1 /*
2 * Copyright (c) 1999, 2017, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "memory/allocation.inline.hpp"
27 #include "opto/addnode.hpp"
28 #include "opto/connode.hpp"
29 #include "opto/divnode.hpp"
30 #include "opto/loopnode.hpp"
31 #include "opto/matcher.hpp"
32 #include "opto/mulnode.hpp"
33 #include "opto/rootnode.hpp"
34 #include "opto/subnode.hpp"
35 #if INCLUDE_ALL_GCS
36 #include "gc_implementation/shenandoah/c2/shenandoahSupport.hpp"
37 #endif
38
39 //=============================================================================
40 //------------------------------split_thru_phi---------------------------------
41 // Split Node 'n' through merge point if there is enough win.
42 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
43 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
44 // ConvI2L may have type information on it which is unsafe to push up
45 // so disable this for now
46 return NULL;
47 }
48
49 // Splitting range check CastIIs through a loop induction Phi can
50 // cause new Phis to be created that are left unrelated to the loop
51 // induction Phi and prevent optimizations (vectorization)
52 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() &&
53 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) {
54 return NULL;
55 }
56
57 int wins = 0;
58 assert(!n->is_CFG(), "");
59 assert(region->is_Region(), "");
60
61 const Type* type = n->bottom_type();
62 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr();
63 Node *phi;
64 if (t_oop != NULL && t_oop->is_known_instance_field()) {
65 int iid = t_oop->instance_id();
66 int index = C->get_alias_index(t_oop);
67 int offset = t_oop->offset();
68 phi = new (C) PhiNode(region, type, NULL, iid, index, offset);
69 } else {
70 phi = PhiNode::make_blank(region, n);
71 }
72 uint old_unique = C->unique();
73 for (uint i = 1; i < region->req(); i++) {
74 Node *x;
75 Node* the_clone = NULL;
76 if (region->in(i) == C->top()) {
77 x = C->top(); // Dead path? Use a dead data op
78 } else {
79 x = n->clone(); // Else clone up the data op
80 the_clone = x; // Remember for possible deletion.
81 // Alter data node to use pre-phi inputs
82 if (n->in(0) == region)
83 x->set_req( 0, region->in(i) );
84 for (uint j = 1; j < n->req(); j++) {
85 Node *in = n->in(j);
86 if (in->is_Phi() && in->in(0) == region)
87 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
88 }
89 }
90 // Check for a 'win' on some paths
91 const Type *t = x->Value(&_igvn);
92
93 bool singleton = t->singleton();
94
95 // A TOP singleton indicates that there are no possible values incoming
96 // along a particular edge. In most cases, this is OK, and the Phi will
97 // be eliminated later in an Ideal call. However, we can't allow this to
98 // happen if the singleton occurs on loop entry, as the elimination of
99 // the PhiNode may cause the resulting node to migrate back to a previous
100 // loop iteration.
101 if (singleton && t == Type::TOP) {
102 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
103 // irreducible loop may not be indicated by an affirmative is_Loop());
104 // therefore, the only top we can split thru a phi is on a backedge of
105 // a loop.
106 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
107 }
108
109 if (singleton) {
110 wins++;
111 x = ((PhaseGVN&)_igvn).makecon(t);
112 } else {
113 // We now call Identity to try to simplify the cloned node.
114 // Note that some Identity methods call phase->type(this).
115 // Make sure that the type array is big enough for
116 // our new node, even though we may throw the node away.
117 // (Note: This tweaking with igvn only works because x is a new node.)
118 _igvn.set_type(x, t);
119 // If x is a TypeNode, capture any more-precise type permanently into Node
120 // otherwise it will be not updated during igvn->transform since
121 // igvn->type(x) is set to x->Value() already.
122 x->raise_bottom_type(t);
123 if (x->Opcode() != Op_ShenandoahLoadReferenceBarrier) {
124 Node *y = x->Identity(&_igvn);
125 if (y != x) {
126 wins++;
127 x = y;
128 } else {
129 y = _igvn.hash_find(x);
130 if (y) {
131 wins++;
132 x = y;
133 } else {
134 // Else x is a new node we are keeping
135 // We do not need register_new_node_with_optimizer
136 // because set_type has already been called.
137 _igvn._worklist.push(x);
138 }
139 }
140 } else {
141 _igvn._worklist.push(x);
142 }
143 }
144 if (x != the_clone && the_clone != NULL)
145 _igvn.remove_dead_node(the_clone);
146 phi->set_req( i, x );
147 }
148 // Too few wins?
149 if (wins <= policy) {
150 _igvn.remove_dead_node(phi);
151 return NULL;
152 }
153
154 // Record Phi
155 register_new_node( phi, region );
156
157 for (uint i2 = 1; i2 < phi->req(); i2++) {
158 Node *x = phi->in(i2);
159 // If we commoned up the cloned 'x' with another existing Node,
160 // the existing Node picks up a new use. We need to make the
161 // existing Node occur higher up so it dominates its uses.
162 Node *old_ctrl;
163 IdealLoopTree *old_loop;
164
165 if (x->is_Con()) {
166 // Constant's control is always root.
167 set_ctrl(x, C->root());
168 continue;
169 }
170 // The occasional new node
171 if (x->_idx >= old_unique) { // Found a new, unplaced node?
172 old_ctrl = NULL;
173 old_loop = NULL; // Not in any prior loop
174 } else {
175 old_ctrl = get_ctrl(x);
176 old_loop = get_loop(old_ctrl); // Get prior loop
177 }
178 // New late point must dominate new use
179 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
180 if (new_ctrl == old_ctrl) // Nothing is changed
181 continue;
182
183 IdealLoopTree *new_loop = get_loop(new_ctrl);
184
185 // Don't move x into a loop if its uses are
186 // outside of loop. Otherwise x will be cloned
187 // for each use outside of this loop.
188 IdealLoopTree *use_loop = get_loop(region);
189 if (!new_loop->is_member(use_loop) &&
190 (old_loop == NULL || !new_loop->is_member(old_loop))) {
191 // Take early control, later control will be recalculated
192 // during next iteration of loop optimizations.
193 new_ctrl = get_early_ctrl(x);
194 new_loop = get_loop(new_ctrl);
195 }
196 // Set new location
197 set_ctrl(x, new_ctrl);
198 // If changing loop bodies, see if we need to collect into new body
199 if (old_loop != new_loop) {
200 if (old_loop && !old_loop->_child)
201 old_loop->_body.yank(x);
202 if (!new_loop->_child)
203 new_loop->_body.push(x); // Collect body info
204 }
205 }
206
207 return phi;
208 }
209
210 //------------------------------dominated_by------------------------------------
211 // Replace the dominated test with an obvious true or false. Place it on the
212 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
213 // live path up to the dominating control.
214 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) {
215 #ifndef PRODUCT
216 if (VerifyLoopOptimizations && PrintOpto) tty->print_cr("dominating test");
217 #endif
218
219
220 // prevdom is the dominating projection of the dominating test.
221 assert( iff->is_If(), "" );
222 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
223 int pop = prevdom->Opcode();
224 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
225 if (flip) {
226 if (pop == Op_IfTrue)
227 pop = Op_IfFalse;
228 else
229 pop = Op_IfTrue;
230 }
231 // 'con' is set to true or false to kill the dominated test.
232 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
233 set_ctrl(con, C->root()); // Constant gets a new use
234 // Hack the dominated test
235 _igvn.replace_input_of(iff, 1, con);
236
237 // If I dont have a reachable TRUE and FALSE path following the IfNode then
238 // I can assume this path reaches an infinite loop. In this case it's not
239 // important to optimize the data Nodes - either the whole compilation will
240 // be tossed or this path (and all data Nodes) will go dead.
241 if (iff->outcnt() != 2) return;
242
243 // Make control-dependent data Nodes on the live path (path that will remain
244 // once the dominated IF is removed) become control-dependent on the
245 // dominating projection.
246 Node* dp = iff->as_If()->proj_out(pop == Op_IfTrue);
247
248 // Loop predicates may have depending checks which should not
249 // be skipped. For example, range check predicate has two checks
250 // for lower and upper bounds.
251 if (dp == NULL)
252 return;
253
254 ProjNode* dp_proj = dp->as_Proj();
255 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj();
256 if (exclude_loop_predicate &&
257 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) ||
258 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check))) {
259 // If this is a range check (IfNode::is_range_check), do not
260 // reorder because Compile::allow_range_check_smearing might have
261 // changed the check.
262 return; // Let IGVN transformation change control dependence.
263 }
264
265 IdealLoopTree *old_loop = get_loop(dp);
266
267 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
268 Node* cd = dp->fast_out(i); // Control-dependent node
269 if (cd->depends_only_on_test()) {
270 assert(cd->in(0) == dp, "");
271 _igvn.replace_input_of(cd, 0, prevdom);
272 set_early_ctrl(cd);
273 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
274 if (old_loop != new_loop) {
275 if (!old_loop->_child) old_loop->_body.yank(cd);
276 if (!new_loop->_child) new_loop->_body.push(cd);
277 }
278 --i;
279 --imax;
280 }
281 }
282 }
283
284 //------------------------------has_local_phi_input----------------------------
285 // Return TRUE if 'n' has Phi inputs from its local block and no other
286 // block-local inputs (all non-local-phi inputs come from earlier blocks)
287 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
288 Node *n_ctrl = get_ctrl(n);
289 // See if some inputs come from a Phi in this block, or from before
290 // this block.
291 uint i;
292 for( i = 1; i < n->req(); i++ ) {
293 Node *phi = n->in(i);
294 if( phi->is_Phi() && phi->in(0) == n_ctrl )
295 break;
296 }
297 if( i >= n->req() )
298 return NULL; // No Phi inputs; nowhere to clone thru
299
300 // Check for inputs created between 'n' and the Phi input. These
301 // must split as well; they have already been given the chance
302 // (courtesy of a post-order visit) and since they did not we must
303 // recover the 'cost' of splitting them by being very profitable
304 // when splitting 'n'. Since this is unlikely we simply give up.
305 for( i = 1; i < n->req(); i++ ) {
306 Node *m = n->in(i);
307 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
308 // We allow the special case of AddP's with no local inputs.
309 // This allows us to split-up address expressions.
310 if (m->is_AddP() &&
311 get_ctrl(m->in(2)) != n_ctrl &&
312 get_ctrl(m->in(3)) != n_ctrl) {
313 // Move the AddP up to dominating point
314 Node* c = find_non_split_ctrl(idom(n_ctrl));
315 set_ctrl_and_loop(m, c);
316 continue;
317 }
318 return NULL;
319 }
320 assert(n->is_Phi() || 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 (C) SubINode( _igvn.intcon(0), add->in(2) );
387 register_new_node( neg, get_ctrl(add->in(2) ) );
388 add = new (C) 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 (C) 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 (C) LShiftINode( add_var, scale );
420 register_new_node( var_scale, n_ctrl );
421 Node *var_add = new (C) 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 (C) 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 (C) 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_AddI ) {
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 (C) 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 (C) 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_FLOAT:
531 case T_DOUBLE: {
532 cost += Matcher::float_cmove_cost(); // Could be very expensive
533 break;
534 }
535 case T_LONG: {
536 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
537 }
538 case T_INT: // These all CMOV fine
539 case T_ADDRESS: { // (RawPtr)
540 cost++;
541 break;
542 }
543 case T_NARROWOOP: // Fall through
544 case T_OBJECT: { // Base oops are OK, but not derived oops
545 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
546 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
547 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
548 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
549 // have a Phi for the base here that we convert to a CMOVE all is well
550 // and good. But if the base is dead, we'll not make a CMOVE. Later
551 // the allocator will have to produce a base by creating a CMOVE of the
552 // relevant bases. This puts the allocator in the business of
553 // manufacturing expensive instructions, generally a bad plan.
554 // Just Say No to Conditionally-Moved Derived Pointers.
555 if (tp && tp->offset() != 0)
556 return NULL;
557 cost++;
558 break;
559 }
560 default:
561 return NULL; // In particular, can't do memory or I/O
562 }
563 // Add in cost any speculative ops
564 for (uint j = 1; j < region->req(); j++) {
565 Node *proj = region->in(j);
566 Node *inp = phi->in(j);
567 if (get_ctrl(inp) == proj) { // Found local op
568 cost++;
569 // Check for a chain of dependent ops; these will all become
570 // speculative in a CMOV.
571 for (uint k = 1; k < inp->req(); k++)
572 if (get_ctrl(inp->in(k)) == proj)
573 cost += ConditionalMoveLimit; // Too much speculative goo
574 }
575 }
576 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
577 // This will likely Split-If, a higher-payoff operation.
578 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
579 Node* use = phi->fast_out(k);
580 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
581 cost += ConditionalMoveLimit;
582 // Is there a use inside the loop?
583 // Note: check only basic types since CMoveP is pinned.
584 if (!used_inside_loop && is_java_primitive(bt)) {
585 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
586 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
587 used_inside_loop = true;
588 }
589 }
590 }
591 }
592 Node* bol = iff->in(1);
593 assert(bol->Opcode() == Op_Bool, "");
594 int cmp_op = bol->in(1)->Opcode();
595 // It is expensive to generate flags from a float compare.
596 // Avoid duplicated float compare.
597 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
598
599 float infrequent_prob = PROB_UNLIKELY_MAG(3);
600 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
601 if (used_inside_loop) {
602 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
603
604 // BlockLayoutByFrequency optimization moves infrequent branch
605 // from hot path. No point in CMOV'ing in such case (110 is used
606 // instead of 100 to take into account not exactness of float value).
607 if (BlockLayoutByFrequency) {
608 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
609 }
610 }
611 // Check for highly predictable branch. No point in CMOV'ing if
612 // we are going to predict accurately all the time.
613 if (iff->_prob < infrequent_prob ||
614 iff->_prob > (1.0f - infrequent_prob))
615 return NULL;
616
617 // --------------
618 // Now replace all Phis with CMOV's
619 Node *cmov_ctrl = iff->in(0);
620 uint flip = (lp->Opcode() == Op_IfTrue);
621 Node_List wq;
622 while (1) {
623 PhiNode* phi = NULL;
624 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
625 Node *out = region->fast_out(i);
626 if (out->is_Phi()) {
627 phi = out->as_Phi();
628 break;
629 }
630 }
631 if (phi == NULL) break;
632 #ifndef PRODUCT
633 if (PrintOpto && VerifyLoopOptimizations) tty->print_cr("CMOV");
634 #endif
635 // Move speculative ops
636 wq.push(phi);
637 while (wq.size() > 0) {
638 Node *n = wq.pop();
639 for (uint j = 1; j < n->req(); j++) {
640 Node* m = n->in(j);
641 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) {
642 #ifndef PRODUCT
643 if (PrintOpto && VerifyLoopOptimizations) {
644 tty->print(" speculate: ");
645 m->dump();
646 }
647 #endif
648 set_ctrl(m, cmov_ctrl);
649 wq.push(m);
650 }
651 }
652 }
653 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) );
654 register_new_node( cmov, cmov_ctrl );
655 _igvn.replace_node( phi, cmov );
656 #ifndef PRODUCT
657 if (TraceLoopOpts) {
658 tty->print("CMOV ");
659 r_loop->dump_head();
660 if (Verbose) {
661 bol->in(1)->dump(1);
662 cmov->dump(1);
663 }
664 }
665 if (VerifyLoopOptimizations) verify();
666 #endif
667 }
668
669 // The useless CFG diamond will fold up later; see the optimization in
670 // RegionNode::Ideal.
671 _igvn._worklist.push(region);
672
673 return iff->in(1);
674 }
675
676 //------------------------------split_if_with_blocks_pre-----------------------
677 // Do the real work in a non-recursive function. Data nodes want to be
678 // cloned in the pre-order so they can feed each other nicely.
679 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
680 // Cloning these guys is unlikely to win
681 int n_op = n->Opcode();
682 if( n_op == Op_MergeMem ) return n;
683 if( n->is_Proj() ) return n;
684 // Do not clone-up CmpFXXX variations, as these are always
685 // followed by a CmpI
686 if( n->is_Cmp() ) return n;
687 // Attempt to use a conditional move instead of a phi/branch
688 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
689 Node *cmov = conditional_move( n );
690 if( cmov ) return cmov;
691 }
692 if( n->is_CFG() || n->is_LoadStore() )
693 return n;
694 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
695 n_op == Op_Opaque2 ) {
696 if( !C->major_progress() ) // If chance of no more loop opts...
697 _igvn._worklist.push(n); // maybe we'll remove them
698 return n;
699 }
700
701 if( n->is_Con() ) return n; // No cloning for Con nodes
702
703 Node *n_ctrl = get_ctrl(n);
704 if( !n_ctrl ) return n; // Dead node
705
706 // Attempt to remix address expressions for loop invariants
707 Node *m = remix_address_expressions( n );
708 if( m ) return m;
709
710 // Determine if the Node has inputs from some local Phi.
711 // Returns the block to clone thru.
712 Node *n_blk = has_local_phi_input( n );
713 if( !n_blk ) return n;
714 // Do not clone the trip counter through on a CountedLoop
715 // (messes up the canonical shape).
716 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
717
718 // Check for having no control input; not pinned. Allow
719 // dominating control.
720 if( n->in(0) ) {
721 Node *dom = idom(n_blk);
722 if( dom_lca( n->in(0), dom ) != n->in(0) )
723 return n;
724 }
725 // Policy: when is it profitable. You must get more wins than
726 // policy before it is considered profitable. Policy is usually 0,
727 // so 1 win is considered profitable. Big merges will require big
728 // cloning, so get a larger policy.
729 int policy = n_blk->req() >> 2;
730
731 // If the loop is a candidate for range check elimination,
732 // delay splitting through it's phi until a later loop optimization
733 if (n_blk->is_CountedLoop()) {
734 IdealLoopTree *lp = get_loop(n_blk);
735 if (lp && lp->_rce_candidate) {
736 return n;
737 }
738 }
739
740 // Use same limit as split_if_with_blocks_post
741 if( C->unique() > 35000 ) return n; // Method too big
742
743 // Split 'n' through the merge point if it is profitable
744 Node *phi = split_thru_phi( n, n_blk, policy );
745 if (!phi) return n;
746
747 // Found a Phi to split thru!
748 // Replace 'n' with the new phi
749 _igvn.replace_node( n, phi );
750 // Moved a load around the loop, 'en-registering' something.
751 if (n_blk->is_Loop() && n->is_Load() &&
752 !phi->in(LoopNode::LoopBackControl)->is_Load())
753 C->set_major_progress();
754
755 // Moved a barrier around the loop, 'en-registering' something.
756 if (n_blk->is_Loop() && n->Opcode() == Op_ShenandoahLoadReferenceBarrier &&
757 phi->in(LoopNode::LoopBackControl)->Opcode() != Op_ShenandoahLoadReferenceBarrier)
758 C->set_major_progress();
759
760 return phi;
761 }
762
763 static bool merge_point_too_heavy(Compile* C, Node* region) {
764 // Bail out if the region and its phis have too many users.
765 int weight = 0;
766 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
767 weight += region->fast_out(i)->outcnt();
768 }
769 int nodes_left = C->max_node_limit() - C->live_nodes();
770 if (weight * 8 > nodes_left) {
771 #ifndef PRODUCT
772 if (PrintOpto)
773 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
774 #endif
775 return true;
776 } else {
777 return false;
778 }
779 }
780
781 static bool merge_point_safe(Node* region) {
782 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
783 // having a PhiNode input. This sidesteps the dangerous case where the split
784 // ConvI2LNode may become TOP if the input Value() does not
785 // overlap the ConvI2L range, leaving a node which may not dominate its
786 // uses.
787 // A better fix for this problem can be found in the BugTraq entry, but
788 // expediency for Mantis demands this hack.
789 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop
790 // split_if_with_blocks from splitting a block because we could not move around
791 // the FastLockNode.
792 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
793 Node* n = region->fast_out(i);
794 if (n->is_Phi()) {
795 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
796 Node* m = n->fast_out(j);
797 if (m->is_FastLock())
798 return false;
799 #ifdef _LP64
800 if (m->Opcode() == Op_ConvI2L)
801 return false;
802 if (m->is_CastII() && m->isa_CastII()->has_range_check()) {
803 return false;
804 }
805 #endif
806 }
807 }
808 }
809 return true;
810 }
811
812
813 //------------------------------place_near_use---------------------------------
814 // Place some computation next to use but not inside inner loops.
815 // For inner loop uses move it to the preheader area.
816 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const {
817 IdealLoopTree *u_loop = get_loop( useblock );
818 return (u_loop->_irreducible || u_loop->_child)
819 ? useblock
820 : u_loop->_head->in(LoopNode::EntryControl);
821 }
822
823
824 //------------------------------split_if_with_blocks_post----------------------
825 // Do the real work in a non-recursive function. CFG hackery wants to be
826 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
827 // info.
828 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) {
829
830 // Cloning Cmp through Phi's involves the split-if transform.
831 // FastLock is not used by an If
832 if( n->is_Cmp() && !n->is_FastLock() ) {
833 if( C->unique() > 35000 ) return; // Method too big
834
835 // Do not do 'split-if' if irreducible loops are present.
836 if( _has_irreducible_loops )
837 return;
838
839 Node *n_ctrl = get_ctrl(n);
840 // Determine if the Node has inputs from some local Phi.
841 // Returns the block to clone thru.
842 Node *n_blk = has_local_phi_input( n );
843 if( n_blk != n_ctrl ) return;
844
845 if( merge_point_too_heavy(C, n_ctrl) )
846 return;
847
848 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare?
849 Node *bol = n->unique_out();
850 assert( bol->is_Bool(), "expect a bool here" );
851 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare?
852 Node *iff = bol->unique_out();
853
854 // Check some safety conditions
855 if( iff->is_If() ) { // Classic split-if?
856 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if
857 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
858 // Can't split CMove with different control edge.
859 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) return;
860 if( get_ctrl(iff->in(2)) == n_ctrl ||
861 get_ctrl(iff->in(3)) == n_ctrl )
862 return; // Inputs not yet split-up
863 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) {
864 return; // Loop-invar test gates loop-varying CMOVE
865 }
866 } else {
867 return; // some other kind of node, such as an Allocate
868 }
869
870 // Do not do 'split-if' if some paths are dead. First do dead code
871 // elimination and then see if its still profitable.
872 for( uint i = 1; i < n_ctrl->req(); i++ )
873 if( n_ctrl->in(i) == C->top() )
874 return;
875
876 // When is split-if profitable? Every 'win' on means some control flow
877 // goes dead, so it's almost always a win.
878 int policy = 0;
879 // If trying to do a 'Split-If' at the loop head, it is only
880 // profitable if the cmp folds up on BOTH paths. Otherwise we
881 // risk peeling a loop forever.
882
883 // CNC - Disabled for now. Requires careful handling of loop
884 // body selection for the cloned code. Also, make sure we check
885 // for any input path not being in the same loop as n_ctrl. For
886 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
887 // because the alternative loop entry points won't be converted
888 // into LoopNodes.
889 IdealLoopTree *n_loop = get_loop(n_ctrl);
890 for( uint j = 1; j < n_ctrl->req(); j++ )
891 if( get_loop(n_ctrl->in(j)) != n_loop )
892 return;
893
894 // Check for safety of the merge point.
895 if( !merge_point_safe(n_ctrl) ) {
896 return;
897 }
898
899 // Split compare 'n' through the merge point if it is profitable
900 Node *phi = split_thru_phi( n, n_ctrl, policy );
901 if( !phi ) return;
902
903 // Found a Phi to split thru!
904 // Replace 'n' with the new phi
905 _igvn.replace_node( n, phi );
906
907 // Now split the bool up thru the phi
908 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 );
909 guarantee(bolphi != NULL, "null boolean phi node");
910
911 _igvn.replace_node( bol, bolphi );
912 assert( iff->in(1) == bolphi, "" );
913
914 if( bolphi->Value(&_igvn)->singleton() )
915 return;
916
917 // Conditional-move? Must split up now
918 if( !iff->is_If() ) {
919 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 );
920 _igvn.replace_node( iff, cmovphi );
921 return;
922 }
923
924 // Now split the IF
925 do_split_if( iff );
926 return;
927 }
928
929 // Check for an IF ready to split; one that has its
930 // condition codes input coming from a Phi at the block start.
931 int n_op = n->Opcode();
932
933 // Check for an IF being dominated by another IF same test
934 if (n_op == Op_If) {
935 Node *bol = n->in(1);
936 uint max = bol->outcnt();
937 // Check for same test used more than once?
938 if (max > 1 && bol->is_Bool()) {
939 // Search up IDOMs to see if this IF is dominated.
940 Node *cutoff = get_ctrl(bol);
941
942 // Now search up IDOMs till cutoff, looking for a dominating test
943 Node *prevdom = n;
944 Node *dom = idom(prevdom);
945 while (dom != cutoff) {
946 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) {
947 // Replace the dominated test with an obvious true or false.
948 // Place it on the IGVN worklist for later cleanup.
949 C->set_major_progress();
950 dominated_by(prevdom, n, false, true);
951 #ifndef PRODUCT
952 if( VerifyLoopOptimizations ) verify();
953 #endif
954 return;
955 }
956 prevdom = dom;
957 dom = idom(prevdom);
958 }
959 }
960 }
961
962 // See if a shared loop-varying computation has no loop-varying uses.
963 // Happens if something is only used for JVM state in uncommon trap exits,
964 // like various versions of induction variable+offset. Clone the
965 // computation per usage to allow it to sink out of the loop.
966 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
967 Node *n_ctrl = get_ctrl(n);
968 IdealLoopTree *n_loop = get_loop(n_ctrl);
969 if( n_loop != _ltree_root ) {
970 DUIterator_Fast imax, i = n->fast_outs(imax);
971 for (; i < imax; i++) {
972 Node* u = n->fast_out(i);
973 if( !has_ctrl(u) ) break; // Found control user
974 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
975 if( u_loop == n_loop ) break; // Found loop-varying use
976 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
977 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
978 }
979 bool did_break = (i < imax); // Did we break out of the previous loop?
980 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
981 Node *late_load_ctrl = NULL;
982 if (n->is_Load()) {
983 // If n is a load, get and save the result from get_late_ctrl(),
984 // to be later used in calculating the control for n's clones.
985 clear_dom_lca_tags();
986 late_load_ctrl = get_late_ctrl(n, n_ctrl);
987 }
988 // If n is a load, and the late control is the same as the current
989 // control, then the cloning of n is a pointless exercise, because
990 // GVN will ensure that we end up where we started.
991 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
992 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
993 Node *u = n->last_out(j); // Clone private computation per use
994 _igvn.rehash_node_delayed(u);
995 Node *x = n->clone(); // Clone computation
996 Node *x_ctrl = NULL;
997 if( u->is_Phi() ) {
998 // Replace all uses of normal nodes. Replace Phi uses
999 // individually, so the separate Nodes can sink down
1000 // different paths.
1001 uint k = 1;
1002 while( u->in(k) != n ) k++;
1003 u->set_req( k, x );
1004 // x goes next to Phi input path
1005 x_ctrl = u->in(0)->in(k);
1006 --j;
1007 } else { // Normal use
1008 // Replace all uses
1009 for( uint k = 0; k < u->req(); k++ ) {
1010 if( u->in(k) == n ) {
1011 u->set_req( k, x );
1012 --j;
1013 }
1014 }
1015 x_ctrl = get_ctrl(u);
1016 }
1017
1018 // Find control for 'x' next to use but not inside inner loops.
1019 // For inner loop uses get the preheader area.
1020 x_ctrl = place_near_use(x_ctrl);
1021
1022 if (n->is_Load()) {
1023 // For loads, add a control edge to a CFG node outside of the loop
1024 // to force them to not combine and return back inside the loop
1025 // during GVN optimization (4641526).
1026 //
1027 // Because we are setting the actual control input, factor in
1028 // the result from get_late_ctrl() so we respect any
1029 // anti-dependences. (6233005).
1030 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
1031
1032 // Don't allow the control input to be a CFG splitting node.
1033 // Such nodes should only have ProjNodes as outs, e.g. IfNode
1034 // should only have IfTrueNode and IfFalseNode (4985384).
1035 x_ctrl = find_non_split_ctrl(x_ctrl);
1036 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1037
1038 x->set_req(0, x_ctrl);
1039 }
1040 register_new_node(x, x_ctrl);
1041
1042 // Some institutional knowledge is needed here: 'x' is
1043 // yanked because if the optimizer runs GVN on it all the
1044 // cloned x's will common up and undo this optimization and
1045 // be forced back in the loop. This is annoying because it
1046 // makes +VerifyOpto report false-positives on progress. I
1047 // tried setting control edges on the x's to force them to
1048 // not combine, but the matching gets worried when it tries
1049 // to fold a StoreP and an AddP together (as part of an
1050 // address expression) and the AddP and StoreP have
1051 // different controls.
1052 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x);
1053 }
1054 _igvn.remove_dead_node(n);
1055 }
1056 }
1057 }
1058 }
1059
1060 // Check for Opaque2's who's loop has disappeared - who's input is in the
1061 // same loop nest as their output. Remove 'em, they are no longer useful.
1062 if( n_op == Op_Opaque2 &&
1063 n->in(1) != NULL &&
1064 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1065 _igvn.replace_node( n, n->in(1) );
1066 }
1067 }
1068
1069 //------------------------------split_if_with_blocks---------------------------
1070 // Check for aggressive application of 'split-if' optimization,
1071 // using basic block level info.
1072 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) {
1073 Node *n = C->root();
1074 visited.set(n->_idx); // first, mark node as visited
1075 // Do pre-visit work for root
1076 n = split_if_with_blocks_pre( n );
1077 uint cnt = n->outcnt();
1078 uint i = 0;
1079 while (true) {
1080 // Visit all children
1081 if (i < cnt) {
1082 Node* use = n->raw_out(i);
1083 ++i;
1084 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1085 // Now do pre-visit work for this use
1086 use = split_if_with_blocks_pre( use );
1087 nstack.push(n, i); // Save parent and next use's index.
1088 n = use; // Process all children of current use.
1089 cnt = use->outcnt();
1090 i = 0;
1091 }
1092 }
1093 else {
1094 // All of n's children have been processed, complete post-processing.
1095 if (cnt != 0 && !n->is_Con()) {
1096 assert(has_node(n), "no dead nodes");
1097 split_if_with_blocks_post( n );
1098 }
1099 if (nstack.is_empty()) {
1100 // Finished all nodes on stack.
1101 break;
1102 }
1103 // Get saved parent node and next use's index. Visit the rest of uses.
1104 n = nstack.node();
1105 cnt = n->outcnt();
1106 i = nstack.index();
1107 nstack.pop();
1108 }
1109 }
1110 }
1111
1112
1113 //=============================================================================
1114 //
1115 // C L O N E A L O O P B O D Y
1116 //
1117
1118 //------------------------------clone_iff--------------------------------------
1119 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1120 // "Nearly" because all Nodes have been cloned from the original in the loop,
1121 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1122 // through the Phi recursively, and return a Bool.
1123 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) {
1124
1125 // Convert this Phi into a Phi merging Bools
1126 uint i;
1127 for( i = 1; i < phi->req(); i++ ) {
1128 Node *b = phi->in(i);
1129 if( b->is_Phi() ) {
1130 _igvn.replace_input_of(phi, i, clone_iff( b->as_Phi(), loop ));
1131 } else {
1132 assert( b->is_Bool(), "" );
1133 }
1134 }
1135
1136 Node *sample_bool = phi->in(1);
1137 Node *sample_cmp = sample_bool->in(1);
1138
1139 // Make Phis to merge the Cmp's inputs.
1140 PhiNode *phi1 = new (C) PhiNode( phi->in(0), Type::TOP );
1141 PhiNode *phi2 = new (C) PhiNode( phi->in(0), Type::TOP );
1142 for( i = 1; i < phi->req(); i++ ) {
1143 Node *n1 = phi->in(i)->in(1)->in(1);
1144 Node *n2 = phi->in(i)->in(1)->in(2);
1145 phi1->set_req( i, n1 );
1146 phi2->set_req( i, n2 );
1147 phi1->set_type( phi1->type()->meet_speculative(n1->bottom_type()));
1148 phi2->set_type( phi2->type()->meet_speculative(n2->bottom_type()));
1149 }
1150 // See if these Phis have been made before.
1151 // Register with optimizer
1152 Node *hit1 = _igvn.hash_find_insert(phi1);
1153 if( hit1 ) { // Hit, toss just made Phi
1154 _igvn.remove_dead_node(phi1); // Remove new phi
1155 assert( hit1->is_Phi(), "" );
1156 phi1 = (PhiNode*)hit1; // Use existing phi
1157 } else { // Miss
1158 _igvn.register_new_node_with_optimizer(phi1);
1159 }
1160 Node *hit2 = _igvn.hash_find_insert(phi2);
1161 if( hit2 ) { // Hit, toss just made Phi
1162 _igvn.remove_dead_node(phi2); // Remove new phi
1163 assert( hit2->is_Phi(), "" );
1164 phi2 = (PhiNode*)hit2; // Use existing phi
1165 } else { // Miss
1166 _igvn.register_new_node_with_optimizer(phi2);
1167 }
1168 // Register Phis with loop/block info
1169 set_ctrl(phi1, phi->in(0));
1170 set_ctrl(phi2, phi->in(0));
1171 // Make a new Cmp
1172 Node *cmp = sample_cmp->clone();
1173 cmp->set_req( 1, phi1 );
1174 cmp->set_req( 2, phi2 );
1175 _igvn.register_new_node_with_optimizer(cmp);
1176 set_ctrl(cmp, phi->in(0));
1177
1178 // Make a new Bool
1179 Node *b = sample_bool->clone();
1180 b->set_req(1,cmp);
1181 _igvn.register_new_node_with_optimizer(b);
1182 set_ctrl(b, phi->in(0));
1183
1184 assert( b->is_Bool(), "" );
1185 return (BoolNode*)b;
1186 }
1187
1188 //------------------------------clone_bool-------------------------------------
1189 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1190 // "Nearly" because all Nodes have been cloned from the original in the loop,
1191 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1192 // through the Phi recursively, and return a Bool.
1193 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1194 uint i;
1195 // Convert this Phi into a Phi merging Bools
1196 for( i = 1; i < phi->req(); i++ ) {
1197 Node *b = phi->in(i);
1198 if( b->is_Phi() ) {
1199 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1200 } else {
1201 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1202 }
1203 }
1204
1205 Node *sample_cmp = phi->in(1);
1206
1207 // Make Phis to merge the Cmp's inputs.
1208 PhiNode *phi1 = new (C) PhiNode( phi->in(0), Type::TOP );
1209 PhiNode *phi2 = new (C) PhiNode( phi->in(0), Type::TOP );
1210 for( uint j = 1; j < phi->req(); j++ ) {
1211 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1212 Node *n1, *n2;
1213 if( cmp_top->is_Cmp() ) {
1214 n1 = cmp_top->in(1);
1215 n2 = cmp_top->in(2);
1216 } else {
1217 n1 = n2 = cmp_top;
1218 }
1219 phi1->set_req( j, n1 );
1220 phi2->set_req( j, n2 );
1221 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1222 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1223 }
1224
1225 // See if these Phis have been made before.
1226 // Register with optimizer
1227 Node *hit1 = _igvn.hash_find_insert(phi1);
1228 if( hit1 ) { // Hit, toss just made Phi
1229 _igvn.remove_dead_node(phi1); // Remove new phi
1230 assert( hit1->is_Phi(), "" );
1231 phi1 = (PhiNode*)hit1; // Use existing phi
1232 } else { // Miss
1233 _igvn.register_new_node_with_optimizer(phi1);
1234 }
1235 Node *hit2 = _igvn.hash_find_insert(phi2);
1236 if( hit2 ) { // Hit, toss just made Phi
1237 _igvn.remove_dead_node(phi2); // Remove new phi
1238 assert( hit2->is_Phi(), "" );
1239 phi2 = (PhiNode*)hit2; // Use existing phi
1240 } else { // Miss
1241 _igvn.register_new_node_with_optimizer(phi2);
1242 }
1243 // Register Phis with loop/block info
1244 set_ctrl(phi1, phi->in(0));
1245 set_ctrl(phi2, phi->in(0));
1246 // Make a new Cmp
1247 Node *cmp = sample_cmp->clone();
1248 cmp->set_req( 1, phi1 );
1249 cmp->set_req( 2, phi2 );
1250 _igvn.register_new_node_with_optimizer(cmp);
1251 set_ctrl(cmp, phi->in(0));
1252
1253 assert( cmp->is_Cmp(), "" );
1254 return (CmpNode*)cmp;
1255 }
1256
1257 //------------------------------sink_use---------------------------------------
1258 // If 'use' was in the loop-exit block, it now needs to be sunk
1259 // below the post-loop merge point.
1260 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1261 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1262 set_ctrl(use, post_loop);
1263 for (DUIterator j = use->outs(); use->has_out(j); j++)
1264 sink_use(use->out(j), post_loop);
1265 }
1266 }
1267
1268 //------------------------------clone_loop-------------------------------------
1269 //
1270 // C L O N E A L O O P B O D Y
1271 //
1272 // This is the basic building block of the loop optimizations. It clones an
1273 // entire loop body. It makes an old_new loop body mapping; with this mapping
1274 // you can find the new-loop equivalent to an old-loop node. All new-loop
1275 // nodes are exactly equal to their old-loop counterparts, all edges are the
1276 // same. All exits from the old-loop now have a RegionNode that merges the
1277 // equivalent new-loop path. This is true even for the normal "loop-exit"
1278 // condition. All uses of loop-invariant old-loop values now come from (one
1279 // or more) Phis that merge their new-loop equivalents.
1280 //
1281 // This operation leaves the graph in an illegal state: there are two valid
1282 // control edges coming from the loop pre-header to both loop bodies. I'll
1283 // definitely have to hack the graph after running this transform.
1284 //
1285 // From this building block I will further edit edges to perform loop peeling
1286 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1287 //
1288 // Parameter side_by_size_idom:
1289 // When side_by_size_idom is NULL, the dominator tree is constructed for
1290 // the clone loop to dominate the original. Used in construction of
1291 // pre-main-post loop sequence.
1292 // When nonnull, the clone and original are side-by-side, both are
1293 // dominated by the side_by_side_idom node. Used in construction of
1294 // unswitched loops.
1295 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1296 Node* side_by_side_idom) {
1297
1298 // Step 1: Clone the loop body. Make the old->new mapping.
1299 uint i;
1300 for( i = 0; i < loop->_body.size(); i++ ) {
1301 Node *old = loop->_body.at(i);
1302 Node *nnn = old->clone();
1303 old_new.map( old->_idx, nnn );
1304 _igvn.register_new_node_with_optimizer(nnn);
1305 }
1306
1307
1308 // Step 2: Fix the edges in the new body. If the old input is outside the
1309 // loop use it. If the old input is INside the loop, use the corresponding
1310 // new node instead.
1311 for( i = 0; i < loop->_body.size(); i++ ) {
1312 Node *old = loop->_body.at(i);
1313 Node *nnn = old_new[old->_idx];
1314 // Fix CFG/Loop controlling the new node
1315 if (has_ctrl(old)) {
1316 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
1317 } else {
1318 set_loop(nnn, loop->_parent);
1319 if (old->outcnt() > 0) {
1320 set_idom( nnn, old_new[idom(old)->_idx], dd );
1321 }
1322 }
1323 // Correct edges to the new node
1324 for( uint j = 0; j < nnn->req(); j++ ) {
1325 Node *n = nnn->in(j);
1326 if( n ) {
1327 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
1328 if( loop->is_member( old_in_loop ) )
1329 nnn->set_req(j, old_new[n->_idx]);
1330 }
1331 }
1332 _igvn.hash_find_insert(nnn);
1333 }
1334 Node *newhead = old_new[loop->_head->_idx];
1335 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1336
1337
1338 // Step 3: Now fix control uses. Loop varying control uses have already
1339 // been fixed up (as part of all input edges in Step 2). Loop invariant
1340 // control uses must be either an IfFalse or an IfTrue. Make a merge
1341 // point to merge the old and new IfFalse/IfTrue nodes; make the use
1342 // refer to this.
1343 ResourceArea *area = Thread::current()->resource_area();
1344 Node_List worklist(area);
1345 uint new_counter = C->unique();
1346 for( i = 0; i < loop->_body.size(); i++ ) {
1347 Node* old = loop->_body.at(i);
1348 if( !old->is_CFG() ) continue;
1349 Node* nnn = old_new[old->_idx];
1350
1351 // Copy uses to a worklist, so I can munge the def-use info
1352 // with impunity.
1353 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1354 worklist.push(old->fast_out(j));
1355
1356 while( worklist.size() ) { // Visit all uses
1357 Node *use = worklist.pop();
1358 if (!has_node(use)) continue; // Ignore dead nodes
1359 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1360 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
1361 // Both OLD and USE are CFG nodes here.
1362 assert( use->is_Proj(), "" );
1363
1364 // Clone the loop exit control projection
1365 Node *newuse = use->clone();
1366 newuse->set_req(0,nnn);
1367 _igvn.register_new_node_with_optimizer(newuse);
1368 set_loop(newuse, use_loop);
1369 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
1370
1371 // We need a Region to merge the exit from the peeled body and the
1372 // exit from the old loop body.
1373 RegionNode *r = new (C) RegionNode(3);
1374 // Map the old use to the new merge point
1375 old_new.map( use->_idx, r );
1376 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
1377 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
1378
1379 // The original user of 'use' uses 'r' instead.
1380 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
1381 Node* useuse = use->last_out(l);
1382 _igvn.rehash_node_delayed(useuse);
1383 uint uses_found = 0;
1384 if( useuse->in(0) == use ) {
1385 useuse->set_req(0, r);
1386 uses_found++;
1387 if( useuse->is_CFG() ) {
1388 assert( dom_depth(useuse) > dd_r, "" );
1389 set_idom(useuse, r, dom_depth(useuse));
1390 }
1391 }
1392 for( uint k = 1; k < useuse->req(); k++ ) {
1393 if( useuse->in(k) == use ) {
1394 useuse->set_req(k, r);
1395 uses_found++;
1396 }
1397 }
1398 l -= uses_found; // we deleted 1 or more copies of this edge
1399 }
1400
1401 // Now finish up 'r'
1402 r->set_req( 1, newuse );
1403 r->set_req( 2, use );
1404 _igvn.register_new_node_with_optimizer(r);
1405 set_loop(r, use_loop);
1406 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
1407 } // End of if a loop-exit test
1408 }
1409 }
1410
1411 // Step 4: If loop-invariant use is not control, it must be dominated by a
1412 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
1413 // there if needed. Make a Phi there merging old and new used values.
1414 Node_List *split_if_set = NULL;
1415 Node_List *split_bool_set = NULL;
1416 Node_List *split_cex_set = NULL;
1417 for( i = 0; i < loop->_body.size(); i++ ) {
1418 Node* old = loop->_body.at(i);
1419 Node* nnn = old_new[old->_idx];
1420 // Copy uses to a worklist, so I can munge the def-use info
1421 // with impunity.
1422 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1423 worklist.push(old->fast_out(j));
1424
1425 while( worklist.size() ) {
1426 Node *use = worklist.pop();
1427 if (!has_node(use)) continue; // Ignore dead nodes
1428 if (use->in(0) == C->top()) continue;
1429 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1430 // Check for data-use outside of loop - at least one of OLD or USE
1431 // must not be a CFG node.
1432 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) {
1433
1434 // If the Data use is an IF, that means we have an IF outside of the
1435 // loop that is switching on a condition that is set inside of the
1436 // loop. Happens if people set a loop-exit flag; then test the flag
1437 // in the loop to break the loop, then test is again outside of the
1438 // loop to determine which way the loop exited.
1439 // Loop predicate If node connects to Bool node through Opaque1 node.
1440 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use)) {
1441 // Since this code is highly unlikely, we lazily build the worklist
1442 // of such Nodes to go split.
1443 if( !split_if_set )
1444 split_if_set = new Node_List(area);
1445 split_if_set->push(use);
1446 }
1447 if( use->is_Bool() ) {
1448 if( !split_bool_set )
1449 split_bool_set = new Node_List(area);
1450 split_bool_set->push(use);
1451 }
1452 if( use->Opcode() == Op_CreateEx ) {
1453 if( !split_cex_set )
1454 split_cex_set = new Node_List(area);
1455 split_cex_set->push(use);
1456 }
1457
1458
1459 // Get "block" use is in
1460 uint idx = 0;
1461 while( use->in(idx) != old ) idx++;
1462 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1463 assert( !loop->is_member( get_loop( prev ) ), "" );
1464 Node *cfg = prev->_idx >= new_counter
1465 ? prev->in(2)
1466 : idom(prev);
1467 if( use->is_Phi() ) // Phi use is in prior block
1468 cfg = prev->in(idx); // NOT in block of Phi itself
1469 if (cfg->is_top()) { // Use is dead?
1470 _igvn.replace_input_of(use, idx, C->top());
1471 continue;
1472 }
1473
1474 while( !loop->is_member( get_loop( cfg ) ) ) {
1475 prev = cfg;
1476 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1477 }
1478 // If the use occurs after merging several exits from the loop, then
1479 // old value must have dominated all those exits. Since the same old
1480 // value was used on all those exits we did not need a Phi at this
1481 // merge point. NOW we do need a Phi here. Each loop exit value
1482 // is now merged with the peeled body exit; each exit gets its own
1483 // private Phi and those Phis need to be merged here.
1484 Node *phi;
1485 if( prev->is_Region() ) {
1486 if( idx == 0 ) { // Updating control edge?
1487 phi = prev; // Just use existing control
1488 } else { // Else need a new Phi
1489 phi = PhiNode::make( prev, old );
1490 // Now recursively fix up the new uses of old!
1491 for( uint i = 1; i < prev->req(); i++ ) {
1492 worklist.push(phi); // Onto worklist once for each 'old' input
1493 }
1494 }
1495 } else {
1496 // Get new RegionNode merging old and new loop exits
1497 prev = old_new[prev->_idx];
1498 assert( prev, "just made this in step 7" );
1499 if( idx == 0 ) { // Updating control edge?
1500 phi = prev; // Just use existing control
1501 } else { // Else need a new Phi
1502 // Make a new Phi merging data values properly
1503 phi = PhiNode::make( prev, old );
1504 phi->set_req( 1, nnn );
1505 }
1506 }
1507 // If inserting a new Phi, check for prior hits
1508 if( idx != 0 ) {
1509 Node *hit = _igvn.hash_find_insert(phi);
1510 if( hit == NULL ) {
1511 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1512 } else { // or
1513 // Remove the new phi from the graph and use the hit
1514 _igvn.remove_dead_node(phi);
1515 phi = hit; // Use existing phi
1516 }
1517 set_ctrl(phi, prev);
1518 }
1519 // Make 'use' use the Phi instead of the old loop body exit value
1520 _igvn.replace_input_of(use, idx, phi);
1521 if( use->_idx >= new_counter ) { // If updating new phis
1522 // Not needed for correctness, but prevents a weak assert
1523 // in AddPNode from tripping (when we end up with different
1524 // base & derived Phis that will become the same after
1525 // IGVN does CSE).
1526 Node *hit = _igvn.hash_find_insert(use);
1527 if( hit ) // Go ahead and re-hash for hits.
1528 _igvn.replace_node( use, hit );
1529 }
1530
1531 // If 'use' was in the loop-exit block, it now needs to be sunk
1532 // below the post-loop merge point.
1533 sink_use( use, prev );
1534 }
1535 }
1536 }
1537
1538 // Check for IFs that need splitting/cloning. Happens if an IF outside of
1539 // the loop uses a condition set in the loop. The original IF probably
1540 // takes control from one or more OLD Regions (which in turn get from NEW
1541 // Regions). In any case, there will be a set of Phis for each merge point
1542 // from the IF up to where the original BOOL def exists the loop.
1543 if( split_if_set ) {
1544 while( split_if_set->size() ) {
1545 Node *iff = split_if_set->pop();
1546 if( iff->in(1)->is_Phi() ) {
1547 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop );
1548 _igvn.replace_input_of(iff, 1, b);
1549 }
1550 }
1551 }
1552 if( split_bool_set ) {
1553 while( split_bool_set->size() ) {
1554 Node *b = split_bool_set->pop();
1555 Node *phi = b->in(1);
1556 assert( phi->is_Phi(), "" );
1557 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop );
1558 _igvn.replace_input_of(b, 1, cmp);
1559 }
1560 }
1561 if( split_cex_set ) {
1562 while( split_cex_set->size() ) {
1563 Node *b = split_cex_set->pop();
1564 assert( b->in(0)->is_Region(), "" );
1565 assert( b->in(1)->is_Phi(), "" );
1566 assert( b->in(0)->in(0) == b->in(1)->in(0), "" );
1567 split_up( b, b->in(0), NULL );
1568 }
1569 }
1570
1571 }
1572
1573
1574 //---------------------- stride_of_possible_iv -------------------------------------
1575 // Looks for an iff/bool/comp with one operand of the compare
1576 // being a cycle involving an add and a phi,
1577 // with an optional truncation (left-shift followed by a right-shift)
1578 // of the add. Returns zero if not an iv.
1579 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
1580 Node* trunc1 = NULL;
1581 Node* trunc2 = NULL;
1582 const TypeInt* ttype = NULL;
1583 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
1584 return 0;
1585 }
1586 BoolNode* bl = iff->in(1)->as_Bool();
1587 Node* cmp = bl->in(1);
1588 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) {
1589 return 0;
1590 }
1591 // Must have an invariant operand
1592 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
1593 return 0;
1594 }
1595 Node* add2 = NULL;
1596 Node* cmp1 = cmp->in(1);
1597 if (cmp1->is_Phi()) {
1598 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
1599 Node* phi = cmp1;
1600 for (uint i = 1; i < phi->req(); i++) {
1601 Node* in = phi->in(i);
1602 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
1603 &trunc1, &trunc2, &ttype);
1604 if (add && add->in(1) == phi) {
1605 add2 = add->in(2);
1606 break;
1607 }
1608 }
1609 } else {
1610 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
1611 Node* addtrunc = cmp1;
1612 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
1613 &trunc1, &trunc2, &ttype);
1614 if (add && add->in(1)->is_Phi()) {
1615 Node* phi = add->in(1);
1616 for (uint i = 1; i < phi->req(); i++) {
1617 if (phi->in(i) == addtrunc) {
1618 add2 = add->in(2);
1619 break;
1620 }
1621 }
1622 }
1623 }
1624 if (add2 != NULL) {
1625 const TypeInt* add2t = _igvn.type(add2)->is_int();
1626 if (add2t->is_con()) {
1627 return add2t->get_con();
1628 }
1629 }
1630 return 0;
1631 }
1632
1633
1634 //---------------------- stay_in_loop -------------------------------------
1635 // Return the (unique) control output node that's in the loop (if it exists.)
1636 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
1637 Node* unique = NULL;
1638 if (!n) return NULL;
1639 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1640 Node* use = n->fast_out(i);
1641 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
1642 if (unique != NULL) {
1643 return NULL;
1644 }
1645 unique = use;
1646 }
1647 }
1648 return unique;
1649 }
1650
1651 //------------------------------ register_node -------------------------------------
1652 // Utility to register node "n" with PhaseIdealLoop
1653 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
1654 _igvn.register_new_node_with_optimizer(n);
1655 loop->_body.push(n);
1656 if (n->is_CFG()) {
1657 set_loop(n, loop);
1658 set_idom(n, pred, ddepth);
1659 } else {
1660 set_ctrl(n, pred);
1661 }
1662 }
1663
1664 //------------------------------ proj_clone -------------------------------------
1665 // Utility to create an if-projection
1666 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
1667 ProjNode* c = p->clone()->as_Proj();
1668 c->set_req(0, iff);
1669 return c;
1670 }
1671
1672 //------------------------------ short_circuit_if -------------------------------------
1673 // Force the iff control output to be the live_proj
1674 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
1675 guarantee(live_proj != NULL, "null projection");
1676 int proj_con = live_proj->_con;
1677 assert(proj_con == 0 || proj_con == 1, "false or true projection");
1678 Node *con = _igvn.intcon(proj_con);
1679 set_ctrl(con, C->root());
1680 if (iff) {
1681 iff->set_req(1, con);
1682 }
1683 return con;
1684 }
1685
1686 //------------------------------ insert_if_before_proj -------------------------------------
1687 // Insert a new if before an if projection (* - new node)
1688 //
1689 // before
1690 // if(test)
1691 // / \
1692 // v v
1693 // other-proj proj (arg)
1694 //
1695 // after
1696 // if(test)
1697 // / \
1698 // / v
1699 // | * proj-clone
1700 // v |
1701 // other-proj v
1702 // * new_if(relop(cmp[IU](left,right)))
1703 // / \
1704 // v v
1705 // * new-proj proj
1706 // (returned)
1707 //
1708 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
1709 IfNode* iff = proj->in(0)->as_If();
1710 IdealLoopTree *loop = get_loop(proj);
1711 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1712 int ddepth = dom_depth(proj);
1713
1714 _igvn.rehash_node_delayed(iff);
1715 _igvn.rehash_node_delayed(proj);
1716
1717 proj->set_req(0, NULL); // temporary disconnect
1718 ProjNode* proj2 = proj_clone(proj, iff);
1719 register_node(proj2, loop, iff, ddepth);
1720
1721 Node* cmp = Signed ? (Node*) new (C)CmpINode(left, right) : (Node*) new (C)CmpUNode(left, right);
1722 register_node(cmp, loop, proj2, ddepth);
1723
1724 BoolNode* bol = new (C)BoolNode(cmp, relop);
1725 register_node(bol, loop, proj2, ddepth);
1726
1727 IfNode* new_if = new (C)IfNode(proj2, bol, iff->_prob, iff->_fcnt);
1728 register_node(new_if, loop, proj2, ddepth);
1729
1730 proj->set_req(0, new_if); // reattach
1731 set_idom(proj, new_if, ddepth);
1732
1733 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
1734 guarantee(new_exit != NULL, "null exit node");
1735 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
1736
1737 return new_exit;
1738 }
1739
1740 //------------------------------ insert_region_before_proj -------------------------------------
1741 // Insert a region before an if projection (* - new node)
1742 //
1743 // before
1744 // if(test)
1745 // / |
1746 // v |
1747 // proj v
1748 // other-proj
1749 //
1750 // after
1751 // if(test)
1752 // / |
1753 // v |
1754 // * proj-clone v
1755 // | other-proj
1756 // v
1757 // * new-region
1758 // |
1759 // v
1760 // * dum_if
1761 // / \
1762 // v \
1763 // * dum-proj v
1764 // proj
1765 //
1766 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
1767 IfNode* iff = proj->in(0)->as_If();
1768 IdealLoopTree *loop = get_loop(proj);
1769 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1770 int ddepth = dom_depth(proj);
1771
1772 _igvn.rehash_node_delayed(iff);
1773 _igvn.rehash_node_delayed(proj);
1774
1775 proj->set_req(0, NULL); // temporary disconnect
1776 ProjNode* proj2 = proj_clone(proj, iff);
1777 register_node(proj2, loop, iff, ddepth);
1778
1779 RegionNode* reg = new (C)RegionNode(2);
1780 reg->set_req(1, proj2);
1781 register_node(reg, loop, iff, ddepth);
1782
1783 IfNode* dum_if = new (C)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
1784 register_node(dum_if, loop, reg, ddepth);
1785
1786 proj->set_req(0, dum_if); // reattach
1787 set_idom(proj, dum_if, ddepth);
1788
1789 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
1790 register_node(dum_proj, loop, dum_if, ddepth);
1791
1792 return reg;
1793 }
1794
1795 //------------------------------ insert_cmpi_loop_exit -------------------------------------
1796 // Clone a signed compare loop exit from an unsigned compare and
1797 // insert it before the unsigned cmp on the stay-in-loop path.
1798 // All new nodes inserted in the dominator tree between the original
1799 // if and it's projections. The original if test is replaced with
1800 // a constant to force the stay-in-loop path.
1801 //
1802 // This is done to make sure that the original if and it's projections
1803 // still dominate the same set of control nodes, that the ctrl() relation
1804 // from data nodes to them is preserved, and that their loop nesting is
1805 // preserved.
1806 //
1807 // before
1808 // if(i <u limit) unsigned compare loop exit
1809 // / |
1810 // v v
1811 // exit-proj stay-in-loop-proj
1812 //
1813 // after
1814 // if(stay-in-loop-const) original if
1815 // / |
1816 // / v
1817 // / if(i < limit) new signed test
1818 // / / |
1819 // / / v
1820 // / / if(i <u limit) new cloned unsigned test
1821 // / / / |
1822 // v v v |
1823 // region |
1824 // | |
1825 // dum-if |
1826 // / | |
1827 // ether | |
1828 // v v
1829 // exit-proj stay-in-loop-proj
1830 //
1831 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
1832 const bool Signed = true;
1833 const bool Unsigned = false;
1834
1835 BoolNode* bol = if_cmpu->in(1)->as_Bool();
1836 if (bol->_test._test != BoolTest::lt) return NULL;
1837 CmpNode* cmpu = bol->in(1)->as_Cmp();
1838 if (cmpu->Opcode() != Op_CmpU) return NULL;
1839 int stride = stride_of_possible_iv(if_cmpu);
1840 if (stride == 0) return NULL;
1841
1842 Node* lp_proj = stay_in_loop(if_cmpu, loop);
1843 guarantee(lp_proj != NULL, "null loop node");
1844
1845 ProjNode* lp_continue = lp_proj->as_Proj();
1846 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
1847
1848 Node* limit = NULL;
1849 if (stride > 0) {
1850 limit = cmpu->in(2);
1851 } else {
1852 limit = _igvn.makecon(TypeInt::ZERO);
1853 set_ctrl(limit, C->root());
1854 }
1855 // Create a new region on the exit path
1856 RegionNode* reg = insert_region_before_proj(lp_exit);
1857 guarantee(reg != NULL, "null region node");
1858
1859 // Clone the if-cmpu-true-false using a signed compare
1860 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
1861 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
1862 reg->add_req(cmpi_exit);
1863
1864 // Clone the if-cmpu-true-false
1865 BoolTest::mask rel_u = bol->_test._test;
1866 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
1867 reg->add_req(cmpu_exit);
1868
1869 // Force original if to stay in loop.
1870 short_circuit_if(if_cmpu, lp_continue);
1871
1872 return cmpi_exit->in(0)->as_If();
1873 }
1874
1875 //------------------------------ remove_cmpi_loop_exit -------------------------------------
1876 // Remove a previously inserted signed compare loop exit.
1877 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
1878 Node* lp_proj = stay_in_loop(if_cmp, loop);
1879 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
1880 stay_in_loop(lp_proj, loop)->is_If() &&
1881 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
1882 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
1883 set_ctrl(con, C->root());
1884 if_cmp->set_req(1, con);
1885 }
1886
1887 //------------------------------ scheduled_nodelist -------------------------------------
1888 // Create a post order schedule of nodes that are in the
1889 // "member" set. The list is returned in "sched".
1890 // The first node in "sched" is the loop head, followed by
1891 // nodes which have no inputs in the "member" set, and then
1892 // followed by the nodes that have an immediate input dependence
1893 // on a node in "sched".
1894 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
1895
1896 assert(member.test(loop->_head->_idx), "loop head must be in member set");
1897 Arena *a = Thread::current()->resource_area();
1898 VectorSet visited(a);
1899 Node_Stack nstack(a, loop->_body.size());
1900
1901 Node* n = loop->_head; // top of stack is cached in "n"
1902 uint idx = 0;
1903 visited.set(n->_idx);
1904
1905 // Initially push all with no inputs from within member set
1906 for(uint i = 0; i < loop->_body.size(); i++ ) {
1907 Node *elt = loop->_body.at(i);
1908 if (member.test(elt->_idx)) {
1909 bool found = false;
1910 for (uint j = 0; j < elt->req(); j++) {
1911 Node* def = elt->in(j);
1912 if (def && member.test(def->_idx) && def != elt) {
1913 found = true;
1914 break;
1915 }
1916 }
1917 if (!found && elt != loop->_head) {
1918 nstack.push(n, idx);
1919 n = elt;
1920 assert(!visited.test(n->_idx), "not seen yet");
1921 visited.set(n->_idx);
1922 }
1923 }
1924 }
1925
1926 // traverse out's that are in the member set
1927 while (true) {
1928 if (idx < n->outcnt()) {
1929 Node* use = n->raw_out(idx);
1930 idx++;
1931 if (!visited.test_set(use->_idx)) {
1932 if (member.test(use->_idx)) {
1933 nstack.push(n, idx);
1934 n = use;
1935 idx = 0;
1936 }
1937 }
1938 } else {
1939 // All outputs processed
1940 sched.push(n);
1941 if (nstack.is_empty()) break;
1942 n = nstack.node();
1943 idx = nstack.index();
1944 nstack.pop();
1945 }
1946 }
1947 }
1948
1949
1950 //------------------------------ has_use_in_set -------------------------------------
1951 // Has a use in the vector set
1952 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
1953 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1954 Node* use = n->fast_out(j);
1955 if (vset.test(use->_idx)) {
1956 return true;
1957 }
1958 }
1959 return false;
1960 }
1961
1962
1963 //------------------------------ has_use_internal_to_set -------------------------------------
1964 // Has use internal to the vector set (ie. not in a phi at the loop head)
1965 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
1966 Node* head = loop->_head;
1967 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1968 Node* use = n->fast_out(j);
1969 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
1970 return true;
1971 }
1972 }
1973 return false;
1974 }
1975
1976
1977 //------------------------------ clone_for_use_outside_loop -------------------------------------
1978 // clone "n" for uses that are outside of loop
1979 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
1980 int cloned = 0;
1981 assert(worklist.size() == 0, "should be empty");
1982 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1983 Node* use = n->fast_out(j);
1984 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
1985 worklist.push(use);
1986 }
1987 }
1988 while( worklist.size() ) {
1989 Node *use = worklist.pop();
1990 if (!has_node(use) || use->in(0) == C->top()) continue;
1991 uint j;
1992 for (j = 0; j < use->req(); j++) {
1993 if (use->in(j) == n) break;
1994 }
1995 assert(j < use->req(), "must be there");
1996
1997 // clone "n" and insert it between the inputs of "n" and the use outside the loop
1998 Node* n_clone = n->clone();
1999 _igvn.replace_input_of(use, j, n_clone);
2000 cloned++;
2001 Node* use_c;
2002 if (!use->is_Phi()) {
2003 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
2004 } else {
2005 // Use in a phi is considered a use in the associated predecessor block
2006 use_c = use->in(0)->in(j);
2007 }
2008 set_ctrl(n_clone, use_c);
2009 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
2010 get_loop(use_c)->_body.push(n_clone);
2011 _igvn.register_new_node_with_optimizer(n_clone);
2012 #if !defined(PRODUCT)
2013 if (TracePartialPeeling) {
2014 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
2015 }
2016 #endif
2017 }
2018 return cloned;
2019 }
2020
2021
2022 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
2023 // clone "n" for special uses that are in the not_peeled region.
2024 // If these def-uses occur in separate blocks, the code generator
2025 // marks the method as not compilable. For example, if a "BoolNode"
2026 // is in a different basic block than the "IfNode" that uses it, then
2027 // the compilation is aborted in the code generator.
2028 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
2029 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
2030 if (n->is_Phi() || n->is_Load()) {
2031 return;
2032 }
2033 assert(worklist.size() == 0, "should be empty");
2034 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2035 Node* use = n->fast_out(j);
2036 if ( not_peel.test(use->_idx) &&
2037 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
2038 use->in(1) == n) {
2039 worklist.push(use);
2040 }
2041 }
2042 if (worklist.size() > 0) {
2043 // clone "n" and insert it between inputs of "n" and the use
2044 Node* n_clone = n->clone();
2045 loop->_body.push(n_clone);
2046 _igvn.register_new_node_with_optimizer(n_clone);
2047 set_ctrl(n_clone, get_ctrl(n));
2048 sink_list.push(n_clone);
2049 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
2050 #if !defined(PRODUCT)
2051 if (TracePartialPeeling) {
2052 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2053 }
2054 #endif
2055 while( worklist.size() ) {
2056 Node *use = worklist.pop();
2057 _igvn.rehash_node_delayed(use);
2058 for (uint j = 1; j < use->req(); j++) {
2059 if (use->in(j) == n) {
2060 use->set_req(j, n_clone);
2061 }
2062 }
2063 }
2064 }
2065 }
2066
2067
2068 //------------------------------ insert_phi_for_loop -------------------------------------
2069 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2070 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2071 Node *phi = PhiNode::make(lp, back_edge_val);
2072 phi->set_req(LoopNode::EntryControl, lp_entry_val);
2073 // Use existing phi if it already exists
2074 Node *hit = _igvn.hash_find_insert(phi);
2075 if( hit == NULL ) {
2076 _igvn.register_new_node_with_optimizer(phi);
2077 set_ctrl(phi, lp);
2078 } else {
2079 // Remove the new phi from the graph and use the hit
2080 _igvn.remove_dead_node(phi);
2081 phi = hit;
2082 }
2083 _igvn.replace_input_of(use, idx, phi);
2084 }
2085
2086 #ifdef ASSERT
2087 //------------------------------ is_valid_loop_partition -------------------------------------
2088 // Validate the loop partition sets: peel and not_peel
2089 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2090 VectorSet& not_peel ) {
2091 uint i;
2092 // Check that peel_list entries are in the peel set
2093 for (i = 0; i < peel_list.size(); i++) {
2094 if (!peel.test(peel_list.at(i)->_idx)) {
2095 return false;
2096 }
2097 }
2098 // Check at loop members are in one of peel set or not_peel set
2099 for (i = 0; i < loop->_body.size(); i++ ) {
2100 Node *def = loop->_body.at(i);
2101 uint di = def->_idx;
2102 // Check that peel set elements are in peel_list
2103 if (peel.test(di)) {
2104 if (not_peel.test(di)) {
2105 return false;
2106 }
2107 // Must be in peel_list also
2108 bool found = false;
2109 for (uint j = 0; j < peel_list.size(); j++) {
2110 if (peel_list.at(j)->_idx == di) {
2111 found = true;
2112 break;
2113 }
2114 }
2115 if (!found) {
2116 return false;
2117 }
2118 } else if (not_peel.test(di)) {
2119 if (peel.test(di)) {
2120 return false;
2121 }
2122 } else {
2123 return false;
2124 }
2125 }
2126 return true;
2127 }
2128
2129 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2130 // Ensure a use outside of loop is of the right form
2131 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2132 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2133 return (use->is_Phi() &&
2134 use_c->is_Region() && use_c->req() == 3 &&
2135 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2136 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2137 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2138 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2139 }
2140
2141 //------------------------------ is_valid_clone_loop_form -------------------------------------
2142 // Ensure that all uses outside of loop are of the right form
2143 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2144 uint orig_exit_idx, uint clone_exit_idx) {
2145 uint len = peel_list.size();
2146 for (uint i = 0; i < len; i++) {
2147 Node *def = peel_list.at(i);
2148
2149 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2150 Node *use = def->fast_out(j);
2151 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2152 if (!loop->is_member(get_loop(use_c))) {
2153 // use is not in the loop, check for correct structure
2154 if (use->in(0) == def) {
2155 // Okay
2156 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2157 return false;
2158 }
2159 }
2160 }
2161 }
2162 return true;
2163 }
2164 #endif
2165
2166 //------------------------------ partial_peel -------------------------------------
2167 // Partially peel (aka loop rotation) the top portion of a loop (called
2168 // the peel section below) by cloning it and placing one copy just before
2169 // the new loop head and the other copy at the bottom of the new loop.
2170 //
2171 // before after where it came from
2172 //
2173 // stmt1 stmt1
2174 // loop: stmt2 clone
2175 // stmt2 if condA goto exitA clone
2176 // if condA goto exitA new_loop: new
2177 // stmt3 stmt3 clone
2178 // if !condB goto loop if condB goto exitB clone
2179 // exitB: stmt2 orig
2180 // stmt4 if !condA goto new_loop orig
2181 // exitA: goto exitA
2182 // exitB:
2183 // stmt4
2184 // exitA:
2185 //
2186 // Step 1: find the cut point: an exit test on probable
2187 // induction variable.
2188 // Step 2: schedule (with cloning) operations in the peel
2189 // section that can be executed after the cut into
2190 // the section that is not peeled. This may need
2191 // to clone operations into exit blocks. For
2192 // instance, a reference to A[i] in the not-peel
2193 // section and a reference to B[i] in an exit block
2194 // may cause a left-shift of i by 2 to be placed
2195 // in the peel block. This step will clone the left
2196 // shift into the exit block and sink the left shift
2197 // from the peel to the not-peel section.
2198 // Step 3: clone the loop, retarget the control, and insert
2199 // phis for values that are live across the new loop
2200 // head. This is very dependent on the graph structure
2201 // from clone_loop. It creates region nodes for
2202 // exit control and associated phi nodes for values
2203 // flow out of the loop through that exit. The region
2204 // node is dominated by the clone's control projection.
2205 // So the clone's peel section is placed before the
2206 // new loop head, and the clone's not-peel section is
2207 // forms the top part of the new loop. The original
2208 // peel section forms the tail of the new loop.
2209 // Step 4: update the dominator tree and recompute the
2210 // dominator depth.
2211 //
2212 // orig
2213 //
2214 // stmt1
2215 // |
2216 // v
2217 // loop predicate
2218 // |
2219 // v
2220 // loop<----+
2221 // | |
2222 // stmt2 |
2223 // | |
2224 // v |
2225 // ifA |
2226 // / | |
2227 // v v |
2228 // false true ^ <-- last_peel
2229 // / | |
2230 // / ===|==cut |
2231 // / stmt3 | <-- first_not_peel
2232 // / | |
2233 // | v |
2234 // v ifB |
2235 // exitA: / \ |
2236 // / \ |
2237 // v v |
2238 // false true |
2239 // / \ |
2240 // / ----+
2241 // |
2242 // v
2243 // exitB:
2244 // stmt4
2245 //
2246 //
2247 // after clone loop
2248 //
2249 // stmt1
2250 // |
2251 // v
2252 // loop predicate
2253 // / \
2254 // clone / \ orig
2255 // / \
2256 // / \
2257 // v v
2258 // +---->loop loop<----+
2259 // | | | |
2260 // | stmt2 stmt2 |
2261 // | | | |
2262 // | v v |
2263 // | ifA ifA |
2264 // | | \ / | |
2265 // | v v v v |
2266 // ^ true false false true ^ <-- last_peel
2267 // | | ^ \ / | |
2268 // | cut==|== \ \ / ===|==cut |
2269 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2270 // | | dom | | | |
2271 // | v \ 1v v2 v |
2272 // | ifB regionA ifB |
2273 // | / \ | / \ |
2274 // | / \ v / \ |
2275 // | v v exitA: v v |
2276 // | true false false true |
2277 // | / ^ \ / \ |
2278 // +---- \ \ / ----+
2279 // dom \ /
2280 // \ 1v v2
2281 // regionB
2282 // |
2283 // v
2284 // exitB:
2285 // stmt4
2286 //
2287 //
2288 // after partial peel
2289 //
2290 // stmt1
2291 // |
2292 // v
2293 // loop predicate
2294 // /
2295 // clone / orig
2296 // / TOP
2297 // / \
2298 // v v
2299 // TOP->loop loop----+
2300 // | | |
2301 // stmt2 stmt2 |
2302 // | | |
2303 // v v |
2304 // ifA ifA |
2305 // | \ / | |
2306 // v v v v |
2307 // true false false true | <-- last_peel
2308 // | ^ \ / +------|---+
2309 // +->newloop \ \ / === ==cut | |
2310 // | stmt3 \ \ / TOP | |
2311 // | | dom | | stmt3 | | <-- first_not_peel
2312 // | v \ 1v v2 v | |
2313 // | ifB regionA ifB ^ v
2314 // | / \ | / \ | |
2315 // | / \ v / \ | |
2316 // | v v exitA: v v | |
2317 // | true false false true | |
2318 // | / ^ \ / \ | |
2319 // | | \ \ / v | |
2320 // | | dom \ / TOP | |
2321 // | | \ 1v v2 | |
2322 // ^ v regionB | |
2323 // | | | | |
2324 // | | v ^ v
2325 // | | exitB: | |
2326 // | | stmt4 | |
2327 // | +------------>-----------------+ |
2328 // | |
2329 // +-----------------<---------------------+
2330 //
2331 //
2332 // final graph
2333 //
2334 // stmt1
2335 // |
2336 // v
2337 // loop predicate
2338 // |
2339 // v
2340 // stmt2 clone
2341 // |
2342 // v
2343 // ........> ifA clone
2344 // : / |
2345 // dom / |
2346 // : v v
2347 // : false true
2348 // : | |
2349 // : | v
2350 // : | newloop<-----+
2351 // : | | |
2352 // : | stmt3 clone |
2353 // : | | |
2354 // : | v |
2355 // : | ifB |
2356 // : | / \ |
2357 // : | v v |
2358 // : | false true |
2359 // : | | | |
2360 // : | v stmt2 |
2361 // : | exitB: | |
2362 // : | stmt4 v |
2363 // : | ifA orig |
2364 // : | / \ |
2365 // : | / \ |
2366 // : | v v |
2367 // : | false true |
2368 // : | / \ |
2369 // : v v -----+
2370 // RegionA
2371 // |
2372 // v
2373 // exitA
2374 //
2375 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
2376
2377 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
2378 if (!loop->_head->is_Loop()) {
2379 return false; }
2380
2381 LoopNode *head = loop->_head->as_Loop();
2382
2383 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
2384 return false;
2385 }
2386
2387 // Check for complex exit control
2388 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
2389 Node *n = loop->_body.at(ii);
2390 int opc = n->Opcode();
2391 if (n->is_Call() ||
2392 opc == Op_Catch ||
2393 opc == Op_CatchProj ||
2394 opc == Op_Jump ||
2395 opc == Op_JumpProj) {
2396 #if !defined(PRODUCT)
2397 if (TracePartialPeeling) {
2398 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
2399 }
2400 #endif
2401 return false;
2402 }
2403 }
2404
2405 int dd = dom_depth(head);
2406
2407 // Step 1: find cut point
2408
2409 // Walk up dominators to loop head looking for first loop exit
2410 // which is executed on every path thru loop.
2411 IfNode *peel_if = NULL;
2412 IfNode *peel_if_cmpu = NULL;
2413
2414 Node *iff = loop->tail();
2415 while( iff != head ) {
2416 if( iff->is_If() ) {
2417 Node *ctrl = get_ctrl(iff->in(1));
2418 if (ctrl->is_top()) return false; // Dead test on live IF.
2419 // If loop-varying exit-test, check for induction variable
2420 if( loop->is_member(get_loop(ctrl)) &&
2421 loop->is_loop_exit(iff) &&
2422 is_possible_iv_test(iff)) {
2423 Node* cmp = iff->in(1)->in(1);
2424 if (cmp->Opcode() == Op_CmpI) {
2425 peel_if = iff->as_If();
2426 } else {
2427 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
2428 peel_if_cmpu = iff->as_If();
2429 }
2430 }
2431 }
2432 iff = idom(iff);
2433 }
2434 // Prefer signed compare over unsigned compare.
2435 IfNode* new_peel_if = NULL;
2436 if (peel_if == NULL) {
2437 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
2438 return false; // No peel point found
2439 }
2440 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
2441 if (new_peel_if == NULL) {
2442 return false; // No peel point found
2443 }
2444 peel_if = new_peel_if;
2445 }
2446 Node* last_peel = stay_in_loop(peel_if, loop);
2447 Node* first_not_peeled = stay_in_loop(last_peel, loop);
2448 if (first_not_peeled == NULL || first_not_peeled == head) {
2449 return false;
2450 }
2451
2452 #if !defined(PRODUCT)
2453 if (TraceLoopOpts) {
2454 tty->print("PartialPeel ");
2455 loop->dump_head();
2456 }
2457
2458 if (TracePartialPeeling) {
2459 tty->print_cr("before partial peel one iteration");
2460 Node_List wl;
2461 Node* t = head->in(2);
2462 while (true) {
2463 wl.push(t);
2464 if (t == head) break;
2465 t = idom(t);
2466 }
2467 while (wl.size() > 0) {
2468 Node* tt = wl.pop();
2469 tt->dump();
2470 if (tt == last_peel) tty->print_cr("-- cut --");
2471 }
2472 }
2473 #endif
2474 ResourceArea *area = Thread::current()->resource_area();
2475 VectorSet peel(area);
2476 VectorSet not_peel(area);
2477 Node_List peel_list(area);
2478 Node_List worklist(area);
2479 Node_List sink_list(area);
2480
2481 // Set of cfg nodes to peel are those that are executable from
2482 // the head through last_peel.
2483 assert(worklist.size() == 0, "should be empty");
2484 worklist.push(head);
2485 peel.set(head->_idx);
2486 while (worklist.size() > 0) {
2487 Node *n = worklist.pop();
2488 if (n != last_peel) {
2489 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2490 Node* use = n->fast_out(j);
2491 if (use->is_CFG() &&
2492 loop->is_member(get_loop(use)) &&
2493 !peel.test_set(use->_idx)) {
2494 worklist.push(use);
2495 }
2496 }
2497 }
2498 }
2499
2500 // Set of non-cfg nodes to peel are those that are control
2501 // dependent on the cfg nodes.
2502 uint i;
2503 for(i = 0; i < loop->_body.size(); i++ ) {
2504 Node *n = loop->_body.at(i);
2505 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
2506 if (peel.test(n_c->_idx)) {
2507 peel.set(n->_idx);
2508 } else {
2509 not_peel.set(n->_idx);
2510 }
2511 }
2512
2513 // Step 2: move operations from the peeled section down into the
2514 // not-peeled section
2515
2516 // Get a post order schedule of nodes in the peel region
2517 // Result in right-most operand.
2518 scheduled_nodelist(loop, peel, peel_list );
2519
2520 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2521
2522 // For future check for too many new phis
2523 uint old_phi_cnt = 0;
2524 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2525 Node* use = head->fast_out(j);
2526 if (use->is_Phi()) old_phi_cnt++;
2527 }
2528
2529 #if !defined(PRODUCT)
2530 if (TracePartialPeeling) {
2531 tty->print_cr("\npeeled list");
2532 }
2533 #endif
2534
2535 // Evacuate nodes in peel region into the not_peeled region if possible
2536 uint new_phi_cnt = 0;
2537 uint cloned_for_outside_use = 0;
2538 for (i = 0; i < peel_list.size();) {
2539 Node* n = peel_list.at(i);
2540 #if !defined(PRODUCT)
2541 if (TracePartialPeeling) n->dump();
2542 #endif
2543 bool incr = true;
2544 if ( !n->is_CFG() ) {
2545
2546 if ( has_use_in_set(n, not_peel) ) {
2547
2548 // If not used internal to the peeled region,
2549 // move "n" from peeled to not_peeled region.
2550
2551 if ( !has_use_internal_to_set(n, peel, loop) ) {
2552
2553 // if not pinned and not a load (which maybe anti-dependent on a store)
2554 // and not a CMove (Matcher expects only bool->cmove).
2555 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
2556 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist );
2557 sink_list.push(n);
2558 peel >>= n->_idx; // delete n from peel set.
2559 not_peel <<= n->_idx; // add n to not_peel set.
2560 peel_list.remove(i);
2561 incr = false;
2562 #if !defined(PRODUCT)
2563 if (TracePartialPeeling) {
2564 tty->print_cr("sink to not_peeled region: %d newbb: %d",
2565 n->_idx, get_ctrl(n)->_idx);
2566 }
2567 #endif
2568 }
2569 } else {
2570 // Otherwise check for special def-use cases that span
2571 // the peel/not_peel boundary such as bool->if
2572 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
2573 new_phi_cnt++;
2574 }
2575 }
2576 }
2577 if (incr) i++;
2578 }
2579
2580 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
2581 #if !defined(PRODUCT)
2582 if (TracePartialPeeling) {
2583 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
2584 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
2585 }
2586 #endif
2587 if (new_peel_if != NULL) {
2588 remove_cmpi_loop_exit(new_peel_if, loop);
2589 }
2590 // Inhibit more partial peeling on this loop
2591 assert(!head->is_partial_peel_loop(), "not partial peeled");
2592 head->mark_partial_peel_failed();
2593 if (cloned_for_outside_use > 0) {
2594 // Terminate this round of loop opts because
2595 // the graph outside this loop was changed.
2596 C->set_major_progress();
2597 return true;
2598 }
2599 return false;
2600 }
2601
2602 // Step 3: clone loop, retarget control, and insert new phis
2603
2604 // Create new loop head for new phis and to hang
2605 // the nodes being moved (sinked) from the peel region.
2606 LoopNode* new_head = new (C) LoopNode(last_peel, last_peel);
2607 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
2608 _igvn.register_new_node_with_optimizer(new_head);
2609 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
2610 first_not_peeled->set_req(0, new_head);
2611 set_loop(new_head, loop);
2612 loop->_body.push(new_head);
2613 not_peel.set(new_head->_idx);
2614 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
2615 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
2616
2617 while (sink_list.size() > 0) {
2618 Node* n = sink_list.pop();
2619 set_ctrl(n, new_head);
2620 }
2621
2622 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2623
2624 clone_loop( loop, old_new, dd );
2625
2626 const uint clone_exit_idx = 1;
2627 const uint orig_exit_idx = 2;
2628 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
2629
2630 Node* head_clone = old_new[head->_idx];
2631 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
2632 Node* orig_tail_clone = head_clone->in(2);
2633
2634 // Add phi if "def" node is in peel set and "use" is not
2635
2636 for(i = 0; i < peel_list.size(); i++ ) {
2637 Node *def = peel_list.at(i);
2638 if (!def->is_CFG()) {
2639 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2640 Node *use = def->fast_out(j);
2641 if (has_node(use) && use->in(0) != C->top() &&
2642 (!peel.test(use->_idx) ||
2643 (use->is_Phi() && use->in(0) == head)) ) {
2644 worklist.push(use);
2645 }
2646 }
2647 while( worklist.size() ) {
2648 Node *use = worklist.pop();
2649 for (uint j = 1; j < use->req(); j++) {
2650 Node* n = use->in(j);
2651 if (n == def) {
2652
2653 // "def" is in peel set, "use" is not in peel set
2654 // or "use" is in the entry boundary (a phi) of the peel set
2655
2656 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
2657
2658 if ( loop->is_member(get_loop( use_c )) ) {
2659 // use is in loop
2660 if (old_new[use->_idx] != NULL) { // null for dead code
2661 Node* use_clone = old_new[use->_idx];
2662 _igvn.replace_input_of(use, j, C->top());
2663 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
2664 }
2665 } else {
2666 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
2667 // use is not in the loop, check if the live range includes the cut
2668 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
2669 if (not_peel.test(lp_if->_idx)) {
2670 assert(j == orig_exit_idx, "use from original loop");
2671 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
2672 }
2673 }
2674 }
2675 }
2676 }
2677 }
2678 }
2679
2680 // Step 3b: retarget control
2681
2682 // Redirect control to the new loop head if a cloned node in
2683 // the not_peeled region has control that points into the peeled region.
2684 // This necessary because the cloned peeled region will be outside
2685 // the loop.
2686 // from to
2687 // cloned-peeled <---+
2688 // new_head_clone: | <--+
2689 // cloned-not_peeled in(0) in(0)
2690 // orig-peeled
2691
2692 for(i = 0; i < loop->_body.size(); i++ ) {
2693 Node *n = loop->_body.at(i);
2694 if (!n->is_CFG() && n->in(0) != NULL &&
2695 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
2696 Node* n_clone = old_new[n->_idx];
2697 _igvn.replace_input_of(n_clone, 0, new_head_clone);
2698 }
2699 }
2700
2701 // Backedge of the surviving new_head (the clone) is original last_peel
2702 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
2703
2704 // Cut first node in original not_peel set
2705 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
2706 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
2707 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
2708
2709 // Copy head_clone back-branch info to original head
2710 // and remove original head's loop entry and
2711 // clone head's back-branch
2712 _igvn.rehash_node_delayed(head); // Multiple edge updates
2713 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
2714 head->set_req(LoopNode::LoopBackControl, C->top());
2715 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
2716
2717 // Similarly modify the phis
2718 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
2719 Node* use = head->fast_out(k);
2720 if (use->is_Phi() && use->outcnt() > 0) {
2721 Node* use_clone = old_new[use->_idx];
2722 _igvn.rehash_node_delayed(use); // Multiple edge updates
2723 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
2724 use->set_req(LoopNode::LoopBackControl, C->top());
2725 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
2726 }
2727 }
2728
2729 // Step 4: update dominator tree and dominator depth
2730
2731 set_idom(head, orig_tail_clone, dd);
2732 recompute_dom_depth();
2733
2734 // Inhibit more partial peeling on this loop
2735 new_head_clone->set_partial_peel_loop();
2736 C->set_major_progress();
2737 loop->record_for_igvn();
2738
2739 #if !defined(PRODUCT)
2740 if (TracePartialPeeling) {
2741 tty->print_cr("\nafter partial peel one iteration");
2742 Node_List wl(area);
2743 Node* t = last_peel;
2744 while (true) {
2745 wl.push(t);
2746 if (t == head_clone) break;
2747 t = idom(t);
2748 }
2749 while (wl.size() > 0) {
2750 Node* tt = wl.pop();
2751 if (tt == head) tty->print_cr("orig head");
2752 else if (tt == new_head_clone) tty->print_cr("new head");
2753 else if (tt == head_clone) tty->print_cr("clone head");
2754 tt->dump();
2755 }
2756 }
2757 #endif
2758 return true;
2759 }
2760
2761 //------------------------------reorg_offsets----------------------------------
2762 // Reorganize offset computations to lower register pressure. Mostly
2763 // prevent loop-fallout uses of the pre-incremented trip counter (which are
2764 // then alive with the post-incremented trip counter forcing an extra
2765 // register move)
2766 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
2767 // Perform it only for canonical counted loops.
2768 // Loop's shape could be messed up by iteration_split_impl.
2769 if (!loop->_head->is_CountedLoop())
2770 return;
2771 if (!loop->_head->as_Loop()->is_valid_counted_loop())
2772 return;
2773
2774 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2775 CountedLoopEndNode *cle = cl->loopexit();
2776 Node *exit = cle->proj_out(false);
2777 Node *phi = cl->phi();
2778
2779 // Check for the special case of folks using the pre-incremented
2780 // trip-counter on the fall-out path (forces the pre-incremented
2781 // and post-incremented trip counter to be live at the same time).
2782 // Fix this by adjusting to use the post-increment trip counter.
2783
2784 bool progress = true;
2785 while (progress) {
2786 progress = false;
2787 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
2788 Node* use = phi->fast_out(i); // User of trip-counter
2789 if (!has_ctrl(use)) continue;
2790 Node *u_ctrl = get_ctrl(use);
2791 if (use->is_Phi()) {
2792 u_ctrl = NULL;
2793 for (uint j = 1; j < use->req(); j++)
2794 if (use->in(j) == phi)
2795 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
2796 }
2797 IdealLoopTree *u_loop = get_loop(u_ctrl);
2798 // Look for loop-invariant use
2799 if (u_loop == loop) continue;
2800 if (loop->is_member(u_loop)) continue;
2801 // Check that use is live out the bottom. Assuming the trip-counter
2802 // update is right at the bottom, uses of of the loop middle are ok.
2803 if (dom_lca(exit, u_ctrl) != exit) continue;
2804 // Hit! Refactor use to use the post-incremented tripcounter.
2805 // Compute a post-increment tripcounter.
2806 Node *opaq = new (C) Opaque2Node( C, cle->incr() );
2807 register_new_node(opaq, exit);
2808 Node *neg_stride = _igvn.intcon(-cle->stride_con());
2809 set_ctrl(neg_stride, C->root());
2810 Node *post = new (C) AddINode( opaq, neg_stride);
2811 register_new_node(post, exit);
2812 _igvn.rehash_node_delayed(use);
2813 for (uint j = 1; j < use->req(); j++) {
2814 if (use->in(j) == phi)
2815 use->set_req(j, post);
2816 }
2817 // Since DU info changed, rerun loop
2818 progress = true;
2819 break;
2820 }
2821 }
2822
2823 }