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