Thu, 20 Sep 2012 16:49:17 +0200
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
Summary: use shorter instruction sequences for atomic add and atomic exchange when possible.
Reviewed-by: kvn, jrose
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.replace_input_of(iff, 1, con);
221 // If I dont have a reachable TRUE and FALSE path following the IfNode then
222 // I can assume this path reaches an infinite loop. In this case it's not
223 // important to optimize the data Nodes - either the whole compilation will
224 // be tossed or this path (and all data Nodes) will go dead.
225 if (iff->outcnt() != 2) return;
227 // Make control-dependent data Nodes on the live path (path that will remain
228 // once the dominated IF is removed) become control-dependent on the
229 // dominating projection.
230 Node* dp = iff->as_If()->proj_out(pop == Op_IfTrue);
232 // Loop predicates may have depending checks which should not
233 // be skipped. For example, range check predicate has two checks
234 // for lower and upper bounds.
235 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp->as_Proj()->_con)->as_Proj();
236 if (exclude_loop_predicate &&
237 is_uncommon_trap_proj(unc_proj, Deoptimization::Reason_predicate))
238 return; // Let IGVN transformation change control dependence.
240 IdealLoopTree *old_loop = get_loop(dp);
242 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
243 Node* cd = dp->fast_out(i); // Control-dependent node
244 if (cd->depends_only_on_test()) {
245 assert(cd->in(0) == dp, "");
246 _igvn.replace_input_of(cd, 0, prevdom);
247 set_early_ctrl(cd);
248 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
249 if (old_loop != new_loop) {
250 if (!old_loop->_child) old_loop->_body.yank(cd);
251 if (!new_loop->_child) new_loop->_body.push(cd);
252 }
253 --i;
254 --imax;
255 }
256 }
257 }
259 //------------------------------has_local_phi_input----------------------------
260 // Return TRUE if 'n' has Phi inputs from its local block and no other
261 // block-local inputs (all non-local-phi inputs come from earlier blocks)
262 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
263 Node *n_ctrl = get_ctrl(n);
264 // See if some inputs come from a Phi in this block, or from before
265 // this block.
266 uint i;
267 for( i = 1; i < n->req(); i++ ) {
268 Node *phi = n->in(i);
269 if( phi->is_Phi() && phi->in(0) == n_ctrl )
270 break;
271 }
272 if( i >= n->req() )
273 return NULL; // No Phi inputs; nowhere to clone thru
275 // Check for inputs created between 'n' and the Phi input. These
276 // must split as well; they have already been given the chance
277 // (courtesy of a post-order visit) and since they did not we must
278 // recover the 'cost' of splitting them by being very profitable
279 // when splitting 'n'. Since this is unlikely we simply give up.
280 for( i = 1; i < n->req(); i++ ) {
281 Node *m = n->in(i);
282 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
283 // We allow the special case of AddP's with no local inputs.
284 // This allows us to split-up address expressions.
285 if (m->is_AddP() &&
286 get_ctrl(m->in(2)) != n_ctrl &&
287 get_ctrl(m->in(3)) != n_ctrl) {
288 // Move the AddP up to dominating point
289 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl)));
290 continue;
291 }
292 return NULL;
293 }
294 }
296 return n_ctrl;
297 }
299 //------------------------------remix_address_expressions----------------------
300 // Rework addressing expressions to get the most loop-invariant stuff
301 // moved out. We'd like to do all associative operators, but it's especially
302 // important (common) to do address expressions.
303 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
304 if (!has_ctrl(n)) return NULL;
305 Node *n_ctrl = get_ctrl(n);
306 IdealLoopTree *n_loop = get_loop(n_ctrl);
308 // See if 'n' mixes loop-varying and loop-invariant inputs and
309 // itself is loop-varying.
311 // Only interested in binary ops (and AddP)
312 if( n->req() < 3 || n->req() > 4 ) return NULL;
314 Node *n1_ctrl = get_ctrl(n->in( 1));
315 Node *n2_ctrl = get_ctrl(n->in( 2));
316 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
317 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
318 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
319 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
321 // Does one of my inputs spin in a tighter loop than self?
322 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
323 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
324 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
325 return NULL; // Leave well enough alone
327 // Is at least one of my inputs loop-invariant?
328 if( n1_loop == n_loop &&
329 n2_loop == n_loop &&
330 n3_loop == n_loop )
331 return NULL; // No loop-invariant inputs
334 int n_op = n->Opcode();
336 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
337 if( n_op == Op_LShiftI ) {
338 // Scale is loop invariant
339 Node *scale = n->in(2);
340 Node *scale_ctrl = get_ctrl(scale);
341 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
342 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
343 return NULL;
344 const TypeInt *scale_t = scale->bottom_type()->isa_int();
345 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
346 return NULL; // Dont bother with byte/short masking
347 // Add must vary with loop (else shift would be loop-invariant)
348 Node *add = n->in(1);
349 Node *add_ctrl = get_ctrl(add);
350 IdealLoopTree *add_loop = get_loop(add_ctrl);
351 //assert( n_loop == add_loop, "" );
352 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
354 // Convert I-V into I+ (0-V); same for V-I
355 if( add->Opcode() == Op_SubI &&
356 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
357 Node *zero = _igvn.intcon(0);
358 set_ctrl(zero, C->root());
359 Node *neg = new (C, 3) SubINode( _igvn.intcon(0), add->in(2) );
360 register_new_node( neg, get_ctrl(add->in(2) ) );
361 add = new (C, 3) AddINode( add->in(1), neg );
362 register_new_node( add, add_ctrl );
363 }
364 if( add->Opcode() != Op_AddI ) return NULL;
365 // See if one add input is loop invariant
366 Node *add_var = add->in(1);
367 Node *add_var_ctrl = get_ctrl(add_var);
368 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
369 Node *add_invar = add->in(2);
370 Node *add_invar_ctrl = get_ctrl(add_invar);
371 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
372 if( add_var_loop == n_loop ) {
373 } else if( add_invar_loop == n_loop ) {
374 // Swap to find the invariant part
375 add_invar = add_var;
376 add_invar_ctrl = add_var_ctrl;
377 add_invar_loop = add_var_loop;
378 add_var = add->in(2);
379 Node *add_var_ctrl = get_ctrl(add_var);
380 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
381 } else // Else neither input is loop invariant
382 return NULL;
383 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
384 return NULL; // No invariant part of the add?
386 // Yes! Reshape address expression!
387 Node *inv_scale = new (C, 3) LShiftINode( add_invar, scale );
388 Node *inv_scale_ctrl =
389 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
390 add_invar_ctrl : scale_ctrl;
391 register_new_node( inv_scale, inv_scale_ctrl );
392 Node *var_scale = new (C, 3) LShiftINode( add_var, scale );
393 register_new_node( var_scale, n_ctrl );
394 Node *var_add = new (C, 3) AddINode( var_scale, inv_scale );
395 register_new_node( var_add, n_ctrl );
396 _igvn.replace_node( n, var_add );
397 return var_add;
398 }
400 // Replace (I+V) with (V+I)
401 if( n_op == Op_AddI ||
402 n_op == Op_AddL ||
403 n_op == Op_AddF ||
404 n_op == Op_AddD ||
405 n_op == Op_MulI ||
406 n_op == Op_MulL ||
407 n_op == Op_MulF ||
408 n_op == Op_MulD ) {
409 if( n2_loop == n_loop ) {
410 assert( n1_loop != n_loop, "" );
411 n->swap_edges(1, 2);
412 }
413 }
415 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
416 // but not if I2 is a constant.
417 if( n_op == Op_AddP ) {
418 if( n2_loop == n_loop && n3_loop != n_loop ) {
419 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
420 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
421 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
422 IdealLoopTree *n22loop = get_loop( n22_ctrl );
423 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
424 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
425 n23_loop == n_loop ) {
426 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
427 // Stuff new AddP in the loop preheader
428 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
429 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) );
430 register_new_node( add2, n_ctrl );
431 _igvn.replace_node( n, add2 );
432 return add2;
433 }
434 }
435 }
437 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
438 if( n2_loop != n_loop && n3_loop == n_loop ) {
439 if( n->in(3)->Opcode() == Op_AddI ) {
440 Node *V = n->in(3)->in(1);
441 Node *I = n->in(3)->in(2);
442 if( is_member(n_loop,get_ctrl(V)) ) {
443 } else {
444 Node *tmp = V; V = I; I = tmp;
445 }
446 if( !is_member(n_loop,get_ctrl(I)) ) {
447 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I );
448 // Stuff new AddP in the loop preheader
449 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
450 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V );
451 register_new_node( add2, n_ctrl );
452 _igvn.replace_node( n, add2 );
453 return add2;
454 }
455 }
456 }
457 }
459 return NULL;
460 }
462 //------------------------------conditional_move-------------------------------
463 // Attempt to replace a Phi with a conditional move. We have some pretty
464 // strict profitability requirements. All Phis at the merge point must
465 // be converted, so we can remove the control flow. We need to limit the
466 // number of c-moves to a small handful. All code that was in the side-arms
467 // of the CFG diamond is now speculatively executed. This code has to be
468 // "cheap enough". We are pretty much limited to CFG diamonds that merge
469 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
470 Node *PhaseIdealLoop::conditional_move( Node *region ) {
472 assert(region->is_Region(), "sanity check");
473 if (region->req() != 3) return NULL;
475 // Check for CFG diamond
476 Node *lp = region->in(1);
477 Node *rp = region->in(2);
478 if (!lp || !rp) return NULL;
479 Node *lp_c = lp->in(0);
480 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
481 IfNode *iff = lp_c->as_If();
483 // Check for ops pinned in an arm of the diamond.
484 // Can't remove the control flow in this case
485 if (lp->outcnt() > 1) return NULL;
486 if (rp->outcnt() > 1) return NULL;
488 IdealLoopTree* r_loop = get_loop(region);
489 assert(r_loop == get_loop(iff), "sanity");
490 // Always convert to CMOVE if all results are used only outside this loop.
491 bool used_inside_loop = (r_loop == _ltree_root);
493 // Check profitability
494 int cost = 0;
495 int phis = 0;
496 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
497 Node *out = region->fast_out(i);
498 if (!out->is_Phi()) continue; // Ignore other control edges, etc
499 phis++;
500 PhiNode* phi = out->as_Phi();
501 BasicType bt = phi->type()->basic_type();
502 switch (bt) {
503 case T_FLOAT:
504 case T_DOUBLE: {
505 cost += Matcher::float_cmove_cost(); // Could be very expensive
506 break;
507 }
508 case T_LONG: {
509 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
510 }
511 case T_INT: // These all CMOV fine
512 case T_ADDRESS: { // (RawPtr)
513 cost++;
514 break;
515 }
516 case T_NARROWOOP: // Fall through
517 case T_OBJECT: { // Base oops are OK, but not derived oops
518 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
519 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
520 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
521 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
522 // have a Phi for the base here that we convert to a CMOVE all is well
523 // and good. But if the base is dead, we'll not make a CMOVE. Later
524 // the allocator will have to produce a base by creating a CMOVE of the
525 // relevant bases. This puts the allocator in the business of
526 // manufacturing expensive instructions, generally a bad plan.
527 // Just Say No to Conditionally-Moved Derived Pointers.
528 if (tp && tp->offset() != 0)
529 return NULL;
530 cost++;
531 break;
532 }
533 default:
534 return NULL; // In particular, can't do memory or I/O
535 }
536 // Add in cost any speculative ops
537 for (uint j = 1; j < region->req(); j++) {
538 Node *proj = region->in(j);
539 Node *inp = phi->in(j);
540 if (get_ctrl(inp) == proj) { // Found local op
541 cost++;
542 // Check for a chain of dependent ops; these will all become
543 // speculative in a CMOV.
544 for (uint k = 1; k < inp->req(); k++)
545 if (get_ctrl(inp->in(k)) == proj)
546 cost += ConditionalMoveLimit; // Too much speculative goo
547 }
548 }
549 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
550 // This will likely Split-If, a higher-payoff operation.
551 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
552 Node* use = phi->fast_out(k);
553 if (use->is_Cmp() || use->is_DecodeN() || use->is_EncodeP())
554 cost += ConditionalMoveLimit;
555 // Is there a use inside the loop?
556 // Note: check only basic types since CMoveP is pinned.
557 if (!used_inside_loop && is_java_primitive(bt)) {
558 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
559 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
560 used_inside_loop = true;
561 }
562 }
563 }
564 }
565 Node* bol = iff->in(1);
566 assert(bol->Opcode() == Op_Bool, "");
567 int cmp_op = bol->in(1)->Opcode();
568 // It is expensive to generate flags from a float compare.
569 // Avoid duplicated float compare.
570 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
572 float infrequent_prob = PROB_UNLIKELY_MAG(3);
573 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
574 if (used_inside_loop) {
575 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
577 // BlockLayoutByFrequency optimization moves infrequent branch
578 // from hot path. No point in CMOV'ing in such case (110 is used
579 // instead of 100 to take into account not exactness of float value).
580 if (BlockLayoutByFrequency) {
581 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
582 }
583 }
584 // Check for highly predictable branch. No point in CMOV'ing if
585 // we are going to predict accurately all the time.
586 if (iff->_prob < infrequent_prob ||
587 iff->_prob > (1.0f - infrequent_prob))
588 return NULL;
590 // --------------
591 // Now replace all Phis with CMOV's
592 Node *cmov_ctrl = iff->in(0);
593 uint flip = (lp->Opcode() == Op_IfTrue);
594 while (1) {
595 PhiNode* phi = NULL;
596 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
597 Node *out = region->fast_out(i);
598 if (out->is_Phi()) {
599 phi = out->as_Phi();
600 break;
601 }
602 }
603 if (phi == NULL) break;
604 #ifndef PRODUCT
605 if (PrintOpto && VerifyLoopOptimizations) tty->print_cr("CMOV");
606 #endif
607 // Move speculative ops
608 for (uint j = 1; j < region->req(); j++) {
609 Node *proj = region->in(j);
610 Node *inp = phi->in(j);
611 if (get_ctrl(inp) == proj) { // Found local op
612 #ifndef PRODUCT
613 if (PrintOpto && VerifyLoopOptimizations) {
614 tty->print(" speculate: ");
615 inp->dump();
616 }
617 #endif
618 set_ctrl(inp, cmov_ctrl);
619 }
620 }
621 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) );
622 register_new_node( cmov, cmov_ctrl );
623 _igvn.replace_node( phi, cmov );
624 #ifndef PRODUCT
625 if (TraceLoopOpts) {
626 tty->print("CMOV ");
627 r_loop->dump_head();
628 if (Verbose) {
629 bol->in(1)->dump(1);
630 cmov->dump(1);
631 }
632 }
633 if (VerifyLoopOptimizations) verify();
634 #endif
635 }
637 // The useless CFG diamond will fold up later; see the optimization in
638 // RegionNode::Ideal.
639 _igvn._worklist.push(region);
641 return iff->in(1);
642 }
644 //------------------------------split_if_with_blocks_pre-----------------------
645 // Do the real work in a non-recursive function. Data nodes want to be
646 // cloned in the pre-order so they can feed each other nicely.
647 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
648 // Cloning these guys is unlikely to win
649 int n_op = n->Opcode();
650 if( n_op == Op_MergeMem ) return n;
651 if( n->is_Proj() ) return n;
652 // Do not clone-up CmpFXXX variations, as these are always
653 // followed by a CmpI
654 if( n->is_Cmp() ) return n;
655 // Attempt to use a conditional move instead of a phi/branch
656 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
657 Node *cmov = conditional_move( n );
658 if( cmov ) return cmov;
659 }
660 if( n->is_CFG() || n->is_LoadStore() )
661 return n;
662 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
663 n_op == Op_Opaque2 ) {
664 if( !C->major_progress() ) // If chance of no more loop opts...
665 _igvn._worklist.push(n); // maybe we'll remove them
666 return n;
667 }
669 if( n->is_Con() ) return n; // No cloning for Con nodes
671 Node *n_ctrl = get_ctrl(n);
672 if( !n_ctrl ) return n; // Dead node
674 // Attempt to remix address expressions for loop invariants
675 Node *m = remix_address_expressions( n );
676 if( m ) return m;
678 // Determine if the Node has inputs from some local Phi.
679 // Returns the block to clone thru.
680 Node *n_blk = has_local_phi_input( n );
681 if( !n_blk ) return n;
682 // Do not clone the trip counter through on a CountedLoop
683 // (messes up the canonical shape).
684 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
686 // Check for having no control input; not pinned. Allow
687 // dominating control.
688 if( n->in(0) ) {
689 Node *dom = idom(n_blk);
690 if( dom_lca( n->in(0), dom ) != n->in(0) )
691 return n;
692 }
693 // Policy: when is it profitable. You must get more wins than
694 // policy before it is considered profitable. Policy is usually 0,
695 // so 1 win is considered profitable. Big merges will require big
696 // cloning, so get a larger policy.
697 int policy = n_blk->req() >> 2;
699 // If the loop is a candidate for range check elimination,
700 // delay splitting through it's phi until a later loop optimization
701 if (n_blk->is_CountedLoop()) {
702 IdealLoopTree *lp = get_loop(n_blk);
703 if (lp && lp->_rce_candidate) {
704 return n;
705 }
706 }
708 // Use same limit as split_if_with_blocks_post
709 if( C->unique() > 35000 ) return n; // Method too big
711 // Split 'n' through the merge point if it is profitable
712 Node *phi = split_thru_phi( n, n_blk, policy );
713 if (!phi) return n;
715 // Found a Phi to split thru!
716 // Replace 'n' with the new phi
717 _igvn.replace_node( n, phi );
718 // Moved a load around the loop, 'en-registering' something.
719 if (n_blk->is_Loop() && n->is_Load() &&
720 !phi->in(LoopNode::LoopBackControl)->is_Load())
721 C->set_major_progress();
723 return phi;
724 }
726 static bool merge_point_too_heavy(Compile* C, Node* region) {
727 // Bail out if the region and its phis have too many users.
728 int weight = 0;
729 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
730 weight += region->fast_out(i)->outcnt();
731 }
732 int nodes_left = MaxNodeLimit - C->unique();
733 if (weight * 8 > nodes_left) {
734 #ifndef PRODUCT
735 if (PrintOpto)
736 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
737 #endif
738 return true;
739 } else {
740 return false;
741 }
742 }
744 static bool merge_point_safe(Node* region) {
745 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
746 // having a PhiNode input. This sidesteps the dangerous case where the split
747 // ConvI2LNode may become TOP if the input Value() does not
748 // overlap the ConvI2L range, leaving a node which may not dominate its
749 // uses.
750 // A better fix for this problem can be found in the BugTraq entry, but
751 // expediency for Mantis demands this hack.
752 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop
753 // split_if_with_blocks from splitting a block because we could not move around
754 // the FastLockNode.
755 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
756 Node* n = region->fast_out(i);
757 if (n->is_Phi()) {
758 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
759 Node* m = n->fast_out(j);
760 if (m->is_FastLock())
761 return false;
762 #ifdef _LP64
763 if (m->Opcode() == Op_ConvI2L)
764 return false;
765 #endif
766 }
767 }
768 }
769 return true;
770 }
773 //------------------------------place_near_use---------------------------------
774 // Place some computation next to use but not inside inner loops.
775 // For inner loop uses move it to the preheader area.
776 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const {
777 IdealLoopTree *u_loop = get_loop( useblock );
778 return (u_loop->_irreducible || u_loop->_child)
779 ? useblock
780 : u_loop->_head->in(LoopNode::EntryControl);
781 }
784 //------------------------------split_if_with_blocks_post----------------------
785 // Do the real work in a non-recursive function. CFG hackery wants to be
786 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
787 // info.
788 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) {
790 // Cloning Cmp through Phi's involves the split-if transform.
791 // FastLock is not used by an If
792 if( n->is_Cmp() && !n->is_FastLock() ) {
793 if( C->unique() > 35000 ) return; // Method too big
795 // Do not do 'split-if' if irreducible loops are present.
796 if( _has_irreducible_loops )
797 return;
799 Node *n_ctrl = get_ctrl(n);
800 // Determine if the Node has inputs from some local Phi.
801 // Returns the block to clone thru.
802 Node *n_blk = has_local_phi_input( n );
803 if( n_blk != n_ctrl ) return;
805 if( merge_point_too_heavy(C, n_ctrl) )
806 return;
808 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare?
809 Node *bol = n->unique_out();
810 assert( bol->is_Bool(), "expect a bool here" );
811 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare?
812 Node *iff = bol->unique_out();
814 // Check some safety conditions
815 if( iff->is_If() ) { // Classic split-if?
816 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if
817 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
818 // Can't split CMove with different control edge.
819 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) return;
820 if( get_ctrl(iff->in(2)) == n_ctrl ||
821 get_ctrl(iff->in(3)) == n_ctrl )
822 return; // Inputs not yet split-up
823 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) {
824 return; // Loop-invar test gates loop-varying CMOVE
825 }
826 } else {
827 return; // some other kind of node, such as an Allocate
828 }
830 // Do not do 'split-if' if some paths are dead. First do dead code
831 // elimination and then see if its still profitable.
832 for( uint i = 1; i < n_ctrl->req(); i++ )
833 if( n_ctrl->in(i) == C->top() )
834 return;
836 // When is split-if profitable? Every 'win' on means some control flow
837 // goes dead, so it's almost always a win.
838 int policy = 0;
839 // If trying to do a 'Split-If' at the loop head, it is only
840 // profitable if the cmp folds up on BOTH paths. Otherwise we
841 // risk peeling a loop forever.
843 // CNC - Disabled for now. Requires careful handling of loop
844 // body selection for the cloned code. Also, make sure we check
845 // for any input path not being in the same loop as n_ctrl. For
846 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
847 // because the alternative loop entry points won't be converted
848 // into LoopNodes.
849 IdealLoopTree *n_loop = get_loop(n_ctrl);
850 for( uint j = 1; j < n_ctrl->req(); j++ )
851 if( get_loop(n_ctrl->in(j)) != n_loop )
852 return;
854 // Check for safety of the merge point.
855 if( !merge_point_safe(n_ctrl) ) {
856 return;
857 }
859 // Split compare 'n' through the merge point if it is profitable
860 Node *phi = split_thru_phi( n, n_ctrl, policy );
861 if( !phi ) return;
863 // Found a Phi to split thru!
864 // Replace 'n' with the new phi
865 _igvn.replace_node( n, phi );
867 // Now split the bool up thru the phi
868 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 );
869 _igvn.replace_node( bol, bolphi );
870 assert( iff->in(1) == bolphi, "" );
871 if( bolphi->Value(&_igvn)->singleton() )
872 return;
874 // Conditional-move? Must split up now
875 if( !iff->is_If() ) {
876 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 );
877 _igvn.replace_node( iff, cmovphi );
878 return;
879 }
881 // Now split the IF
882 do_split_if( iff );
883 return;
884 }
886 // Check for an IF ready to split; one that has its
887 // condition codes input coming from a Phi at the block start.
888 int n_op = n->Opcode();
890 // Check for an IF being dominated by another IF same test
891 if( n_op == Op_If ) {
892 Node *bol = n->in(1);
893 uint max = bol->outcnt();
894 // Check for same test used more than once?
895 if( n_op == Op_If && max > 1 && bol->is_Bool() ) {
896 // Search up IDOMs to see if this IF is dominated.
897 Node *cutoff = get_ctrl(bol);
899 // Now search up IDOMs till cutoff, looking for a dominating test
900 Node *prevdom = n;
901 Node *dom = idom(prevdom);
902 while( dom != cutoff ) {
903 if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) {
904 // Replace the dominated test with an obvious true or false.
905 // Place it on the IGVN worklist for later cleanup.
906 C->set_major_progress();
907 dominated_by( prevdom, n, false, true );
908 #ifndef PRODUCT
909 if( VerifyLoopOptimizations ) verify();
910 #endif
911 return;
912 }
913 prevdom = dom;
914 dom = idom(prevdom);
915 }
916 }
917 }
919 // See if a shared loop-varying computation has no loop-varying uses.
920 // Happens if something is only used for JVM state in uncommon trap exits,
921 // like various versions of induction variable+offset. Clone the
922 // computation per usage to allow it to sink out of the loop.
923 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
924 Node *n_ctrl = get_ctrl(n);
925 IdealLoopTree *n_loop = get_loop(n_ctrl);
926 if( n_loop != _ltree_root ) {
927 DUIterator_Fast imax, i = n->fast_outs(imax);
928 for (; i < imax; i++) {
929 Node* u = n->fast_out(i);
930 if( !has_ctrl(u) ) break; // Found control user
931 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
932 if( u_loop == n_loop ) break; // Found loop-varying use
933 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
934 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
935 }
936 bool did_break = (i < imax); // Did we break out of the previous loop?
937 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
938 Node *late_load_ctrl = NULL;
939 if (n->is_Load()) {
940 // If n is a load, get and save the result from get_late_ctrl(),
941 // to be later used in calculating the control for n's clones.
942 clear_dom_lca_tags();
943 late_load_ctrl = get_late_ctrl(n, n_ctrl);
944 }
945 // If n is a load, and the late control is the same as the current
946 // control, then the cloning of n is a pointless exercise, because
947 // GVN will ensure that we end up where we started.
948 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
949 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
950 Node *u = n->last_out(j); // Clone private computation per use
951 _igvn.rehash_node_delayed(u);
952 Node *x = n->clone(); // Clone computation
953 Node *x_ctrl = NULL;
954 if( u->is_Phi() ) {
955 // Replace all uses of normal nodes. Replace Phi uses
956 // individually, so the separate Nodes can sink down
957 // different paths.
958 uint k = 1;
959 while( u->in(k) != n ) k++;
960 u->set_req( k, x );
961 // x goes next to Phi input path
962 x_ctrl = u->in(0)->in(k);
963 --j;
964 } else { // Normal use
965 // Replace all uses
966 for( uint k = 0; k < u->req(); k++ ) {
967 if( u->in(k) == n ) {
968 u->set_req( k, x );
969 --j;
970 }
971 }
972 x_ctrl = get_ctrl(u);
973 }
975 // Find control for 'x' next to use but not inside inner loops.
976 // For inner loop uses get the preheader area.
977 x_ctrl = place_near_use(x_ctrl);
979 if (n->is_Load()) {
980 // For loads, add a control edge to a CFG node outside of the loop
981 // to force them to not combine and return back inside the loop
982 // during GVN optimization (4641526).
983 //
984 // Because we are setting the actual control input, factor in
985 // the result from get_late_ctrl() so we respect any
986 // anti-dependences. (6233005).
987 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
989 // Don't allow the control input to be a CFG splitting node.
990 // Such nodes should only have ProjNodes as outs, e.g. IfNode
991 // should only have IfTrueNode and IfFalseNode (4985384).
992 x_ctrl = find_non_split_ctrl(x_ctrl);
993 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
995 x->set_req(0, x_ctrl);
996 }
997 register_new_node(x, x_ctrl);
999 // Some institutional knowledge is needed here: 'x' is
1000 // yanked because if the optimizer runs GVN on it all the
1001 // cloned x's will common up and undo this optimization and
1002 // be forced back in the loop. This is annoying because it
1003 // makes +VerifyOpto report false-positives on progress. I
1004 // tried setting control edges on the x's to force them to
1005 // not combine, but the matching gets worried when it tries
1006 // to fold a StoreP and an AddP together (as part of an
1007 // address expression) and the AddP and StoreP have
1008 // different controls.
1009 if( !x->is_Load() && !x->is_DecodeN() ) _igvn._worklist.yank(x);
1010 }
1011 _igvn.remove_dead_node(n);
1012 }
1013 }
1014 }
1015 }
1017 // Check for Opaque2's who's loop has disappeared - who's input is in the
1018 // same loop nest as their output. Remove 'em, they are no longer useful.
1019 if( n_op == Op_Opaque2 &&
1020 n->in(1) != NULL &&
1021 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1022 _igvn.replace_node( n, n->in(1) );
1023 }
1024 }
1026 //------------------------------split_if_with_blocks---------------------------
1027 // Check for aggressive application of 'split-if' optimization,
1028 // using basic block level info.
1029 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) {
1030 Node *n = C->root();
1031 visited.set(n->_idx); // first, mark node as visited
1032 // Do pre-visit work for root
1033 n = split_if_with_blocks_pre( n );
1034 uint cnt = n->outcnt();
1035 uint i = 0;
1036 while (true) {
1037 // Visit all children
1038 if (i < cnt) {
1039 Node* use = n->raw_out(i);
1040 ++i;
1041 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1042 // Now do pre-visit work for this use
1043 use = split_if_with_blocks_pre( use );
1044 nstack.push(n, i); // Save parent and next use's index.
1045 n = use; // Process all children of current use.
1046 cnt = use->outcnt();
1047 i = 0;
1048 }
1049 }
1050 else {
1051 // All of n's children have been processed, complete post-processing.
1052 if (cnt != 0 && !n->is_Con()) {
1053 assert(has_node(n), "no dead nodes");
1054 split_if_with_blocks_post( n );
1055 }
1056 if (nstack.is_empty()) {
1057 // Finished all nodes on stack.
1058 break;
1059 }
1060 // Get saved parent node and next use's index. Visit the rest of uses.
1061 n = nstack.node();
1062 cnt = n->outcnt();
1063 i = nstack.index();
1064 nstack.pop();
1065 }
1066 }
1067 }
1070 //=============================================================================
1071 //
1072 // C L O N E A L O O P B O D Y
1073 //
1075 //------------------------------clone_iff--------------------------------------
1076 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1077 // "Nearly" because all Nodes have been cloned from the original in the loop,
1078 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1079 // through the Phi recursively, and return a Bool.
1080 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) {
1082 // Convert this Phi into a Phi merging Bools
1083 uint i;
1084 for( i = 1; i < phi->req(); i++ ) {
1085 Node *b = phi->in(i);
1086 if( b->is_Phi() ) {
1087 _igvn.replace_input_of(phi, i, clone_iff( b->as_Phi(), loop ));
1088 } else {
1089 assert( b->is_Bool(), "" );
1090 }
1091 }
1093 Node *sample_bool = phi->in(1);
1094 Node *sample_cmp = sample_bool->in(1);
1096 // Make Phis to merge the Cmp's inputs.
1097 int size = phi->in(0)->req();
1098 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1099 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1100 for( i = 1; i < phi->req(); i++ ) {
1101 Node *n1 = phi->in(i)->in(1)->in(1);
1102 Node *n2 = phi->in(i)->in(1)->in(2);
1103 phi1->set_req( i, n1 );
1104 phi2->set_req( i, n2 );
1105 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1106 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1107 }
1108 // See if these Phis have been made before.
1109 // Register with optimizer
1110 Node *hit1 = _igvn.hash_find_insert(phi1);
1111 if( hit1 ) { // Hit, toss just made Phi
1112 _igvn.remove_dead_node(phi1); // Remove new phi
1113 assert( hit1->is_Phi(), "" );
1114 phi1 = (PhiNode*)hit1; // Use existing phi
1115 } else { // Miss
1116 _igvn.register_new_node_with_optimizer(phi1);
1117 }
1118 Node *hit2 = _igvn.hash_find_insert(phi2);
1119 if( hit2 ) { // Hit, toss just made Phi
1120 _igvn.remove_dead_node(phi2); // Remove new phi
1121 assert( hit2->is_Phi(), "" );
1122 phi2 = (PhiNode*)hit2; // Use existing phi
1123 } else { // Miss
1124 _igvn.register_new_node_with_optimizer(phi2);
1125 }
1126 // Register Phis with loop/block info
1127 set_ctrl(phi1, phi->in(0));
1128 set_ctrl(phi2, phi->in(0));
1129 // Make a new Cmp
1130 Node *cmp = sample_cmp->clone();
1131 cmp->set_req( 1, phi1 );
1132 cmp->set_req( 2, phi2 );
1133 _igvn.register_new_node_with_optimizer(cmp);
1134 set_ctrl(cmp, phi->in(0));
1136 // Make a new Bool
1137 Node *b = sample_bool->clone();
1138 b->set_req(1,cmp);
1139 _igvn.register_new_node_with_optimizer(b);
1140 set_ctrl(b, phi->in(0));
1142 assert( b->is_Bool(), "" );
1143 return (BoolNode*)b;
1144 }
1146 //------------------------------clone_bool-------------------------------------
1147 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1148 // "Nearly" because all Nodes have been cloned from the original in the loop,
1149 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1150 // through the Phi recursively, and return a Bool.
1151 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1152 uint i;
1153 // Convert this Phi into a Phi merging Bools
1154 for( i = 1; i < phi->req(); i++ ) {
1155 Node *b = phi->in(i);
1156 if( b->is_Phi() ) {
1157 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1158 } else {
1159 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1160 }
1161 }
1163 Node *sample_cmp = phi->in(1);
1165 // Make Phis to merge the Cmp's inputs.
1166 int size = phi->in(0)->req();
1167 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1168 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1169 for( uint j = 1; j < phi->req(); j++ ) {
1170 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1171 Node *n1, *n2;
1172 if( cmp_top->is_Cmp() ) {
1173 n1 = cmp_top->in(1);
1174 n2 = cmp_top->in(2);
1175 } else {
1176 n1 = n2 = cmp_top;
1177 }
1178 phi1->set_req( j, n1 );
1179 phi2->set_req( j, n2 );
1180 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1181 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1182 }
1184 // See if these Phis have been made before.
1185 // Register with optimizer
1186 Node *hit1 = _igvn.hash_find_insert(phi1);
1187 if( hit1 ) { // Hit, toss just made Phi
1188 _igvn.remove_dead_node(phi1); // Remove new phi
1189 assert( hit1->is_Phi(), "" );
1190 phi1 = (PhiNode*)hit1; // Use existing phi
1191 } else { // Miss
1192 _igvn.register_new_node_with_optimizer(phi1);
1193 }
1194 Node *hit2 = _igvn.hash_find_insert(phi2);
1195 if( hit2 ) { // Hit, toss just made Phi
1196 _igvn.remove_dead_node(phi2); // Remove new phi
1197 assert( hit2->is_Phi(), "" );
1198 phi2 = (PhiNode*)hit2; // Use existing phi
1199 } else { // Miss
1200 _igvn.register_new_node_with_optimizer(phi2);
1201 }
1202 // Register Phis with loop/block info
1203 set_ctrl(phi1, phi->in(0));
1204 set_ctrl(phi2, phi->in(0));
1205 // Make a new Cmp
1206 Node *cmp = sample_cmp->clone();
1207 cmp->set_req( 1, phi1 );
1208 cmp->set_req( 2, phi2 );
1209 _igvn.register_new_node_with_optimizer(cmp);
1210 set_ctrl(cmp, phi->in(0));
1212 assert( cmp->is_Cmp(), "" );
1213 return (CmpNode*)cmp;
1214 }
1216 //------------------------------sink_use---------------------------------------
1217 // If 'use' was in the loop-exit block, it now needs to be sunk
1218 // below the post-loop merge point.
1219 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1220 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1221 set_ctrl(use, post_loop);
1222 for (DUIterator j = use->outs(); use->has_out(j); j++)
1223 sink_use(use->out(j), post_loop);
1224 }
1225 }
1227 //------------------------------clone_loop-------------------------------------
1228 //
1229 // C L O N E A L O O P B O D Y
1230 //
1231 // This is the basic building block of the loop optimizations. It clones an
1232 // entire loop body. It makes an old_new loop body mapping; with this mapping
1233 // you can find the new-loop equivalent to an old-loop node. All new-loop
1234 // nodes are exactly equal to their old-loop counterparts, all edges are the
1235 // same. All exits from the old-loop now have a RegionNode that merges the
1236 // equivalent new-loop path. This is true even for the normal "loop-exit"
1237 // condition. All uses of loop-invariant old-loop values now come from (one
1238 // or more) Phis that merge their new-loop equivalents.
1239 //
1240 // This operation leaves the graph in an illegal state: there are two valid
1241 // control edges coming from the loop pre-header to both loop bodies. I'll
1242 // definitely have to hack the graph after running this transform.
1243 //
1244 // From this building block I will further edit edges to perform loop peeling
1245 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1246 //
1247 // Parameter side_by_size_idom:
1248 // When side_by_size_idom is NULL, the dominator tree is constructed for
1249 // the clone loop to dominate the original. Used in construction of
1250 // pre-main-post loop sequence.
1251 // When nonnull, the clone and original are side-by-side, both are
1252 // dominated by the side_by_side_idom node. Used in construction of
1253 // unswitched loops.
1254 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1255 Node* side_by_side_idom) {
1257 // Step 1: Clone the loop body. Make the old->new mapping.
1258 uint i;
1259 for( i = 0; i < loop->_body.size(); i++ ) {
1260 Node *old = loop->_body.at(i);
1261 Node *nnn = old->clone();
1262 old_new.map( old->_idx, nnn );
1263 _igvn.register_new_node_with_optimizer(nnn);
1264 }
1267 // Step 2: Fix the edges in the new body. If the old input is outside the
1268 // loop use it. If the old input is INside the loop, use the corresponding
1269 // new node instead.
1270 for( i = 0; i < loop->_body.size(); i++ ) {
1271 Node *old = loop->_body.at(i);
1272 Node *nnn = old_new[old->_idx];
1273 // Fix CFG/Loop controlling the new node
1274 if (has_ctrl(old)) {
1275 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
1276 } else {
1277 set_loop(nnn, loop->_parent);
1278 if (old->outcnt() > 0) {
1279 set_idom( nnn, old_new[idom(old)->_idx], dd );
1280 }
1281 }
1282 // Correct edges to the new node
1283 for( uint j = 0; j < nnn->req(); j++ ) {
1284 Node *n = nnn->in(j);
1285 if( n ) {
1286 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
1287 if( loop->is_member( old_in_loop ) )
1288 nnn->set_req(j, old_new[n->_idx]);
1289 }
1290 }
1291 _igvn.hash_find_insert(nnn);
1292 }
1293 Node *newhead = old_new[loop->_head->_idx];
1294 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1297 // Step 3: Now fix control uses. Loop varying control uses have already
1298 // been fixed up (as part of all input edges in Step 2). Loop invariant
1299 // control uses must be either an IfFalse or an IfTrue. Make a merge
1300 // point to merge the old and new IfFalse/IfTrue nodes; make the use
1301 // refer to this.
1302 ResourceArea *area = Thread::current()->resource_area();
1303 Node_List worklist(area);
1304 uint new_counter = C->unique();
1305 for( i = 0; i < loop->_body.size(); i++ ) {
1306 Node* old = loop->_body.at(i);
1307 if( !old->is_CFG() ) continue;
1308 Node* nnn = old_new[old->_idx];
1310 // Copy uses to a worklist, so I can munge the def-use info
1311 // with impunity.
1312 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1313 worklist.push(old->fast_out(j));
1315 while( worklist.size() ) { // Visit all uses
1316 Node *use = worklist.pop();
1317 if (!has_node(use)) continue; // Ignore dead nodes
1318 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1319 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
1320 // Both OLD and USE are CFG nodes here.
1321 assert( use->is_Proj(), "" );
1323 // Clone the loop exit control projection
1324 Node *newuse = use->clone();
1325 newuse->set_req(0,nnn);
1326 _igvn.register_new_node_with_optimizer(newuse);
1327 set_loop(newuse, use_loop);
1328 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
1330 // We need a Region to merge the exit from the peeled body and the
1331 // exit from the old loop body.
1332 RegionNode *r = new (C, 3) RegionNode(3);
1333 // Map the old use to the new merge point
1334 old_new.map( use->_idx, r );
1335 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
1336 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
1338 // The original user of 'use' uses 'r' instead.
1339 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
1340 Node* useuse = use->last_out(l);
1341 _igvn.rehash_node_delayed(useuse);
1342 uint uses_found = 0;
1343 if( useuse->in(0) == use ) {
1344 useuse->set_req(0, r);
1345 uses_found++;
1346 if( useuse->is_CFG() ) {
1347 assert( dom_depth(useuse) > dd_r, "" );
1348 set_idom(useuse, r, dom_depth(useuse));
1349 }
1350 }
1351 for( uint k = 1; k < useuse->req(); k++ ) {
1352 if( useuse->in(k) == use ) {
1353 useuse->set_req(k, r);
1354 uses_found++;
1355 }
1356 }
1357 l -= uses_found; // we deleted 1 or more copies of this edge
1358 }
1360 // Now finish up 'r'
1361 r->set_req( 1, newuse );
1362 r->set_req( 2, use );
1363 _igvn.register_new_node_with_optimizer(r);
1364 set_loop(r, use_loop);
1365 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
1366 } // End of if a loop-exit test
1367 }
1368 }
1370 // Step 4: If loop-invariant use is not control, it must be dominated by a
1371 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
1372 // there if needed. Make a Phi there merging old and new used values.
1373 Node_List *split_if_set = NULL;
1374 Node_List *split_bool_set = NULL;
1375 Node_List *split_cex_set = NULL;
1376 for( i = 0; i < loop->_body.size(); i++ ) {
1377 Node* old = loop->_body.at(i);
1378 Node* nnn = old_new[old->_idx];
1379 // Copy uses to a worklist, so I can munge the def-use info
1380 // with impunity.
1381 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1382 worklist.push(old->fast_out(j));
1384 while( worklist.size() ) {
1385 Node *use = worklist.pop();
1386 if (!has_node(use)) continue; // Ignore dead nodes
1387 if (use->in(0) == C->top()) continue;
1388 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1389 // Check for data-use outside of loop - at least one of OLD or USE
1390 // must not be a CFG node.
1391 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) {
1393 // If the Data use is an IF, that means we have an IF outside of the
1394 // loop that is switching on a condition that is set inside of the
1395 // loop. Happens if people set a loop-exit flag; then test the flag
1396 // in the loop to break the loop, then test is again outside of the
1397 // loop to determine which way the loop exited.
1398 if( use->is_If() || use->is_CMove() ) {
1399 // Since this code is highly unlikely, we lazily build the worklist
1400 // of such Nodes to go split.
1401 if( !split_if_set )
1402 split_if_set = new Node_List(area);
1403 split_if_set->push(use);
1404 }
1405 if( use->is_Bool() ) {
1406 if( !split_bool_set )
1407 split_bool_set = new Node_List(area);
1408 split_bool_set->push(use);
1409 }
1410 if( use->Opcode() == Op_CreateEx ) {
1411 if( !split_cex_set )
1412 split_cex_set = new Node_List(area);
1413 split_cex_set->push(use);
1414 }
1417 // Get "block" use is in
1418 uint idx = 0;
1419 while( use->in(idx) != old ) idx++;
1420 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1421 assert( !loop->is_member( get_loop( prev ) ), "" );
1422 Node *cfg = prev->_idx >= new_counter
1423 ? prev->in(2)
1424 : idom(prev);
1425 if( use->is_Phi() ) // Phi use is in prior block
1426 cfg = prev->in(idx); // NOT in block of Phi itself
1427 if (cfg->is_top()) { // Use is dead?
1428 _igvn.replace_input_of(use, idx, C->top());
1429 continue;
1430 }
1432 while( !loop->is_member( get_loop( cfg ) ) ) {
1433 prev = cfg;
1434 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1435 }
1436 // If the use occurs after merging several exits from the loop, then
1437 // old value must have dominated all those exits. Since the same old
1438 // value was used on all those exits we did not need a Phi at this
1439 // merge point. NOW we do need a Phi here. Each loop exit value
1440 // is now merged with the peeled body exit; each exit gets its own
1441 // private Phi and those Phis need to be merged here.
1442 Node *phi;
1443 if( prev->is_Region() ) {
1444 if( idx == 0 ) { // Updating control edge?
1445 phi = prev; // Just use existing control
1446 } else { // Else need a new Phi
1447 phi = PhiNode::make( prev, old );
1448 // Now recursively fix up the new uses of old!
1449 for( uint i = 1; i < prev->req(); i++ ) {
1450 worklist.push(phi); // Onto worklist once for each 'old' input
1451 }
1452 }
1453 } else {
1454 // Get new RegionNode merging old and new loop exits
1455 prev = old_new[prev->_idx];
1456 assert( prev, "just made this in step 7" );
1457 if( idx == 0 ) { // Updating control edge?
1458 phi = prev; // Just use existing control
1459 } else { // Else need a new Phi
1460 // Make a new Phi merging data values properly
1461 phi = PhiNode::make( prev, old );
1462 phi->set_req( 1, nnn );
1463 }
1464 }
1465 // If inserting a new Phi, check for prior hits
1466 if( idx != 0 ) {
1467 Node *hit = _igvn.hash_find_insert(phi);
1468 if( hit == NULL ) {
1469 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1470 } else { // or
1471 // Remove the new phi from the graph and use the hit
1472 _igvn.remove_dead_node(phi);
1473 phi = hit; // Use existing phi
1474 }
1475 set_ctrl(phi, prev);
1476 }
1477 // Make 'use' use the Phi instead of the old loop body exit value
1478 _igvn.replace_input_of(use, idx, phi);
1479 if( use->_idx >= new_counter ) { // If updating new phis
1480 // Not needed for correctness, but prevents a weak assert
1481 // in AddPNode from tripping (when we end up with different
1482 // base & derived Phis that will become the same after
1483 // IGVN does CSE).
1484 Node *hit = _igvn.hash_find_insert(use);
1485 if( hit ) // Go ahead and re-hash for hits.
1486 _igvn.replace_node( use, hit );
1487 }
1489 // If 'use' was in the loop-exit block, it now needs to be sunk
1490 // below the post-loop merge point.
1491 sink_use( use, prev );
1492 }
1493 }
1494 }
1496 // Check for IFs that need splitting/cloning. Happens if an IF outside of
1497 // the loop uses a condition set in the loop. The original IF probably
1498 // takes control from one or more OLD Regions (which in turn get from NEW
1499 // Regions). In any case, there will be a set of Phis for each merge point
1500 // from the IF up to where the original BOOL def exists the loop.
1501 if( split_if_set ) {
1502 while( split_if_set->size() ) {
1503 Node *iff = split_if_set->pop();
1504 if( iff->in(1)->is_Phi() ) {
1505 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop );
1506 _igvn.replace_input_of(iff, 1, b);
1507 }
1508 }
1509 }
1510 if( split_bool_set ) {
1511 while( split_bool_set->size() ) {
1512 Node *b = split_bool_set->pop();
1513 Node *phi = b->in(1);
1514 assert( phi->is_Phi(), "" );
1515 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop );
1516 _igvn.replace_input_of(b, 1, cmp);
1517 }
1518 }
1519 if( split_cex_set ) {
1520 while( split_cex_set->size() ) {
1521 Node *b = split_cex_set->pop();
1522 assert( b->in(0)->is_Region(), "" );
1523 assert( b->in(1)->is_Phi(), "" );
1524 assert( b->in(0)->in(0) == b->in(1)->in(0), "" );
1525 split_up( b, b->in(0), NULL );
1526 }
1527 }
1529 }
1532 //---------------------- stride_of_possible_iv -------------------------------------
1533 // Looks for an iff/bool/comp with one operand of the compare
1534 // being a cycle involving an add and a phi,
1535 // with an optional truncation (left-shift followed by a right-shift)
1536 // of the add. Returns zero if not an iv.
1537 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
1538 Node* trunc1 = NULL;
1539 Node* trunc2 = NULL;
1540 const TypeInt* ttype = NULL;
1541 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
1542 return 0;
1543 }
1544 BoolNode* bl = iff->in(1)->as_Bool();
1545 Node* cmp = bl->in(1);
1546 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) {
1547 return 0;
1548 }
1549 // Must have an invariant operand
1550 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
1551 return 0;
1552 }
1553 Node* add2 = NULL;
1554 Node* cmp1 = cmp->in(1);
1555 if (cmp1->is_Phi()) {
1556 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
1557 Node* phi = cmp1;
1558 for (uint i = 1; i < phi->req(); i++) {
1559 Node* in = phi->in(i);
1560 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
1561 &trunc1, &trunc2, &ttype);
1562 if (add && add->in(1) == phi) {
1563 add2 = add->in(2);
1564 break;
1565 }
1566 }
1567 } else {
1568 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
1569 Node* addtrunc = cmp1;
1570 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
1571 &trunc1, &trunc2, &ttype);
1572 if (add && add->in(1)->is_Phi()) {
1573 Node* phi = add->in(1);
1574 for (uint i = 1; i < phi->req(); i++) {
1575 if (phi->in(i) == addtrunc) {
1576 add2 = add->in(2);
1577 break;
1578 }
1579 }
1580 }
1581 }
1582 if (add2 != NULL) {
1583 const TypeInt* add2t = _igvn.type(add2)->is_int();
1584 if (add2t->is_con()) {
1585 return add2t->get_con();
1586 }
1587 }
1588 return 0;
1589 }
1592 //---------------------- stay_in_loop -------------------------------------
1593 // Return the (unique) control output node that's in the loop (if it exists.)
1594 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
1595 Node* unique = NULL;
1596 if (!n) return NULL;
1597 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1598 Node* use = n->fast_out(i);
1599 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
1600 if (unique != NULL) {
1601 return NULL;
1602 }
1603 unique = use;
1604 }
1605 }
1606 return unique;
1607 }
1609 //------------------------------ register_node -------------------------------------
1610 // Utility to register node "n" with PhaseIdealLoop
1611 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
1612 _igvn.register_new_node_with_optimizer(n);
1613 loop->_body.push(n);
1614 if (n->is_CFG()) {
1615 set_loop(n, loop);
1616 set_idom(n, pred, ddepth);
1617 } else {
1618 set_ctrl(n, pred);
1619 }
1620 }
1622 //------------------------------ proj_clone -------------------------------------
1623 // Utility to create an if-projection
1624 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
1625 ProjNode* c = p->clone()->as_Proj();
1626 c->set_req(0, iff);
1627 return c;
1628 }
1630 //------------------------------ short_circuit_if -------------------------------------
1631 // Force the iff control output to be the live_proj
1632 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
1633 int proj_con = live_proj->_con;
1634 assert(proj_con == 0 || proj_con == 1, "false or true projection");
1635 Node *con = _igvn.intcon(proj_con);
1636 set_ctrl(con, C->root());
1637 if (iff) {
1638 iff->set_req(1, con);
1639 }
1640 return con;
1641 }
1643 //------------------------------ insert_if_before_proj -------------------------------------
1644 // Insert a new if before an if projection (* - new node)
1645 //
1646 // before
1647 // if(test)
1648 // / \
1649 // v v
1650 // other-proj proj (arg)
1651 //
1652 // after
1653 // if(test)
1654 // / \
1655 // / v
1656 // | * proj-clone
1657 // v |
1658 // other-proj v
1659 // * new_if(relop(cmp[IU](left,right)))
1660 // / \
1661 // v v
1662 // * new-proj proj
1663 // (returned)
1664 //
1665 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
1666 IfNode* iff = proj->in(0)->as_If();
1667 IdealLoopTree *loop = get_loop(proj);
1668 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1669 int ddepth = dom_depth(proj);
1671 _igvn.rehash_node_delayed(iff);
1672 _igvn.rehash_node_delayed(proj);
1674 proj->set_req(0, NULL); // temporary disconnect
1675 ProjNode* proj2 = proj_clone(proj, iff);
1676 register_node(proj2, loop, iff, ddepth);
1678 Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right);
1679 register_node(cmp, loop, proj2, ddepth);
1681 BoolNode* bol = new (C,2)BoolNode(cmp, relop);
1682 register_node(bol, loop, proj2, ddepth);
1684 IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt);
1685 register_node(new_if, loop, proj2, ddepth);
1687 proj->set_req(0, new_if); // reattach
1688 set_idom(proj, new_if, ddepth);
1690 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
1691 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
1693 return new_exit;
1694 }
1696 //------------------------------ insert_region_before_proj -------------------------------------
1697 // Insert a region before an if projection (* - new node)
1698 //
1699 // before
1700 // if(test)
1701 // / |
1702 // v |
1703 // proj v
1704 // other-proj
1705 //
1706 // after
1707 // if(test)
1708 // / |
1709 // v |
1710 // * proj-clone v
1711 // | other-proj
1712 // v
1713 // * new-region
1714 // |
1715 // v
1716 // * dum_if
1717 // / \
1718 // v \
1719 // * dum-proj v
1720 // proj
1721 //
1722 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
1723 IfNode* iff = proj->in(0)->as_If();
1724 IdealLoopTree *loop = get_loop(proj);
1725 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1726 int ddepth = dom_depth(proj);
1728 _igvn.rehash_node_delayed(iff);
1729 _igvn.rehash_node_delayed(proj);
1731 proj->set_req(0, NULL); // temporary disconnect
1732 ProjNode* proj2 = proj_clone(proj, iff);
1733 register_node(proj2, loop, iff, ddepth);
1735 RegionNode* reg = new (C,2)RegionNode(2);
1736 reg->set_req(1, proj2);
1737 register_node(reg, loop, iff, ddepth);
1739 IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
1740 register_node(dum_if, loop, reg, ddepth);
1742 proj->set_req(0, dum_if); // reattach
1743 set_idom(proj, dum_if, ddepth);
1745 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
1746 register_node(dum_proj, loop, dum_if, ddepth);
1748 return reg;
1749 }
1751 //------------------------------ insert_cmpi_loop_exit -------------------------------------
1752 // Clone a signed compare loop exit from an unsigned compare and
1753 // insert it before the unsigned cmp on the stay-in-loop path.
1754 // All new nodes inserted in the dominator tree between the original
1755 // if and it's projections. The original if test is replaced with
1756 // a constant to force the stay-in-loop path.
1757 //
1758 // This is done to make sure that the original if and it's projections
1759 // still dominate the same set of control nodes, that the ctrl() relation
1760 // from data nodes to them is preserved, and that their loop nesting is
1761 // preserved.
1762 //
1763 // before
1764 // if(i <u limit) unsigned compare loop exit
1765 // / |
1766 // v v
1767 // exit-proj stay-in-loop-proj
1768 //
1769 // after
1770 // if(stay-in-loop-const) original if
1771 // / |
1772 // / v
1773 // / if(i < limit) new signed test
1774 // / / |
1775 // / / v
1776 // / / if(i <u limit) new cloned unsigned test
1777 // / / / |
1778 // v v v |
1779 // region |
1780 // | |
1781 // dum-if |
1782 // / | |
1783 // ether | |
1784 // v v
1785 // exit-proj stay-in-loop-proj
1786 //
1787 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
1788 const bool Signed = true;
1789 const bool Unsigned = false;
1791 BoolNode* bol = if_cmpu->in(1)->as_Bool();
1792 if (bol->_test._test != BoolTest::lt) return NULL;
1793 CmpNode* cmpu = bol->in(1)->as_Cmp();
1794 if (cmpu->Opcode() != Op_CmpU) return NULL;
1795 int stride = stride_of_possible_iv(if_cmpu);
1796 if (stride == 0) return NULL;
1798 ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj();
1799 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
1801 Node* limit = NULL;
1802 if (stride > 0) {
1803 limit = cmpu->in(2);
1804 } else {
1805 limit = _igvn.makecon(TypeInt::ZERO);
1806 set_ctrl(limit, C->root());
1807 }
1808 // Create a new region on the exit path
1809 RegionNode* reg = insert_region_before_proj(lp_exit);
1811 // Clone the if-cmpu-true-false using a signed compare
1812 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
1813 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
1814 reg->add_req(cmpi_exit);
1816 // Clone the if-cmpu-true-false
1817 BoolTest::mask rel_u = bol->_test._test;
1818 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
1819 reg->add_req(cmpu_exit);
1821 // Force original if to stay in loop.
1822 short_circuit_if(if_cmpu, lp_continue);
1824 return cmpi_exit->in(0)->as_If();
1825 }
1827 //------------------------------ remove_cmpi_loop_exit -------------------------------------
1828 // Remove a previously inserted signed compare loop exit.
1829 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
1830 Node* lp_proj = stay_in_loop(if_cmp, loop);
1831 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
1832 stay_in_loop(lp_proj, loop)->is_If() &&
1833 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
1834 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
1835 set_ctrl(con, C->root());
1836 if_cmp->set_req(1, con);
1837 }
1839 //------------------------------ scheduled_nodelist -------------------------------------
1840 // Create a post order schedule of nodes that are in the
1841 // "member" set. The list is returned in "sched".
1842 // The first node in "sched" is the loop head, followed by
1843 // nodes which have no inputs in the "member" set, and then
1844 // followed by the nodes that have an immediate input dependence
1845 // on a node in "sched".
1846 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
1848 assert(member.test(loop->_head->_idx), "loop head must be in member set");
1849 Arena *a = Thread::current()->resource_area();
1850 VectorSet visited(a);
1851 Node_Stack nstack(a, loop->_body.size());
1853 Node* n = loop->_head; // top of stack is cached in "n"
1854 uint idx = 0;
1855 visited.set(n->_idx);
1857 // Initially push all with no inputs from within member set
1858 for(uint i = 0; i < loop->_body.size(); i++ ) {
1859 Node *elt = loop->_body.at(i);
1860 if (member.test(elt->_idx)) {
1861 bool found = false;
1862 for (uint j = 0; j < elt->req(); j++) {
1863 Node* def = elt->in(j);
1864 if (def && member.test(def->_idx) && def != elt) {
1865 found = true;
1866 break;
1867 }
1868 }
1869 if (!found && elt != loop->_head) {
1870 nstack.push(n, idx);
1871 n = elt;
1872 assert(!visited.test(n->_idx), "not seen yet");
1873 visited.set(n->_idx);
1874 }
1875 }
1876 }
1878 // traverse out's that are in the member set
1879 while (true) {
1880 if (idx < n->outcnt()) {
1881 Node* use = n->raw_out(idx);
1882 idx++;
1883 if (!visited.test_set(use->_idx)) {
1884 if (member.test(use->_idx)) {
1885 nstack.push(n, idx);
1886 n = use;
1887 idx = 0;
1888 }
1889 }
1890 } else {
1891 // All outputs processed
1892 sched.push(n);
1893 if (nstack.is_empty()) break;
1894 n = nstack.node();
1895 idx = nstack.index();
1896 nstack.pop();
1897 }
1898 }
1899 }
1902 //------------------------------ has_use_in_set -------------------------------------
1903 // Has a use in the vector set
1904 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
1905 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1906 Node* use = n->fast_out(j);
1907 if (vset.test(use->_idx)) {
1908 return true;
1909 }
1910 }
1911 return false;
1912 }
1915 //------------------------------ has_use_internal_to_set -------------------------------------
1916 // Has use internal to the vector set (ie. not in a phi at the loop head)
1917 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
1918 Node* head = loop->_head;
1919 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1920 Node* use = n->fast_out(j);
1921 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
1922 return true;
1923 }
1924 }
1925 return false;
1926 }
1929 //------------------------------ clone_for_use_outside_loop -------------------------------------
1930 // clone "n" for uses that are outside of loop
1931 void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
1933 assert(worklist.size() == 0, "should be empty");
1934 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1935 Node* use = n->fast_out(j);
1936 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
1937 worklist.push(use);
1938 }
1939 }
1940 while( worklist.size() ) {
1941 Node *use = worklist.pop();
1942 if (!has_node(use) || use->in(0) == C->top()) continue;
1943 uint j;
1944 for (j = 0; j < use->req(); j++) {
1945 if (use->in(j) == n) break;
1946 }
1947 assert(j < use->req(), "must be there");
1949 // clone "n" and insert it between the inputs of "n" and the use outside the loop
1950 Node* n_clone = n->clone();
1951 _igvn.replace_input_of(use, j, n_clone);
1952 Node* use_c;
1953 if (!use->is_Phi()) {
1954 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
1955 } else {
1956 // Use in a phi is considered a use in the associated predecessor block
1957 use_c = use->in(0)->in(j);
1958 }
1959 set_ctrl(n_clone, use_c);
1960 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
1961 get_loop(use_c)->_body.push(n_clone);
1962 _igvn.register_new_node_with_optimizer(n_clone);
1963 #if !defined(PRODUCT)
1964 if (TracePartialPeeling) {
1965 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
1966 }
1967 #endif
1968 }
1969 }
1972 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
1973 // clone "n" for special uses that are in the not_peeled region.
1974 // If these def-uses occur in separate blocks, the code generator
1975 // marks the method as not compilable. For example, if a "BoolNode"
1976 // is in a different basic block than the "IfNode" that uses it, then
1977 // the compilation is aborted in the code generator.
1978 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1979 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
1980 if (n->is_Phi() || n->is_Load()) {
1981 return;
1982 }
1983 assert(worklist.size() == 0, "should be empty");
1984 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1985 Node* use = n->fast_out(j);
1986 if ( not_peel.test(use->_idx) &&
1987 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
1988 use->in(1) == n) {
1989 worklist.push(use);
1990 }
1991 }
1992 if (worklist.size() > 0) {
1993 // clone "n" and insert it between inputs of "n" and the use
1994 Node* n_clone = n->clone();
1995 loop->_body.push(n_clone);
1996 _igvn.register_new_node_with_optimizer(n_clone);
1997 set_ctrl(n_clone, get_ctrl(n));
1998 sink_list.push(n_clone);
1999 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
2000 #if !defined(PRODUCT)
2001 if (TracePartialPeeling) {
2002 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2003 }
2004 #endif
2005 while( worklist.size() ) {
2006 Node *use = worklist.pop();
2007 _igvn.rehash_node_delayed(use);
2008 for (uint j = 1; j < use->req(); j++) {
2009 if (use->in(j) == n) {
2010 use->set_req(j, n_clone);
2011 }
2012 }
2013 }
2014 }
2015 }
2018 //------------------------------ insert_phi_for_loop -------------------------------------
2019 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2020 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2021 Node *phi = PhiNode::make(lp, back_edge_val);
2022 phi->set_req(LoopNode::EntryControl, lp_entry_val);
2023 // Use existing phi if it already exists
2024 Node *hit = _igvn.hash_find_insert(phi);
2025 if( hit == NULL ) {
2026 _igvn.register_new_node_with_optimizer(phi);
2027 set_ctrl(phi, lp);
2028 } else {
2029 // Remove the new phi from the graph and use the hit
2030 _igvn.remove_dead_node(phi);
2031 phi = hit;
2032 }
2033 _igvn.replace_input_of(use, idx, phi);
2034 }
2036 #ifdef ASSERT
2037 //------------------------------ is_valid_loop_partition -------------------------------------
2038 // Validate the loop partition sets: peel and not_peel
2039 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2040 VectorSet& not_peel ) {
2041 uint i;
2042 // Check that peel_list entries are in the peel set
2043 for (i = 0; i < peel_list.size(); i++) {
2044 if (!peel.test(peel_list.at(i)->_idx)) {
2045 return false;
2046 }
2047 }
2048 // Check at loop members are in one of peel set or not_peel set
2049 for (i = 0; i < loop->_body.size(); i++ ) {
2050 Node *def = loop->_body.at(i);
2051 uint di = def->_idx;
2052 // Check that peel set elements are in peel_list
2053 if (peel.test(di)) {
2054 if (not_peel.test(di)) {
2055 return false;
2056 }
2057 // Must be in peel_list also
2058 bool found = false;
2059 for (uint j = 0; j < peel_list.size(); j++) {
2060 if (peel_list.at(j)->_idx == di) {
2061 found = true;
2062 break;
2063 }
2064 }
2065 if (!found) {
2066 return false;
2067 }
2068 } else if (not_peel.test(di)) {
2069 if (peel.test(di)) {
2070 return false;
2071 }
2072 } else {
2073 return false;
2074 }
2075 }
2076 return true;
2077 }
2079 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2080 // Ensure a use outside of loop is of the right form
2081 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2082 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2083 return (use->is_Phi() &&
2084 use_c->is_Region() && use_c->req() == 3 &&
2085 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2086 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2087 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2088 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2089 }
2091 //------------------------------ is_valid_clone_loop_form -------------------------------------
2092 // Ensure that all uses outside of loop are of the right form
2093 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2094 uint orig_exit_idx, uint clone_exit_idx) {
2095 uint len = peel_list.size();
2096 for (uint i = 0; i < len; i++) {
2097 Node *def = peel_list.at(i);
2099 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2100 Node *use = def->fast_out(j);
2101 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2102 if (!loop->is_member(get_loop(use_c))) {
2103 // use is not in the loop, check for correct structure
2104 if (use->in(0) == def) {
2105 // Okay
2106 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2107 return false;
2108 }
2109 }
2110 }
2111 }
2112 return true;
2113 }
2114 #endif
2116 //------------------------------ partial_peel -------------------------------------
2117 // Partially peel (aka loop rotation) the top portion of a loop (called
2118 // the peel section below) by cloning it and placing one copy just before
2119 // the new loop head and the other copy at the bottom of the new loop.
2120 //
2121 // before after where it came from
2122 //
2123 // stmt1 stmt1
2124 // loop: stmt2 clone
2125 // stmt2 if condA goto exitA clone
2126 // if condA goto exitA new_loop: new
2127 // stmt3 stmt3 clone
2128 // if !condB goto loop if condB goto exitB clone
2129 // exitB: stmt2 orig
2130 // stmt4 if !condA goto new_loop orig
2131 // exitA: goto exitA
2132 // exitB:
2133 // stmt4
2134 // exitA:
2135 //
2136 // Step 1: find the cut point: an exit test on probable
2137 // induction variable.
2138 // Step 2: schedule (with cloning) operations in the peel
2139 // section that can be executed after the cut into
2140 // the section that is not peeled. This may need
2141 // to clone operations into exit blocks. For
2142 // instance, a reference to A[i] in the not-peel
2143 // section and a reference to B[i] in an exit block
2144 // may cause a left-shift of i by 2 to be placed
2145 // in the peel block. This step will clone the left
2146 // shift into the exit block and sink the left shift
2147 // from the peel to the not-peel section.
2148 // Step 3: clone the loop, retarget the control, and insert
2149 // phis for values that are live across the new loop
2150 // head. This is very dependent on the graph structure
2151 // from clone_loop. It creates region nodes for
2152 // exit control and associated phi nodes for values
2153 // flow out of the loop through that exit. The region
2154 // node is dominated by the clone's control projection.
2155 // So the clone's peel section is placed before the
2156 // new loop head, and the clone's not-peel section is
2157 // forms the top part of the new loop. The original
2158 // peel section forms the tail of the new loop.
2159 // Step 4: update the dominator tree and recompute the
2160 // dominator depth.
2161 //
2162 // orig
2163 //
2164 // stmt1
2165 // |
2166 // v
2167 // loop predicate
2168 // |
2169 // v
2170 // loop<----+
2171 // | |
2172 // stmt2 |
2173 // | |
2174 // v |
2175 // ifA |
2176 // / | |
2177 // v v |
2178 // false true ^ <-- last_peel
2179 // / | |
2180 // / ===|==cut |
2181 // / stmt3 | <-- first_not_peel
2182 // / | |
2183 // | v |
2184 // v ifB |
2185 // exitA: / \ |
2186 // / \ |
2187 // v v |
2188 // false true |
2189 // / \ |
2190 // / ----+
2191 // |
2192 // v
2193 // exitB:
2194 // stmt4
2195 //
2196 //
2197 // after clone loop
2198 //
2199 // stmt1
2200 // |
2201 // v
2202 // loop predicate
2203 // / \
2204 // clone / \ orig
2205 // / \
2206 // / \
2207 // v v
2208 // +---->loop loop<----+
2209 // | | | |
2210 // | stmt2 stmt2 |
2211 // | | | |
2212 // | v v |
2213 // | ifA ifA |
2214 // | | \ / | |
2215 // | v v v v |
2216 // ^ true false false true ^ <-- last_peel
2217 // | | ^ \ / | |
2218 // | cut==|== \ \ / ===|==cut |
2219 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2220 // | | dom | | | |
2221 // | v \ 1v v2 v |
2222 // | ifB regionA ifB |
2223 // | / \ | / \ |
2224 // | / \ v / \ |
2225 // | v v exitA: v v |
2226 // | true false false true |
2227 // | / ^ \ / \ |
2228 // +---- \ \ / ----+
2229 // dom \ /
2230 // \ 1v v2
2231 // regionB
2232 // |
2233 // v
2234 // exitB:
2235 // stmt4
2236 //
2237 //
2238 // after partial peel
2239 //
2240 // stmt1
2241 // |
2242 // v
2243 // loop predicate
2244 // /
2245 // clone / orig
2246 // / TOP
2247 // / \
2248 // v v
2249 // TOP->loop loop----+
2250 // | | |
2251 // stmt2 stmt2 |
2252 // | | |
2253 // v v |
2254 // ifA ifA |
2255 // | \ / | |
2256 // v v v v |
2257 // true false false true | <-- last_peel
2258 // | ^ \ / +------|---+
2259 // +->newloop \ \ / === ==cut | |
2260 // | stmt3 \ \ / TOP | |
2261 // | | dom | | stmt3 | | <-- first_not_peel
2262 // | v \ 1v v2 v | |
2263 // | ifB regionA ifB ^ v
2264 // | / \ | / \ | |
2265 // | / \ v / \ | |
2266 // | v v exitA: v v | |
2267 // | true false false true | |
2268 // | / ^ \ / \ | |
2269 // | | \ \ / v | |
2270 // | | dom \ / TOP | |
2271 // | | \ 1v v2 | |
2272 // ^ v regionB | |
2273 // | | | | |
2274 // | | v ^ v
2275 // | | exitB: | |
2276 // | | stmt4 | |
2277 // | +------------>-----------------+ |
2278 // | |
2279 // +-----------------<---------------------+
2280 //
2281 //
2282 // final graph
2283 //
2284 // stmt1
2285 // |
2286 // v
2287 // loop predicate
2288 // |
2289 // v
2290 // stmt2 clone
2291 // |
2292 // v
2293 // ........> ifA clone
2294 // : / |
2295 // dom / |
2296 // : v v
2297 // : false true
2298 // : | |
2299 // : | v
2300 // : | newloop<-----+
2301 // : | | |
2302 // : | stmt3 clone |
2303 // : | | |
2304 // : | v |
2305 // : | ifB |
2306 // : | / \ |
2307 // : | v v |
2308 // : | false true |
2309 // : | | | |
2310 // : | v stmt2 |
2311 // : | exitB: | |
2312 // : | stmt4 v |
2313 // : | ifA orig |
2314 // : | / \ |
2315 // : | / \ |
2316 // : | v v |
2317 // : | false true |
2318 // : | / \ |
2319 // : v v -----+
2320 // RegionA
2321 // |
2322 // v
2323 // exitA
2324 //
2325 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
2327 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
2328 if (!loop->_head->is_Loop()) {
2329 return false; }
2331 LoopNode *head = loop->_head->as_Loop();
2333 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
2334 return false;
2335 }
2337 // Check for complex exit control
2338 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
2339 Node *n = loop->_body.at(ii);
2340 int opc = n->Opcode();
2341 if (n->is_Call() ||
2342 opc == Op_Catch ||
2343 opc == Op_CatchProj ||
2344 opc == Op_Jump ||
2345 opc == Op_JumpProj) {
2346 #if !defined(PRODUCT)
2347 if (TracePartialPeeling) {
2348 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
2349 }
2350 #endif
2351 return false;
2352 }
2353 }
2355 int dd = dom_depth(head);
2357 // Step 1: find cut point
2359 // Walk up dominators to loop head looking for first loop exit
2360 // which is executed on every path thru loop.
2361 IfNode *peel_if = NULL;
2362 IfNode *peel_if_cmpu = NULL;
2364 Node *iff = loop->tail();
2365 while( iff != head ) {
2366 if( iff->is_If() ) {
2367 Node *ctrl = get_ctrl(iff->in(1));
2368 if (ctrl->is_top()) return false; // Dead test on live IF.
2369 // If loop-varying exit-test, check for induction variable
2370 if( loop->is_member(get_loop(ctrl)) &&
2371 loop->is_loop_exit(iff) &&
2372 is_possible_iv_test(iff)) {
2373 Node* cmp = iff->in(1)->in(1);
2374 if (cmp->Opcode() == Op_CmpI) {
2375 peel_if = iff->as_If();
2376 } else {
2377 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
2378 peel_if_cmpu = iff->as_If();
2379 }
2380 }
2381 }
2382 iff = idom(iff);
2383 }
2384 // Prefer signed compare over unsigned compare.
2385 IfNode* new_peel_if = NULL;
2386 if (peel_if == NULL) {
2387 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
2388 return false; // No peel point found
2389 }
2390 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
2391 if (new_peel_if == NULL) {
2392 return false; // No peel point found
2393 }
2394 peel_if = new_peel_if;
2395 }
2396 Node* last_peel = stay_in_loop(peel_if, loop);
2397 Node* first_not_peeled = stay_in_loop(last_peel, loop);
2398 if (first_not_peeled == NULL || first_not_peeled == head) {
2399 return false;
2400 }
2402 #if !defined(PRODUCT)
2403 if (TraceLoopOpts) {
2404 tty->print("PartialPeel ");
2405 loop->dump_head();
2406 }
2408 if (TracePartialPeeling) {
2409 tty->print_cr("before partial peel one iteration");
2410 Node_List wl;
2411 Node* t = head->in(2);
2412 while (true) {
2413 wl.push(t);
2414 if (t == head) break;
2415 t = idom(t);
2416 }
2417 while (wl.size() > 0) {
2418 Node* tt = wl.pop();
2419 tt->dump();
2420 if (tt == last_peel) tty->print_cr("-- cut --");
2421 }
2422 }
2423 #endif
2424 ResourceArea *area = Thread::current()->resource_area();
2425 VectorSet peel(area);
2426 VectorSet not_peel(area);
2427 Node_List peel_list(area);
2428 Node_List worklist(area);
2429 Node_List sink_list(area);
2431 // Set of cfg nodes to peel are those that are executable from
2432 // the head through last_peel.
2433 assert(worklist.size() == 0, "should be empty");
2434 worklist.push(head);
2435 peel.set(head->_idx);
2436 while (worklist.size() > 0) {
2437 Node *n = worklist.pop();
2438 if (n != last_peel) {
2439 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2440 Node* use = n->fast_out(j);
2441 if (use->is_CFG() &&
2442 loop->is_member(get_loop(use)) &&
2443 !peel.test_set(use->_idx)) {
2444 worklist.push(use);
2445 }
2446 }
2447 }
2448 }
2450 // Set of non-cfg nodes to peel are those that are control
2451 // dependent on the cfg nodes.
2452 uint i;
2453 for(i = 0; i < loop->_body.size(); i++ ) {
2454 Node *n = loop->_body.at(i);
2455 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
2456 if (peel.test(n_c->_idx)) {
2457 peel.set(n->_idx);
2458 } else {
2459 not_peel.set(n->_idx);
2460 }
2461 }
2463 // Step 2: move operations from the peeled section down into the
2464 // not-peeled section
2466 // Get a post order schedule of nodes in the peel region
2467 // Result in right-most operand.
2468 scheduled_nodelist(loop, peel, peel_list );
2470 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2472 // For future check for too many new phis
2473 uint old_phi_cnt = 0;
2474 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2475 Node* use = head->fast_out(j);
2476 if (use->is_Phi()) old_phi_cnt++;
2477 }
2479 #if !defined(PRODUCT)
2480 if (TracePartialPeeling) {
2481 tty->print_cr("\npeeled list");
2482 }
2483 #endif
2485 // Evacuate nodes in peel region into the not_peeled region if possible
2486 uint new_phi_cnt = 0;
2487 for (i = 0; i < peel_list.size();) {
2488 Node* n = peel_list.at(i);
2489 #if !defined(PRODUCT)
2490 if (TracePartialPeeling) n->dump();
2491 #endif
2492 bool incr = true;
2493 if ( !n->is_CFG() ) {
2495 if ( has_use_in_set(n, not_peel) ) {
2497 // If not used internal to the peeled region,
2498 // move "n" from peeled to not_peeled region.
2500 if ( !has_use_internal_to_set(n, peel, loop) ) {
2502 // if not pinned and not a load (which maybe anti-dependent on a store)
2503 // and not a CMove (Matcher expects only bool->cmove).
2504 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
2505 clone_for_use_outside_loop( loop, n, worklist );
2507 sink_list.push(n);
2508 peel >>= n->_idx; // delete n from peel set.
2509 not_peel <<= n->_idx; // add n to not_peel set.
2510 peel_list.remove(i);
2511 incr = false;
2512 #if !defined(PRODUCT)
2513 if (TracePartialPeeling) {
2514 tty->print_cr("sink to not_peeled region: %d newbb: %d",
2515 n->_idx, get_ctrl(n)->_idx);
2516 }
2517 #endif
2518 }
2519 } else {
2520 // Otherwise check for special def-use cases that span
2521 // the peel/not_peel boundary such as bool->if
2522 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
2523 new_phi_cnt++;
2524 }
2525 }
2526 }
2527 if (incr) i++;
2528 }
2530 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
2531 #if !defined(PRODUCT)
2532 if (TracePartialPeeling) {
2533 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
2534 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
2535 }
2536 #endif
2537 if (new_peel_if != NULL) {
2538 remove_cmpi_loop_exit(new_peel_if, loop);
2539 }
2540 // Inhibit more partial peeling on this loop
2541 assert(!head->is_partial_peel_loop(), "not partial peeled");
2542 head->mark_partial_peel_failed();
2543 return false;
2544 }
2546 // Step 3: clone loop, retarget control, and insert new phis
2548 // Create new loop head for new phis and to hang
2549 // the nodes being moved (sinked) from the peel region.
2550 LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel);
2551 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
2552 _igvn.register_new_node_with_optimizer(new_head);
2553 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
2554 first_not_peeled->set_req(0, new_head);
2555 set_loop(new_head, loop);
2556 loop->_body.push(new_head);
2557 not_peel.set(new_head->_idx);
2558 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
2559 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
2561 while (sink_list.size() > 0) {
2562 Node* n = sink_list.pop();
2563 set_ctrl(n, new_head);
2564 }
2566 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2568 clone_loop( loop, old_new, dd );
2570 const uint clone_exit_idx = 1;
2571 const uint orig_exit_idx = 2;
2572 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
2574 Node* head_clone = old_new[head->_idx];
2575 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
2576 Node* orig_tail_clone = head_clone->in(2);
2578 // Add phi if "def" node is in peel set and "use" is not
2580 for(i = 0; i < peel_list.size(); i++ ) {
2581 Node *def = peel_list.at(i);
2582 if (!def->is_CFG()) {
2583 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2584 Node *use = def->fast_out(j);
2585 if (has_node(use) && use->in(0) != C->top() &&
2586 (!peel.test(use->_idx) ||
2587 (use->is_Phi() && use->in(0) == head)) ) {
2588 worklist.push(use);
2589 }
2590 }
2591 while( worklist.size() ) {
2592 Node *use = worklist.pop();
2593 for (uint j = 1; j < use->req(); j++) {
2594 Node* n = use->in(j);
2595 if (n == def) {
2597 // "def" is in peel set, "use" is not in peel set
2598 // or "use" is in the entry boundary (a phi) of the peel set
2600 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
2602 if ( loop->is_member(get_loop( use_c )) ) {
2603 // use is in loop
2604 if (old_new[use->_idx] != NULL) { // null for dead code
2605 Node* use_clone = old_new[use->_idx];
2606 _igvn.replace_input_of(use, j, C->top());
2607 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
2608 }
2609 } else {
2610 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
2611 // use is not in the loop, check if the live range includes the cut
2612 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
2613 if (not_peel.test(lp_if->_idx)) {
2614 assert(j == orig_exit_idx, "use from original loop");
2615 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
2616 }
2617 }
2618 }
2619 }
2620 }
2621 }
2622 }
2624 // Step 3b: retarget control
2626 // Redirect control to the new loop head if a cloned node in
2627 // the not_peeled region has control that points into the peeled region.
2628 // This necessary because the cloned peeled region will be outside
2629 // the loop.
2630 // from to
2631 // cloned-peeled <---+
2632 // new_head_clone: | <--+
2633 // cloned-not_peeled in(0) in(0)
2634 // orig-peeled
2636 for(i = 0; i < loop->_body.size(); i++ ) {
2637 Node *n = loop->_body.at(i);
2638 if (!n->is_CFG() && n->in(0) != NULL &&
2639 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
2640 Node* n_clone = old_new[n->_idx];
2641 _igvn.replace_input_of(n_clone, 0, new_head_clone);
2642 }
2643 }
2645 // Backedge of the surviving new_head (the clone) is original last_peel
2646 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
2648 // Cut first node in original not_peel set
2649 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
2650 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
2651 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
2653 // Copy head_clone back-branch info to original head
2654 // and remove original head's loop entry and
2655 // clone head's back-branch
2656 _igvn.rehash_node_delayed(head); // Multiple edge updates
2657 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
2658 head->set_req(LoopNode::LoopBackControl, C->top());
2659 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
2661 // Similarly modify the phis
2662 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
2663 Node* use = head->fast_out(k);
2664 if (use->is_Phi() && use->outcnt() > 0) {
2665 Node* use_clone = old_new[use->_idx];
2666 _igvn.rehash_node_delayed(use); // Multiple edge updates
2667 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
2668 use->set_req(LoopNode::LoopBackControl, C->top());
2669 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
2670 }
2671 }
2673 // Step 4: update dominator tree and dominator depth
2675 set_idom(head, orig_tail_clone, dd);
2676 recompute_dom_depth();
2678 // Inhibit more partial peeling on this loop
2679 new_head_clone->set_partial_peel_loop();
2680 C->set_major_progress();
2682 #if !defined(PRODUCT)
2683 if (TracePartialPeeling) {
2684 tty->print_cr("\nafter partial peel one iteration");
2685 Node_List wl(area);
2686 Node* t = last_peel;
2687 while (true) {
2688 wl.push(t);
2689 if (t == head_clone) break;
2690 t = idom(t);
2691 }
2692 while (wl.size() > 0) {
2693 Node* tt = wl.pop();
2694 if (tt == head) tty->print_cr("orig head");
2695 else if (tt == new_head_clone) tty->print_cr("new head");
2696 else if (tt == head_clone) tty->print_cr("clone head");
2697 tt->dump();
2698 }
2699 }
2700 #endif
2701 return true;
2702 }
2704 //------------------------------reorg_offsets----------------------------------
2705 // Reorganize offset computations to lower register pressure. Mostly
2706 // prevent loop-fallout uses of the pre-incremented trip counter (which are
2707 // then alive with the post-incremented trip counter forcing an extra
2708 // register move)
2709 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
2710 // Perform it only for canonical counted loops.
2711 // Loop's shape could be messed up by iteration_split_impl.
2712 if (!loop->_head->is_CountedLoop())
2713 return;
2714 if (!loop->_head->as_Loop()->is_valid_counted_loop())
2715 return;
2717 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2718 CountedLoopEndNode *cle = cl->loopexit();
2719 Node *exit = cle->proj_out(false);
2720 Node *phi = cl->phi();
2722 // Check for the special case of folks using the pre-incremented
2723 // trip-counter on the fall-out path (forces the pre-incremented
2724 // and post-incremented trip counter to be live at the same time).
2725 // Fix this by adjusting to use the post-increment trip counter.
2727 bool progress = true;
2728 while (progress) {
2729 progress = false;
2730 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
2731 Node* use = phi->fast_out(i); // User of trip-counter
2732 if (!has_ctrl(use)) continue;
2733 Node *u_ctrl = get_ctrl(use);
2734 if (use->is_Phi()) {
2735 u_ctrl = NULL;
2736 for (uint j = 1; j < use->req(); j++)
2737 if (use->in(j) == phi)
2738 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
2739 }
2740 IdealLoopTree *u_loop = get_loop(u_ctrl);
2741 // Look for loop-invariant use
2742 if (u_loop == loop) continue;
2743 if (loop->is_member(u_loop)) continue;
2744 // Check that use is live out the bottom. Assuming the trip-counter
2745 // update is right at the bottom, uses of of the loop middle are ok.
2746 if (dom_lca(exit, u_ctrl) != exit) continue;
2747 // Hit! Refactor use to use the post-incremented tripcounter.
2748 // Compute a post-increment tripcounter.
2749 Node *opaq = new (C, 2) Opaque2Node( C, cle->incr() );
2750 register_new_node( opaq, u_ctrl );
2751 Node *neg_stride = _igvn.intcon(-cle->stride_con());
2752 set_ctrl(neg_stride, C->root());
2753 Node *post = new (C, 3) AddINode( opaq, neg_stride);
2754 register_new_node( post, u_ctrl );
2755 _igvn.rehash_node_delayed(use);
2756 for (uint j = 1; j < use->req(); j++) {
2757 if (use->in(j) == phi)
2758 use->set_req(j, post);
2759 }
2760 // Since DU info changed, rerun loop
2761 progress = true;
2762 break;
2763 }
2764 }
2766 }