Wed, 16 Nov 2011 09:13:57 -0800
6890673: Eliminate allocations immediately after EA
Summary: Try to eliminate allocations and related locks immediately after escape analysis.
Reviewed-by: never
1 /*
2 * Copyright (c) 1999, 2011, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "memory/allocation.inline.hpp"
27 #include "opto/addnode.hpp"
28 #include "opto/connode.hpp"
29 #include "opto/divnode.hpp"
30 #include "opto/loopnode.hpp"
31 #include "opto/matcher.hpp"
32 #include "opto/mulnode.hpp"
33 #include "opto/rootnode.hpp"
34 #include "opto/subnode.hpp"
36 //=============================================================================
37 //------------------------------split_thru_phi---------------------------------
38 // Split Node 'n' through merge point if there is enough win.
39 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
40 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
41 // ConvI2L may have type information on it which is unsafe to push up
42 // so disable this for now
43 return NULL;
44 }
45 int wins = 0;
46 assert(!n->is_CFG(), "");
47 assert(region->is_Region(), "");
49 const Type* type = n->bottom_type();
50 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr();
51 Node *phi;
52 if (t_oop != NULL && t_oop->is_known_instance_field()) {
53 int iid = t_oop->instance_id();
54 int index = C->get_alias_index(t_oop);
55 int offset = t_oop->offset();
56 phi = new (C,region->req()) PhiNode(region, type, NULL, iid, index, offset);
57 } else {
58 phi = PhiNode::make_blank(region, n);
59 }
60 uint old_unique = C->unique();
61 for (uint i = 1; i < region->req(); i++) {
62 Node *x;
63 Node* the_clone = NULL;
64 if (region->in(i) == C->top()) {
65 x = C->top(); // Dead path? Use a dead data op
66 } else {
67 x = n->clone(); // Else clone up the data op
68 the_clone = x; // Remember for possible deletion.
69 // Alter data node to use pre-phi inputs
70 if (n->in(0) == region)
71 x->set_req( 0, region->in(i) );
72 for (uint j = 1; j < n->req(); j++) {
73 Node *in = n->in(j);
74 if (in->is_Phi() && in->in(0) == region)
75 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
76 }
77 }
78 // Check for a 'win' on some paths
79 const Type *t = x->Value(&_igvn);
81 bool singleton = t->singleton();
83 // A TOP singleton indicates that there are no possible values incoming
84 // along a particular edge. In most cases, this is OK, and the Phi will
85 // be eliminated later in an Ideal call. However, we can't allow this to
86 // happen if the singleton occurs on loop entry, as the elimination of
87 // the PhiNode may cause the resulting node to migrate back to a previous
88 // loop iteration.
89 if (singleton && t == Type::TOP) {
90 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
91 // irreducible loop may not be indicated by an affirmative is_Loop());
92 // therefore, the only top we can split thru a phi is on a backedge of
93 // a loop.
94 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
95 }
97 if (singleton) {
98 wins++;
99 x = ((PhaseGVN&)_igvn).makecon(t);
100 } else {
101 // We now call Identity to try to simplify the cloned node.
102 // Note that some Identity methods call phase->type(this).
103 // Make sure that the type array is big enough for
104 // our new node, even though we may throw the node away.
105 // (Note: This tweaking with igvn only works because x is a new node.)
106 _igvn.set_type(x, t);
107 // If x is a TypeNode, capture any more-precise type permanently into Node
108 // otherwise it will be not updated during igvn->transform since
109 // igvn->type(x) is set to x->Value() already.
110 x->raise_bottom_type(t);
111 Node *y = x->Identity(&_igvn);
112 if (y != x) {
113 wins++;
114 x = y;
115 } else {
116 y = _igvn.hash_find(x);
117 if (y) {
118 wins++;
119 x = y;
120 } else {
121 // Else x is a new node we are keeping
122 // We do not need register_new_node_with_optimizer
123 // because set_type has already been called.
124 _igvn._worklist.push(x);
125 }
126 }
127 }
128 if (x != the_clone && the_clone != NULL)
129 _igvn.remove_dead_node(the_clone);
130 phi->set_req( i, x );
131 }
132 // Too few wins?
133 if (wins <= policy) {
134 _igvn.remove_dead_node(phi);
135 return NULL;
136 }
138 // Record Phi
139 register_new_node( phi, region );
141 for (uint i2 = 1; i2 < phi->req(); i2++) {
142 Node *x = phi->in(i2);
143 // If we commoned up the cloned 'x' with another existing Node,
144 // the existing Node picks up a new use. We need to make the
145 // existing Node occur higher up so it dominates its uses.
146 Node *old_ctrl;
147 IdealLoopTree *old_loop;
149 if (x->is_Con()) {
150 // Constant's control is always root.
151 set_ctrl(x, C->root());
152 continue;
153 }
154 // The occasional new node
155 if (x->_idx >= old_unique) { // Found a new, unplaced node?
156 old_ctrl = NULL;
157 old_loop = NULL; // Not in any prior loop
158 } else {
159 old_ctrl = get_ctrl(x);
160 old_loop = get_loop(old_ctrl); // Get prior loop
161 }
162 // New late point must dominate new use
163 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
164 if (new_ctrl == old_ctrl) // Nothing is changed
165 continue;
167 IdealLoopTree *new_loop = get_loop(new_ctrl);
169 // Don't move x into a loop if its uses are
170 // outside of loop. Otherwise x will be cloned
171 // for each use outside of this loop.
172 IdealLoopTree *use_loop = get_loop(region);
173 if (!new_loop->is_member(use_loop) &&
174 (old_loop == NULL || !new_loop->is_member(old_loop))) {
175 // Take early control, later control will be recalculated
176 // during next iteration of loop optimizations.
177 new_ctrl = get_early_ctrl(x);
178 new_loop = get_loop(new_ctrl);
179 }
180 // Set new location
181 set_ctrl(x, new_ctrl);
182 // If changing loop bodies, see if we need to collect into new body
183 if (old_loop != new_loop) {
184 if (old_loop && !old_loop->_child)
185 old_loop->_body.yank(x);
186 if (!new_loop->_child)
187 new_loop->_body.push(x); // Collect body info
188 }
189 }
191 return phi;
192 }
194 //------------------------------dominated_by------------------------------------
195 // Replace the dominated test with an obvious true or false. Place it on the
196 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
197 // live path up to the dominating control.
198 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) {
199 #ifndef PRODUCT
200 if (VerifyLoopOptimizations && PrintOpto) tty->print_cr("dominating test");
201 #endif
204 // prevdom is the dominating projection of the dominating test.
205 assert( iff->is_If(), "" );
206 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
207 int pop = prevdom->Opcode();
208 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
209 if (flip) {
210 if (pop == Op_IfTrue)
211 pop = Op_IfFalse;
212 else
213 pop = Op_IfTrue;
214 }
215 // 'con' is set to true or false to kill the dominated test.
216 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
217 set_ctrl(con, C->root()); // Constant gets a new use
218 // Hack the dominated test
219 _igvn.hash_delete(iff);
220 iff->set_req(1, con);
221 _igvn._worklist.push(iff);
223 // If I dont have a reachable TRUE and FALSE path following the IfNode then
224 // I can assume this path reaches an infinite loop. In this case it's not
225 // important to optimize the data Nodes - either the whole compilation will
226 // be tossed or this path (and all data Nodes) will go dead.
227 if (iff->outcnt() != 2) return;
229 // Make control-dependent data Nodes on the live path (path that will remain
230 // once the dominated IF is removed) become control-dependent on the
231 // dominating projection.
232 Node* dp = iff->as_If()->proj_out(pop == Op_IfTrue);
234 // Loop predicates may have depending checks which should not
235 // be skipped. For example, range check predicate has two checks
236 // for lower and upper bounds.
237 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp->as_Proj()->_con)->as_Proj();
238 if (exclude_loop_predicate &&
239 is_uncommon_trap_proj(unc_proj, Deoptimization::Reason_predicate))
240 return; // Let IGVN transformation change control dependence.
242 IdealLoopTree *old_loop = get_loop(dp);
244 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
245 Node* cd = dp->fast_out(i); // Control-dependent node
246 if (cd->depends_only_on_test()) {
247 assert(cd->in(0) == dp, "");
248 _igvn.hash_delete(cd);
249 cd->set_req(0, prevdom);
250 set_early_ctrl(cd);
251 _igvn._worklist.push(cd);
252 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
253 if (old_loop != new_loop) {
254 if (!old_loop->_child) old_loop->_body.yank(cd);
255 if (!new_loop->_child) new_loop->_body.push(cd);
256 }
257 --i;
258 --imax;
259 }
260 }
261 }
263 //------------------------------has_local_phi_input----------------------------
264 // Return TRUE if 'n' has Phi inputs from its local block and no other
265 // block-local inputs (all non-local-phi inputs come from earlier blocks)
266 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
267 Node *n_ctrl = get_ctrl(n);
268 // See if some inputs come from a Phi in this block, or from before
269 // this block.
270 uint i;
271 for( i = 1; i < n->req(); i++ ) {
272 Node *phi = n->in(i);
273 if( phi->is_Phi() && phi->in(0) == n_ctrl )
274 break;
275 }
276 if( i >= n->req() )
277 return NULL; // No Phi inputs; nowhere to clone thru
279 // Check for inputs created between 'n' and the Phi input. These
280 // must split as well; they have already been given the chance
281 // (courtesy of a post-order visit) and since they did not we must
282 // recover the 'cost' of splitting them by being very profitable
283 // when splitting 'n'. Since this is unlikely we simply give up.
284 for( i = 1; i < n->req(); i++ ) {
285 Node *m = n->in(i);
286 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
287 // We allow the special case of AddP's with no local inputs.
288 // This allows us to split-up address expressions.
289 if (m->is_AddP() &&
290 get_ctrl(m->in(2)) != n_ctrl &&
291 get_ctrl(m->in(3)) != n_ctrl) {
292 // Move the AddP up to dominating point
293 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl)));
294 continue;
295 }
296 return NULL;
297 }
298 }
300 return n_ctrl;
301 }
303 //------------------------------remix_address_expressions----------------------
304 // Rework addressing expressions to get the most loop-invariant stuff
305 // moved out. We'd like to do all associative operators, but it's especially
306 // important (common) to do address expressions.
307 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
308 if (!has_ctrl(n)) return NULL;
309 Node *n_ctrl = get_ctrl(n);
310 IdealLoopTree *n_loop = get_loop(n_ctrl);
312 // See if 'n' mixes loop-varying and loop-invariant inputs and
313 // itself is loop-varying.
315 // Only interested in binary ops (and AddP)
316 if( n->req() < 3 || n->req() > 4 ) return NULL;
318 Node *n1_ctrl = get_ctrl(n->in( 1));
319 Node *n2_ctrl = get_ctrl(n->in( 2));
320 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
321 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
322 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
323 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
325 // Does one of my inputs spin in a tighter loop than self?
326 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
327 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
328 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
329 return NULL; // Leave well enough alone
331 // Is at least one of my inputs loop-invariant?
332 if( n1_loop == n_loop &&
333 n2_loop == n_loop &&
334 n3_loop == n_loop )
335 return NULL; // No loop-invariant inputs
338 int n_op = n->Opcode();
340 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
341 if( n_op == Op_LShiftI ) {
342 // Scale is loop invariant
343 Node *scale = n->in(2);
344 Node *scale_ctrl = get_ctrl(scale);
345 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
346 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
347 return NULL;
348 const TypeInt *scale_t = scale->bottom_type()->isa_int();
349 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
350 return NULL; // Dont bother with byte/short masking
351 // Add must vary with loop (else shift would be loop-invariant)
352 Node *add = n->in(1);
353 Node *add_ctrl = get_ctrl(add);
354 IdealLoopTree *add_loop = get_loop(add_ctrl);
355 //assert( n_loop == add_loop, "" );
356 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
358 // Convert I-V into I+ (0-V); same for V-I
359 if( add->Opcode() == Op_SubI &&
360 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
361 Node *zero = _igvn.intcon(0);
362 set_ctrl(zero, C->root());
363 Node *neg = new (C, 3) SubINode( _igvn.intcon(0), add->in(2) );
364 register_new_node( neg, get_ctrl(add->in(2) ) );
365 add = new (C, 3) AddINode( add->in(1), neg );
366 register_new_node( add, add_ctrl );
367 }
368 if( add->Opcode() != Op_AddI ) return NULL;
369 // See if one add input is loop invariant
370 Node *add_var = add->in(1);
371 Node *add_var_ctrl = get_ctrl(add_var);
372 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
373 Node *add_invar = add->in(2);
374 Node *add_invar_ctrl = get_ctrl(add_invar);
375 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
376 if( add_var_loop == n_loop ) {
377 } else if( add_invar_loop == n_loop ) {
378 // Swap to find the invariant part
379 add_invar = add_var;
380 add_invar_ctrl = add_var_ctrl;
381 add_invar_loop = add_var_loop;
382 add_var = add->in(2);
383 Node *add_var_ctrl = get_ctrl(add_var);
384 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
385 } else // Else neither input is loop invariant
386 return NULL;
387 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
388 return NULL; // No invariant part of the add?
390 // Yes! Reshape address expression!
391 Node *inv_scale = new (C, 3) LShiftINode( add_invar, scale );
392 Node *inv_scale_ctrl =
393 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
394 add_invar_ctrl : scale_ctrl;
395 register_new_node( inv_scale, inv_scale_ctrl );
396 Node *var_scale = new (C, 3) LShiftINode( add_var, scale );
397 register_new_node( var_scale, n_ctrl );
398 Node *var_add = new (C, 3) AddINode( var_scale, inv_scale );
399 register_new_node( var_add, n_ctrl );
400 _igvn.replace_node( n, var_add );
401 return var_add;
402 }
404 // Replace (I+V) with (V+I)
405 if( n_op == Op_AddI ||
406 n_op == Op_AddL ||
407 n_op == Op_AddF ||
408 n_op == Op_AddD ||
409 n_op == Op_MulI ||
410 n_op == Op_MulL ||
411 n_op == Op_MulF ||
412 n_op == Op_MulD ) {
413 if( n2_loop == n_loop ) {
414 assert( n1_loop != n_loop, "" );
415 n->swap_edges(1, 2);
416 }
417 }
419 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
420 // but not if I2 is a constant.
421 if( n_op == Op_AddP ) {
422 if( n2_loop == n_loop && n3_loop != n_loop ) {
423 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
424 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
425 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
426 IdealLoopTree *n22loop = get_loop( n22_ctrl );
427 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
428 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
429 n23_loop == n_loop ) {
430 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
431 // Stuff new AddP in the loop preheader
432 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
433 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) );
434 register_new_node( add2, n_ctrl );
435 _igvn.replace_node( n, add2 );
436 return add2;
437 }
438 }
439 }
441 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
442 if( n2_loop != n_loop && n3_loop == n_loop ) {
443 if( n->in(3)->Opcode() == Op_AddI ) {
444 Node *V = n->in(3)->in(1);
445 Node *I = n->in(3)->in(2);
446 if( is_member(n_loop,get_ctrl(V)) ) {
447 } else {
448 Node *tmp = V; V = I; I = tmp;
449 }
450 if( !is_member(n_loop,get_ctrl(I)) ) {
451 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I );
452 // Stuff new AddP in the loop preheader
453 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
454 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V );
455 register_new_node( add2, n_ctrl );
456 _igvn.replace_node( n, add2 );
457 return add2;
458 }
459 }
460 }
461 }
463 return NULL;
464 }
466 //------------------------------conditional_move-------------------------------
467 // Attempt to replace a Phi with a conditional move. We have some pretty
468 // strict profitability requirements. All Phis at the merge point must
469 // be converted, so we can remove the control flow. We need to limit the
470 // number of c-moves to a small handful. All code that was in the side-arms
471 // of the CFG diamond is now speculatively executed. This code has to be
472 // "cheap enough". We are pretty much limited to CFG diamonds that merge
473 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
474 Node *PhaseIdealLoop::conditional_move( Node *region ) {
476 assert(region->is_Region(), "sanity check");
477 if (region->req() != 3) return NULL;
479 // Check for CFG diamond
480 Node *lp = region->in(1);
481 Node *rp = region->in(2);
482 if (!lp || !rp) return NULL;
483 Node *lp_c = lp->in(0);
484 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
485 IfNode *iff = lp_c->as_If();
487 // Check for ops pinned in an arm of the diamond.
488 // Can't remove the control flow in this case
489 if (lp->outcnt() > 1) return NULL;
490 if (rp->outcnt() > 1) return NULL;
492 IdealLoopTree* r_loop = get_loop(region);
493 assert(r_loop == get_loop(iff), "sanity");
494 // Always convert to CMOVE if all results are used only outside this loop.
495 bool used_inside_loop = (r_loop == _ltree_root);
497 // Check profitability
498 int cost = 0;
499 int phis = 0;
500 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
501 Node *out = region->fast_out(i);
502 if (!out->is_Phi()) continue; // Ignore other control edges, etc
503 phis++;
504 PhiNode* phi = out->as_Phi();
505 BasicType bt = phi->type()->basic_type();
506 switch (bt) {
507 case T_FLOAT:
508 case T_DOUBLE: {
509 cost += Matcher::float_cmove_cost(); // Could be very expensive
510 break;
511 }
512 case T_LONG: {
513 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
514 }
515 case T_INT: // These all CMOV fine
516 case T_ADDRESS: { // (RawPtr)
517 cost++;
518 break;
519 }
520 case T_NARROWOOP: // Fall through
521 case T_OBJECT: { // Base oops are OK, but not derived oops
522 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
523 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
524 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
525 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
526 // have a Phi for the base here that we convert to a CMOVE all is well
527 // and good. But if the base is dead, we'll not make a CMOVE. Later
528 // the allocator will have to produce a base by creating a CMOVE of the
529 // relevant bases. This puts the allocator in the business of
530 // manufacturing expensive instructions, generally a bad plan.
531 // Just Say No to Conditionally-Moved Derived Pointers.
532 if (tp && tp->offset() != 0)
533 return NULL;
534 cost++;
535 break;
536 }
537 default:
538 return NULL; // In particular, can't do memory or I/O
539 }
540 // Add in cost any speculative ops
541 for (uint j = 1; j < region->req(); j++) {
542 Node *proj = region->in(j);
543 Node *inp = phi->in(j);
544 if (get_ctrl(inp) == proj) { // Found local op
545 cost++;
546 // Check for a chain of dependent ops; these will all become
547 // speculative in a CMOV.
548 for (uint k = 1; k < inp->req(); k++)
549 if (get_ctrl(inp->in(k)) == proj)
550 cost += ConditionalMoveLimit; // Too much speculative goo
551 }
552 }
553 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
554 // This will likely Split-If, a higher-payoff operation.
555 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
556 Node* use = phi->fast_out(k);
557 if (use->is_Cmp() || use->is_DecodeN() || use->is_EncodeP())
558 cost += ConditionalMoveLimit;
559 // Is there a use inside the loop?
560 // Note: check only basic types since CMoveP is pinned.
561 if (!used_inside_loop && is_java_primitive(bt)) {
562 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
563 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
564 used_inside_loop = true;
565 }
566 }
567 }
568 }
569 Node* bol = iff->in(1);
570 assert(bol->Opcode() == Op_Bool, "");
571 int cmp_op = bol->in(1)->Opcode();
572 // It is expensive to generate flags from a float compare.
573 // Avoid duplicated float compare.
574 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
576 float infrequent_prob = PROB_UNLIKELY_MAG(3);
577 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
578 if (used_inside_loop) {
579 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
581 // BlockLayoutByFrequency optimization moves infrequent branch
582 // from hot path. No point in CMOV'ing in such case (110 is used
583 // instead of 100 to take into account not exactness of float value).
584 if (BlockLayoutByFrequency) {
585 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
586 }
587 }
588 // Check for highly predictable branch. No point in CMOV'ing if
589 // we are going to predict accurately all the time.
590 if (iff->_prob < infrequent_prob ||
591 iff->_prob > (1.0f - infrequent_prob))
592 return NULL;
594 // --------------
595 // Now replace all Phis with CMOV's
596 Node *cmov_ctrl = iff->in(0);
597 uint flip = (lp->Opcode() == Op_IfTrue);
598 while (1) {
599 PhiNode* phi = NULL;
600 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
601 Node *out = region->fast_out(i);
602 if (out->is_Phi()) {
603 phi = out->as_Phi();
604 break;
605 }
606 }
607 if (phi == NULL) break;
608 #ifndef PRODUCT
609 if (PrintOpto && VerifyLoopOptimizations) tty->print_cr("CMOV");
610 #endif
611 // Move speculative ops
612 for (uint j = 1; j < region->req(); j++) {
613 Node *proj = region->in(j);
614 Node *inp = phi->in(j);
615 if (get_ctrl(inp) == proj) { // Found local op
616 #ifndef PRODUCT
617 if (PrintOpto && VerifyLoopOptimizations) {
618 tty->print(" speculate: ");
619 inp->dump();
620 }
621 #endif
622 set_ctrl(inp, cmov_ctrl);
623 }
624 }
625 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) );
626 register_new_node( cmov, cmov_ctrl );
627 _igvn.replace_node( phi, cmov );
628 #ifndef PRODUCT
629 if (TraceLoopOpts) {
630 tty->print("CMOV ");
631 r_loop->dump_head();
632 if (Verbose) {
633 bol->in(1)->dump(1);
634 cmov->dump(1);
635 }
636 }
637 if (VerifyLoopOptimizations) verify();
638 #endif
639 }
641 // The useless CFG diamond will fold up later; see the optimization in
642 // RegionNode::Ideal.
643 _igvn._worklist.push(region);
645 return iff->in(1);
646 }
648 //------------------------------split_if_with_blocks_pre-----------------------
649 // Do the real work in a non-recursive function. Data nodes want to be
650 // cloned in the pre-order so they can feed each other nicely.
651 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
652 // Cloning these guys is unlikely to win
653 int n_op = n->Opcode();
654 if( n_op == Op_MergeMem ) return n;
655 if( n->is_Proj() ) return n;
656 // Do not clone-up CmpFXXX variations, as these are always
657 // followed by a CmpI
658 if( n->is_Cmp() ) return n;
659 // Attempt to use a conditional move instead of a phi/branch
660 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
661 Node *cmov = conditional_move( n );
662 if( cmov ) return cmov;
663 }
664 if( n->is_CFG() || n->is_LoadStore() )
665 return n;
666 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
667 n_op == Op_Opaque2 ) {
668 if( !C->major_progress() ) // If chance of no more loop opts...
669 _igvn._worklist.push(n); // maybe we'll remove them
670 return n;
671 }
673 if( n->is_Con() ) return n; // No cloning for Con nodes
675 Node *n_ctrl = get_ctrl(n);
676 if( !n_ctrl ) return n; // Dead node
678 // Attempt to remix address expressions for loop invariants
679 Node *m = remix_address_expressions( n );
680 if( m ) return m;
682 // Determine if the Node has inputs from some local Phi.
683 // Returns the block to clone thru.
684 Node *n_blk = has_local_phi_input( n );
685 if( !n_blk ) return n;
686 // Do not clone the trip counter through on a CountedLoop
687 // (messes up the canonical shape).
688 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
690 // Check for having no control input; not pinned. Allow
691 // dominating control.
692 if( n->in(0) ) {
693 Node *dom = idom(n_blk);
694 if( dom_lca( n->in(0), dom ) != n->in(0) )
695 return n;
696 }
697 // Policy: when is it profitable. You must get more wins than
698 // policy before it is considered profitable. Policy is usually 0,
699 // so 1 win is considered profitable. Big merges will require big
700 // cloning, so get a larger policy.
701 int policy = n_blk->req() >> 2;
703 // If the loop is a candidate for range check elimination,
704 // delay splitting through it's phi until a later loop optimization
705 if (n_blk->is_CountedLoop()) {
706 IdealLoopTree *lp = get_loop(n_blk);
707 if (lp && lp->_rce_candidate) {
708 return n;
709 }
710 }
712 // Use same limit as split_if_with_blocks_post
713 if( C->unique() > 35000 ) return n; // Method too big
715 // Split 'n' through the merge point if it is profitable
716 Node *phi = split_thru_phi( n, n_blk, policy );
717 if (!phi) return n;
719 // Found a Phi to split thru!
720 // Replace 'n' with the new phi
721 _igvn.replace_node( n, phi );
722 // Moved a load around the loop, 'en-registering' something.
723 if (n_blk->is_Loop() && n->is_Load() &&
724 !phi->in(LoopNode::LoopBackControl)->is_Load())
725 C->set_major_progress();
727 return phi;
728 }
730 static bool merge_point_too_heavy(Compile* C, Node* region) {
731 // Bail out if the region and its phis have too many users.
732 int weight = 0;
733 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
734 weight += region->fast_out(i)->outcnt();
735 }
736 int nodes_left = MaxNodeLimit - C->unique();
737 if (weight * 8 > nodes_left) {
738 #ifndef PRODUCT
739 if (PrintOpto)
740 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
741 #endif
742 return true;
743 } else {
744 return false;
745 }
746 }
748 static bool merge_point_safe(Node* region) {
749 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
750 // having a PhiNode input. This sidesteps the dangerous case where the split
751 // ConvI2LNode may become TOP if the input Value() does not
752 // overlap the ConvI2L range, leaving a node which may not dominate its
753 // uses.
754 // A better fix for this problem can be found in the BugTraq entry, but
755 // expediency for Mantis demands this hack.
756 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop
757 // split_if_with_blocks from splitting a block because we could not move around
758 // the FastLockNode.
759 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
760 Node* n = region->fast_out(i);
761 if (n->is_Phi()) {
762 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
763 Node* m = n->fast_out(j);
764 if (m->is_FastLock())
765 return false;
766 #ifdef _LP64
767 if (m->Opcode() == Op_ConvI2L)
768 return false;
769 #endif
770 }
771 }
772 }
773 return true;
774 }
777 //------------------------------place_near_use---------------------------------
778 // Place some computation next to use but not inside inner loops.
779 // For inner loop uses move it to the preheader area.
780 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const {
781 IdealLoopTree *u_loop = get_loop( useblock );
782 return (u_loop->_irreducible || u_loop->_child)
783 ? useblock
784 : u_loop->_head->in(LoopNode::EntryControl);
785 }
788 //------------------------------split_if_with_blocks_post----------------------
789 // Do the real work in a non-recursive function. CFG hackery wants to be
790 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
791 // info.
792 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) {
794 // Cloning Cmp through Phi's involves the split-if transform.
795 // FastLock is not used by an If
796 if( n->is_Cmp() && !n->is_FastLock() ) {
797 if( C->unique() > 35000 ) return; // Method too big
799 // Do not do 'split-if' if irreducible loops are present.
800 if( _has_irreducible_loops )
801 return;
803 Node *n_ctrl = get_ctrl(n);
804 // Determine if the Node has inputs from some local Phi.
805 // Returns the block to clone thru.
806 Node *n_blk = has_local_phi_input( n );
807 if( n_blk != n_ctrl ) return;
809 if( merge_point_too_heavy(C, n_ctrl) )
810 return;
812 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare?
813 Node *bol = n->unique_out();
814 assert( bol->is_Bool(), "expect a bool here" );
815 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare?
816 Node *iff = bol->unique_out();
818 // Check some safety conditions
819 if( iff->is_If() ) { // Classic split-if?
820 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if
821 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
822 if( get_ctrl(iff->in(2)) == n_ctrl ||
823 get_ctrl(iff->in(3)) == n_ctrl )
824 return; // Inputs not yet split-up
825 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) {
826 return; // Loop-invar test gates loop-varying CMOVE
827 }
828 } else {
829 return; // some other kind of node, such as an Allocate
830 }
832 // Do not do 'split-if' if some paths are dead. First do dead code
833 // elimination and then see if its still profitable.
834 for( uint i = 1; i < n_ctrl->req(); i++ )
835 if( n_ctrl->in(i) == C->top() )
836 return;
838 // When is split-if profitable? Every 'win' on means some control flow
839 // goes dead, so it's almost always a win.
840 int policy = 0;
841 // If trying to do a 'Split-If' at the loop head, it is only
842 // profitable if the cmp folds up on BOTH paths. Otherwise we
843 // risk peeling a loop forever.
845 // CNC - Disabled for now. Requires careful handling of loop
846 // body selection for the cloned code. Also, make sure we check
847 // for any input path not being in the same loop as n_ctrl. For
848 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
849 // because the alternative loop entry points won't be converted
850 // into LoopNodes.
851 IdealLoopTree *n_loop = get_loop(n_ctrl);
852 for( uint j = 1; j < n_ctrl->req(); j++ )
853 if( get_loop(n_ctrl->in(j)) != n_loop )
854 return;
856 // Check for safety of the merge point.
857 if( !merge_point_safe(n_ctrl) ) {
858 return;
859 }
861 // Split compare 'n' through the merge point if it is profitable
862 Node *phi = split_thru_phi( n, n_ctrl, policy );
863 if( !phi ) return;
865 // Found a Phi to split thru!
866 // Replace 'n' with the new phi
867 _igvn.replace_node( n, phi );
869 // Now split the bool up thru the phi
870 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 );
871 _igvn.replace_node( bol, bolphi );
872 assert( iff->in(1) == bolphi, "" );
873 if( bolphi->Value(&_igvn)->singleton() )
874 return;
876 // Conditional-move? Must split up now
877 if( !iff->is_If() ) {
878 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 );
879 _igvn.replace_node( iff, cmovphi );
880 return;
881 }
883 // Now split the IF
884 do_split_if( iff );
885 return;
886 }
888 // Check for an IF ready to split; one that has its
889 // condition codes input coming from a Phi at the block start.
890 int n_op = n->Opcode();
892 // Check for an IF being dominated by another IF same test
893 if( n_op == Op_If ) {
894 Node *bol = n->in(1);
895 uint max = bol->outcnt();
896 // Check for same test used more than once?
897 if( n_op == Op_If && max > 1 && bol->is_Bool() ) {
898 // Search up IDOMs to see if this IF is dominated.
899 Node *cutoff = get_ctrl(bol);
901 // Now search up IDOMs till cutoff, looking for a dominating test
902 Node *prevdom = n;
903 Node *dom = idom(prevdom);
904 while( dom != cutoff ) {
905 if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) {
906 // Replace the dominated test with an obvious true or false.
907 // Place it on the IGVN worklist for later cleanup.
908 C->set_major_progress();
909 dominated_by( prevdom, n, false, true );
910 #ifndef PRODUCT
911 if( VerifyLoopOptimizations ) verify();
912 #endif
913 return;
914 }
915 prevdom = dom;
916 dom = idom(prevdom);
917 }
918 }
919 }
921 // See if a shared loop-varying computation has no loop-varying uses.
922 // Happens if something is only used for JVM state in uncommon trap exits,
923 // like various versions of induction variable+offset. Clone the
924 // computation per usage to allow it to sink out of the loop.
925 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
926 Node *n_ctrl = get_ctrl(n);
927 IdealLoopTree *n_loop = get_loop(n_ctrl);
928 if( n_loop != _ltree_root ) {
929 DUIterator_Fast imax, i = n->fast_outs(imax);
930 for (; i < imax; i++) {
931 Node* u = n->fast_out(i);
932 if( !has_ctrl(u) ) break; // Found control user
933 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
934 if( u_loop == n_loop ) break; // Found loop-varying use
935 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
936 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
937 }
938 bool did_break = (i < imax); // Did we break out of the previous loop?
939 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
940 Node *late_load_ctrl;
941 if (n->is_Load()) {
942 // If n is a load, get and save the result from get_late_ctrl(),
943 // to be later used in calculating the control for n's clones.
944 clear_dom_lca_tags();
945 late_load_ctrl = get_late_ctrl(n, n_ctrl);
946 }
947 // If n is a load, and the late control is the same as the current
948 // control, then the cloning of n is a pointless exercise, because
949 // GVN will ensure that we end up where we started.
950 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
951 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
952 Node *u = n->last_out(j); // Clone private computation per use
953 _igvn.hash_delete(u);
954 _igvn._worklist.push(u);
955 Node *x = n->clone(); // Clone computation
956 Node *x_ctrl = NULL;
957 if( u->is_Phi() ) {
958 // Replace all uses of normal nodes. Replace Phi uses
959 // individually, so the separate Nodes can sink down
960 // different paths.
961 uint k = 1;
962 while( u->in(k) != n ) k++;
963 u->set_req( k, x );
964 // x goes next to Phi input path
965 x_ctrl = u->in(0)->in(k);
966 --j;
967 } else { // Normal use
968 // Replace all uses
969 for( uint k = 0; k < u->req(); k++ ) {
970 if( u->in(k) == n ) {
971 u->set_req( k, x );
972 --j;
973 }
974 }
975 x_ctrl = get_ctrl(u);
976 }
978 // Find control for 'x' next to use but not inside inner loops.
979 // For inner loop uses get the preheader area.
980 x_ctrl = place_near_use(x_ctrl);
982 if (n->is_Load()) {
983 // For loads, add a control edge to a CFG node outside of the loop
984 // to force them to not combine and return back inside the loop
985 // during GVN optimization (4641526).
986 //
987 // Because we are setting the actual control input, factor in
988 // the result from get_late_ctrl() so we respect any
989 // anti-dependences. (6233005).
990 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
992 // Don't allow the control input to be a CFG splitting node.
993 // Such nodes should only have ProjNodes as outs, e.g. IfNode
994 // should only have IfTrueNode and IfFalseNode (4985384).
995 x_ctrl = find_non_split_ctrl(x_ctrl);
996 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
998 x->set_req(0, x_ctrl);
999 }
1000 register_new_node(x, x_ctrl);
1002 // Some institutional knowledge is needed here: 'x' is
1003 // yanked because if the optimizer runs GVN on it all the
1004 // cloned x's will common up and undo this optimization and
1005 // be forced back in the loop. This is annoying because it
1006 // makes +VerifyOpto report false-positives on progress. I
1007 // tried setting control edges on the x's to force them to
1008 // not combine, but the matching gets worried when it tries
1009 // to fold a StoreP and an AddP together (as part of an
1010 // address expression) and the AddP and StoreP have
1011 // different controls.
1012 if( !x->is_Load() && !x->is_DecodeN() ) _igvn._worklist.yank(x);
1013 }
1014 _igvn.remove_dead_node(n);
1015 }
1016 }
1017 }
1018 }
1020 // Check for Opaque2's who's loop has disappeared - who's input is in the
1021 // same loop nest as their output. Remove 'em, they are no longer useful.
1022 if( n_op == Op_Opaque2 &&
1023 n->in(1) != NULL &&
1024 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1025 _igvn.replace_node( n, n->in(1) );
1026 }
1027 }
1029 //------------------------------split_if_with_blocks---------------------------
1030 // Check for aggressive application of 'split-if' optimization,
1031 // using basic block level info.
1032 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) {
1033 Node *n = C->root();
1034 visited.set(n->_idx); // first, mark node as visited
1035 // Do pre-visit work for root
1036 n = split_if_with_blocks_pre( n );
1037 uint cnt = n->outcnt();
1038 uint i = 0;
1039 while (true) {
1040 // Visit all children
1041 if (i < cnt) {
1042 Node* use = n->raw_out(i);
1043 ++i;
1044 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1045 // Now do pre-visit work for this use
1046 use = split_if_with_blocks_pre( use );
1047 nstack.push(n, i); // Save parent and next use's index.
1048 n = use; // Process all children of current use.
1049 cnt = use->outcnt();
1050 i = 0;
1051 }
1052 }
1053 else {
1054 // All of n's children have been processed, complete post-processing.
1055 if (cnt != 0 && !n->is_Con()) {
1056 assert(has_node(n), "no dead nodes");
1057 split_if_with_blocks_post( n );
1058 }
1059 if (nstack.is_empty()) {
1060 // Finished all nodes on stack.
1061 break;
1062 }
1063 // Get saved parent node and next use's index. Visit the rest of uses.
1064 n = nstack.node();
1065 cnt = n->outcnt();
1066 i = nstack.index();
1067 nstack.pop();
1068 }
1069 }
1070 }
1073 //=============================================================================
1074 //
1075 // C L O N E A L O O P B O D Y
1076 //
1078 //------------------------------clone_iff--------------------------------------
1079 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1080 // "Nearly" because all Nodes have been cloned from the original in the loop,
1081 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1082 // through the Phi recursively, and return a Bool.
1083 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) {
1085 // Convert this Phi into a Phi merging Bools
1086 uint i;
1087 for( i = 1; i < phi->req(); i++ ) {
1088 Node *b = phi->in(i);
1089 if( b->is_Phi() ) {
1090 _igvn.hash_delete(phi);
1091 _igvn._worklist.push(phi);
1092 phi->set_req(i, clone_iff( b->as_Phi(), loop ));
1093 } else {
1094 assert( b->is_Bool(), "" );
1095 }
1096 }
1098 Node *sample_bool = phi->in(1);
1099 Node *sample_cmp = sample_bool->in(1);
1101 // Make Phis to merge the Cmp's inputs.
1102 int size = phi->in(0)->req();
1103 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1104 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1105 for( i = 1; i < phi->req(); i++ ) {
1106 Node *n1 = phi->in(i)->in(1)->in(1);
1107 Node *n2 = phi->in(i)->in(1)->in(2);
1108 phi1->set_req( i, n1 );
1109 phi2->set_req( i, n2 );
1110 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1111 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1112 }
1113 // See if these Phis have been made before.
1114 // Register with optimizer
1115 Node *hit1 = _igvn.hash_find_insert(phi1);
1116 if( hit1 ) { // Hit, toss just made Phi
1117 _igvn.remove_dead_node(phi1); // Remove new phi
1118 assert( hit1->is_Phi(), "" );
1119 phi1 = (PhiNode*)hit1; // Use existing phi
1120 } else { // Miss
1121 _igvn.register_new_node_with_optimizer(phi1);
1122 }
1123 Node *hit2 = _igvn.hash_find_insert(phi2);
1124 if( hit2 ) { // Hit, toss just made Phi
1125 _igvn.remove_dead_node(phi2); // Remove new phi
1126 assert( hit2->is_Phi(), "" );
1127 phi2 = (PhiNode*)hit2; // Use existing phi
1128 } else { // Miss
1129 _igvn.register_new_node_with_optimizer(phi2);
1130 }
1131 // Register Phis with loop/block info
1132 set_ctrl(phi1, phi->in(0));
1133 set_ctrl(phi2, phi->in(0));
1134 // Make a new Cmp
1135 Node *cmp = sample_cmp->clone();
1136 cmp->set_req( 1, phi1 );
1137 cmp->set_req( 2, phi2 );
1138 _igvn.register_new_node_with_optimizer(cmp);
1139 set_ctrl(cmp, phi->in(0));
1141 // Make a new Bool
1142 Node *b = sample_bool->clone();
1143 b->set_req(1,cmp);
1144 _igvn.register_new_node_with_optimizer(b);
1145 set_ctrl(b, phi->in(0));
1147 assert( b->is_Bool(), "" );
1148 return (BoolNode*)b;
1149 }
1151 //------------------------------clone_bool-------------------------------------
1152 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1153 // "Nearly" because all Nodes have been cloned from the original in the loop,
1154 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1155 // through the Phi recursively, and return a Bool.
1156 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1157 uint i;
1158 // Convert this Phi into a Phi merging Bools
1159 for( i = 1; i < phi->req(); i++ ) {
1160 Node *b = phi->in(i);
1161 if( b->is_Phi() ) {
1162 _igvn.hash_delete(phi);
1163 _igvn._worklist.push(phi);
1164 phi->set_req(i, clone_bool( b->as_Phi(), loop ));
1165 } else {
1166 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1167 }
1168 }
1170 Node *sample_cmp = phi->in(1);
1172 // Make Phis to merge the Cmp's inputs.
1173 int size = phi->in(0)->req();
1174 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1175 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1176 for( uint j = 1; j < phi->req(); j++ ) {
1177 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1178 Node *n1, *n2;
1179 if( cmp_top->is_Cmp() ) {
1180 n1 = cmp_top->in(1);
1181 n2 = cmp_top->in(2);
1182 } else {
1183 n1 = n2 = cmp_top;
1184 }
1185 phi1->set_req( j, n1 );
1186 phi2->set_req( j, n2 );
1187 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1188 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1189 }
1191 // See if these Phis have been made before.
1192 // Register with optimizer
1193 Node *hit1 = _igvn.hash_find_insert(phi1);
1194 if( hit1 ) { // Hit, toss just made Phi
1195 _igvn.remove_dead_node(phi1); // Remove new phi
1196 assert( hit1->is_Phi(), "" );
1197 phi1 = (PhiNode*)hit1; // Use existing phi
1198 } else { // Miss
1199 _igvn.register_new_node_with_optimizer(phi1);
1200 }
1201 Node *hit2 = _igvn.hash_find_insert(phi2);
1202 if( hit2 ) { // Hit, toss just made Phi
1203 _igvn.remove_dead_node(phi2); // Remove new phi
1204 assert( hit2->is_Phi(), "" );
1205 phi2 = (PhiNode*)hit2; // Use existing phi
1206 } else { // Miss
1207 _igvn.register_new_node_with_optimizer(phi2);
1208 }
1209 // Register Phis with loop/block info
1210 set_ctrl(phi1, phi->in(0));
1211 set_ctrl(phi2, phi->in(0));
1212 // Make a new Cmp
1213 Node *cmp = sample_cmp->clone();
1214 cmp->set_req( 1, phi1 );
1215 cmp->set_req( 2, phi2 );
1216 _igvn.register_new_node_with_optimizer(cmp);
1217 set_ctrl(cmp, phi->in(0));
1219 assert( cmp->is_Cmp(), "" );
1220 return (CmpNode*)cmp;
1221 }
1223 //------------------------------sink_use---------------------------------------
1224 // If 'use' was in the loop-exit block, it now needs to be sunk
1225 // below the post-loop merge point.
1226 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1227 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1228 set_ctrl(use, post_loop);
1229 for (DUIterator j = use->outs(); use->has_out(j); j++)
1230 sink_use(use->out(j), post_loop);
1231 }
1232 }
1234 //------------------------------clone_loop-------------------------------------
1235 //
1236 // C L O N E A L O O P B O D Y
1237 //
1238 // This is the basic building block of the loop optimizations. It clones an
1239 // entire loop body. It makes an old_new loop body mapping; with this mapping
1240 // you can find the new-loop equivalent to an old-loop node. All new-loop
1241 // nodes are exactly equal to their old-loop counterparts, all edges are the
1242 // same. All exits from the old-loop now have a RegionNode that merges the
1243 // equivalent new-loop path. This is true even for the normal "loop-exit"
1244 // condition. All uses of loop-invariant old-loop values now come from (one
1245 // or more) Phis that merge their new-loop equivalents.
1246 //
1247 // This operation leaves the graph in an illegal state: there are two valid
1248 // control edges coming from the loop pre-header to both loop bodies. I'll
1249 // definitely have to hack the graph after running this transform.
1250 //
1251 // From this building block I will further edit edges to perform loop peeling
1252 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1253 //
1254 // Parameter side_by_size_idom:
1255 // When side_by_size_idom is NULL, the dominator tree is constructed for
1256 // the clone loop to dominate the original. Used in construction of
1257 // pre-main-post loop sequence.
1258 // When nonnull, the clone and original are side-by-side, both are
1259 // dominated by the side_by_side_idom node. Used in construction of
1260 // unswitched loops.
1261 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1262 Node* side_by_side_idom) {
1264 // Step 1: Clone the loop body. Make the old->new mapping.
1265 uint i;
1266 for( i = 0; i < loop->_body.size(); i++ ) {
1267 Node *old = loop->_body.at(i);
1268 Node *nnn = old->clone();
1269 old_new.map( old->_idx, nnn );
1270 _igvn.register_new_node_with_optimizer(nnn);
1271 }
1274 // Step 2: Fix the edges in the new body. If the old input is outside the
1275 // loop use it. If the old input is INside the loop, use the corresponding
1276 // new node instead.
1277 for( i = 0; i < loop->_body.size(); i++ ) {
1278 Node *old = loop->_body.at(i);
1279 Node *nnn = old_new[old->_idx];
1280 // Fix CFG/Loop controlling the new node
1281 if (has_ctrl(old)) {
1282 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
1283 } else {
1284 set_loop(nnn, loop->_parent);
1285 if (old->outcnt() > 0) {
1286 set_idom( nnn, old_new[idom(old)->_idx], dd );
1287 }
1288 }
1289 // Correct edges to the new node
1290 for( uint j = 0; j < nnn->req(); j++ ) {
1291 Node *n = nnn->in(j);
1292 if( n ) {
1293 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
1294 if( loop->is_member( old_in_loop ) )
1295 nnn->set_req(j, old_new[n->_idx]);
1296 }
1297 }
1298 _igvn.hash_find_insert(nnn);
1299 }
1300 Node *newhead = old_new[loop->_head->_idx];
1301 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1304 // Step 3: Now fix control uses. Loop varying control uses have already
1305 // been fixed up (as part of all input edges in Step 2). Loop invariant
1306 // control uses must be either an IfFalse or an IfTrue. Make a merge
1307 // point to merge the old and new IfFalse/IfTrue nodes; make the use
1308 // refer to this.
1309 ResourceArea *area = Thread::current()->resource_area();
1310 Node_List worklist(area);
1311 uint new_counter = C->unique();
1312 for( i = 0; i < loop->_body.size(); i++ ) {
1313 Node* old = loop->_body.at(i);
1314 if( !old->is_CFG() ) continue;
1315 Node* nnn = old_new[old->_idx];
1317 // Copy uses to a worklist, so I can munge the def-use info
1318 // with impunity.
1319 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1320 worklist.push(old->fast_out(j));
1322 while( worklist.size() ) { // Visit all uses
1323 Node *use = worklist.pop();
1324 if (!has_node(use)) continue; // Ignore dead nodes
1325 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1326 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
1327 // Both OLD and USE are CFG nodes here.
1328 assert( use->is_Proj(), "" );
1330 // Clone the loop exit control projection
1331 Node *newuse = use->clone();
1332 newuse->set_req(0,nnn);
1333 _igvn.register_new_node_with_optimizer(newuse);
1334 set_loop(newuse, use_loop);
1335 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
1337 // We need a Region to merge the exit from the peeled body and the
1338 // exit from the old loop body.
1339 RegionNode *r = new (C, 3) RegionNode(3);
1340 // Map the old use to the new merge point
1341 old_new.map( use->_idx, r );
1342 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
1343 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
1345 // The original user of 'use' uses 'r' instead.
1346 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
1347 Node* useuse = use->last_out(l);
1348 _igvn.hash_delete(useuse);
1349 _igvn._worklist.push(useuse);
1350 uint uses_found = 0;
1351 if( useuse->in(0) == use ) {
1352 useuse->set_req(0, r);
1353 uses_found++;
1354 if( useuse->is_CFG() ) {
1355 assert( dom_depth(useuse) > dd_r, "" );
1356 set_idom(useuse, r, dom_depth(useuse));
1357 }
1358 }
1359 for( uint k = 1; k < useuse->req(); k++ ) {
1360 if( useuse->in(k) == use ) {
1361 useuse->set_req(k, r);
1362 uses_found++;
1363 }
1364 }
1365 l -= uses_found; // we deleted 1 or more copies of this edge
1366 }
1368 // Now finish up 'r'
1369 r->set_req( 1, newuse );
1370 r->set_req( 2, use );
1371 _igvn.register_new_node_with_optimizer(r);
1372 set_loop(r, use_loop);
1373 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
1374 } // End of if a loop-exit test
1375 }
1376 }
1378 // Step 4: If loop-invariant use is not control, it must be dominated by a
1379 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
1380 // there if needed. Make a Phi there merging old and new used values.
1381 Node_List *split_if_set = NULL;
1382 Node_List *split_bool_set = NULL;
1383 Node_List *split_cex_set = NULL;
1384 for( i = 0; i < loop->_body.size(); i++ ) {
1385 Node* old = loop->_body.at(i);
1386 Node* nnn = old_new[old->_idx];
1387 // Copy uses to a worklist, so I can munge the def-use info
1388 // with impunity.
1389 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1390 worklist.push(old->fast_out(j));
1392 while( worklist.size() ) {
1393 Node *use = worklist.pop();
1394 if (!has_node(use)) continue; // Ignore dead nodes
1395 if (use->in(0) == C->top()) continue;
1396 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1397 // Check for data-use outside of loop - at least one of OLD or USE
1398 // must not be a CFG node.
1399 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) {
1401 // If the Data use is an IF, that means we have an IF outside of the
1402 // loop that is switching on a condition that is set inside of the
1403 // loop. Happens if people set a loop-exit flag; then test the flag
1404 // in the loop to break the loop, then test is again outside of the
1405 // loop to determine which way the loop exited.
1406 if( use->is_If() || use->is_CMove() ) {
1407 // Since this code is highly unlikely, we lazily build the worklist
1408 // of such Nodes to go split.
1409 if( !split_if_set )
1410 split_if_set = new Node_List(area);
1411 split_if_set->push(use);
1412 }
1413 if( use->is_Bool() ) {
1414 if( !split_bool_set )
1415 split_bool_set = new Node_List(area);
1416 split_bool_set->push(use);
1417 }
1418 if( use->Opcode() == Op_CreateEx ) {
1419 if( !split_cex_set )
1420 split_cex_set = new Node_List(area);
1421 split_cex_set->push(use);
1422 }
1425 // Get "block" use is in
1426 uint idx = 0;
1427 while( use->in(idx) != old ) idx++;
1428 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1429 assert( !loop->is_member( get_loop( prev ) ), "" );
1430 Node *cfg = prev->_idx >= new_counter
1431 ? prev->in(2)
1432 : idom(prev);
1433 if( use->is_Phi() ) // Phi use is in prior block
1434 cfg = prev->in(idx); // NOT in block of Phi itself
1435 if (cfg->is_top()) { // Use is dead?
1436 _igvn.hash_delete(use);
1437 _igvn._worklist.push(use);
1438 use->set_req(idx, C->top());
1439 continue;
1440 }
1442 while( !loop->is_member( get_loop( cfg ) ) ) {
1443 prev = cfg;
1444 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1445 }
1446 // If the use occurs after merging several exits from the loop, then
1447 // old value must have dominated all those exits. Since the same old
1448 // value was used on all those exits we did not need a Phi at this
1449 // merge point. NOW we do need a Phi here. Each loop exit value
1450 // is now merged with the peeled body exit; each exit gets its own
1451 // private Phi and those Phis need to be merged here.
1452 Node *phi;
1453 if( prev->is_Region() ) {
1454 if( idx == 0 ) { // Updating control edge?
1455 phi = prev; // Just use existing control
1456 } else { // Else need a new Phi
1457 phi = PhiNode::make( prev, old );
1458 // Now recursively fix up the new uses of old!
1459 for( uint i = 1; i < prev->req(); i++ ) {
1460 worklist.push(phi); // Onto worklist once for each 'old' input
1461 }
1462 }
1463 } else {
1464 // Get new RegionNode merging old and new loop exits
1465 prev = old_new[prev->_idx];
1466 assert( prev, "just made this in step 7" );
1467 if( idx == 0 ) { // Updating control edge?
1468 phi = prev; // Just use existing control
1469 } else { // Else need a new Phi
1470 // Make a new Phi merging data values properly
1471 phi = PhiNode::make( prev, old );
1472 phi->set_req( 1, nnn );
1473 }
1474 }
1475 // If inserting a new Phi, check for prior hits
1476 if( idx != 0 ) {
1477 Node *hit = _igvn.hash_find_insert(phi);
1478 if( hit == NULL ) {
1479 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1480 } else { // or
1481 // Remove the new phi from the graph and use the hit
1482 _igvn.remove_dead_node(phi);
1483 phi = hit; // Use existing phi
1484 }
1485 set_ctrl(phi, prev);
1486 }
1487 // Make 'use' use the Phi instead of the old loop body exit value
1488 _igvn.hash_delete(use);
1489 _igvn._worklist.push(use);
1490 use->set_req(idx, phi);
1491 if( use->_idx >= new_counter ) { // If updating new phis
1492 // Not needed for correctness, but prevents a weak assert
1493 // in AddPNode from tripping (when we end up with different
1494 // base & derived Phis that will become the same after
1495 // IGVN does CSE).
1496 Node *hit = _igvn.hash_find_insert(use);
1497 if( hit ) // Go ahead and re-hash for hits.
1498 _igvn.replace_node( use, hit );
1499 }
1501 // If 'use' was in the loop-exit block, it now needs to be sunk
1502 // below the post-loop merge point.
1503 sink_use( use, prev );
1504 }
1505 }
1506 }
1508 // Check for IFs that need splitting/cloning. Happens if an IF outside of
1509 // the loop uses a condition set in the loop. The original IF probably
1510 // takes control from one or more OLD Regions (which in turn get from NEW
1511 // Regions). In any case, there will be a set of Phis for each merge point
1512 // from the IF up to where the original BOOL def exists the loop.
1513 if( split_if_set ) {
1514 while( split_if_set->size() ) {
1515 Node *iff = split_if_set->pop();
1516 if( iff->in(1)->is_Phi() ) {
1517 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop );
1518 _igvn.hash_delete(iff);
1519 _igvn._worklist.push(iff);
1520 iff->set_req(1, b);
1521 }
1522 }
1523 }
1524 if( split_bool_set ) {
1525 while( split_bool_set->size() ) {
1526 Node *b = split_bool_set->pop();
1527 Node *phi = b->in(1);
1528 assert( phi->is_Phi(), "" );
1529 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop );
1530 _igvn.hash_delete(b);
1531 _igvn._worklist.push(b);
1532 b->set_req(1, cmp);
1533 }
1534 }
1535 if( split_cex_set ) {
1536 while( split_cex_set->size() ) {
1537 Node *b = split_cex_set->pop();
1538 assert( b->in(0)->is_Region(), "" );
1539 assert( b->in(1)->is_Phi(), "" );
1540 assert( b->in(0)->in(0) == b->in(1)->in(0), "" );
1541 split_up( b, b->in(0), NULL );
1542 }
1543 }
1545 }
1548 //---------------------- stride_of_possible_iv -------------------------------------
1549 // Looks for an iff/bool/comp with one operand of the compare
1550 // being a cycle involving an add and a phi,
1551 // with an optional truncation (left-shift followed by a right-shift)
1552 // of the add. Returns zero if not an iv.
1553 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
1554 Node* trunc1 = NULL;
1555 Node* trunc2 = NULL;
1556 const TypeInt* ttype = NULL;
1557 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
1558 return 0;
1559 }
1560 BoolNode* bl = iff->in(1)->as_Bool();
1561 Node* cmp = bl->in(1);
1562 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) {
1563 return 0;
1564 }
1565 // Must have an invariant operand
1566 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
1567 return 0;
1568 }
1569 Node* add2 = NULL;
1570 Node* cmp1 = cmp->in(1);
1571 if (cmp1->is_Phi()) {
1572 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
1573 Node* phi = cmp1;
1574 for (uint i = 1; i < phi->req(); i++) {
1575 Node* in = phi->in(i);
1576 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
1577 &trunc1, &trunc2, &ttype);
1578 if (add && add->in(1) == phi) {
1579 add2 = add->in(2);
1580 break;
1581 }
1582 }
1583 } else {
1584 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
1585 Node* addtrunc = cmp1;
1586 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
1587 &trunc1, &trunc2, &ttype);
1588 if (add && add->in(1)->is_Phi()) {
1589 Node* phi = add->in(1);
1590 for (uint i = 1; i < phi->req(); i++) {
1591 if (phi->in(i) == addtrunc) {
1592 add2 = add->in(2);
1593 break;
1594 }
1595 }
1596 }
1597 }
1598 if (add2 != NULL) {
1599 const TypeInt* add2t = _igvn.type(add2)->is_int();
1600 if (add2t->is_con()) {
1601 return add2t->get_con();
1602 }
1603 }
1604 return 0;
1605 }
1608 //---------------------- stay_in_loop -------------------------------------
1609 // Return the (unique) control output node that's in the loop (if it exists.)
1610 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
1611 Node* unique = NULL;
1612 if (!n) return NULL;
1613 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1614 Node* use = n->fast_out(i);
1615 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
1616 if (unique != NULL) {
1617 return NULL;
1618 }
1619 unique = use;
1620 }
1621 }
1622 return unique;
1623 }
1625 //------------------------------ register_node -------------------------------------
1626 // Utility to register node "n" with PhaseIdealLoop
1627 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
1628 _igvn.register_new_node_with_optimizer(n);
1629 loop->_body.push(n);
1630 if (n->is_CFG()) {
1631 set_loop(n, loop);
1632 set_idom(n, pred, ddepth);
1633 } else {
1634 set_ctrl(n, pred);
1635 }
1636 }
1638 //------------------------------ proj_clone -------------------------------------
1639 // Utility to create an if-projection
1640 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
1641 ProjNode* c = p->clone()->as_Proj();
1642 c->set_req(0, iff);
1643 return c;
1644 }
1646 //------------------------------ short_circuit_if -------------------------------------
1647 // Force the iff control output to be the live_proj
1648 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
1649 int proj_con = live_proj->_con;
1650 assert(proj_con == 0 || proj_con == 1, "false or true projection");
1651 Node *con = _igvn.intcon(proj_con);
1652 set_ctrl(con, C->root());
1653 if (iff) {
1654 iff->set_req(1, con);
1655 }
1656 return con;
1657 }
1659 //------------------------------ insert_if_before_proj -------------------------------------
1660 // Insert a new if before an if projection (* - new node)
1661 //
1662 // before
1663 // if(test)
1664 // / \
1665 // v v
1666 // other-proj proj (arg)
1667 //
1668 // after
1669 // if(test)
1670 // / \
1671 // / v
1672 // | * proj-clone
1673 // v |
1674 // other-proj v
1675 // * new_if(relop(cmp[IU](left,right)))
1676 // / \
1677 // v v
1678 // * new-proj proj
1679 // (returned)
1680 //
1681 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
1682 IfNode* iff = proj->in(0)->as_If();
1683 IdealLoopTree *loop = get_loop(proj);
1684 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1685 int ddepth = dom_depth(proj);
1687 _igvn.hash_delete(iff);
1688 _igvn._worklist.push(iff);
1689 _igvn.hash_delete(proj);
1690 _igvn._worklist.push(proj);
1692 proj->set_req(0, NULL); // temporary disconnect
1693 ProjNode* proj2 = proj_clone(proj, iff);
1694 register_node(proj2, loop, iff, ddepth);
1696 Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right);
1697 register_node(cmp, loop, proj2, ddepth);
1699 BoolNode* bol = new (C,2)BoolNode(cmp, relop);
1700 register_node(bol, loop, proj2, ddepth);
1702 IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt);
1703 register_node(new_if, loop, proj2, ddepth);
1705 proj->set_req(0, new_if); // reattach
1706 set_idom(proj, new_if, ddepth);
1708 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
1709 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
1711 return new_exit;
1712 }
1714 //------------------------------ insert_region_before_proj -------------------------------------
1715 // Insert a region before an if projection (* - new node)
1716 //
1717 // before
1718 // if(test)
1719 // / |
1720 // v |
1721 // proj v
1722 // other-proj
1723 //
1724 // after
1725 // if(test)
1726 // / |
1727 // v |
1728 // * proj-clone v
1729 // | other-proj
1730 // v
1731 // * new-region
1732 // |
1733 // v
1734 // * dum_if
1735 // / \
1736 // v \
1737 // * dum-proj v
1738 // proj
1739 //
1740 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
1741 IfNode* iff = proj->in(0)->as_If();
1742 IdealLoopTree *loop = get_loop(proj);
1743 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1744 int ddepth = dom_depth(proj);
1746 _igvn.hash_delete(iff);
1747 _igvn._worklist.push(iff);
1748 _igvn.hash_delete(proj);
1749 _igvn._worklist.push(proj);
1751 proj->set_req(0, NULL); // temporary disconnect
1752 ProjNode* proj2 = proj_clone(proj, iff);
1753 register_node(proj2, loop, iff, ddepth);
1755 RegionNode* reg = new (C,2)RegionNode(2);
1756 reg->set_req(1, proj2);
1757 register_node(reg, loop, iff, ddepth);
1759 IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
1760 register_node(dum_if, loop, reg, ddepth);
1762 proj->set_req(0, dum_if); // reattach
1763 set_idom(proj, dum_if, ddepth);
1765 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
1766 register_node(dum_proj, loop, dum_if, ddepth);
1768 return reg;
1769 }
1771 //------------------------------ insert_cmpi_loop_exit -------------------------------------
1772 // Clone a signed compare loop exit from an unsigned compare and
1773 // insert it before the unsigned cmp on the stay-in-loop path.
1774 // All new nodes inserted in the dominator tree between the original
1775 // if and it's projections. The original if test is replaced with
1776 // a constant to force the stay-in-loop path.
1777 //
1778 // This is done to make sure that the original if and it's projections
1779 // still dominate the same set of control nodes, that the ctrl() relation
1780 // from data nodes to them is preserved, and that their loop nesting is
1781 // preserved.
1782 //
1783 // before
1784 // if(i <u limit) unsigned compare loop exit
1785 // / |
1786 // v v
1787 // exit-proj stay-in-loop-proj
1788 //
1789 // after
1790 // if(stay-in-loop-const) original if
1791 // / |
1792 // / v
1793 // / if(i < limit) new signed test
1794 // / / |
1795 // / / v
1796 // / / if(i <u limit) new cloned unsigned test
1797 // / / / |
1798 // v v v |
1799 // region |
1800 // | |
1801 // dum-if |
1802 // / | |
1803 // ether | |
1804 // v v
1805 // exit-proj stay-in-loop-proj
1806 //
1807 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
1808 const bool Signed = true;
1809 const bool Unsigned = false;
1811 BoolNode* bol = if_cmpu->in(1)->as_Bool();
1812 if (bol->_test._test != BoolTest::lt) return NULL;
1813 CmpNode* cmpu = bol->in(1)->as_Cmp();
1814 if (cmpu->Opcode() != Op_CmpU) return NULL;
1815 int stride = stride_of_possible_iv(if_cmpu);
1816 if (stride == 0) return NULL;
1818 ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj();
1819 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
1821 Node* limit = NULL;
1822 if (stride > 0) {
1823 limit = cmpu->in(2);
1824 } else {
1825 limit = _igvn.makecon(TypeInt::ZERO);
1826 set_ctrl(limit, C->root());
1827 }
1828 // Create a new region on the exit path
1829 RegionNode* reg = insert_region_before_proj(lp_exit);
1831 // Clone the if-cmpu-true-false using a signed compare
1832 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
1833 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
1834 reg->add_req(cmpi_exit);
1836 // Clone the if-cmpu-true-false
1837 BoolTest::mask rel_u = bol->_test._test;
1838 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
1839 reg->add_req(cmpu_exit);
1841 // Force original if to stay in loop.
1842 short_circuit_if(if_cmpu, lp_continue);
1844 return cmpi_exit->in(0)->as_If();
1845 }
1847 //------------------------------ remove_cmpi_loop_exit -------------------------------------
1848 // Remove a previously inserted signed compare loop exit.
1849 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
1850 Node* lp_proj = stay_in_loop(if_cmp, loop);
1851 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
1852 stay_in_loop(lp_proj, loop)->is_If() &&
1853 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
1854 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
1855 set_ctrl(con, C->root());
1856 if_cmp->set_req(1, con);
1857 }
1859 //------------------------------ scheduled_nodelist -------------------------------------
1860 // Create a post order schedule of nodes that are in the
1861 // "member" set. The list is returned in "sched".
1862 // The first node in "sched" is the loop head, followed by
1863 // nodes which have no inputs in the "member" set, and then
1864 // followed by the nodes that have an immediate input dependence
1865 // on a node in "sched".
1866 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
1868 assert(member.test(loop->_head->_idx), "loop head must be in member set");
1869 Arena *a = Thread::current()->resource_area();
1870 VectorSet visited(a);
1871 Node_Stack nstack(a, loop->_body.size());
1873 Node* n = loop->_head; // top of stack is cached in "n"
1874 uint idx = 0;
1875 visited.set(n->_idx);
1877 // Initially push all with no inputs from within member set
1878 for(uint i = 0; i < loop->_body.size(); i++ ) {
1879 Node *elt = loop->_body.at(i);
1880 if (member.test(elt->_idx)) {
1881 bool found = false;
1882 for (uint j = 0; j < elt->req(); j++) {
1883 Node* def = elt->in(j);
1884 if (def && member.test(def->_idx) && def != elt) {
1885 found = true;
1886 break;
1887 }
1888 }
1889 if (!found && elt != loop->_head) {
1890 nstack.push(n, idx);
1891 n = elt;
1892 assert(!visited.test(n->_idx), "not seen yet");
1893 visited.set(n->_idx);
1894 }
1895 }
1896 }
1898 // traverse out's that are in the member set
1899 while (true) {
1900 if (idx < n->outcnt()) {
1901 Node* use = n->raw_out(idx);
1902 idx++;
1903 if (!visited.test_set(use->_idx)) {
1904 if (member.test(use->_idx)) {
1905 nstack.push(n, idx);
1906 n = use;
1907 idx = 0;
1908 }
1909 }
1910 } else {
1911 // All outputs processed
1912 sched.push(n);
1913 if (nstack.is_empty()) break;
1914 n = nstack.node();
1915 idx = nstack.index();
1916 nstack.pop();
1917 }
1918 }
1919 }
1922 //------------------------------ has_use_in_set -------------------------------------
1923 // Has a use in the vector set
1924 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
1925 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1926 Node* use = n->fast_out(j);
1927 if (vset.test(use->_idx)) {
1928 return true;
1929 }
1930 }
1931 return false;
1932 }
1935 //------------------------------ has_use_internal_to_set -------------------------------------
1936 // Has use internal to the vector set (ie. not in a phi at the loop head)
1937 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
1938 Node* head = loop->_head;
1939 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1940 Node* use = n->fast_out(j);
1941 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
1942 return true;
1943 }
1944 }
1945 return false;
1946 }
1949 //------------------------------ clone_for_use_outside_loop -------------------------------------
1950 // clone "n" for uses that are outside of loop
1951 void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
1953 assert(worklist.size() == 0, "should be empty");
1954 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1955 Node* use = n->fast_out(j);
1956 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
1957 worklist.push(use);
1958 }
1959 }
1960 while( worklist.size() ) {
1961 Node *use = worklist.pop();
1962 if (!has_node(use) || use->in(0) == C->top()) continue;
1963 uint j;
1964 for (j = 0; j < use->req(); j++) {
1965 if (use->in(j) == n) break;
1966 }
1967 assert(j < use->req(), "must be there");
1969 // clone "n" and insert it between the inputs of "n" and the use outside the loop
1970 Node* n_clone = n->clone();
1971 _igvn.hash_delete(use);
1972 use->set_req(j, n_clone);
1973 _igvn._worklist.push(use);
1974 Node* use_c;
1975 if (!use->is_Phi()) {
1976 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
1977 } else {
1978 // Use in a phi is considered a use in the associated predecessor block
1979 use_c = use->in(0)->in(j);
1980 }
1981 set_ctrl(n_clone, use_c);
1982 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
1983 get_loop(use_c)->_body.push(n_clone);
1984 _igvn.register_new_node_with_optimizer(n_clone);
1985 #if !defined(PRODUCT)
1986 if (TracePartialPeeling) {
1987 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
1988 }
1989 #endif
1990 }
1991 }
1994 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
1995 // clone "n" for special uses that are in the not_peeled region.
1996 // If these def-uses occur in separate blocks, the code generator
1997 // marks the method as not compilable. For example, if a "BoolNode"
1998 // is in a different basic block than the "IfNode" that uses it, then
1999 // the compilation is aborted in the code generator.
2000 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
2001 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
2002 if (n->is_Phi() || n->is_Load()) {
2003 return;
2004 }
2005 assert(worklist.size() == 0, "should be empty");
2006 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2007 Node* use = n->fast_out(j);
2008 if ( not_peel.test(use->_idx) &&
2009 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
2010 use->in(1) == n) {
2011 worklist.push(use);
2012 }
2013 }
2014 if (worklist.size() > 0) {
2015 // clone "n" and insert it between inputs of "n" and the use
2016 Node* n_clone = n->clone();
2017 loop->_body.push(n_clone);
2018 _igvn.register_new_node_with_optimizer(n_clone);
2019 set_ctrl(n_clone, get_ctrl(n));
2020 sink_list.push(n_clone);
2021 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
2022 #if !defined(PRODUCT)
2023 if (TracePartialPeeling) {
2024 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2025 }
2026 #endif
2027 while( worklist.size() ) {
2028 Node *use = worklist.pop();
2029 _igvn.hash_delete(use);
2030 _igvn._worklist.push(use);
2031 for (uint j = 1; j < use->req(); j++) {
2032 if (use->in(j) == n) {
2033 use->set_req(j, n_clone);
2034 }
2035 }
2036 }
2037 }
2038 }
2041 //------------------------------ insert_phi_for_loop -------------------------------------
2042 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2043 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2044 Node *phi = PhiNode::make(lp, back_edge_val);
2045 phi->set_req(LoopNode::EntryControl, lp_entry_val);
2046 // Use existing phi if it already exists
2047 Node *hit = _igvn.hash_find_insert(phi);
2048 if( hit == NULL ) {
2049 _igvn.register_new_node_with_optimizer(phi);
2050 set_ctrl(phi, lp);
2051 } else {
2052 // Remove the new phi from the graph and use the hit
2053 _igvn.remove_dead_node(phi);
2054 phi = hit;
2055 }
2056 _igvn.hash_delete(use);
2057 _igvn._worklist.push(use);
2058 use->set_req(idx, phi);
2059 }
2061 #ifdef ASSERT
2062 //------------------------------ is_valid_loop_partition -------------------------------------
2063 // Validate the loop partition sets: peel and not_peel
2064 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2065 VectorSet& not_peel ) {
2066 uint i;
2067 // Check that peel_list entries are in the peel set
2068 for (i = 0; i < peel_list.size(); i++) {
2069 if (!peel.test(peel_list.at(i)->_idx)) {
2070 return false;
2071 }
2072 }
2073 // Check at loop members are in one of peel set or not_peel set
2074 for (i = 0; i < loop->_body.size(); i++ ) {
2075 Node *def = loop->_body.at(i);
2076 uint di = def->_idx;
2077 // Check that peel set elements are in peel_list
2078 if (peel.test(di)) {
2079 if (not_peel.test(di)) {
2080 return false;
2081 }
2082 // Must be in peel_list also
2083 bool found = false;
2084 for (uint j = 0; j < peel_list.size(); j++) {
2085 if (peel_list.at(j)->_idx == di) {
2086 found = true;
2087 break;
2088 }
2089 }
2090 if (!found) {
2091 return false;
2092 }
2093 } else if (not_peel.test(di)) {
2094 if (peel.test(di)) {
2095 return false;
2096 }
2097 } else {
2098 return false;
2099 }
2100 }
2101 return true;
2102 }
2104 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2105 // Ensure a use outside of loop is of the right form
2106 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2107 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2108 return (use->is_Phi() &&
2109 use_c->is_Region() && use_c->req() == 3 &&
2110 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2111 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2112 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2113 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2114 }
2116 //------------------------------ is_valid_clone_loop_form -------------------------------------
2117 // Ensure that all uses outside of loop are of the right form
2118 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2119 uint orig_exit_idx, uint clone_exit_idx) {
2120 uint len = peel_list.size();
2121 for (uint i = 0; i < len; i++) {
2122 Node *def = peel_list.at(i);
2124 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2125 Node *use = def->fast_out(j);
2126 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2127 if (!loop->is_member(get_loop(use_c))) {
2128 // use is not in the loop, check for correct structure
2129 if (use->in(0) == def) {
2130 // Okay
2131 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2132 return false;
2133 }
2134 }
2135 }
2136 }
2137 return true;
2138 }
2139 #endif
2141 //------------------------------ partial_peel -------------------------------------
2142 // Partially peel (aka loop rotation) the top portion of a loop (called
2143 // the peel section below) by cloning it and placing one copy just before
2144 // the new loop head and the other copy at the bottom of the new loop.
2145 //
2146 // before after where it came from
2147 //
2148 // stmt1 stmt1
2149 // loop: stmt2 clone
2150 // stmt2 if condA goto exitA clone
2151 // if condA goto exitA new_loop: new
2152 // stmt3 stmt3 clone
2153 // if !condB goto loop if condB goto exitB clone
2154 // exitB: stmt2 orig
2155 // stmt4 if !condA goto new_loop orig
2156 // exitA: goto exitA
2157 // exitB:
2158 // stmt4
2159 // exitA:
2160 //
2161 // Step 1: find the cut point: an exit test on probable
2162 // induction variable.
2163 // Step 2: schedule (with cloning) operations in the peel
2164 // section that can be executed after the cut into
2165 // the section that is not peeled. This may need
2166 // to clone operations into exit blocks. For
2167 // instance, a reference to A[i] in the not-peel
2168 // section and a reference to B[i] in an exit block
2169 // may cause a left-shift of i by 2 to be placed
2170 // in the peel block. This step will clone the left
2171 // shift into the exit block and sink the left shift
2172 // from the peel to the not-peel section.
2173 // Step 3: clone the loop, retarget the control, and insert
2174 // phis for values that are live across the new loop
2175 // head. This is very dependent on the graph structure
2176 // from clone_loop. It creates region nodes for
2177 // exit control and associated phi nodes for values
2178 // flow out of the loop through that exit. The region
2179 // node is dominated by the clone's control projection.
2180 // So the clone's peel section is placed before the
2181 // new loop head, and the clone's not-peel section is
2182 // forms the top part of the new loop. The original
2183 // peel section forms the tail of the new loop.
2184 // Step 4: update the dominator tree and recompute the
2185 // dominator depth.
2186 //
2187 // orig
2188 //
2189 // stmt1
2190 // |
2191 // v
2192 // loop predicate
2193 // |
2194 // v
2195 // loop<----+
2196 // | |
2197 // stmt2 |
2198 // | |
2199 // v |
2200 // ifA |
2201 // / | |
2202 // v v |
2203 // false true ^ <-- last_peel
2204 // / | |
2205 // / ===|==cut |
2206 // / stmt3 | <-- first_not_peel
2207 // / | |
2208 // | v |
2209 // v ifB |
2210 // exitA: / \ |
2211 // / \ |
2212 // v v |
2213 // false true |
2214 // / \ |
2215 // / ----+
2216 // |
2217 // v
2218 // exitB:
2219 // stmt4
2220 //
2221 //
2222 // after clone loop
2223 //
2224 // stmt1
2225 // |
2226 // v
2227 // loop predicate
2228 // / \
2229 // clone / \ orig
2230 // / \
2231 // / \
2232 // v v
2233 // +---->loop loop<----+
2234 // | | | |
2235 // | stmt2 stmt2 |
2236 // | | | |
2237 // | v v |
2238 // | ifA ifA |
2239 // | | \ / | |
2240 // | v v v v |
2241 // ^ true false false true ^ <-- last_peel
2242 // | | ^ \ / | |
2243 // | cut==|== \ \ / ===|==cut |
2244 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2245 // | | dom | | | |
2246 // | v \ 1v v2 v |
2247 // | ifB regionA ifB |
2248 // | / \ | / \ |
2249 // | / \ v / \ |
2250 // | v v exitA: v v |
2251 // | true false false true |
2252 // | / ^ \ / \ |
2253 // +---- \ \ / ----+
2254 // dom \ /
2255 // \ 1v v2
2256 // regionB
2257 // |
2258 // v
2259 // exitB:
2260 // stmt4
2261 //
2262 //
2263 // after partial peel
2264 //
2265 // stmt1
2266 // |
2267 // v
2268 // loop predicate
2269 // /
2270 // clone / orig
2271 // / TOP
2272 // / \
2273 // v v
2274 // TOP->loop loop----+
2275 // | | |
2276 // stmt2 stmt2 |
2277 // | | |
2278 // v v |
2279 // ifA ifA |
2280 // | \ / | |
2281 // v v v v |
2282 // true false false true | <-- last_peel
2283 // | ^ \ / +------|---+
2284 // +->newloop \ \ / === ==cut | |
2285 // | stmt3 \ \ / TOP | |
2286 // | | dom | | stmt3 | | <-- first_not_peel
2287 // | v \ 1v v2 v | |
2288 // | ifB regionA ifB ^ v
2289 // | / \ | / \ | |
2290 // | / \ v / \ | |
2291 // | v v exitA: v v | |
2292 // | true false false true | |
2293 // | / ^ \ / \ | |
2294 // | | \ \ / v | |
2295 // | | dom \ / TOP | |
2296 // | | \ 1v v2 | |
2297 // ^ v regionB | |
2298 // | | | | |
2299 // | | v ^ v
2300 // | | exitB: | |
2301 // | | stmt4 | |
2302 // | +------------>-----------------+ |
2303 // | |
2304 // +-----------------<---------------------+
2305 //
2306 //
2307 // final graph
2308 //
2309 // stmt1
2310 // |
2311 // v
2312 // loop predicate
2313 // |
2314 // v
2315 // stmt2 clone
2316 // |
2317 // v
2318 // ........> ifA clone
2319 // : / |
2320 // dom / |
2321 // : v v
2322 // : false true
2323 // : | |
2324 // : | v
2325 // : | newloop<-----+
2326 // : | | |
2327 // : | stmt3 clone |
2328 // : | | |
2329 // : | v |
2330 // : | ifB |
2331 // : | / \ |
2332 // : | v v |
2333 // : | false true |
2334 // : | | | |
2335 // : | v stmt2 |
2336 // : | exitB: | |
2337 // : | stmt4 v |
2338 // : | ifA orig |
2339 // : | / \ |
2340 // : | / \ |
2341 // : | v v |
2342 // : | false true |
2343 // : | / \ |
2344 // : v v -----+
2345 // RegionA
2346 // |
2347 // v
2348 // exitA
2349 //
2350 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
2352 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
2353 if (!loop->_head->is_Loop()) {
2354 return false; }
2356 LoopNode *head = loop->_head->as_Loop();
2358 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
2359 return false;
2360 }
2362 // Check for complex exit control
2363 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
2364 Node *n = loop->_body.at(ii);
2365 int opc = n->Opcode();
2366 if (n->is_Call() ||
2367 opc == Op_Catch ||
2368 opc == Op_CatchProj ||
2369 opc == Op_Jump ||
2370 opc == Op_JumpProj) {
2371 #if !defined(PRODUCT)
2372 if (TracePartialPeeling) {
2373 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
2374 }
2375 #endif
2376 return false;
2377 }
2378 }
2380 int dd = dom_depth(head);
2382 // Step 1: find cut point
2384 // Walk up dominators to loop head looking for first loop exit
2385 // which is executed on every path thru loop.
2386 IfNode *peel_if = NULL;
2387 IfNode *peel_if_cmpu = NULL;
2389 Node *iff = loop->tail();
2390 while( iff != head ) {
2391 if( iff->is_If() ) {
2392 Node *ctrl = get_ctrl(iff->in(1));
2393 if (ctrl->is_top()) return false; // Dead test on live IF.
2394 // If loop-varying exit-test, check for induction variable
2395 if( loop->is_member(get_loop(ctrl)) &&
2396 loop->is_loop_exit(iff) &&
2397 is_possible_iv_test(iff)) {
2398 Node* cmp = iff->in(1)->in(1);
2399 if (cmp->Opcode() == Op_CmpI) {
2400 peel_if = iff->as_If();
2401 } else {
2402 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
2403 peel_if_cmpu = iff->as_If();
2404 }
2405 }
2406 }
2407 iff = idom(iff);
2408 }
2409 // Prefer signed compare over unsigned compare.
2410 IfNode* new_peel_if = NULL;
2411 if (peel_if == NULL) {
2412 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
2413 return false; // No peel point found
2414 }
2415 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
2416 if (new_peel_if == NULL) {
2417 return false; // No peel point found
2418 }
2419 peel_if = new_peel_if;
2420 }
2421 Node* last_peel = stay_in_loop(peel_if, loop);
2422 Node* first_not_peeled = stay_in_loop(last_peel, loop);
2423 if (first_not_peeled == NULL || first_not_peeled == head) {
2424 return false;
2425 }
2427 #if !defined(PRODUCT)
2428 if (TraceLoopOpts) {
2429 tty->print("PartialPeel ");
2430 loop->dump_head();
2431 }
2433 if (TracePartialPeeling) {
2434 tty->print_cr("before partial peel one iteration");
2435 Node_List wl;
2436 Node* t = head->in(2);
2437 while (true) {
2438 wl.push(t);
2439 if (t == head) break;
2440 t = idom(t);
2441 }
2442 while (wl.size() > 0) {
2443 Node* tt = wl.pop();
2444 tt->dump();
2445 if (tt == last_peel) tty->print_cr("-- cut --");
2446 }
2447 }
2448 #endif
2449 ResourceArea *area = Thread::current()->resource_area();
2450 VectorSet peel(area);
2451 VectorSet not_peel(area);
2452 Node_List peel_list(area);
2453 Node_List worklist(area);
2454 Node_List sink_list(area);
2456 // Set of cfg nodes to peel are those that are executable from
2457 // the head through last_peel.
2458 assert(worklist.size() == 0, "should be empty");
2459 worklist.push(head);
2460 peel.set(head->_idx);
2461 while (worklist.size() > 0) {
2462 Node *n = worklist.pop();
2463 if (n != last_peel) {
2464 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2465 Node* use = n->fast_out(j);
2466 if (use->is_CFG() &&
2467 loop->is_member(get_loop(use)) &&
2468 !peel.test_set(use->_idx)) {
2469 worklist.push(use);
2470 }
2471 }
2472 }
2473 }
2475 // Set of non-cfg nodes to peel are those that are control
2476 // dependent on the cfg nodes.
2477 uint i;
2478 for(i = 0; i < loop->_body.size(); i++ ) {
2479 Node *n = loop->_body.at(i);
2480 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
2481 if (peel.test(n_c->_idx)) {
2482 peel.set(n->_idx);
2483 } else {
2484 not_peel.set(n->_idx);
2485 }
2486 }
2488 // Step 2: move operations from the peeled section down into the
2489 // not-peeled section
2491 // Get a post order schedule of nodes in the peel region
2492 // Result in right-most operand.
2493 scheduled_nodelist(loop, peel, peel_list );
2495 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2497 // For future check for too many new phis
2498 uint old_phi_cnt = 0;
2499 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2500 Node* use = head->fast_out(j);
2501 if (use->is_Phi()) old_phi_cnt++;
2502 }
2504 #if !defined(PRODUCT)
2505 if (TracePartialPeeling) {
2506 tty->print_cr("\npeeled list");
2507 }
2508 #endif
2510 // Evacuate nodes in peel region into the not_peeled region if possible
2511 uint new_phi_cnt = 0;
2512 for (i = 0; i < peel_list.size();) {
2513 Node* n = peel_list.at(i);
2514 #if !defined(PRODUCT)
2515 if (TracePartialPeeling) n->dump();
2516 #endif
2517 bool incr = true;
2518 if ( !n->is_CFG() ) {
2520 if ( has_use_in_set(n, not_peel) ) {
2522 // If not used internal to the peeled region,
2523 // move "n" from peeled to not_peeled region.
2525 if ( !has_use_internal_to_set(n, peel, loop) ) {
2527 // if not pinned and not a load (which maybe anti-dependent on a store)
2528 // and not a CMove (Matcher expects only bool->cmove).
2529 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
2530 clone_for_use_outside_loop( loop, n, worklist );
2532 sink_list.push(n);
2533 peel >>= n->_idx; // delete n from peel set.
2534 not_peel <<= n->_idx; // add n to not_peel set.
2535 peel_list.remove(i);
2536 incr = false;
2537 #if !defined(PRODUCT)
2538 if (TracePartialPeeling) {
2539 tty->print_cr("sink to not_peeled region: %d newbb: %d",
2540 n->_idx, get_ctrl(n)->_idx);
2541 }
2542 #endif
2543 }
2544 } else {
2545 // Otherwise check for special def-use cases that span
2546 // the peel/not_peel boundary such as bool->if
2547 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
2548 new_phi_cnt++;
2549 }
2550 }
2551 }
2552 if (incr) i++;
2553 }
2555 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
2556 #if !defined(PRODUCT)
2557 if (TracePartialPeeling) {
2558 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
2559 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
2560 }
2561 #endif
2562 if (new_peel_if != NULL) {
2563 remove_cmpi_loop_exit(new_peel_if, loop);
2564 }
2565 // Inhibit more partial peeling on this loop
2566 assert(!head->is_partial_peel_loop(), "not partial peeled");
2567 head->mark_partial_peel_failed();
2568 return false;
2569 }
2571 // Step 3: clone loop, retarget control, and insert new phis
2573 // Create new loop head for new phis and to hang
2574 // the nodes being moved (sinked) from the peel region.
2575 LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel);
2576 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
2577 _igvn.register_new_node_with_optimizer(new_head);
2578 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
2579 first_not_peeled->set_req(0, new_head);
2580 set_loop(new_head, loop);
2581 loop->_body.push(new_head);
2582 not_peel.set(new_head->_idx);
2583 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
2584 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
2586 while (sink_list.size() > 0) {
2587 Node* n = sink_list.pop();
2588 set_ctrl(n, new_head);
2589 }
2591 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2593 clone_loop( loop, old_new, dd );
2595 const uint clone_exit_idx = 1;
2596 const uint orig_exit_idx = 2;
2597 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
2599 Node* head_clone = old_new[head->_idx];
2600 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
2601 Node* orig_tail_clone = head_clone->in(2);
2603 // Add phi if "def" node is in peel set and "use" is not
2605 for(i = 0; i < peel_list.size(); i++ ) {
2606 Node *def = peel_list.at(i);
2607 if (!def->is_CFG()) {
2608 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2609 Node *use = def->fast_out(j);
2610 if (has_node(use) && use->in(0) != C->top() &&
2611 (!peel.test(use->_idx) ||
2612 (use->is_Phi() && use->in(0) == head)) ) {
2613 worklist.push(use);
2614 }
2615 }
2616 while( worklist.size() ) {
2617 Node *use = worklist.pop();
2618 for (uint j = 1; j < use->req(); j++) {
2619 Node* n = use->in(j);
2620 if (n == def) {
2622 // "def" is in peel set, "use" is not in peel set
2623 // or "use" is in the entry boundary (a phi) of the peel set
2625 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
2627 if ( loop->is_member(get_loop( use_c )) ) {
2628 // use is in loop
2629 if (old_new[use->_idx] != NULL) { // null for dead code
2630 Node* use_clone = old_new[use->_idx];
2631 _igvn.hash_delete(use);
2632 use->set_req(j, C->top());
2633 _igvn._worklist.push(use);
2634 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
2635 }
2636 } else {
2637 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
2638 // use is not in the loop, check if the live range includes the cut
2639 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
2640 if (not_peel.test(lp_if->_idx)) {
2641 assert(j == orig_exit_idx, "use from original loop");
2642 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
2643 }
2644 }
2645 }
2646 }
2647 }
2648 }
2649 }
2651 // Step 3b: retarget control
2653 // Redirect control to the new loop head if a cloned node in
2654 // the not_peeled region has control that points into the peeled region.
2655 // This necessary because the cloned peeled region will be outside
2656 // the loop.
2657 // from to
2658 // cloned-peeled <---+
2659 // new_head_clone: | <--+
2660 // cloned-not_peeled in(0) in(0)
2661 // orig-peeled
2663 for(i = 0; i < loop->_body.size(); i++ ) {
2664 Node *n = loop->_body.at(i);
2665 if (!n->is_CFG() && n->in(0) != NULL &&
2666 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
2667 Node* n_clone = old_new[n->_idx];
2668 _igvn.hash_delete(n_clone);
2669 n_clone->set_req(0, new_head_clone);
2670 _igvn._worklist.push(n_clone);
2671 }
2672 }
2674 // Backedge of the surviving new_head (the clone) is original last_peel
2675 _igvn.hash_delete(new_head_clone);
2676 new_head_clone->set_req(LoopNode::LoopBackControl, last_peel);
2677 _igvn._worklist.push(new_head_clone);
2679 // Cut first node in original not_peel set
2680 _igvn.hash_delete(new_head);
2681 new_head->set_req(LoopNode::EntryControl, C->top());
2682 new_head->set_req(LoopNode::LoopBackControl, C->top());
2683 _igvn._worklist.push(new_head);
2685 // Copy head_clone back-branch info to original head
2686 // and remove original head's loop entry and
2687 // clone head's back-branch
2688 _igvn.hash_delete(head);
2689 _igvn.hash_delete(head_clone);
2690 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
2691 head->set_req(LoopNode::LoopBackControl, C->top());
2692 head_clone->set_req(LoopNode::LoopBackControl, C->top());
2693 _igvn._worklist.push(head);
2694 _igvn._worklist.push(head_clone);
2696 // Similarly modify the phis
2697 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
2698 Node* use = head->fast_out(k);
2699 if (use->is_Phi() && use->outcnt() > 0) {
2700 Node* use_clone = old_new[use->_idx];
2701 _igvn.hash_delete(use);
2702 _igvn.hash_delete(use_clone);
2703 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
2704 use->set_req(LoopNode::LoopBackControl, C->top());
2705 use_clone->set_req(LoopNode::LoopBackControl, C->top());
2706 _igvn._worklist.push(use);
2707 _igvn._worklist.push(use_clone);
2708 }
2709 }
2711 // Step 4: update dominator tree and dominator depth
2713 set_idom(head, orig_tail_clone, dd);
2714 recompute_dom_depth();
2716 // Inhibit more partial peeling on this loop
2717 new_head_clone->set_partial_peel_loop();
2718 C->set_major_progress();
2720 #if !defined(PRODUCT)
2721 if (TracePartialPeeling) {
2722 tty->print_cr("\nafter partial peel one iteration");
2723 Node_List wl(area);
2724 Node* t = last_peel;
2725 while (true) {
2726 wl.push(t);
2727 if (t == head_clone) break;
2728 t = idom(t);
2729 }
2730 while (wl.size() > 0) {
2731 Node* tt = wl.pop();
2732 if (tt == head) tty->print_cr("orig head");
2733 else if (tt == new_head_clone) tty->print_cr("new head");
2734 else if (tt == head_clone) tty->print_cr("clone head");
2735 tt->dump();
2736 }
2737 }
2738 #endif
2739 return true;
2740 }
2742 //------------------------------reorg_offsets----------------------------------
2743 // Reorganize offset computations to lower register pressure. Mostly
2744 // prevent loop-fallout uses of the pre-incremented trip counter (which are
2745 // then alive with the post-incremented trip counter forcing an extra
2746 // register move)
2747 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
2748 // Perform it only for canonical counted loops.
2749 // Loop's shape could be messed up by iteration_split_impl.
2750 if (!loop->_head->is_CountedLoop())
2751 return;
2752 if (!loop->_head->as_Loop()->is_valid_counted_loop())
2753 return;
2755 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2756 CountedLoopEndNode *cle = cl->loopexit();
2757 Node *exit = cle->proj_out(false);
2758 Node *phi = cl->phi();
2760 // Check for the special case of folks using the pre-incremented
2761 // trip-counter on the fall-out path (forces the pre-incremented
2762 // and post-incremented trip counter to be live at the same time).
2763 // Fix this by adjusting to use the post-increment trip counter.
2765 bool progress = true;
2766 while (progress) {
2767 progress = false;
2768 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
2769 Node* use = phi->fast_out(i); // User of trip-counter
2770 if (!has_ctrl(use)) continue;
2771 Node *u_ctrl = get_ctrl(use);
2772 if (use->is_Phi()) {
2773 u_ctrl = NULL;
2774 for (uint j = 1; j < use->req(); j++)
2775 if (use->in(j) == phi)
2776 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
2777 }
2778 IdealLoopTree *u_loop = get_loop(u_ctrl);
2779 // Look for loop-invariant use
2780 if (u_loop == loop) continue;
2781 if (loop->is_member(u_loop)) continue;
2782 // Check that use is live out the bottom. Assuming the trip-counter
2783 // update is right at the bottom, uses of of the loop middle are ok.
2784 if (dom_lca(exit, u_ctrl) != exit) continue;
2785 // Hit! Refactor use to use the post-incremented tripcounter.
2786 // Compute a post-increment tripcounter.
2787 Node *opaq = new (C, 2) Opaque2Node( C, cle->incr() );
2788 register_new_node( opaq, u_ctrl );
2789 Node *neg_stride = _igvn.intcon(-cle->stride_con());
2790 set_ctrl(neg_stride, C->root());
2791 Node *post = new (C, 3) AddINode( opaq, neg_stride);
2792 register_new_node( post, u_ctrl );
2793 _igvn.hash_delete(use);
2794 _igvn._worklist.push(use);
2795 for (uint j = 1; j < use->req(); j++) {
2796 if (use->in(j) == phi)
2797 use->set_req(j, post);
2798 }
2799 // Since DU info changed, rerun loop
2800 progress = true;
2801 break;
2802 }
2803 }
2805 }