1 /*
2 * Copyright (c) 1999, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "memory/allocation.inline.hpp"
27 #include "memory/resourceArea.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/castnode.hpp"
31 #include "opto/connode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/loopnode.hpp"
35 #include "opto/matcher.hpp"
36 #include "opto/mulnode.hpp"
37 #include "opto/movenode.hpp"
38 #include "opto/opaquenode.hpp"
39 #include "opto/rootnode.hpp"
40 #include "opto/subnode.hpp"
41
42 //=============================================================================
43 //------------------------------split_thru_phi---------------------------------
44 // Split Node 'n' through merge point if there is enough win.
45 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
46 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
47 // ConvI2L may have type information on it which is unsafe to push up
48 // so disable this for now
49 return NULL;
50 }
51
52 // Splitting range check CastIIs through a loop induction Phi can
53 // cause new Phis to be created that are left unrelated to the loop
54 // induction Phi and prevent optimizations (vectorization)
55 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() &&
56 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) {
57 return NULL;
58 }
59
60 int wins = 0;
61 assert(!n->is_CFG(), "");
62 assert(region->is_Region(), "");
63
64 const Type* type = n->bottom_type();
65 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr();
66 Node *phi;
67 if (t_oop != NULL && t_oop->is_known_instance_field()) {
68 int iid = t_oop->instance_id();
69 int index = C->get_alias_index(t_oop);
70 int offset = t_oop->offset();
71 phi = new PhiNode(region, type, NULL, iid, index, offset);
72 } else {
73 phi = PhiNode::make_blank(region, n);
74 }
75 uint old_unique = C->unique();
76 for (uint i = 1; i < region->req(); i++) {
77 Node *x;
78 Node* the_clone = NULL;
79 if (region->in(i) == C->top()) {
80 x = C->top(); // Dead path? Use a dead data op
81 } else {
82 x = n->clone(); // Else clone up the data op
83 the_clone = x; // Remember for possible deletion.
84 // Alter data node to use pre-phi inputs
85 if (n->in(0) == region)
86 x->set_req( 0, region->in(i) );
87 for (uint j = 1; j < n->req(); j++) {
88 Node *in = n->in(j);
89 if (in->is_Phi() && in->in(0) == region)
90 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
91 }
92 }
93 // Check for a 'win' on some paths
94 const Type *t = x->Value(&_igvn);
95
96 bool singleton = t->singleton();
97
98 // A TOP singleton indicates that there are no possible values incoming
99 // along a particular edge. In most cases, this is OK, and the Phi will
100 // be eliminated later in an Ideal call. However, we can't allow this to
101 // happen if the singleton occurs on loop entry, as the elimination of
102 // the PhiNode may cause the resulting node to migrate back to a previous
103 // loop iteration.
104 if (singleton && t == Type::TOP) {
105 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
106 // irreducible loop may not be indicated by an affirmative is_Loop());
107 // therefore, the only top we can split thru a phi is on a backedge of
108 // a loop.
109 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
110 }
111
112 if (singleton) {
113 wins++;
114 x = ((PhaseGVN&)_igvn).makecon(t);
115 } else {
116 // We now call Identity to try to simplify the cloned node.
117 // Note that some Identity methods call phase->type(this).
118 // Make sure that the type array is big enough for
119 // our new node, even though we may throw the node away.
120 // (Note: This tweaking with igvn only works because x is a new node.)
121 _igvn.set_type(x, t);
122 // If x is a TypeNode, capture any more-precise type permanently into Node
123 // otherwise it will be not updated during igvn->transform since
124 // igvn->type(x) is set to x->Value() already.
125 x->raise_bottom_type(t);
126 Node *y = x->Identity(&_igvn);
127 if (y != x) {
128 wins++;
129 x = y;
130 } else {
131 y = _igvn.hash_find(x);
132 if (y) {
133 wins++;
134 x = y;
135 } else {
136 // Else x is a new node we are keeping
137 // We do not need register_new_node_with_optimizer
138 // because set_type has already been called.
139 _igvn._worklist.push(x);
140 }
141 }
142 }
143 if (x != the_clone && the_clone != NULL)
144 _igvn.remove_dead_node(the_clone);
145 phi->set_req( i, x );
146 }
147 // Too few wins?
148 if (wins <= policy) {
149 _igvn.remove_dead_node(phi);
150 return NULL;
151 }
152
153 // Record Phi
154 register_new_node( phi, region );
155
156 for (uint i2 = 1; i2 < phi->req(); i2++) {
157 Node *x = phi->in(i2);
158 // If we commoned up the cloned 'x' with another existing Node,
159 // the existing Node picks up a new use. We need to make the
160 // existing Node occur higher up so it dominates its uses.
161 Node *old_ctrl;
162 IdealLoopTree *old_loop;
163
164 if (x->is_Con()) {
165 // Constant's control is always root.
166 set_ctrl(x, C->root());
167 continue;
168 }
169 // The occasional new node
170 if (x->_idx >= old_unique) { // Found a new, unplaced node?
171 old_ctrl = NULL;
172 old_loop = NULL; // Not in any prior loop
173 } else {
174 old_ctrl = get_ctrl(x);
175 old_loop = get_loop(old_ctrl); // Get prior loop
176 }
177 // New late point must dominate new use
178 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
179 if (new_ctrl == old_ctrl) // Nothing is changed
180 continue;
181
182 IdealLoopTree *new_loop = get_loop(new_ctrl);
183
184 // Don't move x into a loop if its uses are
185 // outside of loop. Otherwise x will be cloned
186 // for each use outside of this loop.
187 IdealLoopTree *use_loop = get_loop(region);
188 if (!new_loop->is_member(use_loop) &&
189 (old_loop == NULL || !new_loop->is_member(old_loop))) {
190 // Take early control, later control will be recalculated
191 // during next iteration of loop optimizations.
192 new_ctrl = get_early_ctrl(x);
193 new_loop = get_loop(new_ctrl);
194 }
195 // Set new location
196 set_ctrl(x, new_ctrl);
197 // If changing loop bodies, see if we need to collect into new body
198 if (old_loop != new_loop) {
199 if (old_loop && !old_loop->_child)
200 old_loop->_body.yank(x);
201 if (!new_loop->_child)
202 new_loop->_body.push(x); // Collect body info
203 }
204 }
205
206 return phi;
207 }
208
209 //------------------------------dominated_by------------------------------------
210 // Replace the dominated test with an obvious true or false. Place it on the
211 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
212 // live path up to the dominating control.
213 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) {
214 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
215
216 // prevdom is the dominating projection of the dominating test.
217 assert( iff->is_If(), "" );
218 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added");
219 int pop = prevdom->Opcode();
220 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
221 if (flip) {
222 if (pop == Op_IfTrue)
223 pop = Op_IfFalse;
224 else
225 pop = Op_IfTrue;
226 }
227 // 'con' is set to true or false to kill the dominated test.
228 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
229 set_ctrl(con, C->root()); // Constant gets a new use
230 // Hack the dominated test
231 _igvn.replace_input_of(iff, 1, con);
232
233 // If I dont have a reachable TRUE and FALSE path following the IfNode then
234 // I can assume this path reaches an infinite loop. In this case it's not
235 // important to optimize the data Nodes - either the whole compilation will
236 // be tossed or this path (and all data Nodes) will go dead.
237 if (iff->outcnt() != 2) return;
238
239 // Make control-dependent data Nodes on the live path (path that will remain
240 // once the dominated IF is removed) become control-dependent on the
241 // dominating projection.
242 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue);
243
244 // Loop predicates may have depending checks which should not
245 // be skipped. For example, range check predicate has two checks
246 // for lower and upper bounds.
247 if (dp == NULL)
248 return;
249
250 ProjNode* dp_proj = dp->as_Proj();
251 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj();
252 if (exclude_loop_predicate &&
253 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL ||
254 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) {
255 // If this is a range check (IfNode::is_range_check), do not
256 // reorder because Compile::allow_range_check_smearing might have
257 // changed the check.
258 return; // Let IGVN transformation change control dependence.
259 }
260
261 IdealLoopTree *old_loop = get_loop(dp);
262
263 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
264 Node* cd = dp->fast_out(i); // Control-dependent node
265 if (cd->depends_only_on_test()) {
266 assert(cd->in(0) == dp, "");
267 _igvn.replace_input_of(cd, 0, prevdom);
268 set_early_ctrl(cd);
269 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
270 if (old_loop != new_loop) {
271 if (!old_loop->_child) old_loop->_body.yank(cd);
272 if (!new_loop->_child) new_loop->_body.push(cd);
273 }
274 --i;
275 --imax;
276 }
277 }
278 }
279
280 //------------------------------has_local_phi_input----------------------------
281 // Return TRUE if 'n' has Phi inputs from its local block and no other
282 // block-local inputs (all non-local-phi inputs come from earlier blocks)
283 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
284 Node *n_ctrl = get_ctrl(n);
285 // See if some inputs come from a Phi in this block, or from before
286 // this block.
287 uint i;
288 for( i = 1; i < n->req(); i++ ) {
289 Node *phi = n->in(i);
290 if( phi->is_Phi() && phi->in(0) == n_ctrl )
291 break;
292 }
293 if( i >= n->req() )
294 return NULL; // No Phi inputs; nowhere to clone thru
295
296 // Check for inputs created between 'n' and the Phi input. These
297 // must split as well; they have already been given the chance
298 // (courtesy of a post-order visit) and since they did not we must
299 // recover the 'cost' of splitting them by being very profitable
300 // when splitting 'n'. Since this is unlikely we simply give up.
301 for( i = 1; i < n->req(); i++ ) {
302 Node *m = n->in(i);
303 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
304 // We allow the special case of AddP's with no local inputs.
305 // This allows us to split-up address expressions.
306 if (m->is_AddP() &&
307 get_ctrl(m->in(2)) != n_ctrl &&
308 get_ctrl(m->in(3)) != n_ctrl) {
309 // Move the AddP up to dominating point
310 Node* c = find_non_split_ctrl(idom(n_ctrl));
311 if (c->is_OuterStripMinedLoop()) {
312 c->as_Loop()->verify_strip_mined(1);
313 c = c->in(LoopNode::EntryControl);
314 }
315 set_ctrl_and_loop(m, c);
316 continue;
317 }
318 return NULL;
319 }
320 assert(m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
321 }
322
323 return n_ctrl;
324 }
325
326 //------------------------------remix_address_expressions----------------------
327 // Rework addressing expressions to get the most loop-invariant stuff
328 // moved out. We'd like to do all associative operators, but it's especially
329 // important (common) to do address expressions.
330 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
331 if (!has_ctrl(n)) return NULL;
332 Node *n_ctrl = get_ctrl(n);
333 IdealLoopTree *n_loop = get_loop(n_ctrl);
334
335 // See if 'n' mixes loop-varying and loop-invariant inputs and
336 // itself is loop-varying.
337
338 // Only interested in binary ops (and AddP)
339 if( n->req() < 3 || n->req() > 4 ) return NULL;
340
341 Node *n1_ctrl = get_ctrl(n->in( 1));
342 Node *n2_ctrl = get_ctrl(n->in( 2));
343 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
344 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
345 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
346 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
347
348 // Does one of my inputs spin in a tighter loop than self?
349 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
350 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
351 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
352 return NULL; // Leave well enough alone
353
354 // Is at least one of my inputs loop-invariant?
355 if( n1_loop == n_loop &&
356 n2_loop == n_loop &&
357 n3_loop == n_loop )
358 return NULL; // No loop-invariant inputs
359
360
361 int n_op = n->Opcode();
362
363 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
364 if( n_op == Op_LShiftI ) {
365 // Scale is loop invariant
366 Node *scale = n->in(2);
367 Node *scale_ctrl = get_ctrl(scale);
368 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
369 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
370 return NULL;
371 const TypeInt *scale_t = scale->bottom_type()->isa_int();
372 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
373 return NULL; // Dont bother with byte/short masking
374 // Add must vary with loop (else shift would be loop-invariant)
375 Node *add = n->in(1);
376 Node *add_ctrl = get_ctrl(add);
377 IdealLoopTree *add_loop = get_loop(add_ctrl);
378 //assert( n_loop == add_loop, "" );
379 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
380
381 // Convert I-V into I+ (0-V); same for V-I
382 if( add->Opcode() == Op_SubI &&
383 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
384 Node *zero = _igvn.intcon(0);
385 set_ctrl(zero, C->root());
386 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) );
387 register_new_node( neg, get_ctrl(add->in(2) ) );
388 add = new AddINode( add->in(1), neg );
389 register_new_node( add, add_ctrl );
390 }
391 if( add->Opcode() != Op_AddI ) return NULL;
392 // See if one add input is loop invariant
393 Node *add_var = add->in(1);
394 Node *add_var_ctrl = get_ctrl(add_var);
395 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
396 Node *add_invar = add->in(2);
397 Node *add_invar_ctrl = get_ctrl(add_invar);
398 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
399 if( add_var_loop == n_loop ) {
400 } else if( add_invar_loop == n_loop ) {
401 // Swap to find the invariant part
402 add_invar = add_var;
403 add_invar_ctrl = add_var_ctrl;
404 add_invar_loop = add_var_loop;
405 add_var = add->in(2);
406 Node *add_var_ctrl = get_ctrl(add_var);
407 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
408 } else // Else neither input is loop invariant
409 return NULL;
410 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
411 return NULL; // No invariant part of the add?
412
413 // Yes! Reshape address expression!
414 Node *inv_scale = new LShiftINode( add_invar, scale );
415 Node *inv_scale_ctrl =
416 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
417 add_invar_ctrl : scale_ctrl;
418 register_new_node( inv_scale, inv_scale_ctrl );
419 Node *var_scale = new LShiftINode( add_var, scale );
420 register_new_node( var_scale, n_ctrl );
421 Node *var_add = new AddINode( var_scale, inv_scale );
422 register_new_node( var_add, n_ctrl );
423 _igvn.replace_node( n, var_add );
424 return var_add;
425 }
426
427 // Replace (I+V) with (V+I)
428 if( n_op == Op_AddI ||
429 n_op == Op_AddL ||
430 n_op == Op_AddF ||
431 n_op == Op_AddD ||
432 n_op == Op_MulI ||
433 n_op == Op_MulL ||
434 n_op == Op_MulF ||
435 n_op == Op_MulD ) {
436 if( n2_loop == n_loop ) {
437 assert( n1_loop != n_loop, "" );
438 n->swap_edges(1, 2);
439 }
440 }
441
442 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
443 // but not if I2 is a constant.
444 if( n_op == Op_AddP ) {
445 if( n2_loop == n_loop && n3_loop != n_loop ) {
446 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
447 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
448 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
449 IdealLoopTree *n22loop = get_loop( n22_ctrl );
450 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
451 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
452 n23_loop == n_loop ) {
453 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
454 // Stuff new AddP in the loop preheader
455 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
456 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) );
457 register_new_node( add2, n_ctrl );
458 _igvn.replace_node( n, add2 );
459 return add2;
460 }
461 }
462 }
463
464 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
465 if (n2_loop != n_loop && n3_loop == n_loop) {
466 if (n->in(3)->Opcode() == Op_AddX) {
467 Node *V = n->in(3)->in(1);
468 Node *I = n->in(3)->in(2);
469 if (is_member(n_loop,get_ctrl(V))) {
470 } else {
471 Node *tmp = V; V = I; I = tmp;
472 }
473 if (!is_member(n_loop,get_ctrl(I))) {
474 Node *add1 = new AddPNode(n->in(1), n->in(2), I);
475 // Stuff new AddP in the loop preheader
476 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl));
477 Node *add2 = new AddPNode(n->in(1), add1, V);
478 register_new_node(add2, n_ctrl);
479 _igvn.replace_node(n, add2);
480 return add2;
481 }
482 }
483 }
484 }
485
486 return NULL;
487 }
488
489 //------------------------------conditional_move-------------------------------
490 // Attempt to replace a Phi with a conditional move. We have some pretty
491 // strict profitability requirements. All Phis at the merge point must
492 // be converted, so we can remove the control flow. We need to limit the
493 // number of c-moves to a small handful. All code that was in the side-arms
494 // of the CFG diamond is now speculatively executed. This code has to be
495 // "cheap enough". We are pretty much limited to CFG diamonds that merge
496 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
497 Node *PhaseIdealLoop::conditional_move( Node *region ) {
498
499 assert(region->is_Region(), "sanity check");
500 if (region->req() != 3) return NULL;
501
502 // Check for CFG diamond
503 Node *lp = region->in(1);
504 Node *rp = region->in(2);
505 if (!lp || !rp) return NULL;
506 Node *lp_c = lp->in(0);
507 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
508 IfNode *iff = lp_c->as_If();
509
510 // Check for ops pinned in an arm of the diamond.
511 // Can't remove the control flow in this case
512 if (lp->outcnt() > 1) return NULL;
513 if (rp->outcnt() > 1) return NULL;
514
515 IdealLoopTree* r_loop = get_loop(region);
516 assert(r_loop == get_loop(iff), "sanity");
517 // Always convert to CMOVE if all results are used only outside this loop.
518 bool used_inside_loop = (r_loop == _ltree_root);
519
520 // Check profitability
521 int cost = 0;
522 int phis = 0;
523 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
524 Node *out = region->fast_out(i);
525 if (!out->is_Phi()) continue; // Ignore other control edges, etc
526 phis++;
527 PhiNode* phi = out->as_Phi();
528 BasicType bt = phi->type()->basic_type();
529 switch (bt) {
530 case T_DOUBLE:
531 case T_FLOAT:
532 if (C->use_cmove()) {
533 continue; //TODO: maybe we want to add some cost
534 }
535 cost += Matcher::float_cmove_cost(); // Could be very expensive
536 break;
537 case T_LONG: {
538 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
539 }
540 case T_INT: // These all CMOV fine
541 case T_ADDRESS: { // (RawPtr)
542 cost++;
543 break;
544 }
545 case T_NARROWOOP: // Fall through
546 case T_OBJECT: { // Base oops are OK, but not derived oops
547 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
548 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
549 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
550 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
551 // have a Phi for the base here that we convert to a CMOVE all is well
552 // and good. But if the base is dead, we'll not make a CMOVE. Later
553 // the allocator will have to produce a base by creating a CMOVE of the
554 // relevant bases. This puts the allocator in the business of
555 // manufacturing expensive instructions, generally a bad plan.
556 // Just Say No to Conditionally-Moved Derived Pointers.
557 if (tp && tp->offset() != 0)
558 return NULL;
559 cost++;
560 break;
561 }
562 default:
563 return NULL; // In particular, can't do memory or I/O
564 }
565 // Add in cost any speculative ops
566 for (uint j = 1; j < region->req(); j++) {
567 Node *proj = region->in(j);
568 Node *inp = phi->in(j);
569 if (get_ctrl(inp) == proj) { // Found local op
570 cost++;
571 // Check for a chain of dependent ops; these will all become
572 // speculative in a CMOV.
573 for (uint k = 1; k < inp->req(); k++)
574 if (get_ctrl(inp->in(k)) == proj)
575 cost += ConditionalMoveLimit; // Too much speculative goo
576 }
577 }
578 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
579 // This will likely Split-If, a higher-payoff operation.
580 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
581 Node* use = phi->fast_out(k);
582 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
583 cost += ConditionalMoveLimit;
584 // Is there a use inside the loop?
585 // Note: check only basic types since CMoveP is pinned.
586 if (!used_inside_loop && is_java_primitive(bt)) {
587 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
588 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
589 used_inside_loop = true;
590 }
591 }
592 }
593 }//for
594 Node* bol = iff->in(1);
595 assert(bol->Opcode() == Op_Bool, "");
596 int cmp_op = bol->in(1)->Opcode();
597 // It is expensive to generate flags from a float compare.
598 // Avoid duplicated float compare.
599 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
600
601 float infrequent_prob = PROB_UNLIKELY_MAG(3);
602 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
603 if (used_inside_loop) {
604 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
605
606 // BlockLayoutByFrequency optimization moves infrequent branch
607 // from hot path. No point in CMOV'ing in such case (110 is used
608 // instead of 100 to take into account not exactness of float value).
609 if (BlockLayoutByFrequency) {
610 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
611 }
612 }
613 // Check for highly predictable branch. No point in CMOV'ing if
614 // we are going to predict accurately all the time.
615 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
616 //keep going
617 } else if (iff->_prob < infrequent_prob ||
618 iff->_prob > (1.0f - infrequent_prob))
619 return NULL;
620
621 // --------------
622 // Now replace all Phis with CMOV's
623 Node *cmov_ctrl = iff->in(0);
624 uint flip = (lp->Opcode() == Op_IfTrue);
625 Node_List wq;
626 while (1) {
627 PhiNode* phi = NULL;
628 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
629 Node *out = region->fast_out(i);
630 if (out->is_Phi()) {
631 phi = out->as_Phi();
632 break;
633 }
634 }
635 if (phi == NULL) break;
636 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
637 // Move speculative ops
638 wq.push(phi);
639 while (wq.size() > 0) {
640 Node *n = wq.pop();
641 for (uint j = 1; j < n->req(); j++) {
642 Node* m = n->in(j);
643 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) {
644 #ifndef PRODUCT
645 if (PrintOpto && VerifyLoopOptimizations) {
646 tty->print(" speculate: ");
647 m->dump();
648 }
649 #endif
650 set_ctrl(m, cmov_ctrl);
651 wq.push(m);
652 }
653 }
654 }
655 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
656 register_new_node( cmov, cmov_ctrl );
657 _igvn.replace_node( phi, cmov );
658 #ifndef PRODUCT
659 if (TraceLoopOpts) {
660 tty->print("CMOV ");
661 r_loop->dump_head();
662 if (Verbose) {
663 bol->in(1)->dump(1);
664 cmov->dump(1);
665 }
666 }
667 if (VerifyLoopOptimizations) verify();
668 #endif
669 }
670
671 // The useless CFG diamond will fold up later; see the optimization in
672 // RegionNode::Ideal.
673 _igvn._worklist.push(region);
674
675 return iff->in(1);
676 }
677
678 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
679 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
680 Node* u = m->fast_out(i);
681 if (u->is_CFG()) {
682 if (u->Opcode() == Op_NeverBranch) {
683 u = ((NeverBranchNode*)u)->proj_out(0);
684 enqueue_cfg_uses(u, wq);
685 } else {
686 wq.push(u);
687 }
688 }
689 }
690 }
691
692 // Try moving a store out of a loop, right before the loop
693 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
694 // Store has to be first in the loop body
695 IdealLoopTree *n_loop = get_loop(n_ctrl);
696 if (n->is_Store() && n_loop != _ltree_root && n_loop->is_loop() && n->in(0) != NULL) {
697 Node* address = n->in(MemNode::Address);
698 Node* value = n->in(MemNode::ValueIn);
699 Node* mem = n->in(MemNode::Memory);
700 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
701 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
702
703 // - address and value must be loop invariant
704 // - memory must be a memory Phi for the loop
705 // - Store must be the only store on this memory slice in the
706 // loop: if there's another store following this one then value
707 // written at iteration i by the second store could be overwritten
708 // at iteration i+n by the first store: it's not safe to move the
709 // first store out of the loop
710 // - nothing must observe the memory Phi: it guarantees no read
711 // before the store, we are also guaranteed the store post
712 // dominates the loop head (ignoring a possible early
713 // exit). Otherwise there would be extra Phi involved between the
714 // loop's Phi and the store.
715 // - there must be no early exit from the loop before the Store
716 // (such an exit most of the time would be an extra use of the
717 // memory Phi but sometimes is a bottom memory Phi that takes the
718 // store as input).
719
720 if (!n_loop->is_member(address_loop) &&
721 !n_loop->is_member(value_loop) &&
722 mem->is_Phi() && mem->in(0) == n_loop->_head &&
723 mem->outcnt() == 1 &&
724 mem->in(LoopNode::LoopBackControl) == n) {
725
726 assert(n_loop->_tail != NULL, "need a tail");
727 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
728
729 // Verify that there's no early exit of the loop before the store.
730 bool ctrl_ok = false;
731 {
732 // Follow control from loop head until n, we exit the loop or
733 // we reach the tail
734 ResourceMark rm;
735 Unique_Node_List wq;
736 wq.push(n_loop->_head);
737
738 for (uint next = 0; next < wq.size(); ++next) {
739 Node *m = wq.at(next);
740 if (m == n->in(0)) {
741 ctrl_ok = true;
742 continue;
743 }
744 assert(!has_ctrl(m), "should be CFG");
745 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
746 ctrl_ok = false;
747 break;
748 }
749 enqueue_cfg_uses(m, wq);
750 if (wq.size() > 10) {
751 ctrl_ok = false;
752 break;
753 }
754 }
755 }
756 if (ctrl_ok) {
757 // move the Store
758 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
759 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
760 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
761 // Disconnect the phi now. An empty phi can confuse other
762 // optimizations in this pass of loop opts.
763 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
764 n_loop->_body.yank(mem);
765
766 set_ctrl_and_loop(n, n->in(0));
767
768 return n;
769 }
770 }
771 }
772 return NULL;
773 }
774
775 // Try moving a store out of a loop, right after the loop
776 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
777 if (n->is_Store() && n->in(0) != NULL) {
778 Node *n_ctrl = get_ctrl(n);
779 IdealLoopTree *n_loop = get_loop(n_ctrl);
780 // Store must be in a loop
781 if (n_loop != _ltree_root && !n_loop->_irreducible) {
782 Node* address = n->in(MemNode::Address);
783 Node* value = n->in(MemNode::ValueIn);
784 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
785 // address must be loop invariant
786 if (!n_loop->is_member(address_loop)) {
787 // Store must be last on this memory slice in the loop and
788 // nothing in the loop must observe it
789 Node* phi = NULL;
790 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
791 Node* u = n->fast_out(i);
792 if (has_ctrl(u)) { // control use?
793 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
794 if (!n_loop->is_member(u_loop)) {
795 continue;
796 }
797 if (u->is_Phi() && u->in(0) == n_loop->_head) {
798 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
799 // multiple phis on the same slice are possible
800 if (phi != NULL) {
801 return;
802 }
803 phi = u;
804 continue;
805 }
806 }
807 return;
808 }
809 if (phi != NULL) {
810 // Nothing in the loop before the store (next iteration)
811 // must observe the stored value
812 bool mem_ok = true;
813 {
814 ResourceMark rm;
815 Unique_Node_List wq;
816 wq.push(phi);
817 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
818 Node *m = wq.at(next);
819 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
820 Node* u = m->fast_out(i);
821 if (u->is_Store() || u->is_Phi()) {
822 if (u != n) {
823 wq.push(u);
824 mem_ok = (wq.size() <= 10);
825 }
826 } else {
827 mem_ok = false;
828 break;
829 }
830 }
831 }
832 }
833 if (mem_ok) {
834 // Move the store out of the loop if the LCA of all
835 // users (except for the phi) is outside the loop.
836 Node* hook = new Node(1);
837 _igvn.rehash_node_delayed(phi);
838 int count = phi->replace_edge(n, hook);
839 assert(count > 0, "inconsistent phi");
840
841 // Compute latest point this store can go
842 Node* lca = get_late_ctrl(n, get_ctrl(n));
843 if (n_loop->is_member(get_loop(lca))) {
844 // LCA is in the loop - bail out
845 _igvn.replace_node(hook, n);
846 return;
847 }
848 #ifdef ASSERT
849 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
850 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
851 n_loop->_head->as_Loop()->verify_strip_mined(1);
852 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
853 IdealLoopTree* outer_loop = get_loop(outer);
854 assert(n_loop->_parent == outer_loop, "broken loop tree");
855 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
856 }
857 #endif
858
859 // Move store out of the loop
860 _igvn.replace_node(hook, n->in(MemNode::Memory));
861 _igvn.replace_input_of(n, 0, lca);
862 set_ctrl_and_loop(n, lca);
863
864 // Disconnect the phi now. An empty phi can confuse other
865 // optimizations in this pass of loop opts..
866 if (phi->in(LoopNode::LoopBackControl) == phi) {
867 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
868 n_loop->_body.yank(phi);
869 }
870 }
871 }
872 }
873 }
874 }
875 }
876
877 //------------------------------split_if_with_blocks_pre-----------------------
878 // Do the real work in a non-recursive function. Data nodes want to be
879 // cloned in the pre-order so they can feed each other nicely.
880 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
881 // Cloning these guys is unlikely to win
882 int n_op = n->Opcode();
883 if( n_op == Op_MergeMem ) return n;
884 if( n->is_Proj() ) return n;
885 // Do not clone-up CmpFXXX variations, as these are always
886 // followed by a CmpI
887 if( n->is_Cmp() ) return n;
888 // Attempt to use a conditional move instead of a phi/branch
889 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
890 Node *cmov = conditional_move( n );
891 if( cmov ) return cmov;
892 }
893 if( n->is_CFG() || n->is_LoadStore() )
894 return n;
895 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
896 n_op == Op_Opaque2 ) {
897 if( !C->major_progress() ) // If chance of no more loop opts...
898 _igvn._worklist.push(n); // maybe we'll remove them
899 return n;
900 }
901
902 if( n->is_Con() ) return n; // No cloning for Con nodes
903
904 Node *n_ctrl = get_ctrl(n);
905 if( !n_ctrl ) return n; // Dead node
906
907 Node* res = try_move_store_before_loop(n, n_ctrl);
908 if (res != NULL) {
909 return n;
910 }
911
912 // Attempt to remix address expressions for loop invariants
913 Node *m = remix_address_expressions( n );
914 if( m ) return m;
915
916 if (n->is_ConstraintCast()) {
917 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
918 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
919 // Node control inputs don't necessarily agree with loop control info (due to
920 // transformations happened in between), thus additional dominance check is needed
921 // to keep loop info valid.
922 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
923 _igvn.replace_node(n, dom_cast);
924 return dom_cast;
925 }
926 }
927
928 // Determine if the Node has inputs from some local Phi.
929 // Returns the block to clone thru.
930 Node *n_blk = has_local_phi_input( n );
931 if( !n_blk ) return n;
932
933 // Do not clone the trip counter through on a CountedLoop
934 // (messes up the canonical shape).
935 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
936
937 // Check for having no control input; not pinned. Allow
938 // dominating control.
939 if (n->in(0)) {
940 Node *dom = idom(n_blk);
941 if (dom_lca(n->in(0), dom) != n->in(0)) {
942 return n;
943 }
944 }
945 // Policy: when is it profitable. You must get more wins than
946 // policy before it is considered profitable. Policy is usually 0,
947 // so 1 win is considered profitable. Big merges will require big
948 // cloning, so get a larger policy.
949 int policy = n_blk->req() >> 2;
950
951 // If the loop is a candidate for range check elimination,
952 // delay splitting through it's phi until a later loop optimization
953 if (n_blk->is_CountedLoop()) {
954 IdealLoopTree *lp = get_loop(n_blk);
955 if (lp && lp->_rce_candidate) {
956 return n;
957 }
958 }
959
960 // Use same limit as split_if_with_blocks_post
961 if( C->live_nodes() > 35000 ) return n; // Method too big
962
963 // Split 'n' through the merge point if it is profitable
964 Node *phi = split_thru_phi( n, n_blk, policy );
965 if (!phi) return n;
966
967 // Found a Phi to split thru!
968 // Replace 'n' with the new phi
969 _igvn.replace_node( n, phi );
970 // Moved a load around the loop, 'en-registering' something.
971 if (n_blk->is_Loop() && n->is_Load() &&
972 !phi->in(LoopNode::LoopBackControl)->is_Load())
973 C->set_major_progress();
974
975 return phi;
976 }
977
978 static bool merge_point_too_heavy(Compile* C, Node* region) {
979 // Bail out if the region and its phis have too many users.
980 int weight = 0;
981 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
982 weight += region->fast_out(i)->outcnt();
983 }
984 int nodes_left = C->max_node_limit() - C->live_nodes();
985 if (weight * 8 > nodes_left) {
986 if (PrintOpto) {
987 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
988 }
989 return true;
990 } else {
991 return false;
992 }
993 }
994
995 static bool merge_point_safe(Node* region) {
996 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
997 // having a PhiNode input. This sidesteps the dangerous case where the split
998 // ConvI2LNode may become TOP if the input Value() does not
999 // overlap the ConvI2L range, leaving a node which may not dominate its
1000 // uses.
1001 // A better fix for this problem can be found in the BugTraq entry, but
1002 // expediency for Mantis demands this hack.
1003 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop
1004 // split_if_with_blocks from splitting a block because we could not move around
1005 // the FastLockNode.
1006 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1007 Node* n = region->fast_out(i);
1008 if (n->is_Phi()) {
1009 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1010 Node* m = n->fast_out(j);
1011 if (m->is_FastLock())
1012 return false;
1013 #ifdef _LP64
1014 if (m->Opcode() == Op_ConvI2L)
1015 return false;
1016 if (m->is_CastII() && m->isa_CastII()->has_range_check()) {
1017 return false;
1018 }
1019 #endif
1020 }
1021 }
1022 }
1023 return true;
1024 }
1025
1026
1027 //------------------------------place_near_use---------------------------------
1028 // Place some computation next to use but not inside inner loops.
1029 // For inner loop uses move it to the preheader area.
1030 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const {
1031 IdealLoopTree *u_loop = get_loop( useblock );
1032 return (u_loop->_irreducible || u_loop->_child)
1033 ? useblock
1034 : u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
1035 }
1036
1037
1038 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1039 if (!n->is_If()) {
1040 return false;
1041 }
1042 if (!n->in(0)->is_Region()) {
1043 return false;
1044 }
1045 Node* region = n->in(0);
1046 Node* dom = idom(region);
1047 if (!dom->is_If() || dom->in(1) != n->in(1)) {
1048 return false;
1049 }
1050 IfNode* dom_if = dom->as_If();
1051 Node* proj_true = dom_if->proj_out(1);
1052 Node* proj_false = dom_if->proj_out(0);
1053
1054 for (uint i = 1; i < region->req(); i++) {
1055 if (is_dominator(proj_true, region->in(i))) {
1056 continue;
1057 }
1058 if (is_dominator(proj_false, region->in(i))) {
1059 continue;
1060 }
1061 return false;
1062 }
1063
1064 return true;
1065 }
1066
1067 bool PhaseIdealLoop::can_split_if(Node *n_ctrl) {
1068 if (C->live_nodes() > 35000) {
1069 return false; // Method too big
1070 }
1071
1072 // Do not do 'split-if' if irreducible loops are present.
1073 if (_has_irreducible_loops) {
1074 return false;
1075 }
1076
1077 if (merge_point_too_heavy(C, n_ctrl)) {
1078 return false;
1079 }
1080
1081 // Do not do 'split-if' if some paths are dead. First do dead code
1082 // elimination and then see if its still profitable.
1083 for (uint i = 1; i < n_ctrl->req(); i++) {
1084 if (n_ctrl->in(i) == C->top()) {
1085 return false;
1086 }
1087 }
1088
1089 // If trying to do a 'Split-If' at the loop head, it is only
1090 // profitable if the cmp folds up on BOTH paths. Otherwise we
1091 // risk peeling a loop forever.
1092
1093 // CNC - Disabled for now. Requires careful handling of loop
1094 // body selection for the cloned code. Also, make sure we check
1095 // for any input path not being in the same loop as n_ctrl. For
1096 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1097 // because the alternative loop entry points won't be converted
1098 // into LoopNodes.
1099 IdealLoopTree *n_loop = get_loop(n_ctrl);
1100 for (uint j = 1; j < n_ctrl->req(); j++) {
1101 if (get_loop(n_ctrl->in(j)) != n_loop) {
1102 return false;
1103 }
1104 }
1105
1106 // Check for safety of the merge point.
1107 if (!merge_point_safe(n_ctrl)) {
1108 return false;
1109 }
1110
1111 return true;
1112 }
1113
1114 //------------------------------split_if_with_blocks_post----------------------
1115 // Do the real work in a non-recursive function. CFG hackery wants to be
1116 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1117 // info.
1118 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1119
1120 // Cloning Cmp through Phi's involves the split-if transform.
1121 // FastLock is not used by an If
1122 if (n->is_Cmp() && !n->is_FastLock()) {
1123 Node *n_ctrl = get_ctrl(n);
1124 // Determine if the Node has inputs from some local Phi.
1125 // Returns the block to clone thru.
1126 Node *n_blk = has_local_phi_input(n);
1127 if (n_blk != n_ctrl) {
1128 return;
1129 }
1130
1131 if (!can_split_if(n_ctrl)) {
1132 return;
1133 }
1134
1135 if (n->outcnt() != 1) {
1136 return; // Multiple bool's from 1 compare?
1137 }
1138 Node *bol = n->unique_out();
1139 assert(bol->is_Bool(), "expect a bool here");
1140 if (bol->outcnt() != 1) {
1141 return;// Multiple branches from 1 compare?
1142 }
1143 Node *iff = bol->unique_out();
1144
1145 // Check some safety conditions
1146 if (iff->is_If()) { // Classic split-if?
1147 if (iff->in(0) != n_ctrl) {
1148 return; // Compare must be in same blk as if
1149 }
1150 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1151 // Can't split CMove with different control edge.
1152 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) {
1153 return;
1154 }
1155 if (get_ctrl(iff->in(2)) == n_ctrl ||
1156 get_ctrl(iff->in(3)) == n_ctrl) {
1157 return; // Inputs not yet split-up
1158 }
1159 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1160 return; // Loop-invar test gates loop-varying CMOVE
1161 }
1162 } else {
1163 return; // some other kind of node, such as an Allocate
1164 }
1165
1166 // When is split-if profitable? Every 'win' on means some control flow
1167 // goes dead, so it's almost always a win.
1168 int policy = 0;
1169 // Split compare 'n' through the merge point if it is profitable
1170 Node *phi = split_thru_phi( n, n_ctrl, policy);
1171 if (!phi) {
1172 return;
1173 }
1174
1175 // Found a Phi to split thru!
1176 // Replace 'n' with the new phi
1177 _igvn.replace_node(n, phi);
1178
1179 // Now split the bool up thru the phi
1180 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1181 guarantee(bolphi != NULL, "null boolean phi node");
1182
1183 _igvn.replace_node(bol, bolphi);
1184 assert(iff->in(1) == bolphi, "");
1185
1186 if (bolphi->Value(&_igvn)->singleton()) {
1187 return;
1188 }
1189
1190 // Conditional-move? Must split up now
1191 if (!iff->is_If()) {
1192 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1193 _igvn.replace_node(iff, cmovphi);
1194 return;
1195 }
1196
1197 // Now split the IF
1198 do_split_if(iff);
1199 return;
1200 }
1201
1202 // Two identical ifs back to back can be merged
1203 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1204 Node *n_ctrl = n->in(0);
1205 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1206 IfNode* dom_if = idom(n_ctrl)->as_If();
1207 Node* proj_true = dom_if->proj_out(1);
1208 Node* proj_false = dom_if->proj_out(0);
1209 Node* con_true = _igvn.makecon(TypeInt::ONE);
1210 Node* con_false = _igvn.makecon(TypeInt::ZERO);
1211
1212 for (uint i = 1; i < n_ctrl->req(); i++) {
1213 if (is_dominator(proj_true, n_ctrl->in(i))) {
1214 bolphi->init_req(i, con_true);
1215 } else {
1216 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1217 bolphi->init_req(i, con_false);
1218 }
1219 }
1220 register_new_node(bolphi, n_ctrl);
1221 _igvn.replace_input_of(n, 1, bolphi);
1222
1223 // Now split the IF
1224 do_split_if(n);
1225 return;
1226 }
1227
1228 // Check for an IF ready to split; one that has its
1229 // condition codes input coming from a Phi at the block start.
1230 int n_op = n->Opcode();
1231
1232 // Check for an IF being dominated by another IF same test
1233 if (n_op == Op_If ||
1234 n_op == Op_RangeCheck) {
1235 Node *bol = n->in(1);
1236 uint max = bol->outcnt();
1237 // Check for same test used more than once?
1238 if (max > 1 && bol->is_Bool()) {
1239 // Search up IDOMs to see if this IF is dominated.
1240 Node *cutoff = get_ctrl(bol);
1241
1242 // Now search up IDOMs till cutoff, looking for a dominating test
1243 Node *prevdom = n;
1244 Node *dom = idom(prevdom);
1245 while (dom != cutoff) {
1246 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) {
1247 // Replace the dominated test with an obvious true or false.
1248 // Place it on the IGVN worklist for later cleanup.
1249 C->set_major_progress();
1250 dominated_by(prevdom, n, false, true);
1251 #ifndef PRODUCT
1252 if( VerifyLoopOptimizations ) verify();
1253 #endif
1254 return;
1255 }
1256 prevdom = dom;
1257 dom = idom(prevdom);
1258 }
1259 }
1260 }
1261
1262 // See if a shared loop-varying computation has no loop-varying uses.
1263 // Happens if something is only used for JVM state in uncommon trap exits,
1264 // like various versions of induction variable+offset. Clone the
1265 // computation per usage to allow it to sink out of the loop.
1266 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
1267 Node *n_ctrl = get_ctrl(n);
1268 IdealLoopTree *n_loop = get_loop(n_ctrl);
1269 if( n_loop != _ltree_root ) {
1270 DUIterator_Fast imax, i = n->fast_outs(imax);
1271 for (; i < imax; i++) {
1272 Node* u = n->fast_out(i);
1273 if( !has_ctrl(u) ) break; // Found control user
1274 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1275 if( u_loop == n_loop ) break; // Found loop-varying use
1276 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
1277 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
1278 }
1279 bool did_break = (i < imax); // Did we break out of the previous loop?
1280 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
1281 Node *late_load_ctrl = NULL;
1282 if (n->is_Load()) {
1283 // If n is a load, get and save the result from get_late_ctrl(),
1284 // to be later used in calculating the control for n's clones.
1285 clear_dom_lca_tags();
1286 late_load_ctrl = get_late_ctrl(n, n_ctrl);
1287 }
1288 // If n is a load, and the late control is the same as the current
1289 // control, then the cloning of n is a pointless exercise, because
1290 // GVN will ensure that we end up where we started.
1291 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
1292 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
1293 Node *u = n->last_out(j); // Clone private computation per use
1294 _igvn.rehash_node_delayed(u);
1295 Node *x = n->clone(); // Clone computation
1296 Node *x_ctrl = NULL;
1297 if( u->is_Phi() ) {
1298 // Replace all uses of normal nodes. Replace Phi uses
1299 // individually, so the separate Nodes can sink down
1300 // different paths.
1301 uint k = 1;
1302 while( u->in(k) != n ) k++;
1303 u->set_req( k, x );
1304 // x goes next to Phi input path
1305 x_ctrl = u->in(0)->in(k);
1306 --j;
1307 } else { // Normal use
1308 // Replace all uses
1309 for( uint k = 0; k < u->req(); k++ ) {
1310 if( u->in(k) == n ) {
1311 u->set_req( k, x );
1312 --j;
1313 }
1314 }
1315 x_ctrl = get_ctrl(u);
1316 }
1317
1318 // Find control for 'x' next to use but not inside inner loops.
1319 // For inner loop uses get the preheader area.
1320 x_ctrl = place_near_use(x_ctrl);
1321
1322 if (n->is_Load()) {
1323 // For loads, add a control edge to a CFG node outside of the loop
1324 // to force them to not combine and return back inside the loop
1325 // during GVN optimization (4641526).
1326 //
1327 // Because we are setting the actual control input, factor in
1328 // the result from get_late_ctrl() so we respect any
1329 // anti-dependences. (6233005).
1330 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
1331
1332 // Don't allow the control input to be a CFG splitting node.
1333 // Such nodes should only have ProjNodes as outs, e.g. IfNode
1334 // should only have IfTrueNode and IfFalseNode (4985384).
1335 x_ctrl = find_non_split_ctrl(x_ctrl);
1336 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1337
1338 x->set_req(0, x_ctrl);
1339 }
1340 register_new_node(x, x_ctrl);
1341
1342 // Some institutional knowledge is needed here: 'x' is
1343 // yanked because if the optimizer runs GVN on it all the
1344 // cloned x's will common up and undo this optimization and
1345 // be forced back in the loop. This is annoying because it
1346 // makes +VerifyOpto report false-positives on progress. I
1347 // tried setting control edges on the x's to force them to
1348 // not combine, but the matching gets worried when it tries
1349 // to fold a StoreP and an AddP together (as part of an
1350 // address expression) and the AddP and StoreP have
1351 // different controls.
1352 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x);
1353 }
1354 _igvn.remove_dead_node(n);
1355 }
1356 }
1357 }
1358 }
1359
1360 try_move_store_after_loop(n);
1361
1362 // Check for Opaque2's who's loop has disappeared - who's input is in the
1363 // same loop nest as their output. Remove 'em, they are no longer useful.
1364 if( n_op == Op_Opaque2 &&
1365 n->in(1) != NULL &&
1366 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1367 _igvn.replace_node( n, n->in(1) );
1368 }
1369 }
1370
1371 //------------------------------split_if_with_blocks---------------------------
1372 // Check for aggressive application of 'split-if' optimization,
1373 // using basic block level info.
1374 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) {
1375 Node *n = C->root();
1376 visited.set(n->_idx); // first, mark node as visited
1377 // Do pre-visit work for root
1378 n = split_if_with_blocks_pre( n );
1379 uint cnt = n->outcnt();
1380 uint i = 0;
1381 while (true) {
1382 // Visit all children
1383 if (i < cnt) {
1384 Node* use = n->raw_out(i);
1385 ++i;
1386 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1387 // Now do pre-visit work for this use
1388 use = split_if_with_blocks_pre( use );
1389 nstack.push(n, i); // Save parent and next use's index.
1390 n = use; // Process all children of current use.
1391 cnt = use->outcnt();
1392 i = 0;
1393 }
1394 }
1395 else {
1396 // All of n's children have been processed, complete post-processing.
1397 if (cnt != 0 && !n->is_Con()) {
1398 assert(has_node(n), "no dead nodes");
1399 split_if_with_blocks_post( n );
1400 }
1401 if (nstack.is_empty()) {
1402 // Finished all nodes on stack.
1403 break;
1404 }
1405 // Get saved parent node and next use's index. Visit the rest of uses.
1406 n = nstack.node();
1407 cnt = n->outcnt();
1408 i = nstack.index();
1409 nstack.pop();
1410 }
1411 }
1412 }
1413
1414
1415 //=============================================================================
1416 //
1417 // C L O N E A L O O P B O D Y
1418 //
1419
1420 //------------------------------clone_iff--------------------------------------
1421 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1422 // "Nearly" because all Nodes have been cloned from the original in the loop,
1423 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1424 // through the Phi recursively, and return a Bool.
1425 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) {
1426
1427 // Convert this Phi into a Phi merging Bools
1428 uint i;
1429 for (i = 1; i < phi->req(); i++) {
1430 Node *b = phi->in(i);
1431 if (b->is_Phi()) {
1432 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop));
1433 } else {
1434 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
1435 }
1436 }
1437
1438 Node* n = phi->in(1);
1439 Node* sample_opaque = NULL;
1440 Node *sample_bool = NULL;
1441 if (n->Opcode() == Op_Opaque4) {
1442 sample_opaque = n;
1443 sample_bool = n->in(1);
1444 assert(sample_bool->is_Bool(), "wrong type");
1445 } else {
1446 sample_bool = n;
1447 }
1448 Node *sample_cmp = sample_bool->in(1);
1449
1450 // Make Phis to merge the Cmp's inputs.
1451 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
1452 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
1453 for (i = 1; i < phi->req(); i++) {
1454 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
1455 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
1456 phi1->set_req(i, n1);
1457 phi2->set_req(i, n2);
1458 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1459 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1460 }
1461 // See if these Phis have been made before.
1462 // Register with optimizer
1463 Node *hit1 = _igvn.hash_find_insert(phi1);
1464 if (hit1) { // Hit, toss just made Phi
1465 _igvn.remove_dead_node(phi1); // Remove new phi
1466 assert(hit1->is_Phi(), "" );
1467 phi1 = (PhiNode*)hit1; // Use existing phi
1468 } else { // Miss
1469 _igvn.register_new_node_with_optimizer(phi1);
1470 }
1471 Node *hit2 = _igvn.hash_find_insert(phi2);
1472 if (hit2) { // Hit, toss just made Phi
1473 _igvn.remove_dead_node(phi2); // Remove new phi
1474 assert(hit2->is_Phi(), "" );
1475 phi2 = (PhiNode*)hit2; // Use existing phi
1476 } else { // Miss
1477 _igvn.register_new_node_with_optimizer(phi2);
1478 }
1479 // Register Phis with loop/block info
1480 set_ctrl(phi1, phi->in(0));
1481 set_ctrl(phi2, phi->in(0));
1482 // Make a new Cmp
1483 Node *cmp = sample_cmp->clone();
1484 cmp->set_req(1, phi1);
1485 cmp->set_req(2, phi2);
1486 _igvn.register_new_node_with_optimizer(cmp);
1487 set_ctrl(cmp, phi->in(0));
1488
1489 // Make a new Bool
1490 Node *b = sample_bool->clone();
1491 b->set_req(1,cmp);
1492 _igvn.register_new_node_with_optimizer(b);
1493 set_ctrl(b, phi->in(0));
1494
1495 if (sample_opaque != NULL) {
1496 Node* opaque = sample_opaque->clone();
1497 opaque->set_req(1, b);
1498 _igvn.register_new_node_with_optimizer(opaque);
1499 set_ctrl(opaque, phi->in(0));
1500 return opaque;
1501 }
1502
1503 assert(b->is_Bool(), "");
1504 return b;
1505 }
1506
1507 //------------------------------clone_bool-------------------------------------
1508 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1509 // "Nearly" because all Nodes have been cloned from the original in the loop,
1510 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1511 // through the Phi recursively, and return a Bool.
1512 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1513 uint i;
1514 // Convert this Phi into a Phi merging Bools
1515 for( i = 1; i < phi->req(); i++ ) {
1516 Node *b = phi->in(i);
1517 if( b->is_Phi() ) {
1518 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1519 } else {
1520 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1521 }
1522 }
1523
1524 Node *sample_cmp = phi->in(1);
1525
1526 // Make Phis to merge the Cmp's inputs.
1527 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1528 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1529 for( uint j = 1; j < phi->req(); j++ ) {
1530 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1531 Node *n1, *n2;
1532 if( cmp_top->is_Cmp() ) {
1533 n1 = cmp_top->in(1);
1534 n2 = cmp_top->in(2);
1535 } else {
1536 n1 = n2 = cmp_top;
1537 }
1538 phi1->set_req( j, n1 );
1539 phi2->set_req( j, n2 );
1540 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1541 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1542 }
1543
1544 // See if these Phis have been made before.
1545 // Register with optimizer
1546 Node *hit1 = _igvn.hash_find_insert(phi1);
1547 if( hit1 ) { // Hit, toss just made Phi
1548 _igvn.remove_dead_node(phi1); // Remove new phi
1549 assert( hit1->is_Phi(), "" );
1550 phi1 = (PhiNode*)hit1; // Use existing phi
1551 } else { // Miss
1552 _igvn.register_new_node_with_optimizer(phi1);
1553 }
1554 Node *hit2 = _igvn.hash_find_insert(phi2);
1555 if( hit2 ) { // Hit, toss just made Phi
1556 _igvn.remove_dead_node(phi2); // Remove new phi
1557 assert( hit2->is_Phi(), "" );
1558 phi2 = (PhiNode*)hit2; // Use existing phi
1559 } else { // Miss
1560 _igvn.register_new_node_with_optimizer(phi2);
1561 }
1562 // Register Phis with loop/block info
1563 set_ctrl(phi1, phi->in(0));
1564 set_ctrl(phi2, phi->in(0));
1565 // Make a new Cmp
1566 Node *cmp = sample_cmp->clone();
1567 cmp->set_req( 1, phi1 );
1568 cmp->set_req( 2, phi2 );
1569 _igvn.register_new_node_with_optimizer(cmp);
1570 set_ctrl(cmp, phi->in(0));
1571
1572 assert( cmp->is_Cmp(), "" );
1573 return (CmpNode*)cmp;
1574 }
1575
1576 //------------------------------sink_use---------------------------------------
1577 // If 'use' was in the loop-exit block, it now needs to be sunk
1578 // below the post-loop merge point.
1579 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1580 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1581 set_ctrl(use, post_loop);
1582 for (DUIterator j = use->outs(); use->has_out(j); j++)
1583 sink_use(use->out(j), post_loop);
1584 }
1585 }
1586
1587 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1588 IdealLoopTree* loop, IdealLoopTree* outer_loop,
1589 Node_List*& split_if_set, Node_List*& split_bool_set,
1590 Node_List*& split_cex_set, Node_List& worklist,
1591 uint new_counter, CloneLoopMode mode) {
1592 Node* nnn = old_new[old->_idx];
1593 // Copy uses to a worklist, so I can munge the def-use info
1594 // with impunity.
1595 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1596 worklist.push(old->fast_out(j));
1597
1598 while( worklist.size() ) {
1599 Node *use = worklist.pop();
1600 if (!has_node(use)) continue; // Ignore dead nodes
1601 if (use->in(0) == C->top()) continue;
1602 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1603 // Check for data-use outside of loop - at least one of OLD or USE
1604 // must not be a CFG node.
1605 #ifdef ASSERT
1606 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) {
1607 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
1608 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node");
1609 }
1610 #endif
1611 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
1612
1613 // If the Data use is an IF, that means we have an IF outside of the
1614 // loop that is switching on a condition that is set inside of the
1615 // loop. Happens if people set a loop-exit flag; then test the flag
1616 // in the loop to break the loop, then test is again outside of the
1617 // loop to determine which way the loop exited.
1618 // Loop predicate If node connects to Bool node through Opaque1 node.
1619 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) {
1620 // Since this code is highly unlikely, we lazily build the worklist
1621 // of such Nodes to go split.
1622 if (!split_if_set) {
1623 ResourceArea *area = Thread::current()->resource_area();
1624 split_if_set = new Node_List(area);
1625 }
1626 split_if_set->push(use);
1627 }
1628 if (use->is_Bool()) {
1629 if (!split_bool_set) {
1630 ResourceArea *area = Thread::current()->resource_area();
1631 split_bool_set = new Node_List(area);
1632 }
1633 split_bool_set->push(use);
1634 }
1635 if (use->Opcode() == Op_CreateEx) {
1636 if (!split_cex_set) {
1637 ResourceArea *area = Thread::current()->resource_area();
1638 split_cex_set = new Node_List(area);
1639 }
1640 split_cex_set->push(use);
1641 }
1642
1643
1644 // Get "block" use is in
1645 uint idx = 0;
1646 while( use->in(idx) != old ) idx++;
1647 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1648 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
1649 Node *cfg = prev->_idx >= new_counter
1650 ? prev->in(2)
1651 : idom(prev);
1652 if( use->is_Phi() ) // Phi use is in prior block
1653 cfg = prev->in(idx); // NOT in block of Phi itself
1654 if (cfg->is_top()) { // Use is dead?
1655 _igvn.replace_input_of(use, idx, C->top());
1656 continue;
1657 }
1658
1659 while(!outer_loop->is_member(get_loop(cfg))) {
1660 prev = cfg;
1661 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1662 }
1663 // If the use occurs after merging several exits from the loop, then
1664 // old value must have dominated all those exits. Since the same old
1665 // value was used on all those exits we did not need a Phi at this
1666 // merge point. NOW we do need a Phi here. Each loop exit value
1667 // is now merged with the peeled body exit; each exit gets its own
1668 // private Phi and those Phis need to be merged here.
1669 Node *phi;
1670 if( prev->is_Region() ) {
1671 if( idx == 0 ) { // Updating control edge?
1672 phi = prev; // Just use existing control
1673 } else { // Else need a new Phi
1674 phi = PhiNode::make( prev, old );
1675 // Now recursively fix up the new uses of old!
1676 for( uint i = 1; i < prev->req(); i++ ) {
1677 worklist.push(phi); // Onto worklist once for each 'old' input
1678 }
1679 }
1680 } else {
1681 // Get new RegionNode merging old and new loop exits
1682 prev = old_new[prev->_idx];
1683 assert( prev, "just made this in step 7" );
1684 if( idx == 0) { // Updating control edge?
1685 phi = prev; // Just use existing control
1686 } else { // Else need a new Phi
1687 // Make a new Phi merging data values properly
1688 phi = PhiNode::make( prev, old );
1689 phi->set_req( 1, nnn );
1690 }
1691 }
1692 // If inserting a new Phi, check for prior hits
1693 if( idx != 0 ) {
1694 Node *hit = _igvn.hash_find_insert(phi);
1695 if( hit == NULL ) {
1696 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1697 } else { // or
1698 // Remove the new phi from the graph and use the hit
1699 _igvn.remove_dead_node(phi);
1700 phi = hit; // Use existing phi
1701 }
1702 set_ctrl(phi, prev);
1703 }
1704 // Make 'use' use the Phi instead of the old loop body exit value
1705 _igvn.replace_input_of(use, idx, phi);
1706 if( use->_idx >= new_counter ) { // If updating new phis
1707 // Not needed for correctness, but prevents a weak assert
1708 // in AddPNode from tripping (when we end up with different
1709 // base & derived Phis that will become the same after
1710 // IGVN does CSE).
1711 Node *hit = _igvn.hash_find_insert(use);
1712 if( hit ) // Go ahead and re-hash for hits.
1713 _igvn.replace_node( use, hit );
1714 }
1715
1716 // If 'use' was in the loop-exit block, it now needs to be sunk
1717 // below the post-loop merge point.
1718 sink_use( use, prev );
1719 }
1720 }
1721 }
1722
1723 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1724 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1725 Node_List& extra_data_nodes) {
1726 if (head->is_strip_mined() && mode != IgnoreStripMined) {
1727 CountedLoopNode* cl = head->as_CountedLoop();
1728 Node* l = cl->outer_loop();
1729 Node* tail = cl->outer_loop_tail();
1730 IfNode* le = cl->outer_loop_end();
1731 Node* sfpt = cl->outer_safepoint();
1732 CountedLoopEndNode* cle = cl->loopexit();
1733 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
1734 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
1735 Node* cle_out = cle->proj_out(false);
1736
1737 Node* new_sfpt = NULL;
1738 Node* new_cle_out = cle_out->clone();
1739 old_new.map(cle_out->_idx, new_cle_out);
1740 if (mode == CloneIncludesStripMined) {
1741 // clone outer loop body
1742 Node* new_l = l->clone();
1743 Node* new_tail = tail->clone();
1744 IfNode* new_le = le->clone()->as_If();
1745 new_sfpt = sfpt->clone();
1746
1747 set_loop(new_l, outer_loop->_parent);
1748 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
1749 set_loop(new_cle_out, outer_loop->_parent);
1750 set_idom(new_cle_out, new_cle, dd);
1751 set_loop(new_sfpt, outer_loop->_parent);
1752 set_idom(new_sfpt, new_cle_out, dd);
1753 set_loop(new_le, outer_loop->_parent);
1754 set_idom(new_le, new_sfpt, dd);
1755 set_loop(new_tail, outer_loop->_parent);
1756 set_idom(new_tail, new_le, dd);
1757 set_idom(new_cl, new_l, dd);
1758
1759 old_new.map(l->_idx, new_l);
1760 old_new.map(tail->_idx, new_tail);
1761 old_new.map(le->_idx, new_le);
1762 old_new.map(sfpt->_idx, new_sfpt);
1763
1764 new_l->set_req(LoopNode::LoopBackControl, new_tail);
1765 new_l->set_req(0, new_l);
1766 new_tail->set_req(0, new_le);
1767 new_le->set_req(0, new_sfpt);
1768 new_sfpt->set_req(0, new_cle_out);
1769 new_cle_out->set_req(0, new_cle);
1770 new_cl->set_req(LoopNode::EntryControl, new_l);
1771
1772 _igvn.register_new_node_with_optimizer(new_l);
1773 _igvn.register_new_node_with_optimizer(new_tail);
1774 _igvn.register_new_node_with_optimizer(new_le);
1775 } else {
1776 Node *newhead = old_new[loop->_head->_idx];
1777 newhead->as_Loop()->clear_strip_mined();
1778 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
1779 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1780 }
1781 // Look at data node that were assigned a control in the outer
1782 // loop: they are kept in the outer loop by the safepoint so start
1783 // from the safepoint node's inputs.
1784 IdealLoopTree* outer_loop = get_loop(l);
1785 Node_Stack stack(2);
1786 stack.push(sfpt, 1);
1787 uint new_counter = C->unique();
1788 while (stack.size() > 0) {
1789 Node* n = stack.node();
1790 uint i = stack.index();
1791 while (i < n->req() &&
1792 (n->in(i) == NULL ||
1793 !has_ctrl(n->in(i)) ||
1794 get_loop(get_ctrl(n->in(i))) != outer_loop ||
1795 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
1796 i++;
1797 }
1798 if (i < n->req()) {
1799 stack.set_index(i+1);
1800 stack.push(n->in(i), 0);
1801 } else {
1802 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
1803 Node* m = n == sfpt ? new_sfpt : n->clone();
1804 if (m != NULL) {
1805 for (uint i = 0; i < n->req(); i++) {
1806 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) {
1807 m->set_req(i, old_new[m->in(i)->_idx]);
1808 }
1809 }
1810 } else {
1811 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
1812 }
1813 if (n != sfpt) {
1814 extra_data_nodes.push(n);
1815 _igvn.register_new_node_with_optimizer(m);
1816 assert(get_ctrl(n) == cle_out, "what other control?");
1817 set_ctrl(m, new_cle_out);
1818 old_new.map(n->_idx, m);
1819 }
1820 stack.pop();
1821 }
1822 }
1823 if (mode == CloneIncludesStripMined) {
1824 _igvn.register_new_node_with_optimizer(new_sfpt);
1825 _igvn.register_new_node_with_optimizer(new_cle_out);
1826 }
1827 } else {
1828 Node *newhead = old_new[loop->_head->_idx];
1829 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1830 }
1831 }
1832
1833 //------------------------------clone_loop-------------------------------------
1834 //
1835 // C L O N E A L O O P B O D Y
1836 //
1837 // This is the basic building block of the loop optimizations. It clones an
1838 // entire loop body. It makes an old_new loop body mapping; with this mapping
1839 // you can find the new-loop equivalent to an old-loop node. All new-loop
1840 // nodes are exactly equal to their old-loop counterparts, all edges are the
1841 // same. All exits from the old-loop now have a RegionNode that merges the
1842 // equivalent new-loop path. This is true even for the normal "loop-exit"
1843 // condition. All uses of loop-invariant old-loop values now come from (one
1844 // or more) Phis that merge their new-loop equivalents.
1845 //
1846 // This operation leaves the graph in an illegal state: there are two valid
1847 // control edges coming from the loop pre-header to both loop bodies. I'll
1848 // definitely have to hack the graph after running this transform.
1849 //
1850 // From this building block I will further edit edges to perform loop peeling
1851 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1852 //
1853 // Parameter side_by_size_idom:
1854 // When side_by_size_idom is NULL, the dominator tree is constructed for
1855 // the clone loop to dominate the original. Used in construction of
1856 // pre-main-post loop sequence.
1857 // When nonnull, the clone and original are side-by-side, both are
1858 // dominated by the side_by_side_idom node. Used in construction of
1859 // unswitched loops.
1860 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1861 CloneLoopMode mode, Node* side_by_side_idom) {
1862
1863 LoopNode* head = loop->_head->as_Loop();
1864 head->verify_strip_mined(1);
1865
1866 if (C->do_vector_loop() && PrintOpto) {
1867 const char* mname = C->method()->name()->as_quoted_ascii();
1868 if (mname != NULL) {
1869 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
1870 }
1871 }
1872
1873 CloneMap& cm = C->clone_map();
1874 Dict* dict = cm.dict();
1875 if (C->do_vector_loop()) {
1876 cm.set_clone_idx(cm.max_gen()+1);
1877 #ifndef PRODUCT
1878 if (PrintOpto) {
1879 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
1880 loop->dump_head();
1881 }
1882 #endif
1883 }
1884
1885 // Step 1: Clone the loop body. Make the old->new mapping.
1886 uint i;
1887 for( i = 0; i < loop->_body.size(); i++ ) {
1888 Node *old = loop->_body.at(i);
1889 Node *nnn = old->clone();
1890 old_new.map( old->_idx, nnn );
1891 if (C->do_vector_loop()) {
1892 cm.verify_insert_and_clone(old, nnn, cm.clone_idx());
1893 }
1894 _igvn.register_new_node_with_optimizer(nnn);
1895 }
1896
1897 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
1898
1899 // Step 2: Fix the edges in the new body. If the old input is outside the
1900 // loop use it. If the old input is INside the loop, use the corresponding
1901 // new node instead.
1902 for( i = 0; i < loop->_body.size(); i++ ) {
1903 Node *old = loop->_body.at(i);
1904 Node *nnn = old_new[old->_idx];
1905 // Fix CFG/Loop controlling the new node
1906 if (has_ctrl(old)) {
1907 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
1908 } else {
1909 set_loop(nnn, outer_loop->_parent);
1910 if (old->outcnt() > 0) {
1911 set_idom( nnn, old_new[idom(old)->_idx], dd );
1912 }
1913 }
1914 // Correct edges to the new node
1915 for( uint j = 0; j < nnn->req(); j++ ) {
1916 Node *n = nnn->in(j);
1917 if( n ) {
1918 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
1919 if( loop->is_member( old_in_loop ) )
1920 nnn->set_req(j, old_new[n->_idx]);
1921 }
1922 }
1923 _igvn.hash_find_insert(nnn);
1924 }
1925
1926 ResourceArea *area = Thread::current()->resource_area();
1927 Node_List extra_data_nodes(area);
1928 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
1929
1930 // Step 3: Now fix control uses. Loop varying control uses have already
1931 // been fixed up (as part of all input edges in Step 2). Loop invariant
1932 // control uses must be either an IfFalse or an IfTrue. Make a merge
1933 // point to merge the old and new IfFalse/IfTrue nodes; make the use
1934 // refer to this.
1935 Node_List worklist(area);
1936 uint new_counter = C->unique();
1937 for( i = 0; i < loop->_body.size(); i++ ) {
1938 Node* old = loop->_body.at(i);
1939 if( !old->is_CFG() ) continue;
1940
1941 // Copy uses to a worklist, so I can munge the def-use info
1942 // with impunity.
1943 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1944 worklist.push(old->fast_out(j));
1945
1946 while( worklist.size() ) { // Visit all uses
1947 Node *use = worklist.pop();
1948 if (!has_node(use)) continue; // Ignore dead nodes
1949 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1950 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
1951 // Both OLD and USE are CFG nodes here.
1952 assert( use->is_Proj(), "" );
1953 Node* nnn = old_new[old->_idx];
1954
1955 Node* newuse = NULL;
1956 if (head->is_strip_mined() && mode != IgnoreStripMined) {
1957 CountedLoopNode* cl = head->as_CountedLoop();
1958 CountedLoopEndNode* cle = cl->loopexit();
1959 Node* cle_out = cle->proj_out_or_null(false);
1960 if (use == cle_out) {
1961 IfNode* le = cl->outer_loop_end();
1962 use = le->proj_out(false);
1963 use_loop = get_loop(use);
1964 if (mode == CloneIncludesStripMined) {
1965 nnn = old_new[le->_idx];
1966 } else {
1967 newuse = old_new[cle_out->_idx];
1968 }
1969 }
1970 }
1971 if (newuse == NULL) {
1972 newuse = use->clone();
1973 }
1974
1975 // Clone the loop exit control projection
1976 if (C->do_vector_loop()) {
1977 cm.verify_insert_and_clone(use, newuse, cm.clone_idx());
1978 }
1979 newuse->set_req(0,nnn);
1980 _igvn.register_new_node_with_optimizer(newuse);
1981 set_loop(newuse, use_loop);
1982 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
1983
1984 // We need a Region to merge the exit from the peeled body and the
1985 // exit from the old loop body.
1986 RegionNode *r = new RegionNode(3);
1987 // Map the old use to the new merge point
1988 old_new.map( use->_idx, r );
1989 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
1990 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
1991
1992 // The original user of 'use' uses 'r' instead.
1993 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
1994 Node* useuse = use->last_out(l);
1995 _igvn.rehash_node_delayed(useuse);
1996 uint uses_found = 0;
1997 if( useuse->in(0) == use ) {
1998 useuse->set_req(0, r);
1999 uses_found++;
2000 if( useuse->is_CFG() ) {
2001 assert( dom_depth(useuse) > dd_r, "" );
2002 set_idom(useuse, r, dom_depth(useuse));
2003 }
2004 }
2005 for( uint k = 1; k < useuse->req(); k++ ) {
2006 if( useuse->in(k) == use ) {
2007 useuse->set_req(k, r);
2008 uses_found++;
2009 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2010 assert(dom_depth(useuse) > dd_r , "");
2011 set_idom(useuse, r, dom_depth(useuse));
2012 }
2013 }
2014 }
2015 l -= uses_found; // we deleted 1 or more copies of this edge
2016 }
2017
2018 // Now finish up 'r'
2019 r->set_req( 1, newuse );
2020 r->set_req( 2, use );
2021 _igvn.register_new_node_with_optimizer(r);
2022 set_loop(r, use_loop);
2023 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
2024 } // End of if a loop-exit test
2025 }
2026 }
2027
2028 // Step 4: If loop-invariant use is not control, it must be dominated by a
2029 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2030 // there if needed. Make a Phi there merging old and new used values.
2031 Node_List *split_if_set = NULL;
2032 Node_List *split_bool_set = NULL;
2033 Node_List *split_cex_set = NULL;
2034 for( i = 0; i < loop->_body.size(); i++ ) {
2035 Node* old = loop->_body.at(i);
2036 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2037 split_bool_set, split_cex_set, worklist, new_counter,
2038 mode);
2039 }
2040
2041 for (i = 0; i < extra_data_nodes.size(); i++) {
2042 Node* old = extra_data_nodes.at(i);
2043 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2044 split_bool_set, split_cex_set, worklist, new_counter,
2045 mode);
2046 }
2047
2048 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2049 // the loop uses a condition set in the loop. The original IF probably
2050 // takes control from one or more OLD Regions (which in turn get from NEW
2051 // Regions). In any case, there will be a set of Phis for each merge point
2052 // from the IF up to where the original BOOL def exists the loop.
2053 if (split_if_set) {
2054 while (split_if_set->size()) {
2055 Node *iff = split_if_set->pop();
2056 if (iff->in(1)->is_Phi()) {
2057 Node *b = clone_iff(iff->in(1)->as_Phi(), loop);
2058 _igvn.replace_input_of(iff, 1, b);
2059 }
2060 }
2061 }
2062 if (split_bool_set) {
2063 while (split_bool_set->size()) {
2064 Node *b = split_bool_set->pop();
2065 Node *phi = b->in(1);
2066 assert(phi->is_Phi(), "");
2067 CmpNode *cmp = clone_bool((PhiNode*)phi, loop);
2068 _igvn.replace_input_of(b, 1, cmp);
2069 }
2070 }
2071 if (split_cex_set) {
2072 while (split_cex_set->size()) {
2073 Node *b = split_cex_set->pop();
2074 assert(b->in(0)->is_Region(), "");
2075 assert(b->in(1)->is_Phi(), "");
2076 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2077 split_up(b, b->in(0), NULL);
2078 }
2079 }
2080
2081 }
2082
2083
2084 //---------------------- stride_of_possible_iv -------------------------------------
2085 // Looks for an iff/bool/comp with one operand of the compare
2086 // being a cycle involving an add and a phi,
2087 // with an optional truncation (left-shift followed by a right-shift)
2088 // of the add. Returns zero if not an iv.
2089 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2090 Node* trunc1 = NULL;
2091 Node* trunc2 = NULL;
2092 const TypeInt* ttype = NULL;
2093 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
2094 return 0;
2095 }
2096 BoolNode* bl = iff->in(1)->as_Bool();
2097 Node* cmp = bl->in(1);
2098 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2099 return 0;
2100 }
2101 // Must have an invariant operand
2102 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2103 return 0;
2104 }
2105 Node* add2 = NULL;
2106 Node* cmp1 = cmp->in(1);
2107 if (cmp1->is_Phi()) {
2108 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2109 Node* phi = cmp1;
2110 for (uint i = 1; i < phi->req(); i++) {
2111 Node* in = phi->in(i);
2112 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2113 &trunc1, &trunc2, &ttype);
2114 if (add && add->in(1) == phi) {
2115 add2 = add->in(2);
2116 break;
2117 }
2118 }
2119 } else {
2120 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2121 Node* addtrunc = cmp1;
2122 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2123 &trunc1, &trunc2, &ttype);
2124 if (add && add->in(1)->is_Phi()) {
2125 Node* phi = add->in(1);
2126 for (uint i = 1; i < phi->req(); i++) {
2127 if (phi->in(i) == addtrunc) {
2128 add2 = add->in(2);
2129 break;
2130 }
2131 }
2132 }
2133 }
2134 if (add2 != NULL) {
2135 const TypeInt* add2t = _igvn.type(add2)->is_int();
2136 if (add2t->is_con()) {
2137 return add2t->get_con();
2138 }
2139 }
2140 return 0;
2141 }
2142
2143
2144 //---------------------- stay_in_loop -------------------------------------
2145 // Return the (unique) control output node that's in the loop (if it exists.)
2146 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2147 Node* unique = NULL;
2148 if (!n) return NULL;
2149 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2150 Node* use = n->fast_out(i);
2151 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2152 if (unique != NULL) {
2153 return NULL;
2154 }
2155 unique = use;
2156 }
2157 }
2158 return unique;
2159 }
2160
2161 //------------------------------ register_node -------------------------------------
2162 // Utility to register node "n" with PhaseIdealLoop
2163 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
2164 _igvn.register_new_node_with_optimizer(n);
2165 loop->_body.push(n);
2166 if (n->is_CFG()) {
2167 set_loop(n, loop);
2168 set_idom(n, pred, ddepth);
2169 } else {
2170 set_ctrl(n, pred);
2171 }
2172 }
2173
2174 //------------------------------ proj_clone -------------------------------------
2175 // Utility to create an if-projection
2176 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2177 ProjNode* c = p->clone()->as_Proj();
2178 c->set_req(0, iff);
2179 return c;
2180 }
2181
2182 //------------------------------ short_circuit_if -------------------------------------
2183 // Force the iff control output to be the live_proj
2184 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2185 guarantee(live_proj != NULL, "null projection");
2186 int proj_con = live_proj->_con;
2187 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2188 Node *con = _igvn.intcon(proj_con);
2189 set_ctrl(con, C->root());
2190 if (iff) {
2191 iff->set_req(1, con);
2192 }
2193 return con;
2194 }
2195
2196 //------------------------------ insert_if_before_proj -------------------------------------
2197 // Insert a new if before an if projection (* - new node)
2198 //
2199 // before
2200 // if(test)
2201 // / \
2202 // v v
2203 // other-proj proj (arg)
2204 //
2205 // after
2206 // if(test)
2207 // / \
2208 // / v
2209 // | * proj-clone
2210 // v |
2211 // other-proj v
2212 // * new_if(relop(cmp[IU](left,right)))
2213 // / \
2214 // v v
2215 // * new-proj proj
2216 // (returned)
2217 //
2218 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2219 IfNode* iff = proj->in(0)->as_If();
2220 IdealLoopTree *loop = get_loop(proj);
2221 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2222 int ddepth = dom_depth(proj);
2223
2224 _igvn.rehash_node_delayed(iff);
2225 _igvn.rehash_node_delayed(proj);
2226
2227 proj->set_req(0, NULL); // temporary disconnect
2228 ProjNode* proj2 = proj_clone(proj, iff);
2229 register_node(proj2, loop, iff, ddepth);
2230
2231 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2232 register_node(cmp, loop, proj2, ddepth);
2233
2234 BoolNode* bol = new BoolNode(cmp, relop);
2235 register_node(bol, loop, proj2, ddepth);
2236
2237 int opcode = iff->Opcode();
2238 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2239 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
2240 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
2241 register_node(new_if, loop, proj2, ddepth);
2242
2243 proj->set_req(0, new_if); // reattach
2244 set_idom(proj, new_if, ddepth);
2245
2246 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2247 guarantee(new_exit != NULL, "null exit node");
2248 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2249
2250 return new_exit;
2251 }
2252
2253 //------------------------------ insert_region_before_proj -------------------------------------
2254 // Insert a region before an if projection (* - new node)
2255 //
2256 // before
2257 // if(test)
2258 // / |
2259 // v |
2260 // proj v
2261 // other-proj
2262 //
2263 // after
2264 // if(test)
2265 // / |
2266 // v |
2267 // * proj-clone v
2268 // | other-proj
2269 // v
2270 // * new-region
2271 // |
2272 // v
2273 // * dum_if
2274 // / \
2275 // v \
2276 // * dum-proj v
2277 // proj
2278 //
2279 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2280 IfNode* iff = proj->in(0)->as_If();
2281 IdealLoopTree *loop = get_loop(proj);
2282 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2283 int ddepth = dom_depth(proj);
2284
2285 _igvn.rehash_node_delayed(iff);
2286 _igvn.rehash_node_delayed(proj);
2287
2288 proj->set_req(0, NULL); // temporary disconnect
2289 ProjNode* proj2 = proj_clone(proj, iff);
2290 register_node(proj2, loop, iff, ddepth);
2291
2292 RegionNode* reg = new RegionNode(2);
2293 reg->set_req(1, proj2);
2294 register_node(reg, loop, iff, ddepth);
2295
2296 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
2297 register_node(dum_if, loop, reg, ddepth);
2298
2299 proj->set_req(0, dum_if); // reattach
2300 set_idom(proj, dum_if, ddepth);
2301
2302 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
2303 register_node(dum_proj, loop, dum_if, ddepth);
2304
2305 return reg;
2306 }
2307
2308 //------------------------------ insert_cmpi_loop_exit -------------------------------------
2309 // Clone a signed compare loop exit from an unsigned compare and
2310 // insert it before the unsigned cmp on the stay-in-loop path.
2311 // All new nodes inserted in the dominator tree between the original
2312 // if and it's projections. The original if test is replaced with
2313 // a constant to force the stay-in-loop path.
2314 //
2315 // This is done to make sure that the original if and it's projections
2316 // still dominate the same set of control nodes, that the ctrl() relation
2317 // from data nodes to them is preserved, and that their loop nesting is
2318 // preserved.
2319 //
2320 // before
2321 // if(i <u limit) unsigned compare loop exit
2322 // / |
2323 // v v
2324 // exit-proj stay-in-loop-proj
2325 //
2326 // after
2327 // if(stay-in-loop-const) original if
2328 // / |
2329 // / v
2330 // / if(i < limit) new signed test
2331 // / / |
2332 // / / v
2333 // / / if(i <u limit) new cloned unsigned test
2334 // / / / |
2335 // v v v |
2336 // region |
2337 // | |
2338 // dum-if |
2339 // / | |
2340 // ether | |
2341 // v v
2342 // exit-proj stay-in-loop-proj
2343 //
2344 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
2345 const bool Signed = true;
2346 const bool Unsigned = false;
2347
2348 BoolNode* bol = if_cmpu->in(1)->as_Bool();
2349 if (bol->_test._test != BoolTest::lt) return NULL;
2350 CmpNode* cmpu = bol->in(1)->as_Cmp();
2351 if (cmpu->Opcode() != Op_CmpU) return NULL;
2352 int stride = stride_of_possible_iv(if_cmpu);
2353 if (stride == 0) return NULL;
2354
2355 Node* lp_proj = stay_in_loop(if_cmpu, loop);
2356 guarantee(lp_proj != NULL, "null loop node");
2357
2358 ProjNode* lp_continue = lp_proj->as_Proj();
2359 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
2360
2361 Node* limit = NULL;
2362 if (stride > 0) {
2363 limit = cmpu->in(2);
2364 } else {
2365 limit = _igvn.makecon(TypeInt::ZERO);
2366 set_ctrl(limit, C->root());
2367 }
2368 // Create a new region on the exit path
2369 RegionNode* reg = insert_region_before_proj(lp_exit);
2370 guarantee(reg != NULL, "null region node");
2371
2372 // Clone the if-cmpu-true-false using a signed compare
2373 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
2374 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
2375 reg->add_req(cmpi_exit);
2376
2377 // Clone the if-cmpu-true-false
2378 BoolTest::mask rel_u = bol->_test._test;
2379 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
2380 reg->add_req(cmpu_exit);
2381
2382 // Force original if to stay in loop.
2383 short_circuit_if(if_cmpu, lp_continue);
2384
2385 return cmpi_exit->in(0)->as_If();
2386 }
2387
2388 //------------------------------ remove_cmpi_loop_exit -------------------------------------
2389 // Remove a previously inserted signed compare loop exit.
2390 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
2391 Node* lp_proj = stay_in_loop(if_cmp, loop);
2392 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
2393 stay_in_loop(lp_proj, loop)->is_If() &&
2394 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
2395 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
2396 set_ctrl(con, C->root());
2397 if_cmp->set_req(1, con);
2398 }
2399
2400 //------------------------------ scheduled_nodelist -------------------------------------
2401 // Create a post order schedule of nodes that are in the
2402 // "member" set. The list is returned in "sched".
2403 // The first node in "sched" is the loop head, followed by
2404 // nodes which have no inputs in the "member" set, and then
2405 // followed by the nodes that have an immediate input dependence
2406 // on a node in "sched".
2407 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
2408
2409 assert(member.test(loop->_head->_idx), "loop head must be in member set");
2410 Arena *a = Thread::current()->resource_area();
2411 VectorSet visited(a);
2412 Node_Stack nstack(a, loop->_body.size());
2413
2414 Node* n = loop->_head; // top of stack is cached in "n"
2415 uint idx = 0;
2416 visited.set(n->_idx);
2417
2418 // Initially push all with no inputs from within member set
2419 for(uint i = 0; i < loop->_body.size(); i++ ) {
2420 Node *elt = loop->_body.at(i);
2421 if (member.test(elt->_idx)) {
2422 bool found = false;
2423 for (uint j = 0; j < elt->req(); j++) {
2424 Node* def = elt->in(j);
2425 if (def && member.test(def->_idx) && def != elt) {
2426 found = true;
2427 break;
2428 }
2429 }
2430 if (!found && elt != loop->_head) {
2431 nstack.push(n, idx);
2432 n = elt;
2433 assert(!visited.test(n->_idx), "not seen yet");
2434 visited.set(n->_idx);
2435 }
2436 }
2437 }
2438
2439 // traverse out's that are in the member set
2440 while (true) {
2441 if (idx < n->outcnt()) {
2442 Node* use = n->raw_out(idx);
2443 idx++;
2444 if (!visited.test_set(use->_idx)) {
2445 if (member.test(use->_idx)) {
2446 nstack.push(n, idx);
2447 n = use;
2448 idx = 0;
2449 }
2450 }
2451 } else {
2452 // All outputs processed
2453 sched.push(n);
2454 if (nstack.is_empty()) break;
2455 n = nstack.node();
2456 idx = nstack.index();
2457 nstack.pop();
2458 }
2459 }
2460 }
2461
2462
2463 //------------------------------ has_use_in_set -------------------------------------
2464 // Has a use in the vector set
2465 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
2466 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2467 Node* use = n->fast_out(j);
2468 if (vset.test(use->_idx)) {
2469 return true;
2470 }
2471 }
2472 return false;
2473 }
2474
2475
2476 //------------------------------ has_use_internal_to_set -------------------------------------
2477 // Has use internal to the vector set (ie. not in a phi at the loop head)
2478 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
2479 Node* head = loop->_head;
2480 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2481 Node* use = n->fast_out(j);
2482 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
2483 return true;
2484 }
2485 }
2486 return false;
2487 }
2488
2489
2490 //------------------------------ clone_for_use_outside_loop -------------------------------------
2491 // clone "n" for uses that are outside of loop
2492 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
2493 int cloned = 0;
2494 assert(worklist.size() == 0, "should be empty");
2495 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2496 Node* use = n->fast_out(j);
2497 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
2498 worklist.push(use);
2499 }
2500 }
2501 while( worklist.size() ) {
2502 Node *use = worklist.pop();
2503 if (!has_node(use) || use->in(0) == C->top()) continue;
2504 uint j;
2505 for (j = 0; j < use->req(); j++) {
2506 if (use->in(j) == n) break;
2507 }
2508 assert(j < use->req(), "must be there");
2509
2510 // clone "n" and insert it between the inputs of "n" and the use outside the loop
2511 Node* n_clone = n->clone();
2512 _igvn.replace_input_of(use, j, n_clone);
2513 cloned++;
2514 Node* use_c;
2515 if (!use->is_Phi()) {
2516 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
2517 } else {
2518 // Use in a phi is considered a use in the associated predecessor block
2519 use_c = use->in(0)->in(j);
2520 }
2521 set_ctrl(n_clone, use_c);
2522 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
2523 get_loop(use_c)->_body.push(n_clone);
2524 _igvn.register_new_node_with_optimizer(n_clone);
2525 #if !defined(PRODUCT)
2526 if (TracePartialPeeling) {
2527 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
2528 }
2529 #endif
2530 }
2531 return cloned;
2532 }
2533
2534
2535 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
2536 // clone "n" for special uses that are in the not_peeled region.
2537 // If these def-uses occur in separate blocks, the code generator
2538 // marks the method as not compilable. For example, if a "BoolNode"
2539 // is in a different basic block than the "IfNode" that uses it, then
2540 // the compilation is aborted in the code generator.
2541 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
2542 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
2543 if (n->is_Phi() || n->is_Load()) {
2544 return;
2545 }
2546 assert(worklist.size() == 0, "should be empty");
2547 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2548 Node* use = n->fast_out(j);
2549 if ( not_peel.test(use->_idx) &&
2550 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
2551 use->in(1) == n) {
2552 worklist.push(use);
2553 }
2554 }
2555 if (worklist.size() > 0) {
2556 // clone "n" and insert it between inputs of "n" and the use
2557 Node* n_clone = n->clone();
2558 loop->_body.push(n_clone);
2559 _igvn.register_new_node_with_optimizer(n_clone);
2560 set_ctrl(n_clone, get_ctrl(n));
2561 sink_list.push(n_clone);
2562 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
2563 #if !defined(PRODUCT)
2564 if (TracePartialPeeling) {
2565 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2566 }
2567 #endif
2568 while( worklist.size() ) {
2569 Node *use = worklist.pop();
2570 _igvn.rehash_node_delayed(use);
2571 for (uint j = 1; j < use->req(); j++) {
2572 if (use->in(j) == n) {
2573 use->set_req(j, n_clone);
2574 }
2575 }
2576 }
2577 }
2578 }
2579
2580
2581 //------------------------------ insert_phi_for_loop -------------------------------------
2582 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2583 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2584 Node *phi = PhiNode::make(lp, back_edge_val);
2585 phi->set_req(LoopNode::EntryControl, lp_entry_val);
2586 // Use existing phi if it already exists
2587 Node *hit = _igvn.hash_find_insert(phi);
2588 if( hit == NULL ) {
2589 _igvn.register_new_node_with_optimizer(phi);
2590 set_ctrl(phi, lp);
2591 } else {
2592 // Remove the new phi from the graph and use the hit
2593 _igvn.remove_dead_node(phi);
2594 phi = hit;
2595 }
2596 _igvn.replace_input_of(use, idx, phi);
2597 }
2598
2599 #ifdef ASSERT
2600 //------------------------------ is_valid_loop_partition -------------------------------------
2601 // Validate the loop partition sets: peel and not_peel
2602 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2603 VectorSet& not_peel ) {
2604 uint i;
2605 // Check that peel_list entries are in the peel set
2606 for (i = 0; i < peel_list.size(); i++) {
2607 if (!peel.test(peel_list.at(i)->_idx)) {
2608 return false;
2609 }
2610 }
2611 // Check at loop members are in one of peel set or not_peel set
2612 for (i = 0; i < loop->_body.size(); i++ ) {
2613 Node *def = loop->_body.at(i);
2614 uint di = def->_idx;
2615 // Check that peel set elements are in peel_list
2616 if (peel.test(di)) {
2617 if (not_peel.test(di)) {
2618 return false;
2619 }
2620 // Must be in peel_list also
2621 bool found = false;
2622 for (uint j = 0; j < peel_list.size(); j++) {
2623 if (peel_list.at(j)->_idx == di) {
2624 found = true;
2625 break;
2626 }
2627 }
2628 if (!found) {
2629 return false;
2630 }
2631 } else if (not_peel.test(di)) {
2632 if (peel.test(di)) {
2633 return false;
2634 }
2635 } else {
2636 return false;
2637 }
2638 }
2639 return true;
2640 }
2641
2642 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2643 // Ensure a use outside of loop is of the right form
2644 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2645 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2646 return (use->is_Phi() &&
2647 use_c->is_Region() && use_c->req() == 3 &&
2648 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2649 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2650 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2651 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2652 }
2653
2654 //------------------------------ is_valid_clone_loop_form -------------------------------------
2655 // Ensure that all uses outside of loop are of the right form
2656 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2657 uint orig_exit_idx, uint clone_exit_idx) {
2658 uint len = peel_list.size();
2659 for (uint i = 0; i < len; i++) {
2660 Node *def = peel_list.at(i);
2661
2662 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2663 Node *use = def->fast_out(j);
2664 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2665 if (!loop->is_member(get_loop(use_c))) {
2666 // use is not in the loop, check for correct structure
2667 if (use->in(0) == def) {
2668 // Okay
2669 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2670 return false;
2671 }
2672 }
2673 }
2674 }
2675 return true;
2676 }
2677 #endif
2678
2679 //------------------------------ partial_peel -------------------------------------
2680 // Partially peel (aka loop rotation) the top portion of a loop (called
2681 // the peel section below) by cloning it and placing one copy just before
2682 // the new loop head and the other copy at the bottom of the new loop.
2683 //
2684 // before after where it came from
2685 //
2686 // stmt1 stmt1
2687 // loop: stmt2 clone
2688 // stmt2 if condA goto exitA clone
2689 // if condA goto exitA new_loop: new
2690 // stmt3 stmt3 clone
2691 // if !condB goto loop if condB goto exitB clone
2692 // exitB: stmt2 orig
2693 // stmt4 if !condA goto new_loop orig
2694 // exitA: goto exitA
2695 // exitB:
2696 // stmt4
2697 // exitA:
2698 //
2699 // Step 1: find the cut point: an exit test on probable
2700 // induction variable.
2701 // Step 2: schedule (with cloning) operations in the peel
2702 // section that can be executed after the cut into
2703 // the section that is not peeled. This may need
2704 // to clone operations into exit blocks. For
2705 // instance, a reference to A[i] in the not-peel
2706 // section and a reference to B[i] in an exit block
2707 // may cause a left-shift of i by 2 to be placed
2708 // in the peel block. This step will clone the left
2709 // shift into the exit block and sink the left shift
2710 // from the peel to the not-peel section.
2711 // Step 3: clone the loop, retarget the control, and insert
2712 // phis for values that are live across the new loop
2713 // head. This is very dependent on the graph structure
2714 // from clone_loop. It creates region nodes for
2715 // exit control and associated phi nodes for values
2716 // flow out of the loop through that exit. The region
2717 // node is dominated by the clone's control projection.
2718 // So the clone's peel section is placed before the
2719 // new loop head, and the clone's not-peel section is
2720 // forms the top part of the new loop. The original
2721 // peel section forms the tail of the new loop.
2722 // Step 4: update the dominator tree and recompute the
2723 // dominator depth.
2724 //
2725 // orig
2726 //
2727 // stmt1
2728 // |
2729 // v
2730 // loop predicate
2731 // |
2732 // v
2733 // loop<----+
2734 // | |
2735 // stmt2 |
2736 // | |
2737 // v |
2738 // ifA |
2739 // / | |
2740 // v v |
2741 // false true ^ <-- last_peel
2742 // / | |
2743 // / ===|==cut |
2744 // / stmt3 | <-- first_not_peel
2745 // / | |
2746 // | v |
2747 // v ifB |
2748 // exitA: / \ |
2749 // / \ |
2750 // v v |
2751 // false true |
2752 // / \ |
2753 // / ----+
2754 // |
2755 // v
2756 // exitB:
2757 // stmt4
2758 //
2759 //
2760 // after clone loop
2761 //
2762 // stmt1
2763 // |
2764 // v
2765 // loop predicate
2766 // / \
2767 // clone / \ orig
2768 // / \
2769 // / \
2770 // v v
2771 // +---->loop loop<----+
2772 // | | | |
2773 // | stmt2 stmt2 |
2774 // | | | |
2775 // | v v |
2776 // | ifA ifA |
2777 // | | \ / | |
2778 // | v v v v |
2779 // ^ true false false true ^ <-- last_peel
2780 // | | ^ \ / | |
2781 // | cut==|== \ \ / ===|==cut |
2782 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2783 // | | dom | | | |
2784 // | v \ 1v v2 v |
2785 // | ifB regionA ifB |
2786 // | / \ | / \ |
2787 // | / \ v / \ |
2788 // | v v exitA: v v |
2789 // | true false false true |
2790 // | / ^ \ / \ |
2791 // +---- \ \ / ----+
2792 // dom \ /
2793 // \ 1v v2
2794 // regionB
2795 // |
2796 // v
2797 // exitB:
2798 // stmt4
2799 //
2800 //
2801 // after partial peel
2802 //
2803 // stmt1
2804 // |
2805 // v
2806 // loop predicate
2807 // /
2808 // clone / orig
2809 // / TOP
2810 // / \
2811 // v v
2812 // TOP->loop loop----+
2813 // | | |
2814 // stmt2 stmt2 |
2815 // | | |
2816 // v v |
2817 // ifA ifA |
2818 // | \ / | |
2819 // v v v v |
2820 // true false false true | <-- last_peel
2821 // | ^ \ / +------|---+
2822 // +->newloop \ \ / === ==cut | |
2823 // | stmt3 \ \ / TOP | |
2824 // | | dom | | stmt3 | | <-- first_not_peel
2825 // | v \ 1v v2 v | |
2826 // | ifB regionA ifB ^ v
2827 // | / \ | / \ | |
2828 // | / \ v / \ | |
2829 // | v v exitA: v v | |
2830 // | true false false true | |
2831 // | / ^ \ / \ | |
2832 // | | \ \ / v | |
2833 // | | dom \ / TOP | |
2834 // | | \ 1v v2 | |
2835 // ^ v regionB | |
2836 // | | | | |
2837 // | | v ^ v
2838 // | | exitB: | |
2839 // | | stmt4 | |
2840 // | +------------>-----------------+ |
2841 // | |
2842 // +-----------------<---------------------+
2843 //
2844 //
2845 // final graph
2846 //
2847 // stmt1
2848 // |
2849 // v
2850 // loop predicate
2851 // |
2852 // v
2853 // stmt2 clone
2854 // |
2855 // v
2856 // ........> ifA clone
2857 // : / |
2858 // dom / |
2859 // : v v
2860 // : false true
2861 // : | |
2862 // : | v
2863 // : | newloop<-----+
2864 // : | | |
2865 // : | stmt3 clone |
2866 // : | | |
2867 // : | v |
2868 // : | ifB |
2869 // : | / \ |
2870 // : | v v |
2871 // : | false true |
2872 // : | | | |
2873 // : | v stmt2 |
2874 // : | exitB: | |
2875 // : | stmt4 v |
2876 // : | ifA orig |
2877 // : | / \ |
2878 // : | / \ |
2879 // : | v v |
2880 // : | false true |
2881 // : | / \ |
2882 // : v v -----+
2883 // RegionA
2884 // |
2885 // v
2886 // exitA
2887 //
2888 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
2889
2890 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
2891 if (!loop->_head->is_Loop()) {
2892 return false; }
2893
2894 LoopNode *head = loop->_head->as_Loop();
2895
2896 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
2897 return false;
2898 }
2899
2900 // Check for complex exit control
2901 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
2902 Node *n = loop->_body.at(ii);
2903 int opc = n->Opcode();
2904 if (n->is_Call() ||
2905 opc == Op_Catch ||
2906 opc == Op_CatchProj ||
2907 opc == Op_Jump ||
2908 opc == Op_JumpProj) {
2909 #if !defined(PRODUCT)
2910 if (TracePartialPeeling) {
2911 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
2912 }
2913 #endif
2914 return false;
2915 }
2916 }
2917
2918 int dd = dom_depth(head);
2919
2920 // Step 1: find cut point
2921
2922 // Walk up dominators to loop head looking for first loop exit
2923 // which is executed on every path thru loop.
2924 IfNode *peel_if = NULL;
2925 IfNode *peel_if_cmpu = NULL;
2926
2927 Node *iff = loop->tail();
2928 while( iff != head ) {
2929 if( iff->is_If() ) {
2930 Node *ctrl = get_ctrl(iff->in(1));
2931 if (ctrl->is_top()) return false; // Dead test on live IF.
2932 // If loop-varying exit-test, check for induction variable
2933 if( loop->is_member(get_loop(ctrl)) &&
2934 loop->is_loop_exit(iff) &&
2935 is_possible_iv_test(iff)) {
2936 Node* cmp = iff->in(1)->in(1);
2937 if (cmp->Opcode() == Op_CmpI) {
2938 peel_if = iff->as_If();
2939 } else {
2940 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
2941 peel_if_cmpu = iff->as_If();
2942 }
2943 }
2944 }
2945 iff = idom(iff);
2946 }
2947 // Prefer signed compare over unsigned compare.
2948 IfNode* new_peel_if = NULL;
2949 if (peel_if == NULL) {
2950 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
2951 return false; // No peel point found
2952 }
2953 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
2954 if (new_peel_if == NULL) {
2955 return false; // No peel point found
2956 }
2957 peel_if = new_peel_if;
2958 }
2959 Node* last_peel = stay_in_loop(peel_if, loop);
2960 Node* first_not_peeled = stay_in_loop(last_peel, loop);
2961 if (first_not_peeled == NULL || first_not_peeled == head) {
2962 return false;
2963 }
2964
2965 #if !defined(PRODUCT)
2966 if (TraceLoopOpts) {
2967 tty->print("PartialPeel ");
2968 loop->dump_head();
2969 }
2970
2971 if (TracePartialPeeling) {
2972 tty->print_cr("before partial peel one iteration");
2973 Node_List wl;
2974 Node* t = head->in(2);
2975 while (true) {
2976 wl.push(t);
2977 if (t == head) break;
2978 t = idom(t);
2979 }
2980 while (wl.size() > 0) {
2981 Node* tt = wl.pop();
2982 tt->dump();
2983 if (tt == last_peel) tty->print_cr("-- cut --");
2984 }
2985 }
2986 #endif
2987 ResourceArea *area = Thread::current()->resource_area();
2988 VectorSet peel(area);
2989 VectorSet not_peel(area);
2990 Node_List peel_list(area);
2991 Node_List worklist(area);
2992 Node_List sink_list(area);
2993
2994 // Set of cfg nodes to peel are those that are executable from
2995 // the head through last_peel.
2996 assert(worklist.size() == 0, "should be empty");
2997 worklist.push(head);
2998 peel.set(head->_idx);
2999 while (worklist.size() > 0) {
3000 Node *n = worklist.pop();
3001 if (n != last_peel) {
3002 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3003 Node* use = n->fast_out(j);
3004 if (use->is_CFG() &&
3005 loop->is_member(get_loop(use)) &&
3006 !peel.test_set(use->_idx)) {
3007 worklist.push(use);
3008 }
3009 }
3010 }
3011 }
3012
3013 // Set of non-cfg nodes to peel are those that are control
3014 // dependent on the cfg nodes.
3015 uint i;
3016 for(i = 0; i < loop->_body.size(); i++ ) {
3017 Node *n = loop->_body.at(i);
3018 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3019 if (peel.test(n_c->_idx)) {
3020 peel.set(n->_idx);
3021 } else {
3022 not_peel.set(n->_idx);
3023 }
3024 }
3025
3026 // Step 2: move operations from the peeled section down into the
3027 // not-peeled section
3028
3029 // Get a post order schedule of nodes in the peel region
3030 // Result in right-most operand.
3031 scheduled_nodelist(loop, peel, peel_list );
3032
3033 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3034
3035 // For future check for too many new phis
3036 uint old_phi_cnt = 0;
3037 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3038 Node* use = head->fast_out(j);
3039 if (use->is_Phi()) old_phi_cnt++;
3040 }
3041
3042 #if !defined(PRODUCT)
3043 if (TracePartialPeeling) {
3044 tty->print_cr("\npeeled list");
3045 }
3046 #endif
3047
3048 // Evacuate nodes in peel region into the not_peeled region if possible
3049 uint new_phi_cnt = 0;
3050 uint cloned_for_outside_use = 0;
3051 for (i = 0; i < peel_list.size();) {
3052 Node* n = peel_list.at(i);
3053 #if !defined(PRODUCT)
3054 if (TracePartialPeeling) n->dump();
3055 #endif
3056 bool incr = true;
3057 if ( !n->is_CFG() ) {
3058
3059 if ( has_use_in_set(n, not_peel) ) {
3060
3061 // If not used internal to the peeled region,
3062 // move "n" from peeled to not_peeled region.
3063
3064 if ( !has_use_internal_to_set(n, peel, loop) ) {
3065
3066 // if not pinned and not a load (which maybe anti-dependent on a store)
3067 // and not a CMove (Matcher expects only bool->cmove).
3068 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
3069 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist );
3070 sink_list.push(n);
3071 peel >>= n->_idx; // delete n from peel set.
3072 not_peel <<= n->_idx; // add n to not_peel set.
3073 peel_list.remove(i);
3074 incr = false;
3075 #if !defined(PRODUCT)
3076 if (TracePartialPeeling) {
3077 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3078 n->_idx, get_ctrl(n)->_idx);
3079 }
3080 #endif
3081 }
3082 } else {
3083 // Otherwise check for special def-use cases that span
3084 // the peel/not_peel boundary such as bool->if
3085 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
3086 new_phi_cnt++;
3087 }
3088 }
3089 }
3090 if (incr) i++;
3091 }
3092
3093 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
3094 #if !defined(PRODUCT)
3095 if (TracePartialPeeling) {
3096 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3097 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
3098 }
3099 #endif
3100 if (new_peel_if != NULL) {
3101 remove_cmpi_loop_exit(new_peel_if, loop);
3102 }
3103 // Inhibit more partial peeling on this loop
3104 assert(!head->is_partial_peel_loop(), "not partial peeled");
3105 head->mark_partial_peel_failed();
3106 if (cloned_for_outside_use > 0) {
3107 // Terminate this round of loop opts because
3108 // the graph outside this loop was changed.
3109 C->set_major_progress();
3110 return true;
3111 }
3112 return false;
3113 }
3114
3115 // Step 3: clone loop, retarget control, and insert new phis
3116
3117 // Create new loop head for new phis and to hang
3118 // the nodes being moved (sinked) from the peel region.
3119 LoopNode* new_head = new LoopNode(last_peel, last_peel);
3120 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3121 _igvn.register_new_node_with_optimizer(new_head);
3122 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3123 _igvn.replace_input_of(first_not_peeled, 0, new_head);
3124 set_loop(new_head, loop);
3125 loop->_body.push(new_head);
3126 not_peel.set(new_head->_idx);
3127 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3128 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3129
3130 while (sink_list.size() > 0) {
3131 Node* n = sink_list.pop();
3132 set_ctrl(n, new_head);
3133 }
3134
3135 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3136
3137 clone_loop(loop, old_new, dd, IgnoreStripMined);
3138
3139 const uint clone_exit_idx = 1;
3140 const uint orig_exit_idx = 2;
3141 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
3142
3143 Node* head_clone = old_new[head->_idx];
3144 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
3145 Node* orig_tail_clone = head_clone->in(2);
3146
3147 // Add phi if "def" node is in peel set and "use" is not
3148
3149 for(i = 0; i < peel_list.size(); i++ ) {
3150 Node *def = peel_list.at(i);
3151 if (!def->is_CFG()) {
3152 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3153 Node *use = def->fast_out(j);
3154 if (has_node(use) && use->in(0) != C->top() &&
3155 (!peel.test(use->_idx) ||
3156 (use->is_Phi() && use->in(0) == head)) ) {
3157 worklist.push(use);
3158 }
3159 }
3160 while( worklist.size() ) {
3161 Node *use = worklist.pop();
3162 for (uint j = 1; j < use->req(); j++) {
3163 Node* n = use->in(j);
3164 if (n == def) {
3165
3166 // "def" is in peel set, "use" is not in peel set
3167 // or "use" is in the entry boundary (a phi) of the peel set
3168
3169 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
3170
3171 if ( loop->is_member(get_loop( use_c )) ) {
3172 // use is in loop
3173 if (old_new[use->_idx] != NULL) { // null for dead code
3174 Node* use_clone = old_new[use->_idx];
3175 _igvn.replace_input_of(use, j, C->top());
3176 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
3177 }
3178 } else {
3179 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
3180 // use is not in the loop, check if the live range includes the cut
3181 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
3182 if (not_peel.test(lp_if->_idx)) {
3183 assert(j == orig_exit_idx, "use from original loop");
3184 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
3185 }
3186 }
3187 }
3188 }
3189 }
3190 }
3191 }
3192
3193 // Step 3b: retarget control
3194
3195 // Redirect control to the new loop head if a cloned node in
3196 // the not_peeled region has control that points into the peeled region.
3197 // This necessary because the cloned peeled region will be outside
3198 // the loop.
3199 // from to
3200 // cloned-peeled <---+
3201 // new_head_clone: | <--+
3202 // cloned-not_peeled in(0) in(0)
3203 // orig-peeled
3204
3205 for(i = 0; i < loop->_body.size(); i++ ) {
3206 Node *n = loop->_body.at(i);
3207 if (!n->is_CFG() && n->in(0) != NULL &&
3208 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
3209 Node* n_clone = old_new[n->_idx];
3210 _igvn.replace_input_of(n_clone, 0, new_head_clone);
3211 }
3212 }
3213
3214 // Backedge of the surviving new_head (the clone) is original last_peel
3215 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
3216
3217 // Cut first node in original not_peel set
3218 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
3219 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
3220 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
3221
3222 // Copy head_clone back-branch info to original head
3223 // and remove original head's loop entry and
3224 // clone head's back-branch
3225 _igvn.rehash_node_delayed(head); // Multiple edge updates
3226 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
3227 head->set_req(LoopNode::LoopBackControl, C->top());
3228 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
3229
3230 // Similarly modify the phis
3231 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
3232 Node* use = head->fast_out(k);
3233 if (use->is_Phi() && use->outcnt() > 0) {
3234 Node* use_clone = old_new[use->_idx];
3235 _igvn.rehash_node_delayed(use); // Multiple edge updates
3236 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
3237 use->set_req(LoopNode::LoopBackControl, C->top());
3238 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
3239 }
3240 }
3241
3242 // Step 4: update dominator tree and dominator depth
3243
3244 set_idom(head, orig_tail_clone, dd);
3245 recompute_dom_depth();
3246
3247 // Inhibit more partial peeling on this loop
3248 new_head_clone->set_partial_peel_loop();
3249 C->set_major_progress();
3250 loop->record_for_igvn();
3251
3252 #if !defined(PRODUCT)
3253 if (TracePartialPeeling) {
3254 tty->print_cr("\nafter partial peel one iteration");
3255 Node_List wl(area);
3256 Node* t = last_peel;
3257 while (true) {
3258 wl.push(t);
3259 if (t == head_clone) break;
3260 t = idom(t);
3261 }
3262 while (wl.size() > 0) {
3263 Node* tt = wl.pop();
3264 if (tt == head) tty->print_cr("orig head");
3265 else if (tt == new_head_clone) tty->print_cr("new head");
3266 else if (tt == head_clone) tty->print_cr("clone head");
3267 tt->dump();
3268 }
3269 }
3270 #endif
3271 return true;
3272 }
3273
3274 //------------------------------reorg_offsets----------------------------------
3275 // Reorganize offset computations to lower register pressure. Mostly
3276 // prevent loop-fallout uses of the pre-incremented trip counter (which are
3277 // then alive with the post-incremented trip counter forcing an extra
3278 // register move)
3279 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
3280 // Perform it only for canonical counted loops.
3281 // Loop's shape could be messed up by iteration_split_impl.
3282 if (!loop->_head->is_CountedLoop())
3283 return;
3284 if (!loop->_head->as_Loop()->is_valid_counted_loop())
3285 return;
3286
3287 CountedLoopNode *cl = loop->_head->as_CountedLoop();
3288 CountedLoopEndNode *cle = cl->loopexit();
3289 Node *exit = cle->proj_out(false);
3290 Node *phi = cl->phi();
3291
3292 // Check for the special case of folks using the pre-incremented
3293 // trip-counter on the fall-out path (forces the pre-incremented
3294 // and post-incremented trip counter to be live at the same time).
3295 // Fix this by adjusting to use the post-increment trip counter.
3296
3297 bool progress = true;
3298 while (progress) {
3299 progress = false;
3300 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3301 Node* use = phi->fast_out(i); // User of trip-counter
3302 if (!has_ctrl(use)) continue;
3303 Node *u_ctrl = get_ctrl(use);
3304 if (use->is_Phi()) {
3305 u_ctrl = NULL;
3306 for (uint j = 1; j < use->req(); j++)
3307 if (use->in(j) == phi)
3308 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
3309 }
3310 IdealLoopTree *u_loop = get_loop(u_ctrl);
3311 // Look for loop-invariant use
3312 if (u_loop == loop) continue;
3313 if (loop->is_member(u_loop)) continue;
3314 // Check that use is live out the bottom. Assuming the trip-counter
3315 // update is right at the bottom, uses of of the loop middle are ok.
3316 if (dom_lca(exit, u_ctrl) != exit) continue;
3317 // Hit! Refactor use to use the post-incremented tripcounter.
3318 // Compute a post-increment tripcounter.
3319 Node *opaq = new Opaque2Node( C, cle->incr() );
3320 register_new_node(opaq, exit);
3321 Node *neg_stride = _igvn.intcon(-cle->stride_con());
3322 set_ctrl(neg_stride, C->root());
3323 Node *post = new AddINode( opaq, neg_stride);
3324 register_new_node(post, exit);
3325 _igvn.rehash_node_delayed(use);
3326 for (uint j = 1; j < use->req(); j++) {
3327 if (use->in(j) == phi)
3328 use->set_req(j, post);
3329 }
3330 // Since DU info changed, rerun loop
3331 progress = true;
3332 break;
3333 }
3334 }
3335
3336 }