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