer
138 // because set_type has already been called.
139 _igvn._worklist.push(x);
140 }
141 }
142 }
143 if (x != the_clone && the_clone != NULL)
144 _igvn.remove_dead_node(the_clone);
145 phi->set_req( i, x );
146 }
147 // Too few wins?
148 if (wins <= policy) {
149 _igvn.remove_dead_node(phi);
150 return NULL;
151 }
152
153 // Record Phi
154 register_new_node( phi, region );
155
156 for (uint i2 = 1; i2 < phi->req(); i2++) {
157 Node *x = phi->in(i2);
158 // If we commoned up the cloned 'x' with another existing Node,
159 // the existing Node picks up a new use. We need to make the
160 // existing Node occur higher up so it dominates its uses.
161 Node *old_ctrl;
162 IdealLoopTree *old_loop;
163
164 if (x->is_Con()) {
165 // Constant's control is always root.
166 set_ctrl(x, C->root());
167 continue;
168 }
169 // The occasional new node
170 if (x->_idx >= old_unique) { // Found a new, unplaced node?
171 old_ctrl = NULL;
172 old_loop = NULL; // Not in any prior loop
173 } else {
174 old_ctrl = get_ctrl(x);
175 old_loop = get_loop(old_ctrl); // Get prior loop
176 }
177 // New late point must dominate new use
178 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
179 if (new_ctrl == old_ctrl) // Nothing is changed
180 continue;
181
182 IdealLoopTree *new_loop = get_loop(new_ctrl);
183
184 // Don't move x into a loop if its uses are
185 // outside of loop. Otherwise x will be cloned
186 // for each use outside of this loop.
187 IdealLoopTree *use_loop = get_loop(region);
188 if (!new_loop->is_member(use_loop) &&
189 (old_loop == NULL || !new_loop->is_member(old_loop))) {
190 // Take early control, later control will be recalculated
191 // during next iteration of loop optimizations.
192 new_ctrl = get_early_ctrl(x);
193 new_loop = get_loop(new_ctrl);
194 }
195 // Set new location
196 set_ctrl(x, new_ctrl);
197 // If changing loop bodies, see if we need to collect into new body
198 if (old_loop != new_loop) {
199 if (old_loop && !old_loop->_child)
200 old_loop->_body.yank(x);
201 if (!new_loop->_child)
202 new_loop->_body.push(x); // Collect body info
203 }
204 }
205
206 return phi;
207 }
208
209 //------------------------------dominated_by------------------------------------
210 // Replace the dominated test with an obvious true or false. Place it on the
211 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
212 // live path up to the dominating control.
213 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) {
214 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
215
216 // prevdom is the dominating projection of the dominating test.
217 assert( iff->is_If(), "" );
218 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added");
219 int pop = prevdom->Opcode();
220 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
221 if (flip) {
222 if (pop == Op_IfTrue)
223 pop = Op_IfFalse;
224 else
225 pop = Op_IfTrue;
226 }
227 // 'con' is set to true or false to kill the dominated test.
228 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
229 set_ctrl(con, C->root()); // Constant gets a new use
230 // Hack the dominated test
231 _igvn.replace_input_of(iff, 1, con);
232
233 // If I dont have a reachable TRUE and FALSE path following the IfNode then
234 // I can assume this path reaches an infinite loop. In this case it's not
235 // important to optimize the data Nodes - either the whole compilation will
236 // be tossed or this path (and all data Nodes) will go dead.
237 if (iff->outcnt() != 2) return;
238
239 // Make control-dependent data Nodes on the live path (path that will remain
240 // once the dominated IF is removed) become control-dependent on the
241 // dominating projection.
242 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue);
243
244 // Loop predicates may have depending checks which should not
245 // be skipped. For example, range check predicate has two checks
246 // for lower and upper bounds.
247 if (dp == NULL)
248 return;
249
250 ProjNode* dp_proj = dp->as_Proj();
251 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj();
252 if (exclude_loop_predicate &&
253 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL ||
254 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) {
255 // If this is a range check (IfNode::is_range_check), do not
256 // reorder because Compile::allow_range_check_smearing might have
257 // changed the check.
258 return; // Let IGVN transformation change control dependence.
259 }
260
261 IdealLoopTree *old_loop = get_loop(dp);
262
263 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
264 Node* cd = dp->fast_out(i); // Control-dependent node
265 if (cd->depends_only_on_test()) {
266 assert(cd->in(0) == dp, "");
267 _igvn.replace_input_of(cd, 0, prevdom);
268 set_early_ctrl(cd);
269 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
270 if (old_loop != new_loop) {
271 if (!old_loop->_child) old_loop->_body.yank(cd);
272 if (!new_loop->_child) new_loop->_body.push(cd);
273 }
274 --i;
275 --imax;
276 }
277 }
278 }
279
280 //------------------------------has_local_phi_input----------------------------
281 // Return TRUE if 'n' has Phi inputs from its local block and no other
282 // block-local inputs (all non-local-phi inputs come from earlier blocks)
283 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
284 Node *n_ctrl = get_ctrl(n);
285 // See if some inputs come from a Phi in this block, or from before
286 // this block.
287 uint i;
288 for( i = 1; i < n->req(); i++ ) {
289 Node *phi = n->in(i);
290 if( phi->is_Phi() && phi->in(0) == n_ctrl )
291 break;
292 }
293 if( i >= n->req() )
294 return NULL; // No Phi inputs; nowhere to clone thru
295
296 // Check for inputs created between 'n' and the Phi input. These
297 // must split as well; they have already been given the chance
298 // (courtesy of a post-order visit) and since they did not we must
299 // recover the 'cost' of splitting them by being very profitable
300 // when splitting 'n'. Since this is unlikely we simply give up.
301 for( i = 1; i < n->req(); i++ ) {
302 Node *m = n->in(i);
303 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
304 // We allow the special case of AddP's with no local inputs.
305 // This allows us to split-up address expressions.
306 if (m->is_AddP() &&
307 get_ctrl(m->in(2)) != n_ctrl &&
308 get_ctrl(m->in(3)) != n_ctrl) {
309 // Move the AddP up to dominating point
310 Node* c = find_non_split_ctrl(idom(n_ctrl));
311 if (c->is_OuterStripMinedLoop()) {
312 c->as_Loop()->verify_strip_mined(1);
313 c = c->in(LoopNode::EntryControl);
314 }
315 set_ctrl_and_loop(m, c);
316 continue;
317 }
318 return NULL;
319 }
320 assert(m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
321 }
322
323 return n_ctrl;
324 }
325
326 //------------------------------remix_address_expressions----------------------
327 // Rework addressing expressions to get the most loop-invariant stuff
328 // moved out. We'd like to do all associative operators, but it's especially
329 // important (common) to do address expressions.
330 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
331 if (!has_ctrl(n)) return NULL;
332 Node *n_ctrl = get_ctrl(n);
333 IdealLoopTree *n_loop = get_loop(n_ctrl);
334
335 // See if 'n' mixes loop-varying and loop-invariant inputs and
336 // itself is loop-varying.
337
338 // Only interested in binary ops (and AddP)
339 if( n->req() < 3 || n->req() > 4 ) return NULL;
340
341 Node *n1_ctrl = get_ctrl(n->in( 1));
342 Node *n2_ctrl = get_ctrl(n->in( 2));
343 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
344 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
345 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
346 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
347
348 // Does one of my inputs spin in a tighter loop than self?
349 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
350 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
351 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
352 return NULL; // Leave well enough alone
353
354 // Is at least one of my inputs loop-invariant?
355 if( n1_loop == n_loop &&
356 n2_loop == n_loop &&
357 n3_loop == n_loop )
358 return NULL; // No loop-invariant inputs
359
360
361 int n_op = n->Opcode();
362
363 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
364 if( n_op == Op_LShiftI ) {
365 // Scale is loop invariant
366 Node *scale = n->in(2);
367 Node *scale_ctrl = get_ctrl(scale);
368 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
369 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
370 return NULL;
371 const TypeInt *scale_t = scale->bottom_type()->isa_int();
372 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
373 return NULL; // Dont bother with byte/short masking
374 // Add must vary with loop (else shift would be loop-invariant)
375 Node *add = n->in(1);
376 Node *add_ctrl = get_ctrl(add);
377 IdealLoopTree *add_loop = get_loop(add_ctrl);
378 //assert( n_loop == add_loop, "" );
379 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
380
381 // Convert I-V into I+ (0-V); same for V-I
382 if( add->Opcode() == Op_SubI &&
383 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
384 Node *zero = _igvn.intcon(0);
385 set_ctrl(zero, C->root());
386 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) );
387 register_new_node( neg, get_ctrl(add->in(2) ) );
388 add = new AddINode( add->in(1), neg );
389 register_new_node( add, add_ctrl );
390 }
391 if( add->Opcode() != Op_AddI ) return NULL;
392 // See if one add input is loop invariant
393 Node *add_var = add->in(1);
394 Node *add_var_ctrl = get_ctrl(add_var);
395 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
396 Node *add_invar = add->in(2);
397 Node *add_invar_ctrl = get_ctrl(add_invar);
398 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
399 if( add_var_loop == n_loop ) {
400 } else if( add_invar_loop == n_loop ) {
401 // Swap to find the invariant part
402 add_invar = add_var;
403 add_invar_ctrl = add_var_ctrl;
404 add_invar_loop = add_var_loop;
405 add_var = add->in(2);
406 Node *add_var_ctrl = get_ctrl(add_var);
407 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
408 } else // Else neither input is loop invariant
409 return NULL;
410 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
411 return NULL; // No invariant part of the add?
412
413 // Yes! Reshape address expression!
414 Node *inv_scale = new LShiftINode( add_invar, scale );
415 Node *inv_scale_ctrl =
416 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
417 add_invar_ctrl : scale_ctrl;
418 register_new_node( inv_scale, inv_scale_ctrl );
419 Node *var_scale = new LShiftINode( add_var, scale );
420 register_new_node( var_scale, n_ctrl );
421 Node *var_add = new AddINode( var_scale, inv_scale );
422 register_new_node( var_add, n_ctrl );
423 _igvn.replace_node( n, var_add );
424 return var_add;
425 }
426
427 // Replace (I+V) with (V+I)
428 if( n_op == Op_AddI ||
429 n_op == Op_AddL ||
430 n_op == Op_AddF ||
431 n_op == Op_AddD ||
432 n_op == Op_MulI ||
433 n_op == Op_MulL ||
434 n_op == Op_MulF ||
435 n_op == Op_MulD ) {
436 if( n2_loop == n_loop ) {
437 assert( n1_loop != n_loop, "" );
438 n->swap_edges(1, 2);
439 }
440 }
441
442 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
443 // but not if I2 is a constant.
444 if( n_op == Op_AddP ) {
445 if( n2_loop == n_loop && n3_loop != n_loop ) {
446 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
447 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
448 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
449 IdealLoopTree *n22loop = get_loop( n22_ctrl );
450 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
451 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
452 n23_loop == n_loop ) {
453 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
454 // Stuff new AddP in the loop preheader
455 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
456 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) );
457 register_new_node( add2, n_ctrl );
458 _igvn.replace_node( n, add2 );
459 return add2;
460 }
461 }
462 }
463
464 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
465 if (n2_loop != n_loop && n3_loop == n_loop) {
466 if (n->in(3)->Opcode() == Op_AddX) {
467 Node *V = n->in(3)->in(1);
468 Node *I = n->in(3)->in(2);
469 if (is_member(n_loop,get_ctrl(V))) {
470 } else {
471 Node *tmp = V; V = I; I = tmp;
472 }
473 if (!is_member(n_loop,get_ctrl(I))) {
474 Node *add1 = new AddPNode(n->in(1), n->in(2), I);
475 // Stuff new AddP in the loop preheader
476 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl));
477 Node *add2 = new AddPNode(n->in(1), add1, V);
478 register_new_node(add2, n_ctrl);
479 _igvn.replace_node(n, add2);
480 return add2;
481 }
482 }
483 }
484 }
485
486 return NULL;
487 }
488
489 //------------------------------conditional_move-------------------------------
490 // Attempt to replace a Phi with a conditional move. We have some pretty
491 // strict profitability requirements. All Phis at the merge point must
492 // be converted, so we can remove the control flow. We need to limit the
493 // number of c-moves to a small handful. All code that was in the side-arms
494 // of the CFG diamond is now speculatively executed. This code has to be
495 // "cheap enough". We are pretty much limited to CFG diamonds that merge
496 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
497 Node *PhaseIdealLoop::conditional_move( Node *region ) {
498
499 assert(region->is_Region(), "sanity check");
500 if (region->req() != 3) return NULL;
501
502 // Check for CFG diamond
503 Node *lp = region->in(1);
504 Node *rp = region->in(2);
505 if (!lp || !rp) return NULL;
506 Node *lp_c = lp->in(0);
507 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
508 IfNode *iff = lp_c->as_If();
509
510 // Check for ops pinned in an arm of the diamond.
511 // Can't remove the control flow in this case
512 if (lp->outcnt() > 1) return NULL;
513 if (rp->outcnt() > 1) return NULL;
514
515 IdealLoopTree* r_loop = get_loop(region);
516 assert(r_loop == get_loop(iff), "sanity");
517 // Always convert to CMOVE if all results are used only outside this loop.
518 bool used_inside_loop = (r_loop == _ltree_root);
519
520 // Check profitability
521 int cost = 0;
522 int phis = 0;
523 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
524 Node *out = region->fast_out(i);
525 if (!out->is_Phi()) continue; // Ignore other control edges, etc
526 phis++;
527 PhiNode* phi = out->as_Phi();
528 BasicType bt = phi->type()->basic_type();
529 switch (bt) {
530 case T_DOUBLE:
531 case T_FLOAT:
532 if (C->use_cmove()) {
533 continue; //TODO: maybe we want to add some cost
534 }
535 cost += Matcher::float_cmove_cost(); // Could be very expensive
536 break;
537 case T_LONG: {
538 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
539 }
540 case T_INT: // These all CMOV fine
541 case T_ADDRESS: { // (RawPtr)
542 cost++;
543 break;
544 }
545 case T_NARROWOOP: // Fall through
546 case T_OBJECT: { // Base oops are OK, but not derived oops
547 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
548 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
549 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
550 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
551 // have a Phi for the base here that we convert to a CMOVE all is well
552 // and good. But if the base is dead, we'll not make a CMOVE. Later
553 // the allocator will have to produce a base by creating a CMOVE of the
554 // relevant bases. This puts the allocator in the business of
555 // manufacturing expensive instructions, generally a bad plan.
556 // Just Say No to Conditionally-Moved Derived Pointers.
557 if (tp && tp->offset() != 0)
558 return NULL;
559 cost++;
560 break;
561 }
562 default:
563 return NULL; // In particular, can't do memory or I/O
564 }
565 // Add in cost any speculative ops
566 for (uint j = 1; j < region->req(); j++) {
567 Node *proj = region->in(j);
568 Node *inp = phi->in(j);
569 if (get_ctrl(inp) == proj) { // Found local op
570 cost++;
571 // Check for a chain of dependent ops; these will all become
572 // speculative in a CMOV.
573 for (uint k = 1; k < inp->req(); k++)
574 if (get_ctrl(inp->in(k)) == proj)
575 cost += ConditionalMoveLimit; // Too much speculative goo
576 }
577 }
578 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
579 // This will likely Split-If, a higher-payoff operation.
580 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
581 Node* use = phi->fast_out(k);
582 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
583 cost += ConditionalMoveLimit;
584 // Is there a use inside the loop?
585 // Note: check only basic types since CMoveP is pinned.
586 if (!used_inside_loop && is_java_primitive(bt)) {
587 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
588 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
589 used_inside_loop = true;
590 }
591 }
592 }
593 }//for
594 Node* bol = iff->in(1);
595 assert(bol->Opcode() == Op_Bool, "");
596 int cmp_op = bol->in(1)->Opcode();
597 // It is expensive to generate flags from a float compare.
598 // Avoid duplicated float compare.
599 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
600
601 float infrequent_prob = PROB_UNLIKELY_MAG(3);
602 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
603 if (used_inside_loop) {
604 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
605
606 // BlockLayoutByFrequency optimization moves infrequent branch
607 // from hot path. No point in CMOV'ing in such case (110 is used
608 // instead of 100 to take into account not exactness of float value).
609 if (BlockLayoutByFrequency) {
610 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
611 }
612 }
613 // Check for highly predictable branch. No point in CMOV'ing if
614 // we are going to predict accurately all the time.
615 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
616 //keep going
617 } else if (iff->_prob < infrequent_prob ||
618 iff->_prob > (1.0f - infrequent_prob))
619 return NULL;
620
621 // --------------
622 // Now replace all Phis with CMOV's
623 Node *cmov_ctrl = iff->in(0);
624 uint flip = (lp->Opcode() == Op_IfTrue);
625 Node_List wq;
626 while (1) {
627 PhiNode* phi = NULL;
628 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
629 Node *out = region->fast_out(i);
630 if (out->is_Phi()) {
631 phi = out->as_Phi();
632 break;
633 }
634 }
635 if (phi == NULL) break;
636 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
637 // Move speculative ops
638 wq.push(phi);
639 while (wq.size() > 0) {
640 Node *n = wq.pop();
641 for (uint j = 1; j < n->req(); j++) {
642 Node* m = n->in(j);
643 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) {
644 #ifndef PRODUCT
645 if (PrintOpto && VerifyLoopOptimizations) {
646 tty->print(" speculate: ");
647 m->dump();
648 }
649 #endif
650 set_ctrl(m, cmov_ctrl);
651 wq.push(m);
652 }
653 }
654 }
655 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
656 register_new_node( cmov, cmov_ctrl );
657 _igvn.replace_node( phi, cmov );
658 #ifndef PRODUCT
659 if (TraceLoopOpts) {
660 tty->print("CMOV ");
661 r_loop->dump_head();
662 if (Verbose) {
663 bol->in(1)->dump(1);
664 cmov->dump(1);
665 }
666 }
667 if (VerifyLoopOptimizations) verify();
668 #endif
669 }
670
671 // The useless CFG diamond will fold up later; see the optimization in
672 // RegionNode::Ideal.
673 _igvn._worklist.push(region);
674
675 return iff->in(1);
676 }
677
678 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
679 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
680 Node* u = m->fast_out(i);
681 if (u->is_CFG()) {
682 if (u->Opcode() == Op_NeverBranch) {
683 u = ((NeverBranchNode*)u)->proj_out(0);
684 enqueue_cfg_uses(u, wq);
685 } else {
686 wq.push(u);
687 }
688 }
689 }
690 }
691
692 // Try moving a store out of a loop, right before the loop
693 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
694 // Store has to be first in the loop body
695 IdealLoopTree *n_loop = get_loop(n_ctrl);
696 if (n->is_Store() && n_loop != _ltree_root &&
697 n_loop->is_loop() && n_loop->_head->is_Loop() &&
698 n->in(0) != NULL) {
699 Node* address = n->in(MemNode::Address);
700 Node* value = n->in(MemNode::ValueIn);
701 Node* mem = n->in(MemNode::Memory);
702 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
703 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
704
705 // - address and value must be loop invariant
706 // - memory must be a memory Phi for the loop
707 // - Store must be the only store on this memory slice in the
708 // loop: if there's another store following this one then value
709 // written at iteration i by the second store could be overwritten
710 // at iteration i+n by the first store: it's not safe to move the
711 // first store out of the loop
712 // - nothing must observe the memory Phi: it guarantees no read
713 // before the store, we are also guaranteed the store post
714 // dominates the loop head (ignoring a possible early
715 // exit). Otherwise there would be extra Phi involved between the
716 // loop's Phi and the store.
717 // - there must be no early exit from the loop before the Store
718 // (such an exit most of the time would be an extra use of the
719 // memory Phi but sometimes is a bottom memory Phi that takes the
720 // store as input).
721
722 if (!n_loop->is_member(address_loop) &&
723 !n_loop->is_member(value_loop) &&
724 mem->is_Phi() && mem->in(0) == n_loop->_head &&
725 mem->outcnt() == 1 &&
726 mem->in(LoopNode::LoopBackControl) == n) {
727
728 assert(n_loop->_tail != NULL, "need a tail");
729 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
730
731 // Verify that there's no early exit of the loop before the store.
732 bool ctrl_ok = false;
733 {
734 // Follow control from loop head until n, we exit the loop or
735 // we reach the tail
736 ResourceMark rm;
737 Unique_Node_List wq;
738 wq.push(n_loop->_head);
739
740 for (uint next = 0; next < wq.size(); ++next) {
741 Node *m = wq.at(next);
742 if (m == n->in(0)) {
743 ctrl_ok = true;
744 continue;
745 }
746 assert(!has_ctrl(m), "should be CFG");
747 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
748 ctrl_ok = false;
749 break;
750 }
751 enqueue_cfg_uses(m, wq);
752 if (wq.size() > 10) {
753 ctrl_ok = false;
754 break;
755 }
756 }
757 }
758 if (ctrl_ok) {
759 // move the Store
760 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
761 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
762 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
763 // Disconnect the phi now. An empty phi can confuse other
764 // optimizations in this pass of loop opts.
765 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
766 n_loop->_body.yank(mem);
767
768 set_ctrl_and_loop(n, n->in(0));
769
770 return n;
771 }
772 }
773 }
774 return NULL;
775 }
776
777 // Try moving a store out of a loop, right after the loop
778 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
779 if (n->is_Store() && n->in(0) != NULL) {
780 Node *n_ctrl = get_ctrl(n);
781 IdealLoopTree *n_loop = get_loop(n_ctrl);
782 // Store must be in a loop
783 if (n_loop != _ltree_root && !n_loop->_irreducible) {
784 Node* address = n->in(MemNode::Address);
785 Node* value = n->in(MemNode::ValueIn);
786 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
787 // address must be loop invariant
788 if (!n_loop->is_member(address_loop)) {
789 // Store must be last on this memory slice in the loop and
790 // nothing in the loop must observe it
791 Node* phi = NULL;
792 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
793 Node* u = n->fast_out(i);
794 if (has_ctrl(u)) { // control use?
795 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
796 if (!n_loop->is_member(u_loop)) {
797 continue;
798 }
799 if (u->is_Phi() && u->in(0) == n_loop->_head) {
800 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
801 // multiple phis on the same slice are possible
802 if (phi != NULL) {
803 return;
804 }
805 phi = u;
806 continue;
807 }
808 }
809 return;
810 }
811 if (phi != NULL) {
812 // Nothing in the loop before the store (next iteration)
813 // must observe the stored value
814 bool mem_ok = true;
815 {
816 ResourceMark rm;
817 Unique_Node_List wq;
818 wq.push(phi);
819 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
820 Node *m = wq.at(next);
821 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
822 Node* u = m->fast_out(i);
823 if (u->is_Store() || u->is_Phi()) {
824 if (u != n) {
825 wq.push(u);
826 mem_ok = (wq.size() <= 10);
827 }
828 } else {
829 mem_ok = false;
830 break;
831 }
832 }
833 }
834 }
835 if (mem_ok) {
836 // Move the store out of the loop if the LCA of all
837 // users (except for the phi) is outside the loop.
838 Node* hook = new Node(1);
839 _igvn.rehash_node_delayed(phi);
840 int count = phi->replace_edge(n, hook);
841 assert(count > 0, "inconsistent phi");
842
843 // Compute latest point this store can go
844 Node* lca = get_late_ctrl(n, get_ctrl(n));
845 if (n_loop->is_member(get_loop(lca))) {
846 // LCA is in the loop - bail out
847 _igvn.replace_node(hook, n);
848 return;
849 }
850 #ifdef ASSERT
851 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
852 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
853 n_loop->_head->as_Loop()->verify_strip_mined(1);
854 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
855 IdealLoopTree* outer_loop = get_loop(outer);
856 assert(n_loop->_parent == outer_loop, "broken loop tree");
857 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
858 }
859 #endif
860
861 // Move store out of the loop
862 _igvn.replace_node(hook, n->in(MemNode::Memory));
863 _igvn.replace_input_of(n, 0, lca);
864 set_ctrl_and_loop(n, lca);
865
866 // Disconnect the phi now. An empty phi can confuse other
867 // optimizations in this pass of loop opts..
868 if (phi->in(LoopNode::LoopBackControl) == phi) {
869 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
870 n_loop->_body.yank(phi);
871 }
872 }
873 }
874 }
875 }
876 }
877 }
878
879 //------------------------------split_if_with_blocks_pre-----------------------
880 // Do the real work in a non-recursive function. Data nodes want to be
881 // cloned in the pre-order so they can feed each other nicely.
882 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
883 // Cloning these guys is unlikely to win
884 int n_op = n->Opcode();
885
886 if( n_op == Op_MergeMem ) return n;
887 if( n->is_Proj() ) return n;
888 // Do not clone-up CmpFXXX variations, as these are always
889 // followed by a CmpI
890 if( n->is_Cmp() ) return n;
891 // Attempt to use a conditional move instead of a phi/branch
892 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
893 Node *cmov = conditional_move( n );
894 if( cmov ) return cmov;
895 }
896
897 if (n->is_CFG() || n->is_LoadStore()) {
898 return n;
899 }
900 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
901 n_op == Op_Opaque2 ) {
902 if( !C->major_progress() ) // If chance of no more loop opts...
903 _igvn._worklist.push(n); // maybe we'll remove them
904 return n;
905 }
906
907 if( n->is_Con() ) return n; // No cloning for Con nodes
908
909 Node *n_ctrl = get_ctrl(n);
910 if( !n_ctrl ) return n; // Dead node
911
912 Node* res = try_move_store_before_loop(n, n_ctrl);
913 if (res != NULL) {
914 return n;
915 }
916
917 // Attempt to remix address expressions for loop invariants
918 Node *m = remix_address_expressions( n );
919 if( m ) return m;
920
921 if (n->is_ConstraintCast()) {
922 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
923 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
924 // Node control inputs don't necessarily agree with loop control info (due to
925 // transformations happened in between), thus additional dominance check is needed
926 // to keep loop info valid.
927 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
928 _igvn.replace_node(n, dom_cast);
929 return dom_cast;
930 }
931 }
932
933 // Determine if the Node has inputs from some local Phi.
934 // Returns the block to clone thru.
935 Node *n_blk = has_local_phi_input( n );
936 if( !n_blk ) return n;
937
938 // Do not clone the trip counter through on a CountedLoop
939 // (messes up the canonical shape).
940 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
941
942 // Check for having no control input; not pinned. Allow
943 // dominating control.
944 if (n->in(0)) {
945 Node *dom = idom(n_blk);
946 if (dom_lca(n->in(0), dom) != n->in(0)) {
947 return n;
948 }
949 }
950 // Policy: when is it profitable. You must get more wins than
951 // policy before it is considered profitable. Policy is usually 0,
952 // so 1 win is considered profitable. Big merges will require big
953 // cloning, so get a larger policy.
954 int policy = n_blk->req() >> 2;
955
956 // If the loop is a candidate for range check elimination,
957 // delay splitting through it's phi until a later loop optimization
958 if (n_blk->is_CountedLoop()) {
959 IdealLoopTree *lp = get_loop(n_blk);
960 if (lp && lp->_rce_candidate) {
961 return n;
962 }
963 }
964
965 // Use same limit as split_if_with_blocks_post
966 if( C->live_nodes() > 35000 ) return n; // Method too big
967
968 // Split 'n' through the merge point if it is profitable
969 Node *phi = split_thru_phi( n, n_blk, policy );
970 if (!phi) return n;
971
972 // Found a Phi to split thru!
973 // Replace 'n' with the new phi
974 _igvn.replace_node( n, phi );
975 // Moved a load around the loop, 'en-registering' something.
976 if (n_blk->is_Loop() && n->is_Load() &&
977 !phi->in(LoopNode::LoopBackControl)->is_Load())
978 C->set_major_progress();
979
980 return phi;
981 }
982
983 static bool merge_point_too_heavy(Compile* C, Node* region) {
984 // Bail out if the region and its phis have too many users.
985 int weight = 0;
986 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
987 weight += region->fast_out(i)->outcnt();
988 }
989 int nodes_left = C->max_node_limit() - C->live_nodes();
990 if (weight * 8 > nodes_left) {
991 if (PrintOpto) {
992 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
993 }
994 return true;
995 } else {
996 return false;
997 }
998 }
999
1000 static bool merge_point_safe(Node* region) {
1001 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1002 // having a PhiNode input. This sidesteps the dangerous case where the split
1003 // ConvI2LNode may become TOP if the input Value() does not
1004 // overlap the ConvI2L range, leaving a node which may not dominate its
1005 // uses.
1006 // A better fix for this problem can be found in the BugTraq entry, but
1007 // expediency for Mantis demands this hack.
1008 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop
1009 // split_if_with_blocks from splitting a block because we could not move around
1010 // the FastLockNode.
1011 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1012 Node* n = region->fast_out(i);
1013 if (n->is_Phi()) {
1014 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1015 Node* m = n->fast_out(j);
1016 if (m->is_FastLock())
1017 return false;
1018 #ifdef _LP64
1019 if (m->Opcode() == Op_ConvI2L)
1020 return false;
1021 if (m->is_CastII() && m->isa_CastII()->has_range_check()) {
1022 return false;
1023 }
1024 #endif
1025 }
1026 }
1027 }
1028 return true;
1029 }
1030
1031
1032 //------------------------------place_near_use---------------------------------
1033 // Place some computation next to use but not inside inner loops.
1034 // For inner loop uses move it to the preheader area.
1035 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const {
1036 IdealLoopTree *u_loop = get_loop( useblock );
1037 return (u_loop->_irreducible || u_loop->_child)
1038 ? useblock
1039 : u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
1040 }
1041
1042
1043 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1044 if (!n->is_If() || n->is_CountedLoopEnd()) {
1045 return false;
1046 }
1047 if (!n->in(0)->is_Region()) {
1048 return false;
1049 }
1050 Node* region = n->in(0);
1051 Node* dom = idom(region);
1052 if (!dom->is_If() || dom->in(1) != n->in(1)) {
1053 return false;
1054 }
1055 IfNode* dom_if = dom->as_If();
1056 Node* proj_true = dom_if->proj_out(1);
1057 Node* proj_false = dom_if->proj_out(0);
1058
1059 for (uint i = 1; i < region->req(); i++) {
1060 if (is_dominator(proj_true, region->in(i))) {
1061 continue;
1062 }
1063 if (is_dominator(proj_false, region->in(i))) {
1064 continue;
1065 }
1066 return false;
1067 }
1068
1069 return true;
1070 }
1071
1072 bool PhaseIdealLoop::can_split_if(Node *n_ctrl) {
1073 if (C->live_nodes() > 35000) {
1074 return false; // Method too big
1075 }
1076
1077 // Do not do 'split-if' if irreducible loops are present.
1078 if (_has_irreducible_loops) {
1079 return false;
1080 }
1081
1082 if (merge_point_too_heavy(C, n_ctrl)) {
1083 return false;
1084 }
1085
1086 // Do not do 'split-if' if some paths are dead. First do dead code
1087 // elimination and then see if its still profitable.
1088 for (uint i = 1; i < n_ctrl->req(); i++) {
1089 if (n_ctrl->in(i) == C->top()) {
1090 return false;
1091 }
1092 }
1093
1094 // If trying to do a 'Split-If' at the loop head, it is only
1095 // profitable if the cmp folds up on BOTH paths. Otherwise we
1096 // risk peeling a loop forever.
1097
1098 // CNC - Disabled for now. Requires careful handling of loop
1099 // body selection for the cloned code. Also, make sure we check
1100 // for any input path not being in the same loop as n_ctrl. For
1101 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1102 // because the alternative loop entry points won't be converted
1103 // into LoopNodes.
1104 IdealLoopTree *n_loop = get_loop(n_ctrl);
1105 for (uint j = 1; j < n_ctrl->req(); j++) {
1106 if (get_loop(n_ctrl->in(j)) != n_loop) {
1107 return false;
1108 }
1109 }
1110
1111 // Check for safety of the merge point.
1112 if (!merge_point_safe(n_ctrl)) {
1113 return false;
1114 }
1115
1116 return true;
1117 }
1118
1119 bool PhaseIdealLoop::replace_with_dominating_barrier(LoadBarrierNode* lb, bool last_round) {
1120 LoadBarrierNode* lb2 = lb->has_dominating_barrier(this, false, last_round);
1121 if (lb2 != NULL) {
1122 if (lb->in(LoadBarrierNode::Oop) != lb2->in(LoadBarrierNode::Oop)) {
1123 assert(lb->in(LoadBarrierNode::Address) == lb2->in(LoadBarrierNode::Address), "");
1124 _igvn.replace_input_of(lb, LoadBarrierNode::Similar, lb2->proj_out(LoadBarrierNode::Oop));
1125 C->set_major_progress();
1126 } else {
1127 // That transformation may cause the Similar edge on dominated load barriers to be invalid
1128 lb->fix_similar_in_uses(&_igvn);
1129
1130 Node* val = lb->proj_out(LoadBarrierNode::Oop);
1131 assert(lb2->has_true_uses(), "");
1132 assert(lb2->in(LoadBarrierNode::Oop) == lb->in(LoadBarrierNode::Oop), "");
1133
1134 lazy_update(lb, lb->in(LoadBarrierNode::Control));
1135 lazy_replace(lb->proj_out(LoadBarrierNode::Control), lb->in(LoadBarrierNode::Control));
1136 _igvn.replace_node(val, lb2->proj_out(LoadBarrierNode::Oop));
1137
1138 return true;
1139 }
1140 }
1141 return false;
1142 }
1143
1144 Node* PhaseIdealLoop::find_dominating_memory(Node* mem, Node* dom, int i) {
1145 assert(dom->is_Region() || i == -1, "");
1146 Node* m = mem;
1147 while(is_dominator(dom, has_ctrl(m) ? get_ctrl(m) : m->in(0))) {
1148 if (m->is_Mem()) {
1149 assert(m->as_Mem()->adr_type() == TypeRawPtr::BOTTOM, "");
1150 m = m->in(MemNode::Memory);
1151 } else if (m->is_MergeMem()) {
1152 m = m->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1153 } else if (m->is_Phi()) {
1154 if (m->in(0) == dom && i != -1) {
1155 m = m->in(i);
1156 break;
1157 } else {
1158 m = m->in(LoopNode::EntryControl);
1159 }
1160 } else if (m->is_Proj()) {
1161 m = m->in(0);
1162 } else if (m->is_SafePoint() || m->is_MemBar()) {
1163 m = m->in(TypeFunc::Memory);
1164 } else {
1165 #ifdef ASSERT
1166 m->dump();
1167 #endif
1168 ShouldNotReachHere();
1169 }
1170 }
1171 return m;
1172 }
1173
1174 LoadBarrierNode* PhaseIdealLoop::clone_load_barrier(LoadBarrierNode* lb, Node* ctl, Node* mem, Node* oop_in) {
1175 Node* the_clone = lb->clone();
1176 the_clone->set_req(LoadBarrierNode::Control, ctl);
1177 the_clone->set_req(LoadBarrierNode::Memory, mem);
1178 if (oop_in != NULL) {
1179 the_clone->set_req(LoadBarrierNode::Oop, oop_in);
1180 }
1181
1182 LoadBarrierNode* new_lb = the_clone->as_LoadBarrier();
1183 _igvn.register_new_node_with_optimizer(new_lb);
1184 IdealLoopTree *loop = get_loop(new_lb->in(0));
1185 set_ctrl(new_lb, new_lb->in(0));
1186 set_loop(new_lb, loop);
1187 set_idom(new_lb, new_lb->in(0), dom_depth(new_lb->in(0))+1);
1188 if (!loop->_child) {
1189 loop->_body.push(new_lb);
1190 }
1191
1192 Node* proj_ctl = new ProjNode(new_lb, LoadBarrierNode::Control);
1193 _igvn.register_new_node_with_optimizer(proj_ctl);
1194 set_ctrl(proj_ctl, proj_ctl->in(0));
1195 set_loop(proj_ctl, loop);
1196 set_idom(proj_ctl, new_lb, dom_depth(new_lb)+1);
1197 if (!loop->_child) {
1198 loop->_body.push(proj_ctl);
1199 }
1200
1201 Node* proj_oop = new ProjNode(new_lb, LoadBarrierNode::Oop);
1202 register_new_node(proj_oop, new_lb);
1203
1204 if (!new_lb->in(LoadBarrierNode::Similar)->is_top()) {
1205 LoadBarrierNode* similar = new_lb->in(LoadBarrierNode::Similar)->in(0)->as_LoadBarrier();
1206 if (!is_dominator(similar, ctl)) {
1207 _igvn.replace_input_of(new_lb, LoadBarrierNode::Similar, C->top());
1208 }
1209 }
1210
1211 return new_lb;
1212 }
1213
1214 void PhaseIdealLoop::replace_barrier(LoadBarrierNode* lb, Node* new_val) {
1215 Node* val = lb->proj_out(LoadBarrierNode::Oop);
1216 _igvn.replace_node(val, new_val);
1217 lazy_update(lb, lb->in(LoadBarrierNode::Control));
1218 lazy_replace(lb->proj_out(LoadBarrierNode::Control), lb->in(LoadBarrierNode::Control));
1219 }
1220
1221 bool PhaseIdealLoop::split_barrier_thru_phi(LoadBarrierNode* lb) {
1222 if (lb->in(LoadBarrierNode::Oop)->is_Phi()) {
1223 Node* oop_phi = lb->in(LoadBarrierNode::Oop);
1224
1225 if (oop_phi->req() == 2) {
1226 // Ignore phis with only one input
1227 return false;
1228 }
1229
1230 if (is_dominator(get_ctrl(lb->in(LoadBarrierNode::Address)), oop_phi->in(0)) && get_ctrl(lb->in(LoadBarrierNode::Address)) != oop_phi->in(0) /*&& (get_ctrl(lb->in(LoadBarrierNode::Memory)) != lb->in(0) || lb->in(LoadBarrierNode::Memory)->is_Phi())*/) {
1231 // That transformation may cause the Similar edge on dominated load barriers to be invalid
1232 lb->fix_similar_in_uses(&_igvn);
1233
1234 RegionNode* region = oop_phi->in(0)->as_Region();
1235
1236 int backedge = LoopNode::LoopBackControl;
1237 if (region->is_Loop() && region->in(backedge)->is_Proj() && region->in(backedge)->in(0)->is_If()) {
1238 Node* c = region->in(backedge)->in(0)->in(0);
1239 assert(c->unique_ctrl_out() == region->in(backedge)->in(0), "");
1240 Node* oop = lb->in(LoadBarrierNode::Oop)->in(backedge);
1241 Node* oop_c = has_ctrl(oop) ? get_ctrl(oop) : oop;
1242 if (!is_dominator(oop_c, c)) {
1243 return false;
1244 }
1245 }
1246
1247 Node *phi = oop_phi->clone();
1248
1249 for (uint i = 1; i < region->req(); i++) {
1250 Node* ctrl = region->in(i);
1251 if (ctrl != C->top()) {
1252 assert(!is_dominator(ctrl, region) || region->is_Loop(), "");
1253
1254 Node* mem = lb->in(LoadBarrierNode::Memory);
1255 Node* m = find_dominating_memory(mem, region, i);
1256
1257 if (region->is_Loop() && i == LoopNode::LoopBackControl && ctrl->is_Proj() && ctrl->in(0)->is_If()) {
1258 ctrl = ctrl->in(0)->in(0);
1259 }
1260
1261 LoadBarrierNode* new_lb = clone_load_barrier(lb, ctrl, m, lb->in(LoadBarrierNode::Oop)->in(i));
1262
1263 Node* out_ctrl = new_lb->proj_out(LoadBarrierNode::Control);
1264 if (ctrl == region->in(i)) {
1265 _igvn.replace_input_of(region, i, new_lb->proj_out(LoadBarrierNode::Control));
1266 } else {
1267 Node* iff = region->in(i)->in(0);
1268 Node* out_ctrl = new_lb->proj_out(LoadBarrierNode::Control);
1269 _igvn.replace_input_of(iff, 0, out_ctrl);
1270 set_idom(iff, out_ctrl, dom_depth(out_ctrl)+1);
1271 }
1272 phi->set_req(i, new_lb->proj_out(LoadBarrierNode::Oop));
1273 }
1274 }
1275 register_new_node(phi, region);
1276
1277 replace_barrier(lb, phi);
1278
1279 if (region->is_Loop()) {
1280 // Load barrier moved to the back edge of the Loop may now
1281 // have a safepoint on the path to the barrier on the Similar
1282 // edge
1283 _igvn.replace_input_of(phi->in(LoopNode::LoopBackControl)->in(0), LoadBarrierNode::Similar, C->top());
1284 Node* head = region->in(LoopNode::EntryControl);
1285 set_idom(region, head, dom_depth(head)+1);
1286 recompute_dom_depth();
1287 if (head->is_CountedLoop() && head->as_CountedLoop()->is_main_loop()) {
1288 head->as_CountedLoop()->set_normal_loop();
1289 }
1290 }
1291 return true;
1292 }
1293 }
1294 return false;
1295 }
1296
1297 bool PhaseIdealLoop::move_out_of_loop(LoadBarrierNode* lb) {
1298 IdealLoopTree *lb_loop = get_loop(lb->in(0));
1299 if (lb_loop != _ltree_root && !lb_loop->_irreducible) {
1300 Node* oop_ctrl = get_ctrl(lb->in(LoadBarrierNode::Oop));
1301 IdealLoopTree *oop_loop = get_loop(oop_ctrl);
1302 IdealLoopTree* adr_loop = get_loop(get_ctrl(lb->in(LoadBarrierNode::Address)));
1303 if (!lb_loop->is_member(oop_loop) && !lb_loop->is_member(adr_loop)) {
1304 // That transformation may cause the Similar edge on dominated load barriers to be invalid
1305 lb->fix_similar_in_uses(&_igvn);
1306
1307 Node* head = lb_loop->_head;
1308 assert(head->is_Loop(), "");
1309
1310 if (is_dominator(head, oop_ctrl)) {
1311 assert(oop_ctrl->Opcode() == Op_CProj && oop_ctrl->in(0)->Opcode() == Op_NeverBranch, "");
1312 assert(lb_loop->is_member(get_loop(oop_ctrl->in(0)->in(0))), "");
1313 return false;
1314 }
1315
1316 if (head->is_CountedLoop() && head->as_CountedLoop()->is_main_loop()) {
1317 head->as_CountedLoop()->set_normal_loop();
1318 }
1319
1320 Node* mem = lb->in(LoadBarrierNode::Memory);
1321 Node* m = find_dominating_memory(mem, head);
1322
1323 LoadBarrierNode* new_lb = clone_load_barrier(lb, head->in(LoopNode::EntryControl), m, NULL);
1324
1325 assert(idom(head) == head->in(LoopNode::EntryControl), "");
1326 Node* proj_ctl = new_lb->proj_out(LoadBarrierNode::Control);
1327 _igvn.replace_input_of(head, LoopNode::EntryControl, proj_ctl);
1328 set_idom(head, proj_ctl, dom_depth(proj_ctl)+1);
1329
1330 replace_barrier(lb, new_lb->proj_out(LoadBarrierNode::Oop));
1331
1332 recompute_dom_depth();
1333
1334 return true;
1335 }
1336 }
1337 return false;
1338 }
1339
1340 bool PhaseIdealLoop::common_barriers(LoadBarrierNode* lb) {
1341 Node* in_val = lb->in(LoadBarrierNode::Oop);
1342 for (DUIterator_Fast imax, i = in_val->fast_outs(imax); i < imax; i++) {
1343 Node* u = in_val->fast_out(i);
1344 if (u != lb && u->is_LoadBarrier() && u->as_LoadBarrier()->has_true_uses()) {
1345 Node* this_ctrl = lb->in(LoadBarrierNode::Control);
1346 Node* other_ctrl = u->in(LoadBarrierNode::Control);
1347
1348 Node* lca = dom_lca(this_ctrl, other_ctrl);
1349 bool ok = true;
1350
1351 Node* proj1 = NULL;
1352 Node* proj2 = NULL;
1353
1354 while (this_ctrl != lca && ok) {
1355 if (this_ctrl->in(0) != NULL &&
1356 this_ctrl->in(0)->is_MultiBranch()) {
1357 if (this_ctrl->in(0)->in(0) == lca) {
1358 assert(proj1 == NULL, "");
1359 assert(this_ctrl->is_Proj(), "");
1360 proj1 = this_ctrl;
1361 } else if (!(this_ctrl->in(0)->is_If() && this_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none))) {
1362 ok = false;
1363 }
1364 }
1365 this_ctrl = idom(this_ctrl);
1366 }
1367 while (other_ctrl != lca && ok) {
1368 if (other_ctrl->in(0) != NULL &&
1369 other_ctrl->in(0)->is_MultiBranch()) {
1370 if (other_ctrl->in(0)->in(0) == lca) {
1371 assert(other_ctrl->is_Proj(), "");
1372 assert(proj2 == NULL, "");
1373 proj2 = other_ctrl;
1374 } else if (!(other_ctrl->in(0)->is_If() && other_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none))) {
1375 ok = false;
1376 }
1377 }
1378 other_ctrl = idom(other_ctrl);
1379 }
1380 assert(proj1 == NULL || proj2 == NULL || proj1->in(0) == proj2->in(0), "");
1381 if (ok && proj1 && proj2 && proj1 != proj2 && proj1->in(0)->is_If()) {
1382 // That transformation may cause the Similar edge on dominated load barriers to be invalid
1383 lb->fix_similar_in_uses(&_igvn);
1384 u->as_LoadBarrier()->fix_similar_in_uses(&_igvn);
1385
1386 Node* split = lca->unique_ctrl_out();
1387 assert(split->in(0) == lca, "");
1388
1389 Node* mem = lb->in(LoadBarrierNode::Memory);
1390 Node* m = find_dominating_memory(mem, split);
1391 LoadBarrierNode* new_lb = clone_load_barrier(lb, lca, m, NULL);
1392
1393 Node* proj_ctl = new_lb->proj_out(LoadBarrierNode::Control);
1394 _igvn.replace_input_of(split, 0, new_lb->proj_out(LoadBarrierNode::Control));
1395 set_idom(split, proj_ctl, dom_depth(proj_ctl)+1);
1396
1397 Node* proj_oop = new_lb->proj_out(LoadBarrierNode::Oop);
1398 replace_barrier(lb, proj_oop);
1399 replace_barrier(u->as_LoadBarrier(), proj_oop);
1400
1401 recompute_dom_depth();
1402
1403 return true;
1404 }
1405 }
1406 }
1407 return false;
1408 }
1409
1410 void PhaseIdealLoop::optimize_load_barrier(LoadBarrierNode* lb, bool last_round) {
1411 if (!C->directive()->OptimizeLoadBarriersOption) {
1412 return;
1413 }
1414
1415 if (lb->has_true_uses()) {
1416
1417 if (replace_with_dominating_barrier(lb, last_round)) {
1418 return;
1419 }
1420
1421 if (split_barrier_thru_phi(lb)) {
1422 return;
1423 }
1424
1425 if (move_out_of_loop(lb)) {
1426 return;
1427 }
1428
1429 if (common_barriers(lb)) {
1430 return;
1431 }
1432 }
1433 }
1434
1435 //------------------------------split_if_with_blocks_post----------------------
1436 // Do the real work in a non-recursive function. CFG hackery wants to be
1437 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1438 // info.
1439 void PhaseIdealLoop::split_if_with_blocks_post(Node *n, bool last_round) {
1440
1441 // Cloning Cmp through Phi's involves the split-if transform.
1442 // FastLock is not used by an If
1443 if (n->is_Cmp() && !n->is_FastLock() && !last_round) {
1444 Node *n_ctrl = get_ctrl(n);
1445 // Determine if the Node has inputs from some local Phi.
1446 // Returns the block to clone thru.
1447 Node *n_blk = has_local_phi_input(n);
1448 if (n_blk != n_ctrl) {
1449 return;
1450 }
1451
1452 if (!can_split_if(n_ctrl)) {
1453 return;
1454 }
1455
1456 if (n->outcnt() != 1) {
1457 return; // Multiple bool's from 1 compare?
1458 }
1459 Node *bol = n->unique_out();
1460 assert(bol->is_Bool(), "expect a bool here");
1461 if (bol->outcnt() != 1) {
1462 return;// Multiple branches from 1 compare?
1463 }
1464 Node *iff = bol->unique_out();
1465
1466 // Check some safety conditions
1467 if (iff->is_If()) { // Classic split-if?
1468 if (iff->in(0) != n_ctrl) {
1469 return; // Compare must be in same blk as if
1470 }
1471 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1472 // Can't split CMove with different control edge.
1473 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) {
1474 return;
1475 }
1476 if (get_ctrl(iff->in(2)) == n_ctrl ||
1477 get_ctrl(iff->in(3)) == n_ctrl) {
1478 return; // Inputs not yet split-up
1479 }
1480 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1481 return; // Loop-invar test gates loop-varying CMOVE
1482 }
1483 } else {
1484 return; // some other kind of node, such as an Allocate
1485 }
1486
1487 // When is split-if profitable? Every 'win' on means some control flow
1488 // goes dead, so it's almost always a win.
1489 int policy = 0;
1490 // Split compare 'n' through the merge point if it is profitable
1491 Node *phi = split_thru_phi( n, n_ctrl, policy);
1492 if (!phi) {
1493 return;
1494 }
1495
1496 // Found a Phi to split thru!
1497 // Replace 'n' with the new phi
1498 _igvn.replace_node(n, phi);
1499
1500 // Now split the bool up thru the phi
1501 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1502 guarantee(bolphi != NULL, "null boolean phi node");
1503
1504 _igvn.replace_node(bol, bolphi);
1505 assert(iff->in(1) == bolphi, "");
1506
1507 if (bolphi->Value(&_igvn)->singleton()) {
1508 return;
1509 }
1510
1511 // Conditional-move? Must split up now
1512 if (!iff->is_If()) {
1513 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1514 _igvn.replace_node(iff, cmovphi);
1515 return;
1516 }
1517
1518 // Now split the IF
1519 do_split_if(iff);
1520 return;
1521 }
1522
1523 // Two identical ifs back to back can be merged
1524 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1525 Node *n_ctrl = n->in(0);
1526 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1527 IfNode* dom_if = idom(n_ctrl)->as_If();
1528 Node* proj_true = dom_if->proj_out(1);
1529 Node* proj_false = dom_if->proj_out(0);
1530 Node* con_true = _igvn.makecon(TypeInt::ONE);
1531 Node* con_false = _igvn.makecon(TypeInt::ZERO);
1532
1533 for (uint i = 1; i < n_ctrl->req(); i++) {
1534 if (is_dominator(proj_true, n_ctrl->in(i))) {
1535 bolphi->init_req(i, con_true);
1536 } else {
1537 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1538 bolphi->init_req(i, con_false);
1539 }
1540 }
1541 register_new_node(bolphi, n_ctrl);
1542 _igvn.replace_input_of(n, 1, bolphi);
1543
1544 // Now split the IF
1545 do_split_if(n);
1546 return;
1547 }
1548
1549 // Check for an IF ready to split; one that has its
1550 // condition codes input coming from a Phi at the block start.
1551 int n_op = n->Opcode();
1552
1553 // Check for an IF being dominated by another IF same test
1554 if (n_op == Op_If ||
1555 n_op == Op_RangeCheck) {
1556 Node *bol = n->in(1);
1557 uint max = bol->outcnt();
1558
1559 // Check for same test used more than once?
1560 if (max > 1 && bol->is_Bool()) {
1561 // Search up IDOMs to see if this IF is dominated.
1562 Node *cutoff = get_ctrl(bol);
1563
1564 // Now search up IDOMs till cutoff, looking for a dominating test
1565 Node *prevdom = n;
1566 Node *dom = idom(prevdom);
1567 while (dom != cutoff) {
1568 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) {
1569 // Replace the dominated test with an obvious true or false.
1570 // Place it on the IGVN worklist for later cleanup.
1571 C->set_major_progress();
1572 dominated_by(prevdom, n, false, true);
1573 #ifndef PRODUCT
1574 if( VerifyLoopOptimizations ) verify();
1575 #endif
1576 return;
1577 }
1578 prevdom = dom;
1579 dom = idom(prevdom);
1580 }
1581 }
1582 }
1583
1584 // See if a shared loop-varying computation has no loop-varying uses.
1585 // Happens if something is only used for JVM state in uncommon trap exits,
1586 // like various versions of induction variable+offset. Clone the
1587 // computation per usage to allow it to sink out of the loop.
1588 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
1589 Node *n_ctrl = get_ctrl(n);
1590 IdealLoopTree *n_loop = get_loop(n_ctrl);
1591 if( n_loop != _ltree_root ) {
1592 DUIterator_Fast imax, i = n->fast_outs(imax);
1593 for (; i < imax; i++) {
1594 Node* u = n->fast_out(i);
1595 if( !has_ctrl(u) ) break; // Found control user
1596 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1597 if( u_loop == n_loop ) break; // Found loop-varying use
1598 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
1599 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
1600 }
1601 bool did_break = (i < imax); // Did we break out of the previous loop?
1602 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
1603 Node *late_load_ctrl = NULL;
1604 if (n->is_Load()) {
1605 // If n is a load, get and save the result from get_late_ctrl(),
1606 // to be later used in calculating the control for n's clones.
1607 clear_dom_lca_tags();
1608 late_load_ctrl = get_late_ctrl(n, n_ctrl);
1609 }
1610 // If n is a load, and the late control is the same as the current
1611 // control, then the cloning of n is a pointless exercise, because
1612 // GVN will ensure that we end up where we started.
1613 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
1614 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
1615 Node *u = n->last_out(j); // Clone private computation per use
1616 _igvn.rehash_node_delayed(u);
1617 Node *x = n->clone(); // Clone computation
1618 Node *x_ctrl = NULL;
1619 if( u->is_Phi() ) {
1620 // Replace all uses of normal nodes. Replace Phi uses
1621 // individually, so the separate Nodes can sink down
1622 // different paths.
1623 uint k = 1;
1624 while( u->in(k) != n ) k++;
1625 u->set_req( k, x );
1626 // x goes next to Phi input path
1627 x_ctrl = u->in(0)->in(k);
1628 --j;
1629 } else { // Normal use
1630 // Replace all uses
1631 for( uint k = 0; k < u->req(); k++ ) {
1632 if( u->in(k) == n ) {
1633 u->set_req( k, x );
1634 --j;
1635 }
1636 }
1637 x_ctrl = get_ctrl(u);
1638 }
1639
1640 // Find control for 'x' next to use but not inside inner loops.
1641 // For inner loop uses get the preheader area.
1642 x_ctrl = place_near_use(x_ctrl);
1643
1644 if (n->is_Load()) {
1645 // For loads, add a control edge to a CFG node outside of the loop
1646 // to force them to not combine and return back inside the loop
1647 // during GVN optimization (4641526).
1648 //
1649 // Because we are setting the actual control input, factor in
1650 // the result from get_late_ctrl() so we respect any
1651 // anti-dependences. (6233005).
1652 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
1653
1654 // Don't allow the control input to be a CFG splitting node.
1655 // Such nodes should only have ProjNodes as outs, e.g. IfNode
1656 // should only have IfTrueNode and IfFalseNode (4985384).
1657 x_ctrl = find_non_split_ctrl(x_ctrl);
1658 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1659
1660 x->set_req(0, x_ctrl);
1661 }
1662 register_new_node(x, x_ctrl);
1663
1664 // Some institutional knowledge is needed here: 'x' is
1665 // yanked because if the optimizer runs GVN on it all the
1666 // cloned x's will common up and undo this optimization and
1667 // be forced back in the loop. This is annoying because it
1668 // makes +VerifyOpto report false-positives on progress. I
1669 // tried setting control edges on the x's to force them to
1670 // not combine, but the matching gets worried when it tries
1671 // to fold a StoreP and an AddP together (as part of an
1672 // address expression) and the AddP and StoreP have
1673 // different controls.
1674 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x);
1675 }
1676 _igvn.remove_dead_node(n);
1677 }
1678 }
1679 }
1680 }
1681
1682 try_move_store_after_loop(n);
1683
1684 // Check for Opaque2's who's loop has disappeared - who's input is in the
1685 // same loop nest as their output. Remove 'em, they are no longer useful.
1686 if( n_op == Op_Opaque2 &&
1687 n->in(1) != NULL &&
1688 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1689 _igvn.replace_node( n, n->in(1) );
1690 }
1691
1692 if (n->is_LoadBarrier()) {
1693 optimize_load_barrier(n->as_LoadBarrier(), last_round);
1694 }
1695 }
1696
1697 //------------------------------split_if_with_blocks---------------------------
1698 // Check for aggressive application of 'split-if' optimization,
1699 // using basic block level info.
1700 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack, bool last_round) {
1701 Node *n = C->root();
1702 visited.set(n->_idx); // first, mark node as visited
1703 // Do pre-visit work for root
1704 n = split_if_with_blocks_pre( n );
1705 uint cnt = n->outcnt();
1706 uint i = 0;
1707 while (true) {
1708 // Visit all children
1709 if (i < cnt) {
1710 Node* use = n->raw_out(i);
1711 ++i;
1712 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1713 // Now do pre-visit work for this use
1714 use = split_if_with_blocks_pre( use );
1715 nstack.push(n, i); // Save parent and next use's index.
1716 n = use; // Process all children of current use.
1717 cnt = use->outcnt();
1718 i = 0;
1719 }
1720 }
1721 else {
1722 // All of n's children have been processed, complete post-processing.
1723 if (cnt != 0 && !n->is_Con()) {
1724 assert(has_node(n), "no dead nodes");
1725 split_if_with_blocks_post(n, last_round);
1726 }
1727 if (nstack.is_empty()) {
1728 // Finished all nodes on stack.
1729 break;
1730 }
1731 // Get saved parent node and next use's index. Visit the rest of uses.
1732 n = nstack.node();
1733 cnt = n->outcnt();
1734 i = nstack.index();
1735 nstack.pop();
1736 }
1737 }
1738 }
1739
1740
1741 //=============================================================================
1742 //
1743 // C L O N E A L O O P B O D Y
1744 //
1745
1746 //------------------------------clone_iff--------------------------------------
1747 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1748 // "Nearly" because all Nodes have been cloned from the original in the loop,
1749 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1750 // through the Phi recursively, and return a Bool.
1751 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) {
1752
1753 // Convert this Phi into a Phi merging Bools
1754 uint i;
1755 for (i = 1; i < phi->req(); i++) {
1756 Node *b = phi->in(i);
1757 if (b->is_Phi()) {
1758 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop));
1759 } else {
1760 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
1761 }
1762 }
1763
1764 Node* n = phi->in(1);
1765 Node* sample_opaque = NULL;
1766 Node *sample_bool = NULL;
1767 if (n->Opcode() == Op_Opaque4) {
1768 sample_opaque = n;
1769 sample_bool = n->in(1);
1770 assert(sample_bool->is_Bool(), "wrong type");
1771 } else {
1772 sample_bool = n;
1773 }
1774 Node *sample_cmp = sample_bool->in(1);
1775
1776 // Make Phis to merge the Cmp's inputs.
1777 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
1778 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
1779 for (i = 1; i < phi->req(); i++) {
1780 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
1781 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
1782 phi1->set_req(i, n1);
1783 phi2->set_req(i, n2);
1784 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1785 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1786 }
1787 // See if these Phis have been made before.
1788 // Register with optimizer
1789 Node *hit1 = _igvn.hash_find_insert(phi1);
1790 if (hit1) { // Hit, toss just made Phi
1791 _igvn.remove_dead_node(phi1); // Remove new phi
1792 assert(hit1->is_Phi(), "" );
1793 phi1 = (PhiNode*)hit1; // Use existing phi
1794 } else { // Miss
1795 _igvn.register_new_node_with_optimizer(phi1);
1796 }
1797 Node *hit2 = _igvn.hash_find_insert(phi2);
1798 if (hit2) { // Hit, toss just made Phi
1799 _igvn.remove_dead_node(phi2); // Remove new phi
1800 assert(hit2->is_Phi(), "" );
1801 phi2 = (PhiNode*)hit2; // Use existing phi
1802 } else { // Miss
1803 _igvn.register_new_node_with_optimizer(phi2);
1804 }
1805 // Register Phis with loop/block info
1806 set_ctrl(phi1, phi->in(0));
1807 set_ctrl(phi2, phi->in(0));
1808 // Make a new Cmp
1809 Node *cmp = sample_cmp->clone();
1810 cmp->set_req(1, phi1);
1811 cmp->set_req(2, phi2);
1812 _igvn.register_new_node_with_optimizer(cmp);
1813 set_ctrl(cmp, phi->in(0));
1814
1815 // Make a new Bool
1816 Node *b = sample_bool->clone();
1817 b->set_req(1,cmp);
1818 _igvn.register_new_node_with_optimizer(b);
1819 set_ctrl(b, phi->in(0));
1820
1821 if (sample_opaque != NULL) {
1822 Node* opaque = sample_opaque->clone();
1823 opaque->set_req(1, b);
1824 _igvn.register_new_node_with_optimizer(opaque);
1825 set_ctrl(opaque, phi->in(0));
1826 return opaque;
1827 }
1828
1829 assert(b->is_Bool(), "");
1830 return b;
1831 }
1832
1833 //------------------------------clone_bool-------------------------------------
1834 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1835 // "Nearly" because all Nodes have been cloned from the original in the loop,
1836 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1837 // through the Phi recursively, and return a Bool.
1838 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1839 uint i;
1840 // Convert this Phi into a Phi merging Bools
1841 for( i = 1; i < phi->req(); i++ ) {
1842 Node *b = phi->in(i);
1843 if( b->is_Phi() ) {
1844 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1845 } else {
1846 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1847 }
1848 }
1849
1850 Node *sample_cmp = phi->in(1);
1851
1852 // Make Phis to merge the Cmp's inputs.
1853 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1854 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1855 for( uint j = 1; j < phi->req(); j++ ) {
1856 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1857 Node *n1, *n2;
1858 if( cmp_top->is_Cmp() ) {
1859 n1 = cmp_top->in(1);
1860 n2 = cmp_top->in(2);
1861 } else {
1862 n1 = n2 = cmp_top;
1863 }
1864 phi1->set_req( j, n1 );
1865 phi2->set_req( j, n2 );
1866 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1867 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1868 }
1869
1870 // See if these Phis have been made before.
1871 // Register with optimizer
1872 Node *hit1 = _igvn.hash_find_insert(phi1);
1873 if( hit1 ) { // Hit, toss just made Phi
1874 _igvn.remove_dead_node(phi1); // Remove new phi
1875 assert( hit1->is_Phi(), "" );
1876 phi1 = (PhiNode*)hit1; // Use existing phi
1877 } else { // Miss
1878 _igvn.register_new_node_with_optimizer(phi1);
1879 }
1880 Node *hit2 = _igvn.hash_find_insert(phi2);
1881 if( hit2 ) { // Hit, toss just made Phi
1882 _igvn.remove_dead_node(phi2); // Remove new phi
1883 assert( hit2->is_Phi(), "" );
1884 phi2 = (PhiNode*)hit2; // Use existing phi
1885 } else { // Miss
1886 _igvn.register_new_node_with_optimizer(phi2);
1887 }
1888 // Register Phis with loop/block info
1889 set_ctrl(phi1, phi->in(0));
1890 set_ctrl(phi2, phi->in(0));
1891 // Make a new Cmp
1892 Node *cmp = sample_cmp->clone();
1893 cmp->set_req( 1, phi1 );
1894 cmp->set_req( 2, phi2 );
1895 _igvn.register_new_node_with_optimizer(cmp);
1896 set_ctrl(cmp, phi->in(0));
1897
1898 assert( cmp->is_Cmp(), "" );
1899 return (CmpNode*)cmp;
1900 }
1901
1902 //------------------------------sink_use---------------------------------------
1903 // If 'use' was in the loop-exit block, it now needs to be sunk
1904 // below the post-loop merge point.
1905 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1906 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1907 set_ctrl(use, post_loop);
1908 for (DUIterator j = use->outs(); use->has_out(j); j++)
1909 sink_use(use->out(j), post_loop);
1910 }
1911 }
1912
1913 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1914 IdealLoopTree* loop, IdealLoopTree* outer_loop,
1915 Node_List*& split_if_set, Node_List*& split_bool_set,
1916 Node_List*& split_cex_set, Node_List& worklist,
1917 uint new_counter, CloneLoopMode mode) {
1918 Node* nnn = old_new[old->_idx];
1919 // Copy uses to a worklist, so I can munge the def-use info
1920 // with impunity.
1921 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1922 worklist.push(old->fast_out(j));
1923
1924 while( worklist.size() ) {
1925 Node *use = worklist.pop();
1926 if (!has_node(use)) continue; // Ignore dead nodes
1927 if (use->in(0) == C->top()) continue;
1928 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1929 // Check for data-use outside of loop - at least one of OLD or USE
1930 // must not be a CFG node.
1931 #ifdef ASSERT
1932 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) {
1933 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
1934 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node");
1935 }
1936 #endif
1937 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
1938
1939 // If the Data use is an IF, that means we have an IF outside of the
1940 // loop that is switching on a condition that is set inside of the
1941 // loop. Happens if people set a loop-exit flag; then test the flag
1942 // in the loop to break the loop, then test is again outside of the
1943 // loop to determine which way the loop exited.
1944 // Loop predicate If node connects to Bool node through Opaque1 node.
1945 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) {
1946 // Since this code is highly unlikely, we lazily build the worklist
1947 // of such Nodes to go split.
1948 if (!split_if_set) {
1949 ResourceArea *area = Thread::current()->resource_area();
1950 split_if_set = new Node_List(area);
1951 }
1952 split_if_set->push(use);
1953 }
1954 if (use->is_Bool()) {
1955 if (!split_bool_set) {
1956 ResourceArea *area = Thread::current()->resource_area();
1957 split_bool_set = new Node_List(area);
1958 }
1959 split_bool_set->push(use);
1960 }
1961 if (use->Opcode() == Op_CreateEx) {
1962 if (!split_cex_set) {
1963 ResourceArea *area = Thread::current()->resource_area();
1964 split_cex_set = new Node_List(area);
1965 }
1966 split_cex_set->push(use);
1967 }
1968
1969
1970 // Get "block" use is in
1971 uint idx = 0;
1972 while( use->in(idx) != old ) idx++;
1973 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1974 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
1975 Node *cfg = prev->_idx >= new_counter
1976 ? prev->in(2)
1977 : idom(prev);
1978 if( use->is_Phi() ) // Phi use is in prior block
1979 cfg = prev->in(idx); // NOT in block of Phi itself
1980 if (cfg->is_top()) { // Use is dead?
1981 _igvn.replace_input_of(use, idx, C->top());
1982 continue;
1983 }
1984
1985 while(!outer_loop->is_member(get_loop(cfg))) {
1986 prev = cfg;
1987 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1988 }
1989 // If the use occurs after merging several exits from the loop, then
1990 // old value must have dominated all those exits. Since the same old
1991 // value was used on all those exits we did not need a Phi at this
1992 // merge point. NOW we do need a Phi here. Each loop exit value
1993 // is now merged with the peeled body exit; each exit gets its own
1994 // private Phi and those Phis need to be merged here.
1995 Node *phi;
1996 if( prev->is_Region() ) {
1997 if( idx == 0 ) { // Updating control edge?
1998 phi = prev; // Just use existing control
1999 } else { // Else need a new Phi
2000 phi = PhiNode::make( prev, old );
2001 // Now recursively fix up the new uses of old!
2002 for( uint i = 1; i < prev->req(); i++ ) {
2003 worklist.push(phi); // Onto worklist once for each 'old' input
2004 }
2005 }
2006 } else {
2007 // Get new RegionNode merging old and new loop exits
2008 prev = old_new[prev->_idx];
2009 assert( prev, "just made this in step 7" );
2010 if( idx == 0) { // Updating control edge?
2011 phi = prev; // Just use existing control
2012 } else { // Else need a new Phi
2013 // Make a new Phi merging data values properly
2014 phi = PhiNode::make( prev, old );
2015 phi->set_req( 1, nnn );
2016 }
2017 }
2018 // If inserting a new Phi, check for prior hits
2019 if( idx != 0 ) {
2020 Node *hit = _igvn.hash_find_insert(phi);
2021 if( hit == NULL ) {
2022 _igvn.register_new_node_with_optimizer(phi); // Register new phi
2023 } else { // or
2024 // Remove the new phi from the graph and use the hit
2025 _igvn.remove_dead_node(phi);
2026 phi = hit; // Use existing phi
2027 }
2028 set_ctrl(phi, prev);
2029 }
2030 // Make 'use' use the Phi instead of the old loop body exit value
2031 _igvn.replace_input_of(use, idx, phi);
2032 if( use->_idx >= new_counter ) { // If updating new phis
2033 // Not needed for correctness, but prevents a weak assert
2034 // in AddPNode from tripping (when we end up with different
2035 // base & derived Phis that will become the same after
2036 // IGVN does CSE).
2037 Node *hit = _igvn.hash_find_insert(use);
2038 if( hit ) // Go ahead and re-hash for hits.
2039 _igvn.replace_node( use, hit );
2040 }
2041
2042 // If 'use' was in the loop-exit block, it now needs to be sunk
2043 // below the post-loop merge point.
2044 sink_use( use, prev );
2045 }
2046 }
2047 }
2048
2049 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2050 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2051 Node_List& extra_data_nodes) {
2052 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2053 CountedLoopNode* cl = head->as_CountedLoop();
2054 Node* l = cl->outer_loop();
2055 Node* tail = cl->outer_loop_tail();
2056 IfNode* le = cl->outer_loop_end();
2057 Node* sfpt = cl->outer_safepoint();
2058 CountedLoopEndNode* cle = cl->loopexit();
2059 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2060 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2061 Node* cle_out = cle->proj_out(false);
2062
2063 Node* new_sfpt = NULL;
2064 Node* new_cle_out = cle_out->clone();
2065 old_new.map(cle_out->_idx, new_cle_out);
2066 if (mode == CloneIncludesStripMined) {
2067 // clone outer loop body
2068 Node* new_l = l->clone();
2069 Node* new_tail = tail->clone();
2070 IfNode* new_le = le->clone()->as_If();
2071 new_sfpt = sfpt->clone();
2072
2073 set_loop(new_l, outer_loop->_parent);
2074 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2075 set_loop(new_cle_out, outer_loop->_parent);
2076 set_idom(new_cle_out, new_cle, dd);
2077 set_loop(new_sfpt, outer_loop->_parent);
2078 set_idom(new_sfpt, new_cle_out, dd);
2079 set_loop(new_le, outer_loop->_parent);
2080 set_idom(new_le, new_sfpt, dd);
2081 set_loop(new_tail, outer_loop->_parent);
2082 set_idom(new_tail, new_le, dd);
2083 set_idom(new_cl, new_l, dd);
2084
2085 old_new.map(l->_idx, new_l);
2086 old_new.map(tail->_idx, new_tail);
2087 old_new.map(le->_idx, new_le);
2088 old_new.map(sfpt->_idx, new_sfpt);
2089
2090 new_l->set_req(LoopNode::LoopBackControl, new_tail);
2091 new_l->set_req(0, new_l);
2092 new_tail->set_req(0, new_le);
2093 new_le->set_req(0, new_sfpt);
2094 new_sfpt->set_req(0, new_cle_out);
2095 new_cle_out->set_req(0, new_cle);
2096 new_cl->set_req(LoopNode::EntryControl, new_l);
2097
2098 _igvn.register_new_node_with_optimizer(new_l);
2099 _igvn.register_new_node_with_optimizer(new_tail);
2100 _igvn.register_new_node_with_optimizer(new_le);
2101 } else {
2102 Node *newhead = old_new[loop->_head->_idx];
2103 newhead->as_Loop()->clear_strip_mined();
2104 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2105 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2106 }
2107 // Look at data node that were assigned a control in the outer
2108 // loop: they are kept in the outer loop by the safepoint so start
2109 // from the safepoint node's inputs.
2110 IdealLoopTree* outer_loop = get_loop(l);
2111 Node_Stack stack(2);
2112 stack.push(sfpt, 1);
2113 uint new_counter = C->unique();
2114 while (stack.size() > 0) {
2115 Node* n = stack.node();
2116 uint i = stack.index();
2117 while (i < n->req() &&
2118 (n->in(i) == NULL ||
2119 !has_ctrl(n->in(i)) ||
2120 get_loop(get_ctrl(n->in(i))) != outer_loop ||
2121 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2122 i++;
2123 }
2124 if (i < n->req()) {
2125 stack.set_index(i+1);
2126 stack.push(n->in(i), 0);
2127 } else {
2128 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2129 Node* m = n == sfpt ? new_sfpt : n->clone();
2130 if (m != NULL) {
2131 for (uint i = 0; i < n->req(); i++) {
2132 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) {
2133 m->set_req(i, old_new[m->in(i)->_idx]);
2134 }
2135 }
2136 } else {
2137 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2138 }
2139 if (n != sfpt) {
2140 extra_data_nodes.push(n);
2141 _igvn.register_new_node_with_optimizer(m);
2142 assert(get_ctrl(n) == cle_out, "what other control?");
2143 set_ctrl(m, new_cle_out);
2144 old_new.map(n->_idx, m);
2145 }
2146 stack.pop();
2147 }
2148 }
2149 if (mode == CloneIncludesStripMined) {
2150 _igvn.register_new_node_with_optimizer(new_sfpt);
2151 _igvn.register_new_node_with_optimizer(new_cle_out);
2152 }
2153 } else {
2154 Node *newhead = old_new[loop->_head->_idx];
2155 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2156 }
2157 }
2158
2159 //------------------------------clone_loop-------------------------------------
2160 //
2161 // C L O N E A L O O P B O D Y
2162 //
2163 // This is the basic building block of the loop optimizations. It clones an
2164 // entire loop body. It makes an old_new loop body mapping; with this mapping
2165 // you can find the new-loop equivalent to an old-loop node. All new-loop
2166 // nodes are exactly equal to their old-loop counterparts, all edges are the
2167 // same. All exits from the old-loop now have a RegionNode that merges the
2168 // equivalent new-loop path. This is true even for the normal "loop-exit"
2169 // condition. All uses of loop-invariant old-loop values now come from (one
2170 // or more) Phis that merge their new-loop equivalents.
2171 //
2172 // This operation leaves the graph in an illegal state: there are two valid
2173 // control edges coming from the loop pre-header to both loop bodies. I'll
2174 // definitely have to hack the graph after running this transform.
2175 //
2176 // From this building block I will further edit edges to perform loop peeling
2177 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2178 //
2179 // Parameter side_by_size_idom:
2180 // When side_by_size_idom is NULL, the dominator tree is constructed for
2181 // the clone loop to dominate the original. Used in construction of
2182 // pre-main-post loop sequence.
2183 // When nonnull, the clone and original are side-by-side, both are
2184 // dominated by the side_by_side_idom node. Used in construction of
2185 // unswitched loops.
2186 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2187 CloneLoopMode mode, Node* side_by_side_idom) {
2188
2189 LoopNode* head = loop->_head->as_Loop();
2190 head->verify_strip_mined(1);
2191
2192 if (C->do_vector_loop() && PrintOpto) {
2193 const char* mname = C->method()->name()->as_quoted_ascii();
2194 if (mname != NULL) {
2195 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2196 }
2197 }
2198
2199 CloneMap& cm = C->clone_map();
2200 Dict* dict = cm.dict();
2201 if (C->do_vector_loop()) {
2202 cm.set_clone_idx(cm.max_gen()+1);
2203 #ifndef PRODUCT
2204 if (PrintOpto) {
2205 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2206 loop->dump_head();
2207 }
2208 #endif
2209 }
2210
2211 // Step 1: Clone the loop body. Make the old->new mapping.
2212 uint i;
2213 for( i = 0; i < loop->_body.size(); i++ ) {
2214 Node *old = loop->_body.at(i);
2215 Node *nnn = old->clone();
2216 old_new.map( old->_idx, nnn );
2217 if (C->do_vector_loop()) {
2218 cm.verify_insert_and_clone(old, nnn, cm.clone_idx());
2219 }
2220 _igvn.register_new_node_with_optimizer(nnn);
2221 }
2222
2223 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2224
2225 // Step 2: Fix the edges in the new body. If the old input is outside the
2226 // loop use it. If the old input is INside the loop, use the corresponding
2227 // new node instead.
2228 for( i = 0; i < loop->_body.size(); i++ ) {
2229 Node *old = loop->_body.at(i);
2230 Node *nnn = old_new[old->_idx];
2231 // Fix CFG/Loop controlling the new node
2232 if (has_ctrl(old)) {
2233 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2234 } else {
2235 set_loop(nnn, outer_loop->_parent);
2236 if (old->outcnt() > 0) {
2237 set_idom( nnn, old_new[idom(old)->_idx], dd );
2238 }
2239 }
2240 // Correct edges to the new node
2241 for( uint j = 0; j < nnn->req(); j++ ) {
2242 Node *n = nnn->in(j);
2243 if( n ) {
2244 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
2245 if( loop->is_member( old_in_loop ) )
2246 nnn->set_req(j, old_new[n->_idx]);
2247 }
2248 }
2249 _igvn.hash_find_insert(nnn);
2250 }
2251
2252 ResourceArea *area = Thread::current()->resource_area();
2253 Node_List extra_data_nodes(area);
2254 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2255
2256 // Step 3: Now fix control uses. Loop varying control uses have already
2257 // been fixed up (as part of all input edges in Step 2). Loop invariant
2258 // control uses must be either an IfFalse or an IfTrue. Make a merge
2259 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2260 // refer to this.
2261 Node_List worklist(area);
2262 uint new_counter = C->unique();
2263 for( i = 0; i < loop->_body.size(); i++ ) {
2264 Node* old = loop->_body.at(i);
2265 if( !old->is_CFG() ) continue;
2266
2267 // Copy uses to a worklist, so I can munge the def-use info
2268 // with impunity.
2269 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2270 worklist.push(old->fast_out(j));
2271
2272 while( worklist.size() ) { // Visit all uses
2273 Node *use = worklist.pop();
2274 if (!has_node(use)) continue; // Ignore dead nodes
2275 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2276 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
2277 // Both OLD and USE are CFG nodes here.
2278 assert( use->is_Proj(), "" );
2279 Node* nnn = old_new[old->_idx];
2280
2281 Node* newuse = NULL;
2282 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2283 CountedLoopNode* cl = head->as_CountedLoop();
2284 CountedLoopEndNode* cle = cl->loopexit();
2285 Node* cle_out = cle->proj_out_or_null(false);
2286 if (use == cle_out) {
2287 IfNode* le = cl->outer_loop_end();
2288 use = le->proj_out(false);
2289 use_loop = get_loop(use);
2290 if (mode == CloneIncludesStripMined) {
2291 nnn = old_new[le->_idx];
2292 } else {
2293 newuse = old_new[cle_out->_idx];
2294 }
2295 }
2296 }
2297 if (newuse == NULL) {
2298 newuse = use->clone();
2299 }
2300
2301 // Clone the loop exit control projection
2302 if (C->do_vector_loop()) {
2303 cm.verify_insert_and_clone(use, newuse, cm.clone_idx());
2304 }
2305 newuse->set_req(0,nnn);
2306 _igvn.register_new_node_with_optimizer(newuse);
2307 set_loop(newuse, use_loop);
2308 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2309
2310 // We need a Region to merge the exit from the peeled body and the
2311 // exit from the old loop body.
2312 RegionNode *r = new RegionNode(3);
2313 // Map the old use to the new merge point
2314 old_new.map( use->_idx, r );
2315 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
2316 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
2317
2318 // The original user of 'use' uses 'r' instead.
2319 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2320 Node* useuse = use->last_out(l);
2321 _igvn.rehash_node_delayed(useuse);
2322 uint uses_found = 0;
2323 if( useuse->in(0) == use ) {
2324 useuse->set_req(0, r);
2325 uses_found++;
2326 if( useuse->is_CFG() ) {
2327 assert( dom_depth(useuse) > dd_r, "" );
2328 set_idom(useuse, r, dom_depth(useuse));
2329 }
2330 }
2331 for( uint k = 1; k < useuse->req(); k++ ) {
2332 if( useuse->in(k) == use ) {
2333 useuse->set_req(k, r);
2334 uses_found++;
2335 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2336 assert(dom_depth(useuse) > dd_r , "");
2337 set_idom(useuse, r, dom_depth(useuse));
2338 }
2339 }
2340 }
2341 l -= uses_found; // we deleted 1 or more copies of this edge
2342 }
2343
2344 // Now finish up 'r'
2345 r->set_req( 1, newuse );
2346 r->set_req( 2, use );
2347 _igvn.register_new_node_with_optimizer(r);
2348 set_loop(r, use_loop);
2349 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
2350 } // End of if a loop-exit test
2351 }
2352 }
2353
2354 // Step 4: If loop-invariant use is not control, it must be dominated by a
2355 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2356 // there if needed. Make a Phi there merging old and new used values.
2357 Node_List *split_if_set = NULL;
2358 Node_List *split_bool_set = NULL;
2359 Node_List *split_cex_set = NULL;
2360 for( i = 0; i < loop->_body.size(); i++ ) {
2361 Node* old = loop->_body.at(i);
2362 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2363 split_bool_set, split_cex_set, worklist, new_counter,
2364 mode);
2365 }
2366
2367 for (i = 0; i < extra_data_nodes.size(); i++) {
2368 Node* old = extra_data_nodes.at(i);
2369 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2370 split_bool_set, split_cex_set, worklist, new_counter,
2371 mode);
2372 }
2373
2374 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2375 // the loop uses a condition set in the loop. The original IF probably
2376 // takes control from one or more OLD Regions (which in turn get from NEW
2377 // Regions). In any case, there will be a set of Phis for each merge point
2378 // from the IF up to where the original BOOL def exists the loop.
2379 if (split_if_set) {
2380 while (split_if_set->size()) {
2381 Node *iff = split_if_set->pop();
2382 if (iff->in(1)->is_Phi()) {
2383 Node *b = clone_iff(iff->in(1)->as_Phi(), loop);
2384 _igvn.replace_input_of(iff, 1, b);
2385 }
2386 }
2387 }
2388 if (split_bool_set) {
2389 while (split_bool_set->size()) {
2390 Node *b = split_bool_set->pop();
2391 Node *phi = b->in(1);
2392 assert(phi->is_Phi(), "");
2393 CmpNode *cmp = clone_bool((PhiNode*)phi, loop);
2394 _igvn.replace_input_of(b, 1, cmp);
2395 }
2396 }
2397 if (split_cex_set) {
2398 while (split_cex_set->size()) {
2399 Node *b = split_cex_set->pop();
2400 assert(b->in(0)->is_Region(), "");
2401 assert(b->in(1)->is_Phi(), "");
2402 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2403 split_up(b, b->in(0), NULL);
2404 }
2405 }
2406
2407 }
2408
2409
2410 //---------------------- stride_of_possible_iv -------------------------------------
2411 // Looks for an iff/bool/comp with one operand of the compare
2412 // being a cycle involving an add and a phi,
2413 // with an optional truncation (left-shift followed by a right-shift)
2414 // of the add. Returns zero if not an iv.
2415 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2416 Node* trunc1 = NULL;
2417 Node* trunc2 = NULL;
2418 const TypeInt* ttype = NULL;
2419 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
2420 return 0;
2421 }
2422 BoolNode* bl = iff->in(1)->as_Bool();
2423 Node* cmp = bl->in(1);
2424 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2425 return 0;
2426 }
2427 // Must have an invariant operand
2428 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2429 return 0;
2430 }
2431 Node* add2 = NULL;
2432 Node* cmp1 = cmp->in(1);
2433 if (cmp1->is_Phi()) {
2434 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2435 Node* phi = cmp1;
2436 for (uint i = 1; i < phi->req(); i++) {
2437 Node* in = phi->in(i);
2438 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2439 &trunc1, &trunc2, &ttype);
2440 if (add && add->in(1) == phi) {
2441 add2 = add->in(2);
2442 break;
2443 }
2444 }
2445 } else {
2446 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2447 Node* addtrunc = cmp1;
2448 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2449 &trunc1, &trunc2, &ttype);
2450 if (add && add->in(1)->is_Phi()) {
2451 Node* phi = add->in(1);
2452 for (uint i = 1; i < phi->req(); i++) {
2453 if (phi->in(i) == addtrunc) {
2454 add2 = add->in(2);
2455 break;
2456 }
2457 }
2458 }
2459 }
2460 if (add2 != NULL) {
2461 const TypeInt* add2t = _igvn.type(add2)->is_int();
2462 if (add2t->is_con()) {
2463 return add2t->get_con();
2464 }
2465 }
2466 return 0;
2467 }
2468
2469
2470 //---------------------- stay_in_loop -------------------------------------
2471 // Return the (unique) control output node that's in the loop (if it exists.)
2472 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2473 Node* unique = NULL;
2474 if (!n) return NULL;
2475 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2476 Node* use = n->fast_out(i);
2477 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2478 if (unique != NULL) {
2479 return NULL;
2480 }
2481 unique = use;
2482 }
2483 }
2484 return unique;
2485 }
2486
2487 //------------------------------ register_node -------------------------------------
2488 // Utility to register node "n" with PhaseIdealLoop
2489 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
2490 _igvn.register_new_node_with_optimizer(n);
2491 loop->_body.push(n);
2492 if (n->is_CFG()) {
2493 set_loop(n, loop);
2494 set_idom(n, pred, ddepth);
2495 } else {
2496 set_ctrl(n, pred);
2497 }
2498 }
2499
2500 //------------------------------ proj_clone -------------------------------------
2501 // Utility to create an if-projection
2502 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2503 ProjNode* c = p->clone()->as_Proj();
2504 c->set_req(0, iff);
2505 return c;
2506 }
2507
2508 //------------------------------ short_circuit_if -------------------------------------
2509 // Force the iff control output to be the live_proj
2510 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2511 guarantee(live_proj != NULL, "null projection");
2512 int proj_con = live_proj->_con;
2513 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2514 Node *con = _igvn.intcon(proj_con);
2515 set_ctrl(con, C->root());
2516 if (iff) {
2517 iff->set_req(1, con);
2518 }
2519 return con;
2520 }
2521
2522 //------------------------------ insert_if_before_proj -------------------------------------
2523 // Insert a new if before an if projection (* - new node)
2524 //
2525 // before
2526 // if(test)
2527 // / \
2528 // v v
2529 // other-proj proj (arg)
2530 //
2531 // after
2532 // if(test)
2533 // / \
2534 // / v
2535 // | * proj-clone
2536 // v |
2537 // other-proj v
2538 // * new_if(relop(cmp[IU](left,right)))
2539 // / \
2540 // v v
2541 // * new-proj proj
2542 // (returned)
2543 //
2544 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2545 IfNode* iff = proj->in(0)->as_If();
2546 IdealLoopTree *loop = get_loop(proj);
2547 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2548 int ddepth = dom_depth(proj);
2549
2550 _igvn.rehash_node_delayed(iff);
2551 _igvn.rehash_node_delayed(proj);
2552
2553 proj->set_req(0, NULL); // temporary disconnect
2554 ProjNode* proj2 = proj_clone(proj, iff);
2555 register_node(proj2, loop, iff, ddepth);
2556
2557 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2558 register_node(cmp, loop, proj2, ddepth);
2559
2560 BoolNode* bol = new BoolNode(cmp, relop);
2561 register_node(bol, loop, proj2, ddepth);
2562
2563 int opcode = iff->Opcode();
2564 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2565 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
2566 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
2567 register_node(new_if, loop, proj2, ddepth);
2568
2569 proj->set_req(0, new_if); // reattach
2570 set_idom(proj, new_if, ddepth);
2571
2572 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2573 guarantee(new_exit != NULL, "null exit node");
2574 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2575
2576 return new_exit;
2577 }
2578
2579 //------------------------------ insert_region_before_proj -------------------------------------
2580 // Insert a region before an if projection (* - new node)
2581 //
2582 // before
2583 // if(test)
2584 // / |
2585 // v |
2586 // proj v
2587 // other-proj
2588 //
2589 // after
2590 // if(test)
2591 // / |
2592 // v |
2593 // * proj-clone v
2594 // | other-proj
2595 // v
2596 // * new-region
2597 // |
2598 // v
2599 // * dum_if
2600 // / \
2601 // v \
2602 // * dum-proj v
2603 // proj
2604 //
2605 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2606 IfNode* iff = proj->in(0)->as_If();
2607 IdealLoopTree *loop = get_loop(proj);
2608 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2609 int ddepth = dom_depth(proj);
2610
2611 _igvn.rehash_node_delayed(iff);
2612 _igvn.rehash_node_delayed(proj);
2613
2614 proj->set_req(0, NULL); // temporary disconnect
2615 ProjNode* proj2 = proj_clone(proj, iff);
2616 register_node(proj2, loop, iff, ddepth);
2617
2618 RegionNode* reg = new RegionNode(2);
2619 reg->set_req(1, proj2);
2620 register_node(reg, loop, iff, ddepth);
2621
2622 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
2623 register_node(dum_if, loop, reg, ddepth);
2624
2625 proj->set_req(0, dum_if); // reattach
2626 set_idom(proj, dum_if, ddepth);
2627
2628 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
2629 register_node(dum_proj, loop, dum_if, ddepth);
2630
2631 return reg;
2632 }
2633
2634 //------------------------------ insert_cmpi_loop_exit -------------------------------------
2635 // Clone a signed compare loop exit from an unsigned compare and
2636 // insert it before the unsigned cmp on the stay-in-loop path.
2637 // All new nodes inserted in the dominator tree between the original
2638 // if and it's projections. The original if test is replaced with
2639 // a constant to force the stay-in-loop path.
2640 //
2641 // This is done to make sure that the original if and it's projections
2642 // still dominate the same set of control nodes, that the ctrl() relation
2643 // from data nodes to them is preserved, and that their loop nesting is
2644 // preserved.
2645 //
2646 // before
2647 // if(i <u limit) unsigned compare loop exit
2648 // / |
2649 // v v
2650 // exit-proj stay-in-loop-proj
2651 //
2652 // after
2653 // if(stay-in-loop-const) original if
2654 // / |
2655 // / v
2656 // / if(i < limit) new signed test
2657 // / / |
2658 // / / v
2659 // / / if(i <u limit) new cloned unsigned test
2660 // / / / |
2661 // v v v |
2662 // region |
2663 // | |
2664 // dum-if |
2665 // / | |
2666 // ether | |
2667 // v v
2668 // exit-proj stay-in-loop-proj
2669 //
2670 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
2671 const bool Signed = true;
2672 const bool Unsigned = false;
2673
2674 BoolNode* bol = if_cmpu->in(1)->as_Bool();
2675 if (bol->_test._test != BoolTest::lt) return NULL;
2676 CmpNode* cmpu = bol->in(1)->as_Cmp();
2677 if (cmpu->Opcode() != Op_CmpU) return NULL;
2678 int stride = stride_of_possible_iv(if_cmpu);
2679 if (stride == 0) return NULL;
2680
2681 Node* lp_proj = stay_in_loop(if_cmpu, loop);
2682 guarantee(lp_proj != NULL, "null loop node");
2683
2684 ProjNode* lp_continue = lp_proj->as_Proj();
2685 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
2686
2687 Node* limit = NULL;
2688 if (stride > 0) {
2689 limit = cmpu->in(2);
2690 } else {
2691 limit = _igvn.makecon(TypeInt::ZERO);
2692 set_ctrl(limit, C->root());
2693 }
2694 // Create a new region on the exit path
2695 RegionNode* reg = insert_region_before_proj(lp_exit);
2696 guarantee(reg != NULL, "null region node");
2697
2698 // Clone the if-cmpu-true-false using a signed compare
2699 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
2700 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
2701 reg->add_req(cmpi_exit);
2702
2703 // Clone the if-cmpu-true-false
2704 BoolTest::mask rel_u = bol->_test._test;
2705 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
2706 reg->add_req(cmpu_exit);
2707
2708 // Force original if to stay in loop.
2709 short_circuit_if(if_cmpu, lp_continue);
2710
2711 return cmpi_exit->in(0)->as_If();
2712 }
2713
2714 //------------------------------ remove_cmpi_loop_exit -------------------------------------
2715 // Remove a previously inserted signed compare loop exit.
2716 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
2717 Node* lp_proj = stay_in_loop(if_cmp, loop);
2718 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
2719 stay_in_loop(lp_proj, loop)->is_If() &&
2720 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
2721 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
2722 set_ctrl(con, C->root());
2723 if_cmp->set_req(1, con);
2724 }
2725
2726 //------------------------------ scheduled_nodelist -------------------------------------
2727 // Create a post order schedule of nodes that are in the
2728 // "member" set. The list is returned in "sched".
2729 // The first node in "sched" is the loop head, followed by
2730 // nodes which have no inputs in the "member" set, and then
2731 // followed by the nodes that have an immediate input dependence
2732 // on a node in "sched".
2733 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
2734
2735 assert(member.test(loop->_head->_idx), "loop head must be in member set");
2736 Arena *a = Thread::current()->resource_area();
2737 VectorSet visited(a);
2738 Node_Stack nstack(a, loop->_body.size());
2739
2740 Node* n = loop->_head; // top of stack is cached in "n"
2741 uint idx = 0;
2742 visited.set(n->_idx);
2743
2744 // Initially push all with no inputs from within member set
2745 for(uint i = 0; i < loop->_body.size(); i++ ) {
2746 Node *elt = loop->_body.at(i);
2747 if (member.test(elt->_idx)) {
2748 bool found = false;
2749 for (uint j = 0; j < elt->req(); j++) {
2750 Node* def = elt->in(j);
2751 if (def && member.test(def->_idx) && def != elt) {
2752 found = true;
2753 break;
2754 }
2755 }
2756 if (!found && elt != loop->_head) {
2757 nstack.push(n, idx);
2758 n = elt;
2759 assert(!visited.test(n->_idx), "not seen yet");
2760 visited.set(n->_idx);
2761 }
2762 }
2763 }
2764
2765 // traverse out's that are in the member set
2766 while (true) {
2767 if (idx < n->outcnt()) {
2768 Node* use = n->raw_out(idx);
2769 idx++;
2770 if (!visited.test_set(use->_idx)) {
2771 if (member.test(use->_idx)) {
2772 nstack.push(n, idx);
2773 n = use;
2774 idx = 0;
2775 }
2776 }
2777 } else {
2778 // All outputs processed
2779 sched.push(n);
2780 if (nstack.is_empty()) break;
2781 n = nstack.node();
2782 idx = nstack.index();
2783 nstack.pop();
2784 }
2785 }
2786 }
2787
2788
2789 //------------------------------ has_use_in_set -------------------------------------
2790 // Has a use in the vector set
2791 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
2792 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2793 Node* use = n->fast_out(j);
2794 if (vset.test(use->_idx)) {
2795 return true;
2796 }
2797 }
2798 return false;
2799 }
2800
2801
2802 //------------------------------ has_use_internal_to_set -------------------------------------
2803 // Has use internal to the vector set (ie. not in a phi at the loop head)
2804 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
2805 Node* head = loop->_head;
2806 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2807 Node* use = n->fast_out(j);
2808 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
2809 return true;
2810 }
2811 }
2812 return false;
2813 }
2814
2815
2816 //------------------------------ clone_for_use_outside_loop -------------------------------------
2817 // clone "n" for uses that are outside of loop
2818 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
2819 int cloned = 0;
2820 assert(worklist.size() == 0, "should be empty");
2821 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2822 Node* use = n->fast_out(j);
2823 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
2824 worklist.push(use);
2825 }
2826 }
2827 while( worklist.size() ) {
2828 Node *use = worklist.pop();
2829 if (!has_node(use) || use->in(0) == C->top()) continue;
2830 uint j;
2831 for (j = 0; j < use->req(); j++) {
2832 if (use->in(j) == n) break;
2833 }
2834 assert(j < use->req(), "must be there");
2835
2836 // clone "n" and insert it between the inputs of "n" and the use outside the loop
2837 Node* n_clone = n->clone();
2838 _igvn.replace_input_of(use, j, n_clone);
2839 cloned++;
2840 Node* use_c;
2841 if (!use->is_Phi()) {
2842 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
2843 } else {
2844 // Use in a phi is considered a use in the associated predecessor block
2845 use_c = use->in(0)->in(j);
2846 }
2847 set_ctrl(n_clone, use_c);
2848 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
2849 get_loop(use_c)->_body.push(n_clone);
2850 _igvn.register_new_node_with_optimizer(n_clone);
2851 #if !defined(PRODUCT)
2852 if (TracePartialPeeling) {
2853 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
2854 }
2855 #endif
2856 }
2857 return cloned;
2858 }
2859
2860
2861 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
2862 // clone "n" for special uses that are in the not_peeled region.
2863 // If these def-uses occur in separate blocks, the code generator
2864 // marks the method as not compilable. For example, if a "BoolNode"
2865 // is in a different basic block than the "IfNode" that uses it, then
2866 // the compilation is aborted in the code generator.
2867 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
2868 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
2869 if (n->is_Phi() || n->is_Load()) {
2870 return;
2871 }
2872 assert(worklist.size() == 0, "should be empty");
2873 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2874 Node* use = n->fast_out(j);
2875 if ( not_peel.test(use->_idx) &&
2876 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
2877 use->in(1) == n) {
2878 worklist.push(use);
2879 }
2880 }
2881 if (worklist.size() > 0) {
2882 // clone "n" and insert it between inputs of "n" and the use
2883 Node* n_clone = n->clone();
2884 loop->_body.push(n_clone);
2885 _igvn.register_new_node_with_optimizer(n_clone);
2886 set_ctrl(n_clone, get_ctrl(n));
2887 sink_list.push(n_clone);
2888 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
2889 #if !defined(PRODUCT)
2890 if (TracePartialPeeling) {
2891 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2892 }
2893 #endif
2894 while( worklist.size() ) {
2895 Node *use = worklist.pop();
2896 _igvn.rehash_node_delayed(use);
2897 for (uint j = 1; j < use->req(); j++) {
2898 if (use->in(j) == n) {
2899 use->set_req(j, n_clone);
2900 }
2901 }
2902 }
2903 }
2904 }
2905
2906
2907 //------------------------------ insert_phi_for_loop -------------------------------------
2908 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2909 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2910 Node *phi = PhiNode::make(lp, back_edge_val);
2911 phi->set_req(LoopNode::EntryControl, lp_entry_val);
2912 // Use existing phi if it already exists
2913 Node *hit = _igvn.hash_find_insert(phi);
2914 if( hit == NULL ) {
2915 _igvn.register_new_node_with_optimizer(phi);
2916 set_ctrl(phi, lp);
2917 } else {
2918 // Remove the new phi from the graph and use the hit
2919 _igvn.remove_dead_node(phi);
2920 phi = hit;
2921 }
2922 _igvn.replace_input_of(use, idx, phi);
2923 }
2924
2925 #ifdef ASSERT
2926 //------------------------------ is_valid_loop_partition -------------------------------------
2927 // Validate the loop partition sets: peel and not_peel
2928 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2929 VectorSet& not_peel ) {
2930 uint i;
2931 // Check that peel_list entries are in the peel set
2932 for (i = 0; i < peel_list.size(); i++) {
2933 if (!peel.test(peel_list.at(i)->_idx)) {
2934 return false;
2935 }
2936 }
2937 // Check at loop members are in one of peel set or not_peel set
2938 for (i = 0; i < loop->_body.size(); i++ ) {
2939 Node *def = loop->_body.at(i);
2940 uint di = def->_idx;
2941 // Check that peel set elements are in peel_list
2942 if (peel.test(di)) {
2943 if (not_peel.test(di)) {
2944 return false;
2945 }
2946 // Must be in peel_list also
2947 bool found = false;
2948 for (uint j = 0; j < peel_list.size(); j++) {
2949 if (peel_list.at(j)->_idx == di) {
2950 found = true;
2951 break;
2952 }
2953 }
2954 if (!found) {
2955 return false;
2956 }
2957 } else if (not_peel.test(di)) {
2958 if (peel.test(di)) {
2959 return false;
2960 }
2961 } else {
2962 return false;
2963 }
2964 }
2965 return true;
2966 }
2967
2968 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2969 // Ensure a use outside of loop is of the right form
2970 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2971 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2972 return (use->is_Phi() &&
2973 use_c->is_Region() && use_c->req() == 3 &&
2974 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2975 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2976 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2977 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2978 }
2979
2980 //------------------------------ is_valid_clone_loop_form -------------------------------------
2981 // Ensure that all uses outside of loop are of the right form
2982 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2983 uint orig_exit_idx, uint clone_exit_idx) {
2984 uint len = peel_list.size();
2985 for (uint i = 0; i < len; i++) {
2986 Node *def = peel_list.at(i);
2987
2988 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2989 Node *use = def->fast_out(j);
2990 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2991 if (!loop->is_member(get_loop(use_c))) {
2992 // use is not in the loop, check for correct structure
2993 if (use->in(0) == def) {
2994 // Okay
2995 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2996 return false;
2997 }
2998 }
2999 }
3000 }
3001 return true;
3002 }
3003 #endif
3004
3005 //------------------------------ partial_peel -------------------------------------
3006 // Partially peel (aka loop rotation) the top portion of a loop (called
3007 // the peel section below) by cloning it and placing one copy just before
3008 // the new loop head and the other copy at the bottom of the new loop.
3009 //
3010 // before after where it came from
3011 //
3012 // stmt1 stmt1
3013 // loop: stmt2 clone
3014 // stmt2 if condA goto exitA clone
3015 // if condA goto exitA new_loop: new
3016 // stmt3 stmt3 clone
3017 // if !condB goto loop if condB goto exitB clone
3018 // exitB: stmt2 orig
3019 // stmt4 if !condA goto new_loop orig
3020 // exitA: goto exitA
3021 // exitB:
3022 // stmt4
3023 // exitA:
3024 //
3025 // Step 1: find the cut point: an exit test on probable
3026 // induction variable.
3027 // Step 2: schedule (with cloning) operations in the peel
3028 // section that can be executed after the cut into
3029 // the section that is not peeled. This may need
3030 // to clone operations into exit blocks. For
3031 // instance, a reference to A[i] in the not-peel
3032 // section and a reference to B[i] in an exit block
3033 // may cause a left-shift of i by 2 to be placed
3034 // in the peel block. This step will clone the left
3035 // shift into the exit block and sink the left shift
3036 // from the peel to the not-peel section.
3037 // Step 3: clone the loop, retarget the control, and insert
3038 // phis for values that are live across the new loop
3039 // head. This is very dependent on the graph structure
3040 // from clone_loop. It creates region nodes for
3041 // exit control and associated phi nodes for values
3042 // flow out of the loop through that exit. The region
3043 // node is dominated by the clone's control projection.
3044 // So the clone's peel section is placed before the
3045 // new loop head, and the clone's not-peel section is
3046 // forms the top part of the new loop. The original
3047 // peel section forms the tail of the new loop.
3048 // Step 4: update the dominator tree and recompute the
3049 // dominator depth.
3050 //
3051 // orig
3052 //
3053 // stmt1
3054 // |
3055 // v
3056 // loop predicate
3057 // |
3058 // v
3059 // loop<----+
3060 // | |
3061 // stmt2 |
3062 // | |
3063 // v |
3064 // ifA |
3065 // / | |
3066 // v v |
3067 // false true ^ <-- last_peel
3068 // / | |
3069 // / ===|==cut |
3070 // / stmt3 | <-- first_not_peel
3071 // / | |
3072 // | v |
3073 // v ifB |
3074 // exitA: / \ |
3075 // / \ |
3076 // v v |
3077 // false true |
3078 // / \ |
3079 // / ----+
3080 // |
3081 // v
3082 // exitB:
3083 // stmt4
3084 //
3085 //
3086 // after clone loop
3087 //
3088 // stmt1
3089 // |
3090 // v
3091 // loop predicate
3092 // / \
3093 // clone / \ orig
3094 // / \
3095 // / \
3096 // v v
3097 // +---->loop loop<----+
3098 // | | | |
3099 // | stmt2 stmt2 |
3100 // | | | |
3101 // | v v |
3102 // | ifA ifA |
3103 // | | \ / | |
3104 // | v v v v |
3105 // ^ true false false true ^ <-- last_peel
3106 // | | ^ \ / | |
3107 // | cut==|== \ \ / ===|==cut |
3108 // | stmt3 \ \ / stmt3 | <-- first_not_peel
3109 // | | dom | | | |
3110 // | v \ 1v v2 v |
3111 // | ifB regionA ifB |
3112 // | / \ | / \ |
3113 // | / \ v / \ |
3114 // | v v exitA: v v |
3115 // | true false false true |
3116 // | / ^ \ / \ |
3117 // +---- \ \ / ----+
3118 // dom \ /
3119 // \ 1v v2
3120 // regionB
3121 // |
3122 // v
3123 // exitB:
3124 // stmt4
3125 //
3126 //
3127 // after partial peel
3128 //
3129 // stmt1
3130 // |
3131 // v
3132 // loop predicate
3133 // /
3134 // clone / orig
3135 // / TOP
3136 // / \
3137 // v v
3138 // TOP->loop loop----+
3139 // | | |
3140 // stmt2 stmt2 |
3141 // | | |
3142 // v v |
3143 // ifA ifA |
3144 // | \ / | |
3145 // v v v v |
3146 // true false false true | <-- last_peel
3147 // | ^ \ / +------|---+
3148 // +->newloop \ \ / === ==cut | |
3149 // | stmt3 \ \ / TOP | |
3150 // | | dom | | stmt3 | | <-- first_not_peel
3151 // | v \ 1v v2 v | |
3152 // | ifB regionA ifB ^ v
3153 // | / \ | / \ | |
3154 // | / \ v / \ | |
3155 // | v v exitA: v v | |
3156 // | true false false true | |
3157 // | / ^ \ / \ | |
3158 // | | \ \ / v | |
3159 // | | dom \ / TOP | |
3160 // | | \ 1v v2 | |
3161 // ^ v regionB | |
3162 // | | | | |
3163 // | | v ^ v
3164 // | | exitB: | |
3165 // | | stmt4 | |
3166 // | +------------>-----------------+ |
3167 // | |
3168 // +-----------------<---------------------+
3169 //
3170 //
3171 // final graph
3172 //
3173 // stmt1
3174 // |
3175 // v
3176 // loop predicate
3177 // |
3178 // v
3179 // stmt2 clone
3180 // |
3181 // v
3182 // ........> ifA clone
3183 // : / |
3184 // dom / |
3185 // : v v
3186 // : false true
3187 // : | |
3188 // : | v
3189 // : | newloop<-----+
3190 // : | | |
3191 // : | stmt3 clone |
3192 // : | | |
3193 // : | v |
3194 // : | ifB |
3195 // : | / \ |
3196 // : | v v |
3197 // : | false true |
3198 // : | | | |
3199 // : | v stmt2 |
3200 // : | exitB: | |
3201 // : | stmt4 v |
3202 // : | ifA orig |
3203 // : | / \ |
3204 // : | / \ |
3205 // : | v v |
3206 // : | false true |
3207 // : | / \ |
3208 // : v v -----+
3209 // RegionA
3210 // |
3211 // v
3212 // exitA
3213 //
3214 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3215
3216 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3217 if (!loop->_head->is_Loop()) {
3218 return false; }
3219
3220 LoopNode *head = loop->_head->as_Loop();
3221
3222 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3223 return false;
3224 }
3225
3226 // Check for complex exit control
3227 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
3228 Node *n = loop->_body.at(ii);
3229 int opc = n->Opcode();
3230 if (n->is_Call() ||
3231 opc == Op_Catch ||
3232 opc == Op_CatchProj ||
3233 opc == Op_Jump ||
3234 opc == Op_JumpProj) {
3235 #if !defined(PRODUCT)
3236 if (TracePartialPeeling) {
3237 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3238 }
3239 #endif
3240 return false;
3241 }
3242 }
3243
3244 int dd = dom_depth(head);
3245
3246 // Step 1: find cut point
3247
3248 // Walk up dominators to loop head looking for first loop exit
3249 // which is executed on every path thru loop.
3250 IfNode *peel_if = NULL;
3251 IfNode *peel_if_cmpu = NULL;
3252
3253 Node *iff = loop->tail();
3254 while( iff != head ) {
3255 if( iff->is_If() ) {
3256 Node *ctrl = get_ctrl(iff->in(1));
3257 if (ctrl->is_top()) return false; // Dead test on live IF.
3258 // If loop-varying exit-test, check for induction variable
3259 if( loop->is_member(get_loop(ctrl)) &&
3260 loop->is_loop_exit(iff) &&
3261 is_possible_iv_test(iff)) {
3262 Node* cmp = iff->in(1)->in(1);
3263 if (cmp->Opcode() == Op_CmpI) {
3264 peel_if = iff->as_If();
3265 } else {
3266 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3267 peel_if_cmpu = iff->as_If();
3268 }
3269 }
3270 }
3271 iff = idom(iff);
3272 }
3273 // Prefer signed compare over unsigned compare.
3274 IfNode* new_peel_if = NULL;
3275 if (peel_if == NULL) {
3276 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
3277 return false; // No peel point found
3278 }
3279 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3280 if (new_peel_if == NULL) {
3281 return false; // No peel point found
3282 }
3283 peel_if = new_peel_if;
3284 }
3285 Node* last_peel = stay_in_loop(peel_if, loop);
3286 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3287 if (first_not_peeled == NULL || first_not_peeled == head) {
3288 return false;
3289 }
3290
3291 #if !defined(PRODUCT)
3292 if (TraceLoopOpts) {
3293 tty->print("PartialPeel ");
3294 loop->dump_head();
3295 }
3296
3297 if (TracePartialPeeling) {
3298 tty->print_cr("before partial peel one iteration");
3299 Node_List wl;
3300 Node* t = head->in(2);
3301 while (true) {
3302 wl.push(t);
3303 if (t == head) break;
3304 t = idom(t);
3305 }
3306 while (wl.size() > 0) {
3307 Node* tt = wl.pop();
3308 tt->dump();
3309 if (tt == last_peel) tty->print_cr("-- cut --");
3310 }
3311 }
3312 #endif
3313 ResourceArea *area = Thread::current()->resource_area();
3314 VectorSet peel(area);
3315 VectorSet not_peel(area);
3316 Node_List peel_list(area);
3317 Node_List worklist(area);
3318 Node_List sink_list(area);
3319
3320 // Set of cfg nodes to peel are those that are executable from
3321 // the head through last_peel.
3322 assert(worklist.size() == 0, "should be empty");
3323 worklist.push(head);
3324 peel.set(head->_idx);
3325 while (worklist.size() > 0) {
3326 Node *n = worklist.pop();
3327 if (n != last_peel) {
3328 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3329 Node* use = n->fast_out(j);
3330 if (use->is_CFG() &&
3331 loop->is_member(get_loop(use)) &&
3332 !peel.test_set(use->_idx)) {
3333 worklist.push(use);
3334 }
3335 }
3336 }
3337 }
3338
3339 // Set of non-cfg nodes to peel are those that are control
3340 // dependent on the cfg nodes.
3341 uint i;
3342 for(i = 0; i < loop->_body.size(); i++ ) {
3343 Node *n = loop->_body.at(i);
3344 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3345 if (peel.test(n_c->_idx)) {
3346 peel.set(n->_idx);
3347 } else {
3348 not_peel.set(n->_idx);
3349 }
3350 }
3351
3352 // Step 2: move operations from the peeled section down into the
3353 // not-peeled section
3354
3355 // Get a post order schedule of nodes in the peel region
3356 // Result in right-most operand.
3357 scheduled_nodelist(loop, peel, peel_list );
3358
3359 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3360
3361 // For future check for too many new phis
3362 uint old_phi_cnt = 0;
3363 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3364 Node* use = head->fast_out(j);
3365 if (use->is_Phi()) old_phi_cnt++;
3366 }
3367
3368 #if !defined(PRODUCT)
3369 if (TracePartialPeeling) {
3370 tty->print_cr("\npeeled list");
3371 }
3372 #endif
3373
3374 // Evacuate nodes in peel region into the not_peeled region if possible
3375 uint new_phi_cnt = 0;
3376 uint cloned_for_outside_use = 0;
3377 for (i = 0; i < peel_list.size();) {
3378 Node* n = peel_list.at(i);
3379 #if !defined(PRODUCT)
3380 if (TracePartialPeeling) n->dump();
3381 #endif
3382 bool incr = true;
3383 if ( !n->is_CFG() ) {
3384
3385 if ( has_use_in_set(n, not_peel) ) {
3386
3387 // If not used internal to the peeled region,
3388 // move "n" from peeled to not_peeled region.
3389
3390 if ( !has_use_internal_to_set(n, peel, loop) ) {
3391
3392 // if not pinned and not a load (which maybe anti-dependent on a store)
3393 // and not a CMove (Matcher expects only bool->cmove).
3394 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
3395 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist );
3396 sink_list.push(n);
3397 peel >>= n->_idx; // delete n from peel set.
3398 not_peel <<= n->_idx; // add n to not_peel set.
3399 peel_list.remove(i);
3400 incr = false;
3401 #if !defined(PRODUCT)
3402 if (TracePartialPeeling) {
3403 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3404 n->_idx, get_ctrl(n)->_idx);
3405 }
3406 #endif
3407 }
3408 } else {
3409 // Otherwise check for special def-use cases that span
3410 // the peel/not_peel boundary such as bool->if
3411 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
3412 new_phi_cnt++;
3413 }
3414 }
3415 }
3416 if (incr) i++;
3417 }
3418
3419 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
3420 #if !defined(PRODUCT)
3421 if (TracePartialPeeling) {
3422 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3423 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
3424 }
3425 #endif
3426 if (new_peel_if != NULL) {
3427 remove_cmpi_loop_exit(new_peel_if, loop);
3428 }
3429 // Inhibit more partial peeling on this loop
3430 assert(!head->is_partial_peel_loop(), "not partial peeled");
3431 head->mark_partial_peel_failed();
3432 if (cloned_for_outside_use > 0) {
3433 // Terminate this round of loop opts because
3434 // the graph outside this loop was changed.
3435 C->set_major_progress();
3436 return true;
3437 }
3438 return false;
3439 }
3440
3441 // Step 3: clone loop, retarget control, and insert new phis
3442
3443 // Create new loop head for new phis and to hang
3444 // the nodes being moved (sinked) from the peel region.
3445 LoopNode* new_head = new LoopNode(last_peel, last_peel);
3446 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3447 _igvn.register_new_node_with_optimizer(new_head);
3448 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3449 _igvn.replace_input_of(first_not_peeled, 0, new_head);
3450 set_loop(new_head, loop);
3451 loop->_body.push(new_head);
3452 not_peel.set(new_head->_idx);
3453 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3454 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3455
3456 while (sink_list.size() > 0) {
3457 Node* n = sink_list.pop();
3458 set_ctrl(n, new_head);
3459 }
3460
3461 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3462
3463 clone_loop(loop, old_new, dd, IgnoreStripMined);
3464
3465 const uint clone_exit_idx = 1;
3466 const uint orig_exit_idx = 2;
3467 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
3468
3469 Node* head_clone = old_new[head->_idx];
3470 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
3471 Node* orig_tail_clone = head_clone->in(2);
3472
3473 // Add phi if "def" node is in peel set and "use" is not
3474
3475 for(i = 0; i < peel_list.size(); i++ ) {
3476 Node *def = peel_list.at(i);
3477 if (!def->is_CFG()) {
3478 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3479 Node *use = def->fast_out(j);
3480 if (has_node(use) && use->in(0) != C->top() &&
3481 (!peel.test(use->_idx) ||
3482 (use->is_Phi() && use->in(0) == head)) ) {
3483 worklist.push(use);
3484 }
3485 }
3486 while( worklist.size() ) {
3487 Node *use = worklist.pop();
3488 for (uint j = 1; j < use->req(); j++) {
3489 Node* n = use->in(j);
3490 if (n == def) {
3491
3492 // "def" is in peel set, "use" is not in peel set
3493 // or "use" is in the entry boundary (a phi) of the peel set
3494
3495 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
3496
3497 if ( loop->is_member(get_loop( use_c )) ) {
3498 // use is in loop
3499 if (old_new[use->_idx] != NULL) { // null for dead code
3500 Node* use_clone = old_new[use->_idx];
3501 _igvn.replace_input_of(use, j, C->top());
3502 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
3503 }
3504 } else {
3505 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
3506 // use is not in the loop, check if the live range includes the cut
3507 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
3508 if (not_peel.test(lp_if->_idx)) {
3509 assert(j == orig_exit_idx, "use from original loop");
3510 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
3511 }
3512 }
3513 }
3514 }
3515 }
3516 }
3517 }
3518
3519 // Step 3b: retarget control
3520
3521 // Redirect control to the new loop head if a cloned node in
3522 // the not_peeled region has control that points into the peeled region.
3523 // This necessary because the cloned peeled region will be outside
3524 // the loop.
3525 // from to
3526 // cloned-peeled <---+
3527 // new_head_clone: | <--+
3528 // cloned-not_peeled in(0) in(0)
3529 // orig-peeled
3530
3531 for(i = 0; i < loop->_body.size(); i++ ) {
3532 Node *n = loop->_body.at(i);
3533 if (!n->is_CFG() && n->in(0) != NULL &&
3534 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
3535 Node* n_clone = old_new[n->_idx];
3536 _igvn.replace_input_of(n_clone, 0, new_head_clone);
3537 }
3538 }
3539
3540 // Backedge of the surviving new_head (the clone) is original last_peel
3541 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
3542
3543 // Cut first node in original not_peel set
3544 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
3545 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
3546 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
3547
3548 // Copy head_clone back-branch info to original head
3549 // and remove original head's loop entry and
3550 // clone head's back-branch
3551 _igvn.rehash_node_delayed(head); // Multiple edge updates
3552 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
3553 head->set_req(LoopNode::LoopBackControl, C->top());
3554 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
3555
3556 // Similarly modify the phis
3557 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
3558 Node* use = head->fast_out(k);
3559 if (use->is_Phi() && use->outcnt() > 0) {
3560 Node* use_clone = old_new[use->_idx];
3561 _igvn.rehash_node_delayed(use); // Multiple edge updates
3562 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
3563 use->set_req(LoopNode::LoopBackControl, C->top());
3564 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
3565 }
3566 }
3567
3568 // Step 4: update dominator tree and dominator depth
3569
3570 set_idom(head, orig_tail_clone, dd);
3571 recompute_dom_depth();
3572
3573 // Inhibit more partial peeling on this loop
3574 new_head_clone->set_partial_peel_loop();
3575 C->set_major_progress();
3576 loop->record_for_igvn();
3577
3578 #if !defined(PRODUCT)
3579 if (TracePartialPeeling) {
3580 tty->print_cr("\nafter partial peel one iteration");
3581 Node_List wl(area);
3582 Node* t = last_peel;
3583 while (true) {
3584 wl.push(t);
3585 if (t == head_clone) break;
3586 t = idom(t);
3587 }
3588 while (wl.size() > 0) {
3589 Node* tt = wl.pop();
3590 if (tt == head) tty->print_cr("orig head");
3591 else if (tt == new_head_clone) tty->print_cr("new head");
3592 else if (tt == head_clone) tty->print_cr("clone head");
3593 tt->dump();
3594 }
3595 }
3596 #endif
3597 return true;
3598 }
3599
3600 //------------------------------reorg_offsets----------------------------------
3601 // Reorganize offset computations to lower register pressure. Mostly
3602 // prevent loop-fallout uses of the pre-incremented trip counter (which are
3603 // then alive with the post-incremented trip counter forcing an extra
3604 // register move)
3605 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
3606 // Perform it only for canonical counted loops.
3607 // Loop's shape could be messed up by iteration_split_impl.
3608 if (!loop->_head->is_CountedLoop())
3609 return;
3610 if (!loop->_head->as_Loop()->is_valid_counted_loop())
3611 return;
3612
3613 CountedLoopNode *cl = loop->_head->as_CountedLoop();
3614 CountedLoopEndNode *cle = cl->loopexit();
3615 Node *exit = cle->proj_out(false);
3616 Node *phi = cl->phi();
3617
3618 // Check for the special case of folks using the pre-incremented
3619 // trip-counter on the fall-out path (forces the pre-incremented
3620 // and post-incremented trip counter to be live at the same time).
3621 // Fix this by adjusting to use the post-increment trip counter.
3622
3623 bool progress = true;
3624 while (progress) {
3625 progress = false;
3626 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3627 Node* use = phi->fast_out(i); // User of trip-counter
3628 if (!has_ctrl(use)) continue;
3629 Node *u_ctrl = get_ctrl(use);
3630 if (use->is_Phi()) {
3631 u_ctrl = NULL;
3632 for (uint j = 1; j < use->req(); j++)
3633 if (use->in(j) == phi)
3634 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
3635 }
3636 IdealLoopTree *u_loop = get_loop(u_ctrl);
3637 // Look for loop-invariant use
3638 if (u_loop == loop) continue;
3639 if (loop->is_member(u_loop)) continue;
3640 // Check that use is live out the bottom. Assuming the trip-counter
3641 // update is right at the bottom, uses of of the loop middle are ok.
3642 if (dom_lca(exit, u_ctrl) != exit) continue;
3643 // Hit! Refactor use to use the post-incremented tripcounter.
3644 // Compute a post-increment tripcounter.
3645 Node *opaq = new Opaque2Node( C, cle->incr() );
3646 register_new_node(opaq, exit);
3647 Node *neg_stride = _igvn.intcon(-cle->stride_con());
3648 set_ctrl(neg_stride, C->root());
3649 Node *post = new AddINode( opaq, neg_stride);
3650 register_new_node(post, exit);
3651 _igvn.rehash_node_delayed(use);
3652 for (uint j = 1; j < use->req(); j++) {
3653 if (use->in(j) == phi)
3654 use->set_req(j, post);
3655 }
3656 // Since DU info changed, rerun loop
3657 progress = true;
3658 break;
3659 }
3660 }
3661
3662 }