1 /*
2 * Copyright (c) 1999, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "memory/allocation.inline.hpp"
27 #include "memory/resourceArea.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/castnode.hpp"
31 #include "opto/connode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/loopnode.hpp"
35 #include "opto/matcher.hpp"
36 #include "opto/mulnode.hpp"
37 #include "opto/movenode.hpp"
38 #include "opto/opaquenode.hpp"
39 #include "opto/rootnode.hpp"
40 #include "opto/subnode.hpp"
41
42 //=============================================================================
43 //------------------------------split_thru_phi---------------------------------
44 // Split Node 'n' through merge point if there is enough win.
45 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
46 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
47 // ConvI2L may have type information on it which is unsafe to push up
48 // so disable this for now
49 return NULL;
50 }
51
52 // Splitting range check CastIIs through a loop induction Phi can
53 // cause new Phis to be created that are left unrelated to the loop
54 // induction Phi and prevent optimizations (vectorization)
55 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() &&
56 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) {
57 return NULL;
58 }
59
60 int wins = 0;
61 assert(!n->is_CFG(), "");
62 assert(region->is_Region(), "");
63
64 const Type* type = n->bottom_type();
65 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr();
66 Node *phi;
67 if (t_oop != NULL && t_oop->is_known_instance_field()) {
68 int iid = t_oop->instance_id();
69 int index = C->get_alias_index(t_oop);
70 int offset = t_oop->offset();
71 phi = new PhiNode(region, type, NULL, iid, index, offset);
72 } else {
73 phi = PhiNode::make_blank(region, n);
74 }
75 uint old_unique = C->unique();
76 for (uint i = 1; i < region->req(); i++) {
77 Node *x;
78 Node* the_clone = NULL;
79 if (region->in(i) == C->top()) {
80 x = C->top(); // Dead path? Use a dead data op
81 } else {
82 x = n->clone(); // Else clone up the data op
83 the_clone = x; // Remember for possible deletion.
84 // Alter data node to use pre-phi inputs
85 if (n->in(0) == region)
86 x->set_req( 0, region->in(i) );
87 for (uint j = 1; j < n->req(); j++) {
88 Node *in = n->in(j);
89 if (in->is_Phi() && in->in(0) == region)
90 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
91 }
92 }
93 // Check for a 'win' on some paths
94 const Type *t = x->Value(&_igvn);
95
96 bool singleton = t->singleton();
97
98 // A TOP singleton indicates that there are no possible values incoming
99 // along a particular edge. In most cases, this is OK, and the Phi will
100 // be eliminated later in an Ideal call. However, we can't allow this to
101 // happen if the singleton occurs on loop entry, as the elimination of
102 // the PhiNode may cause the resulting node to migrate back to a previous
103 // loop iteration.
104 if (singleton && t == Type::TOP) {
105 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
106 // irreducible loop may not be indicated by an affirmative is_Loop());
107 // therefore, the only top we can split thru a phi is on a backedge of
108 // a loop.
109 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
110 }
111
112 if (singleton) {
113 wins++;
114 x = ((PhaseGVN&)_igvn).makecon(t);
115 } else {
116 // We now call Identity to try to simplify the cloned node.
117 // Note that some Identity methods call phase->type(this).
118 // Make sure that the type array is big enough for
119 // our new node, even though we may throw the node away.
120 // (Note: This tweaking with igvn only works because x is a new node.)
121 _igvn.set_type(x, t);
122 // If x is a TypeNode, capture any more-precise type permanently into Node
123 // otherwise it will be not updated during igvn->transform since
124 // igvn->type(x) is set to x->Value() already.
125 x->raise_bottom_type(t);
126 Node *y = x->Identity(&_igvn);
127 if (y != x) {
128 wins++;
129 x = y;
130 } else {
131 y = _igvn.hash_find(x);
132 if (y) {
133 wins++;
134 x = y;
135 } else {
136 // Else x is a new node we are keeping
137 // We do not need register_new_node_with_optimizer
138 // because set_type has already been called.
139 _igvn._worklist.push(x);
140 }
141 }
142 }
143 if (x != the_clone && the_clone != NULL)
144 _igvn.remove_dead_node(the_clone);
145 phi->set_req( i, x );
146 }
147 // Too few wins?
148 if (wins <= policy) {
149 _igvn.remove_dead_node(phi);
150 return NULL;
151 }
152
153 // Record Phi
154 register_new_node( phi, region );
155
156 for (uint i2 = 1; i2 < phi->req(); i2++) {
157 Node *x = phi->in(i2);
158 // If we commoned up the cloned 'x' with another existing Node,
159 // the existing Node picks up a new use. We need to make the
160 // existing Node occur higher up so it dominates its uses.
161 Node *old_ctrl;
162 IdealLoopTree *old_loop;
163
164 if (x->is_Con()) {
165 // Constant's control is always root.
166 set_ctrl(x, C->root());
167 continue;
168 }
169 // The occasional new node
170 if (x->_idx >= old_unique) { // Found a new, unplaced node?
171 old_ctrl = NULL;
172 old_loop = NULL; // Not in any prior loop
173 } else {
174 old_ctrl = get_ctrl(x);
175 old_loop = get_loop(old_ctrl); // Get prior loop
176 }
177 // New late point must dominate new use
178 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
179 if (new_ctrl == old_ctrl) // Nothing is changed
180 continue;
181
182 IdealLoopTree *new_loop = get_loop(new_ctrl);
183
184 // Don't move x into a loop if its uses are
185 // outside of loop. Otherwise x will be cloned
186 // for each use outside of this loop.
187 IdealLoopTree *use_loop = get_loop(region);
188 if (!new_loop->is_member(use_loop) &&
189 (old_loop == NULL || !new_loop->is_member(old_loop))) {
190 // Take early control, later control will be recalculated
191 // during next iteration of loop optimizations.
192 new_ctrl = get_early_ctrl(x);
193 new_loop = get_loop(new_ctrl);
194 }
195 // Set new location
196 set_ctrl(x, new_ctrl);
197 // If changing loop bodies, see if we need to collect into new body
198 if (old_loop != new_loop) {
199 if (old_loop && !old_loop->_child)
200 old_loop->_body.yank(x);
201 if (!new_loop->_child)
202 new_loop->_body.push(x); // Collect body info
203 }
204 }
205
206 return phi;
207 }
208
209 //------------------------------dominated_by------------------------------------
210 // Replace the dominated test with an obvious true or false. Place it on the
211 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
212 // live path up to the dominating control.
213 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) {
214 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
215
216 // prevdom is the dominating projection of the dominating test.
217 assert( iff->is_If(), "" );
218 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added");
219 int pop = prevdom->Opcode();
220 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
221 if (flip) {
222 if (pop == Op_IfTrue)
223 pop = Op_IfFalse;
224 else
225 pop = Op_IfTrue;
226 }
227 // 'con' is set to true or false to kill the dominated test.
228 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
229 set_ctrl(con, C->root()); // Constant gets a new use
230 // Hack the dominated test
231 _igvn.replace_input_of(iff, 1, con);
232
233 // If I dont have a reachable TRUE and FALSE path following the IfNode then
234 // I can assume this path reaches an infinite loop. In this case it's not
235 // important to optimize the data Nodes - either the whole compilation will
236 // be tossed or this path (and all data Nodes) will go dead.
237 if (iff->outcnt() != 2) return;
238
239 // Make control-dependent data Nodes on the live path (path that will remain
240 // once the dominated IF is removed) become control-dependent on the
241 // dominating projection.
242 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue);
243
244 // Loop predicates may have depending checks which should not
245 // be skipped. For example, range check predicate has two checks
246 // for lower and upper bounds.
247 if (dp == NULL)
248 return;
249
250 ProjNode* dp_proj = dp->as_Proj();
251 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj();
252 if (exclude_loop_predicate &&
253 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL ||
254 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) {
255 // If this is a range check (IfNode::is_range_check), do not
256 // reorder because Compile::allow_range_check_smearing might have
257 // changed the check.
258 return; // Let IGVN transformation change control dependence.
259 }
260
261 IdealLoopTree *old_loop = get_loop(dp);
262
263 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
264 Node* cd = dp->fast_out(i); // Control-dependent node
265 if (cd->depends_only_on_test()) {
266 assert(cd->in(0) == dp, "");
267 _igvn.replace_input_of(cd, 0, prevdom);
268 set_early_ctrl(cd);
269 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
270 if (old_loop != new_loop) {
271 if (!old_loop->_child) old_loop->_body.yank(cd);
272 if (!new_loop->_child) new_loop->_body.push(cd);
273 }
274 --i;
275 --imax;
276 }
277 }
278 }
279
280 //------------------------------has_local_phi_input----------------------------
281 // Return TRUE if 'n' has Phi inputs from its local block and no other
282 // block-local inputs (all non-local-phi inputs come from earlier blocks)
283 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
284 Node *n_ctrl = get_ctrl(n);
285 // See if some inputs come from a Phi in this block, or from before
286 // this block.
287 uint i;
288 for( i = 1; i < n->req(); i++ ) {
289 Node *phi = n->in(i);
290 if( phi->is_Phi() && phi->in(0) == n_ctrl )
291 break;
292 }
293 if( i >= n->req() )
294 return NULL; // No Phi inputs; nowhere to clone thru
295
296 // Check for inputs created between 'n' and the Phi input. These
297 // must split as well; they have already been given the chance
298 // (courtesy of a post-order visit) and since they did not we must
299 // recover the 'cost' of splitting them by being very profitable
300 // when splitting 'n'. Since this is unlikely we simply give up.
301 for( i = 1; i < n->req(); i++ ) {
302 Node *m = n->in(i);
303 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
304 // We allow the special case of AddP's with no local inputs.
305 // This allows us to split-up address expressions.
306 if (m->is_AddP() &&
307 get_ctrl(m->in(2)) != n_ctrl &&
308 get_ctrl(m->in(3)) != n_ctrl) {
309 // Move the AddP up to dominating point
310 Node* c = find_non_split_ctrl(idom(n_ctrl));
311 if (c->is_OuterStripMinedLoop()) {
312 c->as_Loop()->verify_strip_mined(1);
313 c = c->in(LoopNode::EntryControl);
314 }
315 set_ctrl_and_loop(m, c);
316 continue;
317 }
318 return NULL;
319 }
320 assert(m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
321 }
322
323 return n_ctrl;
324 }
325
326 //------------------------------remix_address_expressions----------------------
327 // Rework addressing expressions to get the most loop-invariant stuff
328 // moved out. We'd like to do all associative operators, but it's especially
329 // important (common) to do address expressions.
330 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
331 if (!has_ctrl(n)) return NULL;
332 Node *n_ctrl = get_ctrl(n);
333 IdealLoopTree *n_loop = get_loop(n_ctrl);
334
335 // See if 'n' mixes loop-varying and loop-invariant inputs and
336 // itself is loop-varying.
337
338 // Only interested in binary ops (and AddP)
339 if( n->req() < 3 || n->req() > 4 ) return NULL;
340
341 Node *n1_ctrl = get_ctrl(n->in( 1));
342 Node *n2_ctrl = get_ctrl(n->in( 2));
343 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
344 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
345 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
346 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
347
348 // Does one of my inputs spin in a tighter loop than self?
349 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
350 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
351 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
352 return NULL; // Leave well enough alone
353
354 // Is at least one of my inputs loop-invariant?
355 if( n1_loop == n_loop &&
356 n2_loop == n_loop &&
357 n3_loop == n_loop )
358 return NULL; // No loop-invariant inputs
359
360
361 int n_op = n->Opcode();
362
363 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
364 if( n_op == Op_LShiftI ) {
365 // Scale is loop invariant
366 Node *scale = n->in(2);
367 Node *scale_ctrl = get_ctrl(scale);
368 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
369 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
370 return NULL;
371 const TypeInt *scale_t = scale->bottom_type()->isa_int();
372 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
373 return NULL; // Dont bother with byte/short masking
374 // Add must vary with loop (else shift would be loop-invariant)
375 Node *add = n->in(1);
376 Node *add_ctrl = get_ctrl(add);
377 IdealLoopTree *add_loop = get_loop(add_ctrl);
378 //assert( n_loop == add_loop, "" );
379 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
380
381 // Convert I-V into I+ (0-V); same for V-I
382 if( add->Opcode() == Op_SubI &&
383 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
384 Node *zero = _igvn.intcon(0);
385 set_ctrl(zero, C->root());
386 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) );
387 register_new_node( neg, get_ctrl(add->in(2) ) );
388 add = new AddINode( add->in(1), neg );
389 register_new_node( add, add_ctrl );
390 }
391 if( add->Opcode() != Op_AddI ) return NULL;
392 // See if one add input is loop invariant
393 Node *add_var = add->in(1);
394 Node *add_var_ctrl = get_ctrl(add_var);
395 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
396 Node *add_invar = add->in(2);
397 Node *add_invar_ctrl = get_ctrl(add_invar);
398 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
399 if( add_var_loop == n_loop ) {
400 } else if( add_invar_loop == n_loop ) {
401 // Swap to find the invariant part
402 add_invar = add_var;
403 add_invar_ctrl = add_var_ctrl;
404 add_invar_loop = add_var_loop;
405 add_var = add->in(2);
406 Node *add_var_ctrl = get_ctrl(add_var);
407 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
408 } else // Else neither input is loop invariant
409 return NULL;
410 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
411 return NULL; // No invariant part of the add?
412
413 // Yes! Reshape address expression!
414 Node *inv_scale = new LShiftINode( add_invar, scale );
415 Node *inv_scale_ctrl =
416 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
417 add_invar_ctrl : scale_ctrl;
418 register_new_node( inv_scale, inv_scale_ctrl );
419 Node *var_scale = new LShiftINode( add_var, scale );
420 register_new_node( var_scale, n_ctrl );
421 Node *var_add = new AddINode( var_scale, inv_scale );
422 register_new_node( var_add, n_ctrl );
423 _igvn.replace_node( n, var_add );
424 return var_add;
425 }
426
427 // Replace (I+V) with (V+I)
428 if( n_op == Op_AddI ||
429 n_op == Op_AddL ||
430 n_op == Op_AddF ||
431 n_op == Op_AddD ||
432 n_op == Op_MulI ||
433 n_op == Op_MulL ||
434 n_op == Op_MulF ||
435 n_op == Op_MulD ) {
436 if( n2_loop == n_loop ) {
437 assert( n1_loop != n_loop, "" );
438 n->swap_edges(1, 2);
439 }
440 }
441
442 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
443 // but not if I2 is a constant.
444 if( n_op == Op_AddP ) {
445 if( n2_loop == n_loop && n3_loop != n_loop ) {
446 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
447 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
448 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
449 IdealLoopTree *n22loop = get_loop( n22_ctrl );
450 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
451 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
452 n23_loop == n_loop ) {
453 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
454 // Stuff new AddP in the loop preheader
455 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
456 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) );
457 register_new_node( add2, n_ctrl );
458 _igvn.replace_node( n, add2 );
459 return add2;
460 }
461 }
462 }
463
464 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
465 if (n2_loop != n_loop && n3_loop == n_loop) {
466 if (n->in(3)->Opcode() == Op_AddX) {
467 Node *V = n->in(3)->in(1);
468 Node *I = n->in(3)->in(2);
469 if (is_member(n_loop,get_ctrl(V))) {
470 } else {
471 Node *tmp = V; V = I; I = tmp;
472 }
473 if (!is_member(n_loop,get_ctrl(I))) {
474 Node *add1 = new AddPNode(n->in(1), n->in(2), I);
475 // Stuff new AddP in the loop preheader
476 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl));
477 Node *add2 = new AddPNode(n->in(1), add1, V);
478 register_new_node(add2, n_ctrl);
479 _igvn.replace_node(n, add2);
480 return add2;
481 }
482 }
483 }
484 }
485
486 return NULL;
487 }
488
489 //------------------------------conditional_move-------------------------------
490 // Attempt to replace a Phi with a conditional move. We have some pretty
491 // strict profitability requirements. All Phis at the merge point must
492 // be converted, so we can remove the control flow. We need to limit the
493 // number of c-moves to a small handful. All code that was in the side-arms
494 // of the CFG diamond is now speculatively executed. This code has to be
495 // "cheap enough". We are pretty much limited to CFG diamonds that merge
496 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
497 Node *PhaseIdealLoop::conditional_move( Node *region ) {
498
499 assert(region->is_Region(), "sanity check");
500 if (region->req() != 3) return NULL;
501
502 // Check for CFG diamond
503 Node *lp = region->in(1);
504 Node *rp = region->in(2);
505 if (!lp || !rp) return NULL;
506 Node *lp_c = lp->in(0);
507 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
508 IfNode *iff = lp_c->as_If();
509
510 // Check for ops pinned in an arm of the diamond.
511 // Can't remove the control flow in this case
512 if (lp->outcnt() > 1) return NULL;
513 if (rp->outcnt() > 1) return NULL;
514
515 IdealLoopTree* r_loop = get_loop(region);
516 assert(r_loop == get_loop(iff), "sanity");
517 // Always convert to CMOVE if all results are used only outside this loop.
518 bool used_inside_loop = (r_loop == _ltree_root);
519
520 // Check profitability
521 int cost = 0;
522 int phis = 0;
523 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
524 Node *out = region->fast_out(i);
525 if (!out->is_Phi()) continue; // Ignore other control edges, etc
526 phis++;
527 PhiNode* phi = out->as_Phi();
528 BasicType bt = phi->type()->basic_type();
529 switch (bt) {
530 case T_DOUBLE:
531 case T_FLOAT:
532 if (C->use_cmove()) {
533 continue; //TODO: maybe we want to add some cost
534 }
535 cost += Matcher::float_cmove_cost(); // Could be very expensive
536 break;
537 case T_LONG: {
538 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
539 }
540 case T_INT: // These all CMOV fine
541 case T_ADDRESS: { // (RawPtr)
542 cost++;
543 break;
544 }
545 case T_NARROWOOP: // Fall through
546 case T_OBJECT: { // Base oops are OK, but not derived oops
547 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
548 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
549 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
550 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
551 // have a Phi for the base here that we convert to a CMOVE all is well
552 // and good. But if the base is dead, we'll not make a CMOVE. Later
553 // the allocator will have to produce a base by creating a CMOVE of the
554 // relevant bases. This puts the allocator in the business of
555 // manufacturing expensive instructions, generally a bad plan.
556 // Just Say No to Conditionally-Moved Derived Pointers.
557 if (tp && tp->offset() != 0)
558 return NULL;
559 cost++;
560 break;
561 }
562 default:
563 return NULL; // In particular, can't do memory or I/O
564 }
565 // Add in cost any speculative ops
566 for (uint j = 1; j < region->req(); j++) {
567 Node *proj = region->in(j);
568 Node *inp = phi->in(j);
569 if (get_ctrl(inp) == proj) { // Found local op
570 cost++;
571 // Check for a chain of dependent ops; these will all become
572 // speculative in a CMOV.
573 for (uint k = 1; k < inp->req(); k++)
574 if (get_ctrl(inp->in(k)) == proj)
575 cost += ConditionalMoveLimit; // Too much speculative goo
576 }
577 }
578 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
579 // This will likely Split-If, a higher-payoff operation.
580 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
581 Node* use = phi->fast_out(k);
582 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
583 cost += ConditionalMoveLimit;
584 // Is there a use inside the loop?
585 // Note: check only basic types since CMoveP is pinned.
586 if (!used_inside_loop && is_java_primitive(bt)) {
587 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
588 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
589 used_inside_loop = true;
590 }
591 }
592 }
593 }//for
594 Node* bol = iff->in(1);
595 assert(bol->Opcode() == Op_Bool, "");
596 int cmp_op = bol->in(1)->Opcode();
597 // It is expensive to generate flags from a float compare.
598 // Avoid duplicated float compare.
599 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
600
601 float infrequent_prob = PROB_UNLIKELY_MAG(3);
602 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
603 if (used_inside_loop) {
604 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
605
606 // BlockLayoutByFrequency optimization moves infrequent branch
607 // from hot path. No point in CMOV'ing in such case (110 is used
608 // instead of 100 to take into account not exactness of float value).
609 if (BlockLayoutByFrequency) {
610 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
611 }
612 }
613 // Check for highly predictable branch. No point in CMOV'ing if
614 // we are going to predict accurately all the time.
615 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
616 //keep going
617 } else if (iff->_prob < infrequent_prob ||
618 iff->_prob > (1.0f - infrequent_prob))
619 return NULL;
620
621 // --------------
622 // Now replace all Phis with CMOV's
623 Node *cmov_ctrl = iff->in(0);
624 uint flip = (lp->Opcode() == Op_IfTrue);
625 Node_List wq;
626 while (1) {
627 PhiNode* phi = NULL;
628 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
629 Node *out = region->fast_out(i);
630 if (out->is_Phi()) {
631 phi = out->as_Phi();
632 break;
633 }
634 }
635 if (phi == NULL) break;
636 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
637 // Move speculative ops
638 wq.push(phi);
639 while (wq.size() > 0) {
640 Node *n = wq.pop();
641 for (uint j = 1; j < n->req(); j++) {
642 Node* m = n->in(j);
643 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) {
644 #ifndef PRODUCT
645 if (PrintOpto && VerifyLoopOptimizations) {
646 tty->print(" speculate: ");
647 m->dump();
648 }
649 #endif
650 set_ctrl(m, cmov_ctrl);
651 wq.push(m);
652 }
653 }
654 }
655 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
656 register_new_node( cmov, cmov_ctrl );
657 _igvn.replace_node( phi, cmov );
658 #ifndef PRODUCT
659 if (TraceLoopOpts) {
660 tty->print("CMOV ");
661 r_loop->dump_head();
662 if (Verbose) {
663 bol->in(1)->dump(1);
664 cmov->dump(1);
665 }
666 }
667 if (VerifyLoopOptimizations) verify();
668 #endif
669 }
670
671 // The useless CFG diamond will fold up later; see the optimization in
672 // RegionNode::Ideal.
673 _igvn._worklist.push(region);
674
675 return iff->in(1);
676 }
677
678 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
679 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
680 Node* u = m->fast_out(i);
681 if (u->is_CFG()) {
682 if (u->Opcode() == Op_NeverBranch) {
683 u = ((NeverBranchNode*)u)->proj_out(0);
684 enqueue_cfg_uses(u, wq);
685 } else {
686 wq.push(u);
687 }
688 }
689 }
690 }
691
692 // Try moving a store out of a loop, right before the loop
693 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
694 // Store has to be first in the loop body
695 IdealLoopTree *n_loop = get_loop(n_ctrl);
696 if (n->is_Store() && n_loop != _ltree_root &&
697 n_loop->is_loop() && n_loop->_head->is_Loop() &&
698 n->in(0) != NULL) {
699 Node* address = n->in(MemNode::Address);
700 Node* value = n->in(MemNode::ValueIn);
701 Node* mem = n->in(MemNode::Memory);
702 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
703 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
704
705 // - address and value must be loop invariant
706 // - memory must be a memory Phi for the loop
707 // - Store must be the only store on this memory slice in the
708 // loop: if there's another store following this one then value
709 // written at iteration i by the second store could be overwritten
710 // at iteration i+n by the first store: it's not safe to move the
711 // first store out of the loop
712 // - nothing must observe the memory Phi: it guarantees no read
713 // before the store, we are also guaranteed the store post
714 // dominates the loop head (ignoring a possible early
715 // exit). Otherwise there would be extra Phi involved between the
716 // loop's Phi and the store.
717 // - there must be no early exit from the loop before the Store
718 // (such an exit most of the time would be an extra use of the
719 // memory Phi but sometimes is a bottom memory Phi that takes the
720 // store as input).
721
722 if (!n_loop->is_member(address_loop) &&
723 !n_loop->is_member(value_loop) &&
724 mem->is_Phi() && mem->in(0) == n_loop->_head &&
725 mem->outcnt() == 1 &&
726 mem->in(LoopNode::LoopBackControl) == n) {
727
728 assert(n_loop->_tail != NULL, "need a tail");
729 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
730
731 // Verify that there's no early exit of the loop before the store.
732 bool ctrl_ok = false;
733 {
734 // Follow control from loop head until n, we exit the loop or
735 // we reach the tail
736 ResourceMark rm;
737 Unique_Node_List wq;
738 wq.push(n_loop->_head);
739
740 for (uint next = 0; next < wq.size(); ++next) {
741 Node *m = wq.at(next);
742 if (m == n->in(0)) {
743 ctrl_ok = true;
744 continue;
745 }
746 assert(!has_ctrl(m), "should be CFG");
747 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
748 ctrl_ok = false;
749 break;
750 }
751 enqueue_cfg_uses(m, wq);
752 if (wq.size() > 10) {
753 ctrl_ok = false;
754 break;
755 }
756 }
757 }
758 if (ctrl_ok) {
759 // move the Store
760 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
761 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
762 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
763 // Disconnect the phi now. An empty phi can confuse other
764 // optimizations in this pass of loop opts.
765 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
766 n_loop->_body.yank(mem);
767
768 set_ctrl_and_loop(n, n->in(0));
769
770 return n;
771 }
772 }
773 }
774 return NULL;
775 }
776
777 // Try moving a store out of a loop, right after the loop
778 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
779 if (n->is_Store() && n->in(0) != NULL) {
780 Node *n_ctrl = get_ctrl(n);
781 IdealLoopTree *n_loop = get_loop(n_ctrl);
782 // Store must be in a loop
783 if (n_loop != _ltree_root && !n_loop->_irreducible) {
784 Node* address = n->in(MemNode::Address);
785 Node* value = n->in(MemNode::ValueIn);
786 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
787 // address must be loop invariant
788 if (!n_loop->is_member(address_loop)) {
789 // Store must be last on this memory slice in the loop and
790 // nothing in the loop must observe it
791 Node* phi = NULL;
792 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
793 Node* u = n->fast_out(i);
794 if (has_ctrl(u)) { // control use?
795 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
796 if (!n_loop->is_member(u_loop)) {
797 continue;
798 }
799 if (u->is_Phi() && u->in(0) == n_loop->_head) {
800 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
801 // multiple phis on the same slice are possible
802 if (phi != NULL) {
803 return;
804 }
805 phi = u;
806 continue;
807 }
808 }
809 return;
810 }
811 if (phi != NULL) {
812 // Nothing in the loop before the store (next iteration)
813 // must observe the stored value
814 bool mem_ok = true;
815 {
816 ResourceMark rm;
817 Unique_Node_List wq;
818 wq.push(phi);
819 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
820 Node *m = wq.at(next);
821 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
822 Node* u = m->fast_out(i);
823 if (u->is_Store() || u->is_Phi()) {
824 if (u != n) {
825 wq.push(u);
826 mem_ok = (wq.size() <= 10);
827 }
828 } else {
829 mem_ok = false;
830 break;
831 }
832 }
833 }
834 }
835 if (mem_ok) {
836 // Move the store out of the loop if the LCA of all
837 // users (except for the phi) is outside the loop.
838 Node* hook = new Node(1);
839 _igvn.rehash_node_delayed(phi);
840 int count = phi->replace_edge(n, hook);
841 assert(count > 0, "inconsistent phi");
842
843 // Compute latest point this store can go
844 Node* lca = get_late_ctrl(n, get_ctrl(n));
845 if (n_loop->is_member(get_loop(lca))) {
846 // LCA is in the loop - bail out
847 _igvn.replace_node(hook, n);
848 return;
849 }
850 #ifdef ASSERT
851 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
852 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
853 n_loop->_head->as_Loop()->verify_strip_mined(1);
854 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
855 IdealLoopTree* outer_loop = get_loop(outer);
856 assert(n_loop->_parent == outer_loop, "broken loop tree");
857 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
858 }
859 #endif
860
861 // Move store out of the loop
862 _igvn.replace_node(hook, n->in(MemNode::Memory));
863 _igvn.replace_input_of(n, 0, lca);
864 set_ctrl_and_loop(n, lca);
865
866 // Disconnect the phi now. An empty phi can confuse other
867 // optimizations in this pass of loop opts..
868 if (phi->in(LoopNode::LoopBackControl) == phi) {
869 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
870 n_loop->_body.yank(phi);
871 }
872 }
873 }
874 }
875 }
876 }
877 }
878
879 //------------------------------split_if_with_blocks_pre-----------------------
880 // Do the real work in a non-recursive function. Data nodes want to be
881 // cloned in the pre-order so they can feed each other nicely.
882 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
883 // Cloning these guys is unlikely to win
884 int n_op = n->Opcode();
885 if( n_op == Op_MergeMem ) return n;
886 if( n->is_Proj() ) return n;
887 // Do not clone-up CmpFXXX variations, as these are always
888 // followed by a CmpI
889 if( n->is_Cmp() ) return n;
890 // Attempt to use a conditional move instead of a phi/branch
891 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
892 Node *cmov = conditional_move( n );
893 if( cmov ) return cmov;
894 }
895 if( n->is_CFG() || n->is_LoadStore() )
896 return n;
897 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
898 n_op == Op_Opaque2 ) {
899 if( !C->major_progress() ) // If chance of no more loop opts...
900 _igvn._worklist.push(n); // maybe we'll remove them
901 return n;
902 }
903
904 if( n->is_Con() ) return n; // No cloning for Con nodes
905
906 Node *n_ctrl = get_ctrl(n);
907 if( !n_ctrl ) return n; // Dead node
908
909 Node* res = try_move_store_before_loop(n, n_ctrl);
910 if (res != NULL) {
911 return n;
912 }
913
914 // Attempt to remix address expressions for loop invariants
915 Node *m = remix_address_expressions( n );
916 if( m ) return m;
917
918 if (n->is_ConstraintCast()) {
919 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
920 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
921 // Node control inputs don't necessarily agree with loop control info (due to
922 // transformations happened in between), thus additional dominance check is needed
923 // to keep loop info valid.
924 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
925 _igvn.replace_node(n, dom_cast);
926 return dom_cast;
927 }
928 }
929
930 // Determine if the Node has inputs from some local Phi.
931 // Returns the block to clone thru.
932 Node *n_blk = has_local_phi_input( n );
933 if( !n_blk ) return n;
934
935 // Do not clone the trip counter through on a CountedLoop
936 // (messes up the canonical shape).
937 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
938
939 // Check for having no control input; not pinned. Allow
940 // dominating control.
941 if (n->in(0)) {
942 Node *dom = idom(n_blk);
943 if (dom_lca(n->in(0), dom) != n->in(0)) {
944 return n;
945 }
946 }
947 // Policy: when is it profitable. You must get more wins than
948 // policy before it is considered profitable. Policy is usually 0,
949 // so 1 win is considered profitable. Big merges will require big
950 // cloning, so get a larger policy.
951 int policy = n_blk->req() >> 2;
952
953 // If the loop is a candidate for range check elimination,
954 // delay splitting through it's phi until a later loop optimization
955 if (n_blk->is_CountedLoop()) {
956 IdealLoopTree *lp = get_loop(n_blk);
957 if (lp && lp->_rce_candidate) {
958 return n;
959 }
960 }
961
962 // Use same limit as split_if_with_blocks_post
963 if( C->live_nodes() > 35000 ) return n; // Method too big
964
965 // Split 'n' through the merge point if it is profitable
966 Node *phi = split_thru_phi( n, n_blk, policy );
967 if (!phi) return n;
968
969 // Found a Phi to split thru!
970 // Replace 'n' with the new phi
971 _igvn.replace_node( n, phi );
972 // Moved a load around the loop, 'en-registering' something.
973 if (n_blk->is_Loop() && n->is_Load() &&
974 !phi->in(LoopNode::LoopBackControl)->is_Load())
975 C->set_major_progress();
976
977 return phi;
978 }
979
980 static bool merge_point_too_heavy(Compile* C, Node* region) {
981 // Bail out if the region and its phis have too many users.
982 int weight = 0;
983 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
984 weight += region->fast_out(i)->outcnt();
985 }
986 int nodes_left = C->max_node_limit() - C->live_nodes();
987 if (weight * 8 > nodes_left) {
988 if (PrintOpto) {
989 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
990 }
991 return true;
992 } else {
993 return false;
994 }
995 }
996
997 static bool merge_point_safe(Node* region) {
998 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
999 // having a PhiNode input. This sidesteps the dangerous case where the split
1000 // ConvI2LNode may become TOP if the input Value() does not
1001 // overlap the ConvI2L range, leaving a node which may not dominate its
1002 // uses.
1003 // A better fix for this problem can be found in the BugTraq entry, but
1004 // expediency for Mantis demands this hack.
1005 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop
1006 // split_if_with_blocks from splitting a block because we could not move around
1007 // the FastLockNode.
1008 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1009 Node* n = region->fast_out(i);
1010 if (n->is_Phi()) {
1011 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1012 Node* m = n->fast_out(j);
1013 if (m->is_FastLock())
1014 return false;
1015 #ifdef _LP64
1016 if (m->Opcode() == Op_ConvI2L)
1017 return false;
1018 if (m->is_CastII() && m->isa_CastII()->has_range_check()) {
1019 return false;
1020 }
1021 #endif
1022 }
1023 }
1024 }
1025 return true;
1026 }
1027
1028
1029 //------------------------------place_near_use---------------------------------
1030 // Place some computation next to use but not inside inner loops.
1031 // For inner loop uses move it to the preheader area.
1032 Node *PhaseIdealLoop::place_near_use(Node *useblock) const {
1033 IdealLoopTree *u_loop = get_loop( useblock );
1034 if (u_loop->_irreducible) {
1035 return useblock;
1036 }
1037 if (u_loop->_child) {
1038 if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) {
1039 return u_loop->_head->in(LoopNode::EntryControl);
1040 }
1041 return useblock;
1042 }
1043 return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
1044 }
1045
1046
1047 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1048 if (!n->is_If() || n->is_CountedLoopEnd()) {
1049 return false;
1050 }
1051 if (!n->in(0)->is_Region()) {
1052 return false;
1053 }
1054 Node* region = n->in(0);
1055 Node* dom = idom(region);
1056 if (!dom->is_If() || dom->in(1) != n->in(1)) {
1057 return false;
1058 }
1059 IfNode* dom_if = dom->as_If();
1060 Node* proj_true = dom_if->proj_out(1);
1061 Node* proj_false = dom_if->proj_out(0);
1062
1063 for (uint i = 1; i < region->req(); i++) {
1064 if (is_dominator(proj_true, region->in(i))) {
1065 continue;
1066 }
1067 if (is_dominator(proj_false, region->in(i))) {
1068 continue;
1069 }
1070 return false;
1071 }
1072
1073 return true;
1074 }
1075
1076 bool PhaseIdealLoop::can_split_if(Node *n_ctrl) {
1077 if (C->live_nodes() > 35000) {
1078 return false; // Method too big
1079 }
1080
1081 // Do not do 'split-if' if irreducible loops are present.
1082 if (_has_irreducible_loops) {
1083 return false;
1084 }
1085
1086 if (merge_point_too_heavy(C, n_ctrl)) {
1087 return false;
1088 }
1089
1090 // Do not do 'split-if' if some paths are dead. First do dead code
1091 // elimination and then see if its still profitable.
1092 for (uint i = 1; i < n_ctrl->req(); i++) {
1093 if (n_ctrl->in(i) == C->top()) {
1094 return false;
1095 }
1096 }
1097
1098 // If trying to do a 'Split-If' at the loop head, it is only
1099 // profitable if the cmp folds up on BOTH paths. Otherwise we
1100 // risk peeling a loop forever.
1101
1102 // CNC - Disabled for now. Requires careful handling of loop
1103 // body selection for the cloned code. Also, make sure we check
1104 // for any input path not being in the same loop as n_ctrl. For
1105 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1106 // because the alternative loop entry points won't be converted
1107 // into LoopNodes.
1108 IdealLoopTree *n_loop = get_loop(n_ctrl);
1109 for (uint j = 1; j < n_ctrl->req(); j++) {
1110 if (get_loop(n_ctrl->in(j)) != n_loop) {
1111 return false;
1112 }
1113 }
1114
1115 // Check for safety of the merge point.
1116 if (!merge_point_safe(n_ctrl)) {
1117 return false;
1118 }
1119
1120 return true;
1121 }
1122
1123 //------------------------------split_if_with_blocks_post----------------------
1124 // Do the real work in a non-recursive function. CFG hackery wants to be
1125 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1126 // info.
1127 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1128
1129 // Cloning Cmp through Phi's involves the split-if transform.
1130 // FastLock is not used by an If
1131 if (n->is_Cmp() && !n->is_FastLock()) {
1132 Node *n_ctrl = get_ctrl(n);
1133 // Determine if the Node has inputs from some local Phi.
1134 // Returns the block to clone thru.
1135 Node *n_blk = has_local_phi_input(n);
1136 if (n_blk != n_ctrl) {
1137 return;
1138 }
1139
1140 if (!can_split_if(n_ctrl)) {
1141 return;
1142 }
1143
1144 if (n->outcnt() != 1) {
1145 return; // Multiple bool's from 1 compare?
1146 }
1147 Node *bol = n->unique_out();
1148 assert(bol->is_Bool(), "expect a bool here");
1149 if (bol->outcnt() != 1) {
1150 return;// Multiple branches from 1 compare?
1151 }
1152 Node *iff = bol->unique_out();
1153
1154 // Check some safety conditions
1155 if (iff->is_If()) { // Classic split-if?
1156 if (iff->in(0) != n_ctrl) {
1157 return; // Compare must be in same blk as if
1158 }
1159 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1160 // Can't split CMove with different control edge.
1161 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) {
1162 return;
1163 }
1164 if (get_ctrl(iff->in(2)) == n_ctrl ||
1165 get_ctrl(iff->in(3)) == n_ctrl) {
1166 return; // Inputs not yet split-up
1167 }
1168 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1169 return; // Loop-invar test gates loop-varying CMOVE
1170 }
1171 } else {
1172 return; // some other kind of node, such as an Allocate
1173 }
1174
1175 // When is split-if profitable? Every 'win' on means some control flow
1176 // goes dead, so it's almost always a win.
1177 int policy = 0;
1178 // Split compare 'n' through the merge point if it is profitable
1179 Node *phi = split_thru_phi( n, n_ctrl, policy);
1180 if (!phi) {
1181 return;
1182 }
1183
1184 // Found a Phi to split thru!
1185 // Replace 'n' with the new phi
1186 _igvn.replace_node(n, phi);
1187
1188 // Now split the bool up thru the phi
1189 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1190 guarantee(bolphi != NULL, "null boolean phi node");
1191
1192 _igvn.replace_node(bol, bolphi);
1193 assert(iff->in(1) == bolphi, "");
1194
1195 if (bolphi->Value(&_igvn)->singleton()) {
1196 return;
1197 }
1198
1199 // Conditional-move? Must split up now
1200 if (!iff->is_If()) {
1201 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1202 _igvn.replace_node(iff, cmovphi);
1203 return;
1204 }
1205
1206 // Now split the IF
1207 do_split_if(iff);
1208 return;
1209 }
1210
1211 // Two identical ifs back to back can be merged
1212 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1213 Node *n_ctrl = n->in(0);
1214 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1215 IfNode* dom_if = idom(n_ctrl)->as_If();
1216 Node* proj_true = dom_if->proj_out(1);
1217 Node* proj_false = dom_if->proj_out(0);
1218 Node* con_true = _igvn.makecon(TypeInt::ONE);
1219 Node* con_false = _igvn.makecon(TypeInt::ZERO);
1220
1221 for (uint i = 1; i < n_ctrl->req(); i++) {
1222 if (is_dominator(proj_true, n_ctrl->in(i))) {
1223 bolphi->init_req(i, con_true);
1224 } else {
1225 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1226 bolphi->init_req(i, con_false);
1227 }
1228 }
1229 register_new_node(bolphi, n_ctrl);
1230 _igvn.replace_input_of(n, 1, bolphi);
1231
1232 // Now split the IF
1233 do_split_if(n);
1234 return;
1235 }
1236
1237 // Check for an IF ready to split; one that has its
1238 // condition codes input coming from a Phi at the block start.
1239 int n_op = n->Opcode();
1240
1241 // Check for an IF being dominated by another IF same test
1242 if (n_op == Op_If ||
1243 n_op == Op_RangeCheck) {
1244 Node *bol = n->in(1);
1245 uint max = bol->outcnt();
1246 // Check for same test used more than once?
1247 if (max > 1 && bol->is_Bool()) {
1248 // Search up IDOMs to see if this IF is dominated.
1249 Node *cutoff = get_ctrl(bol);
1250
1251 // Now search up IDOMs till cutoff, looking for a dominating test
1252 Node *prevdom = n;
1253 Node *dom = idom(prevdom);
1254 while (dom != cutoff) {
1255 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) {
1256 // Replace the dominated test with an obvious true or false.
1257 // Place it on the IGVN worklist for later cleanup.
1258 C->set_major_progress();
1259 dominated_by(prevdom, n, false, true);
1260 #ifndef PRODUCT
1261 if( VerifyLoopOptimizations ) verify();
1262 #endif
1263 return;
1264 }
1265 prevdom = dom;
1266 dom = idom(prevdom);
1267 }
1268 }
1269 }
1270
1271 // See if a shared loop-varying computation has no loop-varying uses.
1272 // Happens if something is only used for JVM state in uncommon trap exits,
1273 // like various versions of induction variable+offset. Clone the
1274 // computation per usage to allow it to sink out of the loop.
1275 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
1276 Node *n_ctrl = get_ctrl(n);
1277 IdealLoopTree *n_loop = get_loop(n_ctrl);
1278 if( n_loop != _ltree_root ) {
1279 DUIterator_Fast imax, i = n->fast_outs(imax);
1280 for (; i < imax; i++) {
1281 Node* u = n->fast_out(i);
1282 if( !has_ctrl(u) ) break; // Found control user
1283 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1284 if( u_loop == n_loop ) break; // Found loop-varying use
1285 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
1286 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
1287 }
1288 bool did_break = (i < imax); // Did we break out of the previous loop?
1289 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
1290 Node *late_load_ctrl = NULL;
1291 if (n->is_Load()) {
1292 // If n is a load, get and save the result from get_late_ctrl(),
1293 // to be later used in calculating the control for n's clones.
1294 clear_dom_lca_tags();
1295 late_load_ctrl = get_late_ctrl(n, n_ctrl);
1296 }
1297 // If n is a load, and the late control is the same as the current
1298 // control, then the cloning of n is a pointless exercise, because
1299 // GVN will ensure that we end up where we started.
1300 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
1301 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
1302 Node *u = n->last_out(j); // Clone private computation per use
1303 _igvn.rehash_node_delayed(u);
1304 Node *x = n->clone(); // Clone computation
1305 Node *x_ctrl = NULL;
1306 if( u->is_Phi() ) {
1307 // Replace all uses of normal nodes. Replace Phi uses
1308 // individually, so the separate Nodes can sink down
1309 // different paths.
1310 uint k = 1;
1311 while( u->in(k) != n ) k++;
1312 u->set_req( k, x );
1313 // x goes next to Phi input path
1314 x_ctrl = u->in(0)->in(k);
1315 --j;
1316 } else { // Normal use
1317 // Replace all uses
1318 for( uint k = 0; k < u->req(); k++ ) {
1319 if( u->in(k) == n ) {
1320 u->set_req( k, x );
1321 --j;
1322 }
1323 }
1324 x_ctrl = get_ctrl(u);
1325 }
1326
1327 // Find control for 'x' next to use but not inside inner loops.
1328 // For inner loop uses get the preheader area.
1329 x_ctrl = place_near_use(x_ctrl);
1330
1331 if (n->is_Load()) {
1332 // For loads, add a control edge to a CFG node outside of the loop
1333 // to force them to not combine and return back inside the loop
1334 // during GVN optimization (4641526).
1335 //
1336 // Because we are setting the actual control input, factor in
1337 // the result from get_late_ctrl() so we respect any
1338 // anti-dependences. (6233005).
1339 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
1340
1341 // Don't allow the control input to be a CFG splitting node.
1342 // Such nodes should only have ProjNodes as outs, e.g. IfNode
1343 // should only have IfTrueNode and IfFalseNode (4985384).
1344 x_ctrl = find_non_split_ctrl(x_ctrl);
1345 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1346
1347 x->set_req(0, x_ctrl);
1348 }
1349 register_new_node(x, x_ctrl);
1350
1351 // Some institutional knowledge is needed here: 'x' is
1352 // yanked because if the optimizer runs GVN on it all the
1353 // cloned x's will common up and undo this optimization and
1354 // be forced back in the loop. This is annoying because it
1355 // makes +VerifyOpto report false-positives on progress. I
1356 // tried setting control edges on the x's to force them to
1357 // not combine, but the matching gets worried when it tries
1358 // to fold a StoreP and an AddP together (as part of an
1359 // address expression) and the AddP and StoreP have
1360 // different controls.
1361 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x);
1362 }
1363 _igvn.remove_dead_node(n);
1364 }
1365 }
1366 }
1367 }
1368
1369 try_move_store_after_loop(n);
1370
1371 // Check for Opaque2's who's loop has disappeared - who's input is in the
1372 // same loop nest as their output. Remove 'em, they are no longer useful.
1373 if( n_op == Op_Opaque2 &&
1374 n->in(1) != NULL &&
1375 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1376 _igvn.replace_node( n, n->in(1) );
1377 }
1378 }
1379
1380 //------------------------------split_if_with_blocks---------------------------
1381 // Check for aggressive application of 'split-if' optimization,
1382 // using basic block level info.
1383 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) {
1384 Node *n = C->root();
1385 visited.set(n->_idx); // first, mark node as visited
1386 // Do pre-visit work for root
1387 n = split_if_with_blocks_pre( n );
1388 uint cnt = n->outcnt();
1389 uint i = 0;
1390 while (true) {
1391 // Visit all children
1392 if (i < cnt) {
1393 Node* use = n->raw_out(i);
1394 ++i;
1395 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1396 // Now do pre-visit work for this use
1397 use = split_if_with_blocks_pre( use );
1398 nstack.push(n, i); // Save parent and next use's index.
1399 n = use; // Process all children of current use.
1400 cnt = use->outcnt();
1401 i = 0;
1402 }
1403 }
1404 else {
1405 // All of n's children have been processed, complete post-processing.
1406 if (cnt != 0 && !n->is_Con()) {
1407 assert(has_node(n), "no dead nodes");
1408 split_if_with_blocks_post( n );
1409 }
1410 if (nstack.is_empty()) {
1411 // Finished all nodes on stack.
1412 break;
1413 }
1414 // Get saved parent node and next use's index. Visit the rest of uses.
1415 n = nstack.node();
1416 cnt = n->outcnt();
1417 i = nstack.index();
1418 nstack.pop();
1419 }
1420 }
1421 }
1422
1423
1424 //=============================================================================
1425 //
1426 // C L O N E A L O O P B O D Y
1427 //
1428
1429 //------------------------------clone_iff--------------------------------------
1430 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1431 // "Nearly" because all Nodes have been cloned from the original in the loop,
1432 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1433 // through the Phi recursively, and return a Bool.
1434 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) {
1435
1436 // Convert this Phi into a Phi merging Bools
1437 uint i;
1438 for (i = 1; i < phi->req(); i++) {
1439 Node *b = phi->in(i);
1440 if (b->is_Phi()) {
1441 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop));
1442 } else {
1443 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
1444 }
1445 }
1446
1447 Node* n = phi->in(1);
1448 Node* sample_opaque = NULL;
1449 Node *sample_bool = NULL;
1450 if (n->Opcode() == Op_Opaque4) {
1451 sample_opaque = n;
1452 sample_bool = n->in(1);
1453 assert(sample_bool->is_Bool(), "wrong type");
1454 } else {
1455 sample_bool = n;
1456 }
1457 Node *sample_cmp = sample_bool->in(1);
1458
1459 // Make Phis to merge the Cmp's inputs.
1460 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
1461 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
1462 for (i = 1; i < phi->req(); i++) {
1463 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
1464 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
1465 phi1->set_req(i, n1);
1466 phi2->set_req(i, n2);
1467 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1468 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1469 }
1470 // See if these Phis have been made before.
1471 // Register with optimizer
1472 Node *hit1 = _igvn.hash_find_insert(phi1);
1473 if (hit1) { // Hit, toss just made Phi
1474 _igvn.remove_dead_node(phi1); // Remove new phi
1475 assert(hit1->is_Phi(), "" );
1476 phi1 = (PhiNode*)hit1; // Use existing phi
1477 } else { // Miss
1478 _igvn.register_new_node_with_optimizer(phi1);
1479 }
1480 Node *hit2 = _igvn.hash_find_insert(phi2);
1481 if (hit2) { // Hit, toss just made Phi
1482 _igvn.remove_dead_node(phi2); // Remove new phi
1483 assert(hit2->is_Phi(), "" );
1484 phi2 = (PhiNode*)hit2; // Use existing phi
1485 } else { // Miss
1486 _igvn.register_new_node_with_optimizer(phi2);
1487 }
1488 // Register Phis with loop/block info
1489 set_ctrl(phi1, phi->in(0));
1490 set_ctrl(phi2, phi->in(0));
1491 // Make a new Cmp
1492 Node *cmp = sample_cmp->clone();
1493 cmp->set_req(1, phi1);
1494 cmp->set_req(2, phi2);
1495 _igvn.register_new_node_with_optimizer(cmp);
1496 set_ctrl(cmp, phi->in(0));
1497
1498 // Make a new Bool
1499 Node *b = sample_bool->clone();
1500 b->set_req(1,cmp);
1501 _igvn.register_new_node_with_optimizer(b);
1502 set_ctrl(b, phi->in(0));
1503
1504 if (sample_opaque != NULL) {
1505 Node* opaque = sample_opaque->clone();
1506 opaque->set_req(1, b);
1507 _igvn.register_new_node_with_optimizer(opaque);
1508 set_ctrl(opaque, phi->in(0));
1509 return opaque;
1510 }
1511
1512 assert(b->is_Bool(), "");
1513 return b;
1514 }
1515
1516 //------------------------------clone_bool-------------------------------------
1517 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1518 // "Nearly" because all Nodes have been cloned from the original in the loop,
1519 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1520 // through the Phi recursively, and return a Bool.
1521 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1522 uint i;
1523 // Convert this Phi into a Phi merging Bools
1524 for( i = 1; i < phi->req(); i++ ) {
1525 Node *b = phi->in(i);
1526 if( b->is_Phi() ) {
1527 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1528 } else {
1529 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1530 }
1531 }
1532
1533 Node *sample_cmp = phi->in(1);
1534
1535 // Make Phis to merge the Cmp's inputs.
1536 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1537 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1538 for( uint j = 1; j < phi->req(); j++ ) {
1539 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1540 Node *n1, *n2;
1541 if( cmp_top->is_Cmp() ) {
1542 n1 = cmp_top->in(1);
1543 n2 = cmp_top->in(2);
1544 } else {
1545 n1 = n2 = cmp_top;
1546 }
1547 phi1->set_req( j, n1 );
1548 phi2->set_req( j, n2 );
1549 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1550 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1551 }
1552
1553 // See if these Phis have been made before.
1554 // Register with optimizer
1555 Node *hit1 = _igvn.hash_find_insert(phi1);
1556 if( hit1 ) { // Hit, toss just made Phi
1557 _igvn.remove_dead_node(phi1); // Remove new phi
1558 assert( hit1->is_Phi(), "" );
1559 phi1 = (PhiNode*)hit1; // Use existing phi
1560 } else { // Miss
1561 _igvn.register_new_node_with_optimizer(phi1);
1562 }
1563 Node *hit2 = _igvn.hash_find_insert(phi2);
1564 if( hit2 ) { // Hit, toss just made Phi
1565 _igvn.remove_dead_node(phi2); // Remove new phi
1566 assert( hit2->is_Phi(), "" );
1567 phi2 = (PhiNode*)hit2; // Use existing phi
1568 } else { // Miss
1569 _igvn.register_new_node_with_optimizer(phi2);
1570 }
1571 // Register Phis with loop/block info
1572 set_ctrl(phi1, phi->in(0));
1573 set_ctrl(phi2, phi->in(0));
1574 // Make a new Cmp
1575 Node *cmp = sample_cmp->clone();
1576 cmp->set_req( 1, phi1 );
1577 cmp->set_req( 2, phi2 );
1578 _igvn.register_new_node_with_optimizer(cmp);
1579 set_ctrl(cmp, phi->in(0));
1580
1581 assert( cmp->is_Cmp(), "" );
1582 return (CmpNode*)cmp;
1583 }
1584
1585 //------------------------------sink_use---------------------------------------
1586 // If 'use' was in the loop-exit block, it now needs to be sunk
1587 // below the post-loop merge point.
1588 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1589 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1590 set_ctrl(use, post_loop);
1591 for (DUIterator j = use->outs(); use->has_out(j); j++)
1592 sink_use(use->out(j), post_loop);
1593 }
1594 }
1595
1596 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1597 IdealLoopTree* loop, IdealLoopTree* outer_loop,
1598 Node_List*& split_if_set, Node_List*& split_bool_set,
1599 Node_List*& split_cex_set, Node_List& worklist,
1600 uint new_counter, CloneLoopMode mode) {
1601 Node* nnn = old_new[old->_idx];
1602 // Copy uses to a worklist, so I can munge the def-use info
1603 // with impunity.
1604 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1605 worklist.push(old->fast_out(j));
1606
1607 while( worklist.size() ) {
1608 Node *use = worklist.pop();
1609 if (!has_node(use)) continue; // Ignore dead nodes
1610 if (use->in(0) == C->top()) continue;
1611 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1612 // Check for data-use outside of loop - at least one of OLD or USE
1613 // must not be a CFG node.
1614 #ifdef ASSERT
1615 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) {
1616 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
1617 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node");
1618 }
1619 #endif
1620 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
1621
1622 // If the Data use is an IF, that means we have an IF outside of the
1623 // loop that is switching on a condition that is set inside of the
1624 // loop. Happens if people set a loop-exit flag; then test the flag
1625 // in the loop to break the loop, then test is again outside of the
1626 // loop to determine which way the loop exited.
1627 // Loop predicate If node connects to Bool node through Opaque1 node.
1628 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) {
1629 // Since this code is highly unlikely, we lazily build the worklist
1630 // of such Nodes to go split.
1631 if (!split_if_set) {
1632 ResourceArea *area = Thread::current()->resource_area();
1633 split_if_set = new Node_List(area);
1634 }
1635 split_if_set->push(use);
1636 }
1637 if (use->is_Bool()) {
1638 if (!split_bool_set) {
1639 ResourceArea *area = Thread::current()->resource_area();
1640 split_bool_set = new Node_List(area);
1641 }
1642 split_bool_set->push(use);
1643 }
1644 if (use->Opcode() == Op_CreateEx) {
1645 if (!split_cex_set) {
1646 ResourceArea *area = Thread::current()->resource_area();
1647 split_cex_set = new Node_List(area);
1648 }
1649 split_cex_set->push(use);
1650 }
1651
1652
1653 // Get "block" use is in
1654 uint idx = 0;
1655 while( use->in(idx) != old ) idx++;
1656 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1657 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
1658 Node *cfg = prev->_idx >= new_counter
1659 ? prev->in(2)
1660 : idom(prev);
1661 if( use->is_Phi() ) // Phi use is in prior block
1662 cfg = prev->in(idx); // NOT in block of Phi itself
1663 if (cfg->is_top()) { // Use is dead?
1664 _igvn.replace_input_of(use, idx, C->top());
1665 continue;
1666 }
1667
1668 if (mode == IgnoreStripMined && idx == 0) {
1669 LoopNode *head = loop->_head->as_Loop();
1670 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
1671 // That node is outside the inner loop, leave it outside the
1672 // outer loop as well to not confuse verification code.
1673 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
1674 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
1675 continue;
1676 }
1677 }
1678
1679 while(!outer_loop->is_member(get_loop(cfg))) {
1680 prev = cfg;
1681 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1682 }
1683 // If the use occurs after merging several exits from the loop, then
1684 // old value must have dominated all those exits. Since the same old
1685 // value was used on all those exits we did not need a Phi at this
1686 // merge point. NOW we do need a Phi here. Each loop exit value
1687 // is now merged with the peeled body exit; each exit gets its own
1688 // private Phi and those Phis need to be merged here.
1689 Node *phi;
1690 if( prev->is_Region() ) {
1691 if( idx == 0 ) { // Updating control edge?
1692 phi = prev; // Just use existing control
1693 } else { // Else need a new Phi
1694 phi = PhiNode::make( prev, old );
1695 // Now recursively fix up the new uses of old!
1696 for( uint i = 1; i < prev->req(); i++ ) {
1697 worklist.push(phi); // Onto worklist once for each 'old' input
1698 }
1699 }
1700 } else {
1701 // Get new RegionNode merging old and new loop exits
1702 prev = old_new[prev->_idx];
1703 assert( prev, "just made this in step 7" );
1704 if( idx == 0) { // Updating control edge?
1705 phi = prev; // Just use existing control
1706 } else { // Else need a new Phi
1707 // Make a new Phi merging data values properly
1708 phi = PhiNode::make( prev, old );
1709 phi->set_req( 1, nnn );
1710 }
1711 }
1712 // If inserting a new Phi, check for prior hits
1713 if( idx != 0 ) {
1714 Node *hit = _igvn.hash_find_insert(phi);
1715 if( hit == NULL ) {
1716 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1717 } else { // or
1718 // Remove the new phi from the graph and use the hit
1719 _igvn.remove_dead_node(phi);
1720 phi = hit; // Use existing phi
1721 }
1722 set_ctrl(phi, prev);
1723 }
1724 // Make 'use' use the Phi instead of the old loop body exit value
1725 _igvn.replace_input_of(use, idx, phi);
1726 if( use->_idx >= new_counter ) { // If updating new phis
1727 // Not needed for correctness, but prevents a weak assert
1728 // in AddPNode from tripping (when we end up with different
1729 // base & derived Phis that will become the same after
1730 // IGVN does CSE).
1731 Node *hit = _igvn.hash_find_insert(use);
1732 if( hit ) // Go ahead and re-hash for hits.
1733 _igvn.replace_node( use, hit );
1734 }
1735
1736 // If 'use' was in the loop-exit block, it now needs to be sunk
1737 // below the post-loop merge point.
1738 sink_use( use, prev );
1739 }
1740 }
1741 }
1742
1743 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
1744 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
1745 bool check_old_new) {
1746 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1747 Node* u = n->fast_out(j);
1748 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned");
1749 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) {
1750 Node* c = phase->get_ctrl(u);
1751 IdealLoopTree* u_loop = phase->get_loop(c);
1752 assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only");
1753 if (outer_loop->is_member(u_loop)) {
1754 wq.push(u);
1755 }
1756 }
1757 }
1758 }
1759
1760 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1761 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1762 Node_List& extra_data_nodes) {
1763 if (head->is_strip_mined() && mode != IgnoreStripMined) {
1764 CountedLoopNode* cl = head->as_CountedLoop();
1765 Node* l = cl->outer_loop();
1766 Node* tail = cl->outer_loop_tail();
1767 IfNode* le = cl->outer_loop_end();
1768 Node* sfpt = cl->outer_safepoint();
1769 CountedLoopEndNode* cle = cl->loopexit();
1770 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
1771 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
1772 Node* cle_out = cle->proj_out(false);
1773
1774 Node* new_sfpt = NULL;
1775 Node* new_cle_out = cle_out->clone();
1776 old_new.map(cle_out->_idx, new_cle_out);
1777 if (mode == CloneIncludesStripMined) {
1778 // clone outer loop body
1779 Node* new_l = l->clone();
1780 Node* new_tail = tail->clone();
1781 IfNode* new_le = le->clone()->as_If();
1782 new_sfpt = sfpt->clone();
1783
1784 set_loop(new_l, outer_loop->_parent);
1785 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
1786 set_loop(new_cle_out, outer_loop->_parent);
1787 set_idom(new_cle_out, new_cle, dd);
1788 set_loop(new_sfpt, outer_loop->_parent);
1789 set_idom(new_sfpt, new_cle_out, dd);
1790 set_loop(new_le, outer_loop->_parent);
1791 set_idom(new_le, new_sfpt, dd);
1792 set_loop(new_tail, outer_loop->_parent);
1793 set_idom(new_tail, new_le, dd);
1794 set_idom(new_cl, new_l, dd);
1795
1796 old_new.map(l->_idx, new_l);
1797 old_new.map(tail->_idx, new_tail);
1798 old_new.map(le->_idx, new_le);
1799 old_new.map(sfpt->_idx, new_sfpt);
1800
1801 new_l->set_req(LoopNode::LoopBackControl, new_tail);
1802 new_l->set_req(0, new_l);
1803 new_tail->set_req(0, new_le);
1804 new_le->set_req(0, new_sfpt);
1805 new_sfpt->set_req(0, new_cle_out);
1806 new_cle_out->set_req(0, new_cle);
1807 new_cl->set_req(LoopNode::EntryControl, new_l);
1808
1809 _igvn.register_new_node_with_optimizer(new_l);
1810 _igvn.register_new_node_with_optimizer(new_tail);
1811 _igvn.register_new_node_with_optimizer(new_le);
1812 } else {
1813 Node *newhead = old_new[loop->_head->_idx];
1814 newhead->as_Loop()->clear_strip_mined();
1815 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
1816 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1817 }
1818 // Look at data node that were assigned a control in the outer
1819 // loop: they are kept in the outer loop by the safepoint so start
1820 // from the safepoint node's inputs.
1821 IdealLoopTree* outer_loop = get_loop(l);
1822 Node_Stack stack(2);
1823 stack.push(sfpt, 1);
1824 uint new_counter = C->unique();
1825 while (stack.size() > 0) {
1826 Node* n = stack.node();
1827 uint i = stack.index();
1828 while (i < n->req() &&
1829 (n->in(i) == NULL ||
1830 !has_ctrl(n->in(i)) ||
1831 get_loop(get_ctrl(n->in(i))) != outer_loop ||
1832 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
1833 i++;
1834 }
1835 if (i < n->req()) {
1836 stack.set_index(i+1);
1837 stack.push(n->in(i), 0);
1838 } else {
1839 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
1840 Node* m = n == sfpt ? new_sfpt : n->clone();
1841 if (m != NULL) {
1842 for (uint i = 0; i < n->req(); i++) {
1843 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) {
1844 m->set_req(i, old_new[m->in(i)->_idx]);
1845 }
1846 }
1847 } else {
1848 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
1849 }
1850 if (n != sfpt) {
1851 extra_data_nodes.push(n);
1852 _igvn.register_new_node_with_optimizer(m);
1853 assert(get_ctrl(n) == cle_out, "what other control?");
1854 set_ctrl(m, new_cle_out);
1855 old_new.map(n->_idx, m);
1856 }
1857 stack.pop();
1858 }
1859 }
1860 if (mode == CloneIncludesStripMined) {
1861 _igvn.register_new_node_with_optimizer(new_sfpt);
1862 _igvn.register_new_node_with_optimizer(new_cle_out);
1863 }
1864 // Some other transformation may have pessimistically assign some
1865 // data nodes to the outer loop. Set their control so they are out
1866 // of the outer loop.
1867 ResourceMark rm;
1868 Unique_Node_List wq;
1869 for (uint i = 0; i < extra_data_nodes.size(); i++) {
1870 Node* old = extra_data_nodes.at(i);
1871 clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true);
1872 }
1873 Node* new_ctrl = cl->outer_loop_exit();
1874 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
1875 for (uint i = 0; i < wq.size(); i++) {
1876 Node* n = wq.at(i);
1877 set_ctrl(n, new_ctrl);
1878 clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false);
1879 }
1880 } else {
1881 Node *newhead = old_new[loop->_head->_idx];
1882 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1883 }
1884 }
1885
1886 //------------------------------clone_loop-------------------------------------
1887 //
1888 // C L O N E A L O O P B O D Y
1889 //
1890 // This is the basic building block of the loop optimizations. It clones an
1891 // entire loop body. It makes an old_new loop body mapping; with this mapping
1892 // you can find the new-loop equivalent to an old-loop node. All new-loop
1893 // nodes are exactly equal to their old-loop counterparts, all edges are the
1894 // same. All exits from the old-loop now have a RegionNode that merges the
1895 // equivalent new-loop path. This is true even for the normal "loop-exit"
1896 // condition. All uses of loop-invariant old-loop values now come from (one
1897 // or more) Phis that merge their new-loop equivalents.
1898 //
1899 // This operation leaves the graph in an illegal state: there are two valid
1900 // control edges coming from the loop pre-header to both loop bodies. I'll
1901 // definitely have to hack the graph after running this transform.
1902 //
1903 // From this building block I will further edit edges to perform loop peeling
1904 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1905 //
1906 // Parameter side_by_size_idom:
1907 // When side_by_size_idom is NULL, the dominator tree is constructed for
1908 // the clone loop to dominate the original. Used in construction of
1909 // pre-main-post loop sequence.
1910 // When nonnull, the clone and original are side-by-side, both are
1911 // dominated by the side_by_side_idom node. Used in construction of
1912 // unswitched loops.
1913 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1914 CloneLoopMode mode, Node* side_by_side_idom) {
1915
1916 LoopNode* head = loop->_head->as_Loop();
1917 head->verify_strip_mined(1);
1918
1919 if (C->do_vector_loop() && PrintOpto) {
1920 const char* mname = C->method()->name()->as_quoted_ascii();
1921 if (mname != NULL) {
1922 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
1923 }
1924 }
1925
1926 CloneMap& cm = C->clone_map();
1927 Dict* dict = cm.dict();
1928 if (C->do_vector_loop()) {
1929 cm.set_clone_idx(cm.max_gen()+1);
1930 #ifndef PRODUCT
1931 if (PrintOpto) {
1932 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
1933 loop->dump_head();
1934 }
1935 #endif
1936 }
1937
1938 // Step 1: Clone the loop body. Make the old->new mapping.
1939 uint i;
1940 for( i = 0; i < loop->_body.size(); i++ ) {
1941 Node *old = loop->_body.at(i);
1942 Node *nnn = old->clone();
1943 old_new.map( old->_idx, nnn );
1944 if (C->do_vector_loop()) {
1945 cm.verify_insert_and_clone(old, nnn, cm.clone_idx());
1946 }
1947 _igvn.register_new_node_with_optimizer(nnn);
1948 }
1949
1950 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
1951
1952 // Step 2: Fix the edges in the new body. If the old input is outside the
1953 // loop use it. If the old input is INside the loop, use the corresponding
1954 // new node instead.
1955 for( i = 0; i < loop->_body.size(); i++ ) {
1956 Node *old = loop->_body.at(i);
1957 Node *nnn = old_new[old->_idx];
1958 // Fix CFG/Loop controlling the new node
1959 if (has_ctrl(old)) {
1960 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
1961 } else {
1962 set_loop(nnn, outer_loop->_parent);
1963 if (old->outcnt() > 0) {
1964 set_idom( nnn, old_new[idom(old)->_idx], dd );
1965 }
1966 }
1967 // Correct edges to the new node
1968 for( uint j = 0; j < nnn->req(); j++ ) {
1969 Node *n = nnn->in(j);
1970 if( n ) {
1971 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
1972 if( loop->is_member( old_in_loop ) )
1973 nnn->set_req(j, old_new[n->_idx]);
1974 }
1975 }
1976 _igvn.hash_find_insert(nnn);
1977 }
1978
1979 ResourceArea *area = Thread::current()->resource_area();
1980 Node_List extra_data_nodes(area);
1981 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
1982
1983 // Step 3: Now fix control uses. Loop varying control uses have already
1984 // been fixed up (as part of all input edges in Step 2). Loop invariant
1985 // control uses must be either an IfFalse or an IfTrue. Make a merge
1986 // point to merge the old and new IfFalse/IfTrue nodes; make the use
1987 // refer to this.
1988 Node_List worklist(area);
1989 uint new_counter = C->unique();
1990 for( i = 0; i < loop->_body.size(); i++ ) {
1991 Node* old = loop->_body.at(i);
1992 if( !old->is_CFG() ) continue;
1993
1994 // Copy uses to a worklist, so I can munge the def-use info
1995 // with impunity.
1996 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1997 worklist.push(old->fast_out(j));
1998
1999 while( worklist.size() ) { // Visit all uses
2000 Node *use = worklist.pop();
2001 if (!has_node(use)) continue; // Ignore dead nodes
2002 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2003 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
2004 // Both OLD and USE are CFG nodes here.
2005 assert( use->is_Proj(), "" );
2006 Node* nnn = old_new[old->_idx];
2007
2008 Node* newuse = NULL;
2009 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2010 CountedLoopNode* cl = head->as_CountedLoop();
2011 CountedLoopEndNode* cle = cl->loopexit();
2012 Node* cle_out = cle->proj_out_or_null(false);
2013 if (use == cle_out) {
2014 IfNode* le = cl->outer_loop_end();
2015 use = le->proj_out(false);
2016 use_loop = get_loop(use);
2017 if (mode == CloneIncludesStripMined) {
2018 nnn = old_new[le->_idx];
2019 } else {
2020 newuse = old_new[cle_out->_idx];
2021 }
2022 }
2023 }
2024 if (newuse == NULL) {
2025 newuse = use->clone();
2026 }
2027
2028 // Clone the loop exit control projection
2029 if (C->do_vector_loop()) {
2030 cm.verify_insert_and_clone(use, newuse, cm.clone_idx());
2031 }
2032 newuse->set_req(0,nnn);
2033 _igvn.register_new_node_with_optimizer(newuse);
2034 set_loop(newuse, use_loop);
2035 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2036
2037 // We need a Region to merge the exit from the peeled body and the
2038 // exit from the old loop body.
2039 RegionNode *r = new RegionNode(3);
2040 // Map the old use to the new merge point
2041 old_new.map( use->_idx, r );
2042 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
2043 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
2044
2045 // The original user of 'use' uses 'r' instead.
2046 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2047 Node* useuse = use->last_out(l);
2048 _igvn.rehash_node_delayed(useuse);
2049 uint uses_found = 0;
2050 if( useuse->in(0) == use ) {
2051 useuse->set_req(0, r);
2052 uses_found++;
2053 if( useuse->is_CFG() ) {
2054 assert( dom_depth(useuse) > dd_r, "" );
2055 set_idom(useuse, r, dom_depth(useuse));
2056 }
2057 }
2058 for( uint k = 1; k < useuse->req(); k++ ) {
2059 if( useuse->in(k) == use ) {
2060 useuse->set_req(k, r);
2061 uses_found++;
2062 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2063 assert(dom_depth(useuse) > dd_r , "");
2064 set_idom(useuse, r, dom_depth(useuse));
2065 }
2066 }
2067 }
2068 l -= uses_found; // we deleted 1 or more copies of this edge
2069 }
2070
2071 // Now finish up 'r'
2072 r->set_req( 1, newuse );
2073 r->set_req( 2, use );
2074 _igvn.register_new_node_with_optimizer(r);
2075 set_loop(r, use_loop);
2076 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
2077 } // End of if a loop-exit test
2078 }
2079 }
2080
2081 // Step 4: If loop-invariant use is not control, it must be dominated by a
2082 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2083 // there if needed. Make a Phi there merging old and new used values.
2084 Node_List *split_if_set = NULL;
2085 Node_List *split_bool_set = NULL;
2086 Node_List *split_cex_set = NULL;
2087 for( i = 0; i < loop->_body.size(); i++ ) {
2088 Node* old = loop->_body.at(i);
2089 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2090 split_bool_set, split_cex_set, worklist, new_counter,
2091 mode);
2092 }
2093
2094 for (i = 0; i < extra_data_nodes.size(); i++) {
2095 Node* old = extra_data_nodes.at(i);
2096 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2097 split_bool_set, split_cex_set, worklist, new_counter,
2098 mode);
2099 }
2100
2101 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2102 // the loop uses a condition set in the loop. The original IF probably
2103 // takes control from one or more OLD Regions (which in turn get from NEW
2104 // Regions). In any case, there will be a set of Phis for each merge point
2105 // from the IF up to where the original BOOL def exists the loop.
2106 if (split_if_set) {
2107 while (split_if_set->size()) {
2108 Node *iff = split_if_set->pop();
2109 if (iff->in(1)->is_Phi()) {
2110 Node *b = clone_iff(iff->in(1)->as_Phi(), loop);
2111 _igvn.replace_input_of(iff, 1, b);
2112 }
2113 }
2114 }
2115 if (split_bool_set) {
2116 while (split_bool_set->size()) {
2117 Node *b = split_bool_set->pop();
2118 Node *phi = b->in(1);
2119 assert(phi->is_Phi(), "");
2120 CmpNode *cmp = clone_bool((PhiNode*)phi, loop);
2121 _igvn.replace_input_of(b, 1, cmp);
2122 }
2123 }
2124 if (split_cex_set) {
2125 while (split_cex_set->size()) {
2126 Node *b = split_cex_set->pop();
2127 assert(b->in(0)->is_Region(), "");
2128 assert(b->in(1)->is_Phi(), "");
2129 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2130 split_up(b, b->in(0), NULL);
2131 }
2132 }
2133
2134 }
2135
2136
2137 //---------------------- stride_of_possible_iv -------------------------------------
2138 // Looks for an iff/bool/comp with one operand of the compare
2139 // being a cycle involving an add and a phi,
2140 // with an optional truncation (left-shift followed by a right-shift)
2141 // of the add. Returns zero if not an iv.
2142 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2143 Node* trunc1 = NULL;
2144 Node* trunc2 = NULL;
2145 const TypeInt* ttype = NULL;
2146 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
2147 return 0;
2148 }
2149 BoolNode* bl = iff->in(1)->as_Bool();
2150 Node* cmp = bl->in(1);
2151 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2152 return 0;
2153 }
2154 // Must have an invariant operand
2155 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2156 return 0;
2157 }
2158 Node* add2 = NULL;
2159 Node* cmp1 = cmp->in(1);
2160 if (cmp1->is_Phi()) {
2161 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2162 Node* phi = cmp1;
2163 for (uint i = 1; i < phi->req(); i++) {
2164 Node* in = phi->in(i);
2165 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2166 &trunc1, &trunc2, &ttype);
2167 if (add && add->in(1) == phi) {
2168 add2 = add->in(2);
2169 break;
2170 }
2171 }
2172 } else {
2173 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2174 Node* addtrunc = cmp1;
2175 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2176 &trunc1, &trunc2, &ttype);
2177 if (add && add->in(1)->is_Phi()) {
2178 Node* phi = add->in(1);
2179 for (uint i = 1; i < phi->req(); i++) {
2180 if (phi->in(i) == addtrunc) {
2181 add2 = add->in(2);
2182 break;
2183 }
2184 }
2185 }
2186 }
2187 if (add2 != NULL) {
2188 const TypeInt* add2t = _igvn.type(add2)->is_int();
2189 if (add2t->is_con()) {
2190 return add2t->get_con();
2191 }
2192 }
2193 return 0;
2194 }
2195
2196
2197 //---------------------- stay_in_loop -------------------------------------
2198 // Return the (unique) control output node that's in the loop (if it exists.)
2199 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2200 Node* unique = NULL;
2201 if (!n) return NULL;
2202 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2203 Node* use = n->fast_out(i);
2204 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2205 if (unique != NULL) {
2206 return NULL;
2207 }
2208 unique = use;
2209 }
2210 }
2211 return unique;
2212 }
2213
2214 //------------------------------ register_node -------------------------------------
2215 // Utility to register node "n" with PhaseIdealLoop
2216 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
2217 _igvn.register_new_node_with_optimizer(n);
2218 loop->_body.push(n);
2219 if (n->is_CFG()) {
2220 set_loop(n, loop);
2221 set_idom(n, pred, ddepth);
2222 } else {
2223 set_ctrl(n, pred);
2224 }
2225 }
2226
2227 //------------------------------ proj_clone -------------------------------------
2228 // Utility to create an if-projection
2229 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2230 ProjNode* c = p->clone()->as_Proj();
2231 c->set_req(0, iff);
2232 return c;
2233 }
2234
2235 //------------------------------ short_circuit_if -------------------------------------
2236 // Force the iff control output to be the live_proj
2237 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2238 guarantee(live_proj != NULL, "null projection");
2239 int proj_con = live_proj->_con;
2240 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2241 Node *con = _igvn.intcon(proj_con);
2242 set_ctrl(con, C->root());
2243 if (iff) {
2244 iff->set_req(1, con);
2245 }
2246 return con;
2247 }
2248
2249 //------------------------------ insert_if_before_proj -------------------------------------
2250 // Insert a new if before an if projection (* - new node)
2251 //
2252 // before
2253 // if(test)
2254 // / \
2255 // v v
2256 // other-proj proj (arg)
2257 //
2258 // after
2259 // if(test)
2260 // / \
2261 // / v
2262 // | * proj-clone
2263 // v |
2264 // other-proj v
2265 // * new_if(relop(cmp[IU](left,right)))
2266 // / \
2267 // v v
2268 // * new-proj proj
2269 // (returned)
2270 //
2271 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2272 IfNode* iff = proj->in(0)->as_If();
2273 IdealLoopTree *loop = get_loop(proj);
2274 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2275 int ddepth = dom_depth(proj);
2276
2277 _igvn.rehash_node_delayed(iff);
2278 _igvn.rehash_node_delayed(proj);
2279
2280 proj->set_req(0, NULL); // temporary disconnect
2281 ProjNode* proj2 = proj_clone(proj, iff);
2282 register_node(proj2, loop, iff, ddepth);
2283
2284 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2285 register_node(cmp, loop, proj2, ddepth);
2286
2287 BoolNode* bol = new BoolNode(cmp, relop);
2288 register_node(bol, loop, proj2, ddepth);
2289
2290 int opcode = iff->Opcode();
2291 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2292 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
2293 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
2294 register_node(new_if, loop, proj2, ddepth);
2295
2296 proj->set_req(0, new_if); // reattach
2297 set_idom(proj, new_if, ddepth);
2298
2299 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2300 guarantee(new_exit != NULL, "null exit node");
2301 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2302
2303 return new_exit;
2304 }
2305
2306 //------------------------------ insert_region_before_proj -------------------------------------
2307 // Insert a region before an if projection (* - new node)
2308 //
2309 // before
2310 // if(test)
2311 // / |
2312 // v |
2313 // proj v
2314 // other-proj
2315 //
2316 // after
2317 // if(test)
2318 // / |
2319 // v |
2320 // * proj-clone v
2321 // | other-proj
2322 // v
2323 // * new-region
2324 // |
2325 // v
2326 // * dum_if
2327 // / \
2328 // v \
2329 // * dum-proj v
2330 // proj
2331 //
2332 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2333 IfNode* iff = proj->in(0)->as_If();
2334 IdealLoopTree *loop = get_loop(proj);
2335 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2336 int ddepth = dom_depth(proj);
2337
2338 _igvn.rehash_node_delayed(iff);
2339 _igvn.rehash_node_delayed(proj);
2340
2341 proj->set_req(0, NULL); // temporary disconnect
2342 ProjNode* proj2 = proj_clone(proj, iff);
2343 register_node(proj2, loop, iff, ddepth);
2344
2345 RegionNode* reg = new RegionNode(2);
2346 reg->set_req(1, proj2);
2347 register_node(reg, loop, iff, ddepth);
2348
2349 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
2350 register_node(dum_if, loop, reg, ddepth);
2351
2352 proj->set_req(0, dum_if); // reattach
2353 set_idom(proj, dum_if, ddepth);
2354
2355 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
2356 register_node(dum_proj, loop, dum_if, ddepth);
2357
2358 return reg;
2359 }
2360
2361 //------------------------------ insert_cmpi_loop_exit -------------------------------------
2362 // Clone a signed compare loop exit from an unsigned compare and
2363 // insert it before the unsigned cmp on the stay-in-loop path.
2364 // All new nodes inserted in the dominator tree between the original
2365 // if and it's projections. The original if test is replaced with
2366 // a constant to force the stay-in-loop path.
2367 //
2368 // This is done to make sure that the original if and it's projections
2369 // still dominate the same set of control nodes, that the ctrl() relation
2370 // from data nodes to them is preserved, and that their loop nesting is
2371 // preserved.
2372 //
2373 // before
2374 // if(i <u limit) unsigned compare loop exit
2375 // / |
2376 // v v
2377 // exit-proj stay-in-loop-proj
2378 //
2379 // after
2380 // if(stay-in-loop-const) original if
2381 // / |
2382 // / v
2383 // / if(i < limit) new signed test
2384 // / / |
2385 // / / v
2386 // / / if(i <u limit) new cloned unsigned test
2387 // / / / |
2388 // v v v |
2389 // region |
2390 // | |
2391 // dum-if |
2392 // / | |
2393 // ether | |
2394 // v v
2395 // exit-proj stay-in-loop-proj
2396 //
2397 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
2398 const bool Signed = true;
2399 const bool Unsigned = false;
2400
2401 BoolNode* bol = if_cmpu->in(1)->as_Bool();
2402 if (bol->_test._test != BoolTest::lt) return NULL;
2403 CmpNode* cmpu = bol->in(1)->as_Cmp();
2404 if (cmpu->Opcode() != Op_CmpU) return NULL;
2405 int stride = stride_of_possible_iv(if_cmpu);
2406 if (stride == 0) return NULL;
2407
2408 Node* lp_proj = stay_in_loop(if_cmpu, loop);
2409 guarantee(lp_proj != NULL, "null loop node");
2410
2411 ProjNode* lp_continue = lp_proj->as_Proj();
2412 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
2413
2414 Node* limit = NULL;
2415 if (stride > 0) {
2416 limit = cmpu->in(2);
2417 } else {
2418 limit = _igvn.makecon(TypeInt::ZERO);
2419 set_ctrl(limit, C->root());
2420 }
2421 // Create a new region on the exit path
2422 RegionNode* reg = insert_region_before_proj(lp_exit);
2423 guarantee(reg != NULL, "null region node");
2424
2425 // Clone the if-cmpu-true-false using a signed compare
2426 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
2427 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
2428 reg->add_req(cmpi_exit);
2429
2430 // Clone the if-cmpu-true-false
2431 BoolTest::mask rel_u = bol->_test._test;
2432 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
2433 reg->add_req(cmpu_exit);
2434
2435 // Force original if to stay in loop.
2436 short_circuit_if(if_cmpu, lp_continue);
2437
2438 return cmpi_exit->in(0)->as_If();
2439 }
2440
2441 //------------------------------ remove_cmpi_loop_exit -------------------------------------
2442 // Remove a previously inserted signed compare loop exit.
2443 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
2444 Node* lp_proj = stay_in_loop(if_cmp, loop);
2445 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
2446 stay_in_loop(lp_proj, loop)->is_If() &&
2447 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
2448 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
2449 set_ctrl(con, C->root());
2450 if_cmp->set_req(1, con);
2451 }
2452
2453 //------------------------------ scheduled_nodelist -------------------------------------
2454 // Create a post order schedule of nodes that are in the
2455 // "member" set. The list is returned in "sched".
2456 // The first node in "sched" is the loop head, followed by
2457 // nodes which have no inputs in the "member" set, and then
2458 // followed by the nodes that have an immediate input dependence
2459 // on a node in "sched".
2460 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
2461
2462 assert(member.test(loop->_head->_idx), "loop head must be in member set");
2463 Arena *a = Thread::current()->resource_area();
2464 VectorSet visited(a);
2465 Node_Stack nstack(a, loop->_body.size());
2466
2467 Node* n = loop->_head; // top of stack is cached in "n"
2468 uint idx = 0;
2469 visited.set(n->_idx);
2470
2471 // Initially push all with no inputs from within member set
2472 for(uint i = 0; i < loop->_body.size(); i++ ) {
2473 Node *elt = loop->_body.at(i);
2474 if (member.test(elt->_idx)) {
2475 bool found = false;
2476 for (uint j = 0; j < elt->req(); j++) {
2477 Node* def = elt->in(j);
2478 if (def && member.test(def->_idx) && def != elt) {
2479 found = true;
2480 break;
2481 }
2482 }
2483 if (!found && elt != loop->_head) {
2484 nstack.push(n, idx);
2485 n = elt;
2486 assert(!visited.test(n->_idx), "not seen yet");
2487 visited.set(n->_idx);
2488 }
2489 }
2490 }
2491
2492 // traverse out's that are in the member set
2493 while (true) {
2494 if (idx < n->outcnt()) {
2495 Node* use = n->raw_out(idx);
2496 idx++;
2497 if (!visited.test_set(use->_idx)) {
2498 if (member.test(use->_idx)) {
2499 nstack.push(n, idx);
2500 n = use;
2501 idx = 0;
2502 }
2503 }
2504 } else {
2505 // All outputs processed
2506 sched.push(n);
2507 if (nstack.is_empty()) break;
2508 n = nstack.node();
2509 idx = nstack.index();
2510 nstack.pop();
2511 }
2512 }
2513 }
2514
2515
2516 //------------------------------ has_use_in_set -------------------------------------
2517 // Has a use in the vector set
2518 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
2519 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2520 Node* use = n->fast_out(j);
2521 if (vset.test(use->_idx)) {
2522 return true;
2523 }
2524 }
2525 return false;
2526 }
2527
2528
2529 //------------------------------ has_use_internal_to_set -------------------------------------
2530 // Has use internal to the vector set (ie. not in a phi at the loop head)
2531 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
2532 Node* head = loop->_head;
2533 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2534 Node* use = n->fast_out(j);
2535 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
2536 return true;
2537 }
2538 }
2539 return false;
2540 }
2541
2542
2543 //------------------------------ clone_for_use_outside_loop -------------------------------------
2544 // clone "n" for uses that are outside of loop
2545 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
2546 int cloned = 0;
2547 assert(worklist.size() == 0, "should be empty");
2548 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2549 Node* use = n->fast_out(j);
2550 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
2551 worklist.push(use);
2552 }
2553 }
2554 while( worklist.size() ) {
2555 Node *use = worklist.pop();
2556 if (!has_node(use) || use->in(0) == C->top()) continue;
2557 uint j;
2558 for (j = 0; j < use->req(); j++) {
2559 if (use->in(j) == n) break;
2560 }
2561 assert(j < use->req(), "must be there");
2562
2563 // clone "n" and insert it between the inputs of "n" and the use outside the loop
2564 Node* n_clone = n->clone();
2565 _igvn.replace_input_of(use, j, n_clone);
2566 cloned++;
2567 Node* use_c;
2568 if (!use->is_Phi()) {
2569 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
2570 } else {
2571 // Use in a phi is considered a use in the associated predecessor block
2572 use_c = use->in(0)->in(j);
2573 }
2574 set_ctrl(n_clone, use_c);
2575 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
2576 get_loop(use_c)->_body.push(n_clone);
2577 _igvn.register_new_node_with_optimizer(n_clone);
2578 #if !defined(PRODUCT)
2579 if (TracePartialPeeling) {
2580 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
2581 }
2582 #endif
2583 }
2584 return cloned;
2585 }
2586
2587
2588 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
2589 // clone "n" for special uses that are in the not_peeled region.
2590 // If these def-uses occur in separate blocks, the code generator
2591 // marks the method as not compilable. For example, if a "BoolNode"
2592 // is in a different basic block than the "IfNode" that uses it, then
2593 // the compilation is aborted in the code generator.
2594 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
2595 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
2596 if (n->is_Phi() || n->is_Load()) {
2597 return;
2598 }
2599 assert(worklist.size() == 0, "should be empty");
2600 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2601 Node* use = n->fast_out(j);
2602 if ( not_peel.test(use->_idx) &&
2603 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
2604 use->in(1) == n) {
2605 worklist.push(use);
2606 }
2607 }
2608 if (worklist.size() > 0) {
2609 // clone "n" and insert it between inputs of "n" and the use
2610 Node* n_clone = n->clone();
2611 loop->_body.push(n_clone);
2612 _igvn.register_new_node_with_optimizer(n_clone);
2613 set_ctrl(n_clone, get_ctrl(n));
2614 sink_list.push(n_clone);
2615 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
2616 #if !defined(PRODUCT)
2617 if (TracePartialPeeling) {
2618 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2619 }
2620 #endif
2621 while( worklist.size() ) {
2622 Node *use = worklist.pop();
2623 _igvn.rehash_node_delayed(use);
2624 for (uint j = 1; j < use->req(); j++) {
2625 if (use->in(j) == n) {
2626 use->set_req(j, n_clone);
2627 }
2628 }
2629 }
2630 }
2631 }
2632
2633
2634 //------------------------------ insert_phi_for_loop -------------------------------------
2635 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2636 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2637 Node *phi = PhiNode::make(lp, back_edge_val);
2638 phi->set_req(LoopNode::EntryControl, lp_entry_val);
2639 // Use existing phi if it already exists
2640 Node *hit = _igvn.hash_find_insert(phi);
2641 if( hit == NULL ) {
2642 _igvn.register_new_node_with_optimizer(phi);
2643 set_ctrl(phi, lp);
2644 } else {
2645 // Remove the new phi from the graph and use the hit
2646 _igvn.remove_dead_node(phi);
2647 phi = hit;
2648 }
2649 _igvn.replace_input_of(use, idx, phi);
2650 }
2651
2652 #ifdef ASSERT
2653 //------------------------------ is_valid_loop_partition -------------------------------------
2654 // Validate the loop partition sets: peel and not_peel
2655 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2656 VectorSet& not_peel ) {
2657 uint i;
2658 // Check that peel_list entries are in the peel set
2659 for (i = 0; i < peel_list.size(); i++) {
2660 if (!peel.test(peel_list.at(i)->_idx)) {
2661 return false;
2662 }
2663 }
2664 // Check at loop members are in one of peel set or not_peel set
2665 for (i = 0; i < loop->_body.size(); i++ ) {
2666 Node *def = loop->_body.at(i);
2667 uint di = def->_idx;
2668 // Check that peel set elements are in peel_list
2669 if (peel.test(di)) {
2670 if (not_peel.test(di)) {
2671 return false;
2672 }
2673 // Must be in peel_list also
2674 bool found = false;
2675 for (uint j = 0; j < peel_list.size(); j++) {
2676 if (peel_list.at(j)->_idx == di) {
2677 found = true;
2678 break;
2679 }
2680 }
2681 if (!found) {
2682 return false;
2683 }
2684 } else if (not_peel.test(di)) {
2685 if (peel.test(di)) {
2686 return false;
2687 }
2688 } else {
2689 return false;
2690 }
2691 }
2692 return true;
2693 }
2694
2695 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2696 // Ensure a use outside of loop is of the right form
2697 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2698 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2699 return (use->is_Phi() &&
2700 use_c->is_Region() && use_c->req() == 3 &&
2701 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2702 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2703 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2704 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2705 }
2706
2707 //------------------------------ is_valid_clone_loop_form -------------------------------------
2708 // Ensure that all uses outside of loop are of the right form
2709 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2710 uint orig_exit_idx, uint clone_exit_idx) {
2711 uint len = peel_list.size();
2712 for (uint i = 0; i < len; i++) {
2713 Node *def = peel_list.at(i);
2714
2715 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2716 Node *use = def->fast_out(j);
2717 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2718 if (!loop->is_member(get_loop(use_c))) {
2719 // use is not in the loop, check for correct structure
2720 if (use->in(0) == def) {
2721 // Okay
2722 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2723 return false;
2724 }
2725 }
2726 }
2727 }
2728 return true;
2729 }
2730 #endif
2731
2732 //------------------------------ partial_peel -------------------------------------
2733 // Partially peel (aka loop rotation) the top portion of a loop (called
2734 // the peel section below) by cloning it and placing one copy just before
2735 // the new loop head and the other copy at the bottom of the new loop.
2736 //
2737 // before after where it came from
2738 //
2739 // stmt1 stmt1
2740 // loop: stmt2 clone
2741 // stmt2 if condA goto exitA clone
2742 // if condA goto exitA new_loop: new
2743 // stmt3 stmt3 clone
2744 // if !condB goto loop if condB goto exitB clone
2745 // exitB: stmt2 orig
2746 // stmt4 if !condA goto new_loop orig
2747 // exitA: goto exitA
2748 // exitB:
2749 // stmt4
2750 // exitA:
2751 //
2752 // Step 1: find the cut point: an exit test on probable
2753 // induction variable.
2754 // Step 2: schedule (with cloning) operations in the peel
2755 // section that can be executed after the cut into
2756 // the section that is not peeled. This may need
2757 // to clone operations into exit blocks. For
2758 // instance, a reference to A[i] in the not-peel
2759 // section and a reference to B[i] in an exit block
2760 // may cause a left-shift of i by 2 to be placed
2761 // in the peel block. This step will clone the left
2762 // shift into the exit block and sink the left shift
2763 // from the peel to the not-peel section.
2764 // Step 3: clone the loop, retarget the control, and insert
2765 // phis for values that are live across the new loop
2766 // head. This is very dependent on the graph structure
2767 // from clone_loop. It creates region nodes for
2768 // exit control and associated phi nodes for values
2769 // flow out of the loop through that exit. The region
2770 // node is dominated by the clone's control projection.
2771 // So the clone's peel section is placed before the
2772 // new loop head, and the clone's not-peel section is
2773 // forms the top part of the new loop. The original
2774 // peel section forms the tail of the new loop.
2775 // Step 4: update the dominator tree and recompute the
2776 // dominator depth.
2777 //
2778 // orig
2779 //
2780 // stmt1
2781 // |
2782 // v
2783 // loop predicate
2784 // |
2785 // v
2786 // loop<----+
2787 // | |
2788 // stmt2 |
2789 // | |
2790 // v |
2791 // ifA |
2792 // / | |
2793 // v v |
2794 // false true ^ <-- last_peel
2795 // / | |
2796 // / ===|==cut |
2797 // / stmt3 | <-- first_not_peel
2798 // / | |
2799 // | v |
2800 // v ifB |
2801 // exitA: / \ |
2802 // / \ |
2803 // v v |
2804 // false true |
2805 // / \ |
2806 // / ----+
2807 // |
2808 // v
2809 // exitB:
2810 // stmt4
2811 //
2812 //
2813 // after clone loop
2814 //
2815 // stmt1
2816 // |
2817 // v
2818 // loop predicate
2819 // / \
2820 // clone / \ orig
2821 // / \
2822 // / \
2823 // v v
2824 // +---->loop loop<----+
2825 // | | | |
2826 // | stmt2 stmt2 |
2827 // | | | |
2828 // | v v |
2829 // | ifA ifA |
2830 // | | \ / | |
2831 // | v v v v |
2832 // ^ true false false true ^ <-- last_peel
2833 // | | ^ \ / | |
2834 // | cut==|== \ \ / ===|==cut |
2835 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2836 // | | dom | | | |
2837 // | v \ 1v v2 v |
2838 // | ifB regionA ifB |
2839 // | / \ | / \ |
2840 // | / \ v / \ |
2841 // | v v exitA: v v |
2842 // | true false false true |
2843 // | / ^ \ / \ |
2844 // +---- \ \ / ----+
2845 // dom \ /
2846 // \ 1v v2
2847 // regionB
2848 // |
2849 // v
2850 // exitB:
2851 // stmt4
2852 //
2853 //
2854 // after partial peel
2855 //
2856 // stmt1
2857 // |
2858 // v
2859 // loop predicate
2860 // /
2861 // clone / orig
2862 // / TOP
2863 // / \
2864 // v v
2865 // TOP->loop loop----+
2866 // | | |
2867 // stmt2 stmt2 |
2868 // | | |
2869 // v v |
2870 // ifA ifA |
2871 // | \ / | |
2872 // v v v v |
2873 // true false false true | <-- last_peel
2874 // | ^ \ / +------|---+
2875 // +->newloop \ \ / === ==cut | |
2876 // | stmt3 \ \ / TOP | |
2877 // | | dom | | stmt3 | | <-- first_not_peel
2878 // | v \ 1v v2 v | |
2879 // | ifB regionA ifB ^ v
2880 // | / \ | / \ | |
2881 // | / \ v / \ | |
2882 // | v v exitA: v v | |
2883 // | true false false true | |
2884 // | / ^ \ / \ | |
2885 // | | \ \ / v | |
2886 // | | dom \ / TOP | |
2887 // | | \ 1v v2 | |
2888 // ^ v regionB | |
2889 // | | | | |
2890 // | | v ^ v
2891 // | | exitB: | |
2892 // | | stmt4 | |
2893 // | +------------>-----------------+ |
2894 // | |
2895 // +-----------------<---------------------+
2896 //
2897 //
2898 // final graph
2899 //
2900 // stmt1
2901 // |
2902 // v
2903 // loop predicate
2904 // |
2905 // v
2906 // stmt2 clone
2907 // |
2908 // v
2909 // ........> ifA clone
2910 // : / |
2911 // dom / |
2912 // : v v
2913 // : false true
2914 // : | |
2915 // : | v
2916 // : | newloop<-----+
2917 // : | | |
2918 // : | stmt3 clone |
2919 // : | | |
2920 // : | v |
2921 // : | ifB |
2922 // : | / \ |
2923 // : | v v |
2924 // : | false true |
2925 // : | | | |
2926 // : | v stmt2 |
2927 // : | exitB: | |
2928 // : | stmt4 v |
2929 // : | ifA orig |
2930 // : | / \ |
2931 // : | / \ |
2932 // : | v v |
2933 // : | false true |
2934 // : | / \ |
2935 // : v v -----+
2936 // RegionA
2937 // |
2938 // v
2939 // exitA
2940 //
2941 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
2942
2943 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
2944 if (!loop->_head->is_Loop()) {
2945 return false; }
2946
2947 LoopNode *head = loop->_head->as_Loop();
2948
2949 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
2950 return false;
2951 }
2952
2953 // Check for complex exit control
2954 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
2955 Node *n = loop->_body.at(ii);
2956 int opc = n->Opcode();
2957 if (n->is_Call() ||
2958 opc == Op_Catch ||
2959 opc == Op_CatchProj ||
2960 opc == Op_Jump ||
2961 opc == Op_JumpProj) {
2962 #if !defined(PRODUCT)
2963 if (TracePartialPeeling) {
2964 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
2965 }
2966 #endif
2967 return false;
2968 }
2969 }
2970
2971 int dd = dom_depth(head);
2972
2973 // Step 1: find cut point
2974
2975 // Walk up dominators to loop head looking for first loop exit
2976 // which is executed on every path thru loop.
2977 IfNode *peel_if = NULL;
2978 IfNode *peel_if_cmpu = NULL;
2979
2980 Node *iff = loop->tail();
2981 while( iff != head ) {
2982 if( iff->is_If() ) {
2983 Node *ctrl = get_ctrl(iff->in(1));
2984 if (ctrl->is_top()) return false; // Dead test on live IF.
2985 // If loop-varying exit-test, check for induction variable
2986 if( loop->is_member(get_loop(ctrl)) &&
2987 loop->is_loop_exit(iff) &&
2988 is_possible_iv_test(iff)) {
2989 Node* cmp = iff->in(1)->in(1);
2990 if (cmp->Opcode() == Op_CmpI) {
2991 peel_if = iff->as_If();
2992 } else {
2993 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
2994 peel_if_cmpu = iff->as_If();
2995 }
2996 }
2997 }
2998 iff = idom(iff);
2999 }
3000 // Prefer signed compare over unsigned compare.
3001 IfNode* new_peel_if = NULL;
3002 if (peel_if == NULL) {
3003 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
3004 return false; // No peel point found
3005 }
3006 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3007 if (new_peel_if == NULL) {
3008 return false; // No peel point found
3009 }
3010 peel_if = new_peel_if;
3011 }
3012 Node* last_peel = stay_in_loop(peel_if, loop);
3013 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3014 if (first_not_peeled == NULL || first_not_peeled == head) {
3015 return false;
3016 }
3017
3018 #if !defined(PRODUCT)
3019 if (TraceLoopOpts) {
3020 tty->print("PartialPeel ");
3021 loop->dump_head();
3022 }
3023
3024 if (TracePartialPeeling) {
3025 tty->print_cr("before partial peel one iteration");
3026 Node_List wl;
3027 Node* t = head->in(2);
3028 while (true) {
3029 wl.push(t);
3030 if (t == head) break;
3031 t = idom(t);
3032 }
3033 while (wl.size() > 0) {
3034 Node* tt = wl.pop();
3035 tt->dump();
3036 if (tt == last_peel) tty->print_cr("-- cut --");
3037 }
3038 }
3039 #endif
3040 ResourceArea *area = Thread::current()->resource_area();
3041 VectorSet peel(area);
3042 VectorSet not_peel(area);
3043 Node_List peel_list(area);
3044 Node_List worklist(area);
3045 Node_List sink_list(area);
3046
3047 // Set of cfg nodes to peel are those that are executable from
3048 // the head through last_peel.
3049 assert(worklist.size() == 0, "should be empty");
3050 worklist.push(head);
3051 peel.set(head->_idx);
3052 while (worklist.size() > 0) {
3053 Node *n = worklist.pop();
3054 if (n != last_peel) {
3055 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3056 Node* use = n->fast_out(j);
3057 if (use->is_CFG() &&
3058 loop->is_member(get_loop(use)) &&
3059 !peel.test_set(use->_idx)) {
3060 worklist.push(use);
3061 }
3062 }
3063 }
3064 }
3065
3066 // Set of non-cfg nodes to peel are those that are control
3067 // dependent on the cfg nodes.
3068 uint i;
3069 for(i = 0; i < loop->_body.size(); i++ ) {
3070 Node *n = loop->_body.at(i);
3071 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3072 if (peel.test(n_c->_idx)) {
3073 peel.set(n->_idx);
3074 } else {
3075 not_peel.set(n->_idx);
3076 }
3077 }
3078
3079 // Step 2: move operations from the peeled section down into the
3080 // not-peeled section
3081
3082 // Get a post order schedule of nodes in the peel region
3083 // Result in right-most operand.
3084 scheduled_nodelist(loop, peel, peel_list );
3085
3086 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3087
3088 // For future check for too many new phis
3089 uint old_phi_cnt = 0;
3090 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3091 Node* use = head->fast_out(j);
3092 if (use->is_Phi()) old_phi_cnt++;
3093 }
3094
3095 #if !defined(PRODUCT)
3096 if (TracePartialPeeling) {
3097 tty->print_cr("\npeeled list");
3098 }
3099 #endif
3100
3101 // Evacuate nodes in peel region into the not_peeled region if possible
3102 uint new_phi_cnt = 0;
3103 uint cloned_for_outside_use = 0;
3104 for (i = 0; i < peel_list.size();) {
3105 Node* n = peel_list.at(i);
3106 #if !defined(PRODUCT)
3107 if (TracePartialPeeling) n->dump();
3108 #endif
3109 bool incr = true;
3110 if ( !n->is_CFG() ) {
3111
3112 if ( has_use_in_set(n, not_peel) ) {
3113
3114 // If not used internal to the peeled region,
3115 // move "n" from peeled to not_peeled region.
3116
3117 if ( !has_use_internal_to_set(n, peel, loop) ) {
3118
3119 // if not pinned and not a load (which maybe anti-dependent on a store)
3120 // and not a CMove (Matcher expects only bool->cmove).
3121 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
3122 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist );
3123 sink_list.push(n);
3124 peel >>= n->_idx; // delete n from peel set.
3125 not_peel <<= n->_idx; // add n to not_peel set.
3126 peel_list.remove(i);
3127 incr = false;
3128 #if !defined(PRODUCT)
3129 if (TracePartialPeeling) {
3130 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3131 n->_idx, get_ctrl(n)->_idx);
3132 }
3133 #endif
3134 }
3135 } else {
3136 // Otherwise check for special def-use cases that span
3137 // the peel/not_peel boundary such as bool->if
3138 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
3139 new_phi_cnt++;
3140 }
3141 }
3142 }
3143 if (incr) i++;
3144 }
3145
3146 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
3147 #if !defined(PRODUCT)
3148 if (TracePartialPeeling) {
3149 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3150 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
3151 }
3152 #endif
3153 if (new_peel_if != NULL) {
3154 remove_cmpi_loop_exit(new_peel_if, loop);
3155 }
3156 // Inhibit more partial peeling on this loop
3157 assert(!head->is_partial_peel_loop(), "not partial peeled");
3158 head->mark_partial_peel_failed();
3159 if (cloned_for_outside_use > 0) {
3160 // Terminate this round of loop opts because
3161 // the graph outside this loop was changed.
3162 C->set_major_progress();
3163 return true;
3164 }
3165 return false;
3166 }
3167
3168 // Step 3: clone loop, retarget control, and insert new phis
3169
3170 // Create new loop head for new phis and to hang
3171 // the nodes being moved (sinked) from the peel region.
3172 LoopNode* new_head = new LoopNode(last_peel, last_peel);
3173 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3174 _igvn.register_new_node_with_optimizer(new_head);
3175 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3176 _igvn.replace_input_of(first_not_peeled, 0, new_head);
3177 set_loop(new_head, loop);
3178 loop->_body.push(new_head);
3179 not_peel.set(new_head->_idx);
3180 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3181 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3182
3183 while (sink_list.size() > 0) {
3184 Node* n = sink_list.pop();
3185 set_ctrl(n, new_head);
3186 }
3187
3188 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3189
3190 clone_loop(loop, old_new, dd, IgnoreStripMined);
3191
3192 const uint clone_exit_idx = 1;
3193 const uint orig_exit_idx = 2;
3194 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
3195
3196 Node* head_clone = old_new[head->_idx];
3197 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
3198 Node* orig_tail_clone = head_clone->in(2);
3199
3200 // Add phi if "def" node is in peel set and "use" is not
3201
3202 for(i = 0; i < peel_list.size(); i++ ) {
3203 Node *def = peel_list.at(i);
3204 if (!def->is_CFG()) {
3205 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3206 Node *use = def->fast_out(j);
3207 if (has_node(use) && use->in(0) != C->top() &&
3208 (!peel.test(use->_idx) ||
3209 (use->is_Phi() && use->in(0) == head)) ) {
3210 worklist.push(use);
3211 }
3212 }
3213 while( worklist.size() ) {
3214 Node *use = worklist.pop();
3215 for (uint j = 1; j < use->req(); j++) {
3216 Node* n = use->in(j);
3217 if (n == def) {
3218
3219 // "def" is in peel set, "use" is not in peel set
3220 // or "use" is in the entry boundary (a phi) of the peel set
3221
3222 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
3223
3224 if ( loop->is_member(get_loop( use_c )) ) {
3225 // use is in loop
3226 if (old_new[use->_idx] != NULL) { // null for dead code
3227 Node* use_clone = old_new[use->_idx];
3228 _igvn.replace_input_of(use, j, C->top());
3229 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
3230 }
3231 } else {
3232 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
3233 // use is not in the loop, check if the live range includes the cut
3234 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
3235 if (not_peel.test(lp_if->_idx)) {
3236 assert(j == orig_exit_idx, "use from original loop");
3237 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
3238 }
3239 }
3240 }
3241 }
3242 }
3243 }
3244 }
3245
3246 // Step 3b: retarget control
3247
3248 // Redirect control to the new loop head if a cloned node in
3249 // the not_peeled region has control that points into the peeled region.
3250 // This necessary because the cloned peeled region will be outside
3251 // the loop.
3252 // from to
3253 // cloned-peeled <---+
3254 // new_head_clone: | <--+
3255 // cloned-not_peeled in(0) in(0)
3256 // orig-peeled
3257
3258 for(i = 0; i < loop->_body.size(); i++ ) {
3259 Node *n = loop->_body.at(i);
3260 if (!n->is_CFG() && n->in(0) != NULL &&
3261 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
3262 Node* n_clone = old_new[n->_idx];
3263 _igvn.replace_input_of(n_clone, 0, new_head_clone);
3264 }
3265 }
3266
3267 // Backedge of the surviving new_head (the clone) is original last_peel
3268 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
3269
3270 // Cut first node in original not_peel set
3271 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
3272 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
3273 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
3274
3275 // Copy head_clone back-branch info to original head
3276 // and remove original head's loop entry and
3277 // clone head's back-branch
3278 _igvn.rehash_node_delayed(head); // Multiple edge updates
3279 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
3280 head->set_req(LoopNode::LoopBackControl, C->top());
3281 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
3282
3283 // Similarly modify the phis
3284 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
3285 Node* use = head->fast_out(k);
3286 if (use->is_Phi() && use->outcnt() > 0) {
3287 Node* use_clone = old_new[use->_idx];
3288 _igvn.rehash_node_delayed(use); // Multiple edge updates
3289 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
3290 use->set_req(LoopNode::LoopBackControl, C->top());
3291 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
3292 }
3293 }
3294
3295 // Step 4: update dominator tree and dominator depth
3296
3297 set_idom(head, orig_tail_clone, dd);
3298 recompute_dom_depth();
3299
3300 // Inhibit more partial peeling on this loop
3301 new_head_clone->set_partial_peel_loop();
3302 C->set_major_progress();
3303 loop->record_for_igvn();
3304
3305 #if !defined(PRODUCT)
3306 if (TracePartialPeeling) {
3307 tty->print_cr("\nafter partial peel one iteration");
3308 Node_List wl(area);
3309 Node* t = last_peel;
3310 while (true) {
3311 wl.push(t);
3312 if (t == head_clone) break;
3313 t = idom(t);
3314 }
3315 while (wl.size() > 0) {
3316 Node* tt = wl.pop();
3317 if (tt == head) tty->print_cr("orig head");
3318 else if (tt == new_head_clone) tty->print_cr("new head");
3319 else if (tt == head_clone) tty->print_cr("clone head");
3320 tt->dump();
3321 }
3322 }
3323 #endif
3324 return true;
3325 }
3326
3327 //------------------------------reorg_offsets----------------------------------
3328 // Reorganize offset computations to lower register pressure. Mostly
3329 // prevent loop-fallout uses of the pre-incremented trip counter (which are
3330 // then alive with the post-incremented trip counter forcing an extra
3331 // register move)
3332 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
3333 // Perform it only for canonical counted loops.
3334 // Loop's shape could be messed up by iteration_split_impl.
3335 if (!loop->_head->is_CountedLoop())
3336 return;
3337 if (!loop->_head->as_Loop()->is_valid_counted_loop())
3338 return;
3339
3340 CountedLoopNode *cl = loop->_head->as_CountedLoop();
3341 CountedLoopEndNode *cle = cl->loopexit();
3342 Node *exit = cle->proj_out(false);
3343 Node *phi = cl->phi();
3344
3345 // Check for the special case of folks using the pre-incremented
3346 // trip-counter on the fall-out path (forces the pre-incremented
3347 // and post-incremented trip counter to be live at the same time).
3348 // Fix this by adjusting to use the post-increment trip counter.
3349
3350 bool progress = true;
3351 while (progress) {
3352 progress = false;
3353 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3354 Node* use = phi->fast_out(i); // User of trip-counter
3355 if (!has_ctrl(use)) continue;
3356 Node *u_ctrl = get_ctrl(use);
3357 if (use->is_Phi()) {
3358 u_ctrl = NULL;
3359 for (uint j = 1; j < use->req(); j++)
3360 if (use->in(j) == phi)
3361 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
3362 }
3363 IdealLoopTree *u_loop = get_loop(u_ctrl);
3364 // Look for loop-invariant use
3365 if (u_loop == loop) continue;
3366 if (loop->is_member(u_loop)) continue;
3367 // Check that use is live out the bottom. Assuming the trip-counter
3368 // update is right at the bottom, uses of of the loop middle are ok.
3369 if (dom_lca(exit, u_ctrl) != exit) continue;
3370 // Hit! Refactor use to use the post-incremented tripcounter.
3371 // Compute a post-increment tripcounter.
3372 Node *opaq = new Opaque2Node( C, cle->incr() );
3373 register_new_node(opaq, exit);
3374 Node *neg_stride = _igvn.intcon(-cle->stride_con());
3375 set_ctrl(neg_stride, C->root());
3376 Node *post = new AddINode( opaq, neg_stride);
3377 register_new_node(post, exit);
3378 _igvn.rehash_node_delayed(use);
3379 for (uint j = 1; j < use->req(); j++) {
3380 if (use->in(j) == phi)
3381 use->set_req(j, post);
3382 }
3383 // Since DU info changed, rerun loop
3384 progress = true;
3385 break;
3386 }
3387 }
3388
3389 }