Wed, 10 Aug 2016 14:59:21 +0200
8152172: PPC64: Support AES intrinsics
Summary: Add support for AES intrinsics on PPC64.
Reviewed-by: kvn, mdoerr, simonis, zmajo
Contributed-by: Hiroshi H Horii <horii@jp.ibm.com>
1 /*
2 * Copyright (c) 1999, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "memory/allocation.inline.hpp"
27 #include "opto/addnode.hpp"
28 #include "opto/connode.hpp"
29 #include "opto/divnode.hpp"
30 #include "opto/loopnode.hpp"
31 #include "opto/matcher.hpp"
32 #include "opto/mulnode.hpp"
33 #include "opto/rootnode.hpp"
34 #include "opto/subnode.hpp"
36 //=============================================================================
37 //------------------------------split_thru_phi---------------------------------
38 // Split Node 'n' through merge point if there is enough win.
39 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
40 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
41 // ConvI2L may have type information on it which is unsafe to push up
42 // so disable this for now
43 return NULL;
44 }
46 // Splitting range check CastIIs through a loop induction Phi can
47 // cause new Phis to be created that are left unrelated to the loop
48 // induction Phi and prevent optimizations (vectorization)
49 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() &&
50 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) {
51 return NULL;
52 }
54 int wins = 0;
55 assert(!n->is_CFG(), "");
56 assert(region->is_Region(), "");
58 const Type* type = n->bottom_type();
59 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr();
60 Node *phi;
61 if (t_oop != NULL && t_oop->is_known_instance_field()) {
62 int iid = t_oop->instance_id();
63 int index = C->get_alias_index(t_oop);
64 int offset = t_oop->offset();
65 phi = new (C) PhiNode(region, type, NULL, iid, index, offset);
66 } else {
67 phi = PhiNode::make_blank(region, n);
68 }
69 uint old_unique = C->unique();
70 for (uint i = 1; i < region->req(); i++) {
71 Node *x;
72 Node* the_clone = NULL;
73 if (region->in(i) == C->top()) {
74 x = C->top(); // Dead path? Use a dead data op
75 } else {
76 x = n->clone(); // Else clone up the data op
77 the_clone = x; // Remember for possible deletion.
78 // Alter data node to use pre-phi inputs
79 if (n->in(0) == region)
80 x->set_req( 0, region->in(i) );
81 for (uint j = 1; j < n->req(); j++) {
82 Node *in = n->in(j);
83 if (in->is_Phi() && in->in(0) == region)
84 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
85 }
86 }
87 // Check for a 'win' on some paths
88 const Type *t = x->Value(&_igvn);
90 bool singleton = t->singleton();
92 // A TOP singleton indicates that there are no possible values incoming
93 // along a particular edge. In most cases, this is OK, and the Phi will
94 // be eliminated later in an Ideal call. However, we can't allow this to
95 // happen if the singleton occurs on loop entry, as the elimination of
96 // the PhiNode may cause the resulting node to migrate back to a previous
97 // loop iteration.
98 if (singleton && t == Type::TOP) {
99 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
100 // irreducible loop may not be indicated by an affirmative is_Loop());
101 // therefore, the only top we can split thru a phi is on a backedge of
102 // a loop.
103 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
104 }
106 if (singleton) {
107 wins++;
108 x = ((PhaseGVN&)_igvn).makecon(t);
109 } else {
110 // We now call Identity to try to simplify the cloned node.
111 // Note that some Identity methods call phase->type(this).
112 // Make sure that the type array is big enough for
113 // our new node, even though we may throw the node away.
114 // (Note: This tweaking with igvn only works because x is a new node.)
115 _igvn.set_type(x, t);
116 // If x is a TypeNode, capture any more-precise type permanently into Node
117 // otherwise it will be not updated during igvn->transform since
118 // igvn->type(x) is set to x->Value() already.
119 x->raise_bottom_type(t);
120 Node *y = x->Identity(&_igvn);
121 if (y != x) {
122 wins++;
123 x = y;
124 } else {
125 y = _igvn.hash_find(x);
126 if (y) {
127 wins++;
128 x = y;
129 } else {
130 // Else x is a new node we are keeping
131 // We do not need register_new_node_with_optimizer
132 // because set_type has already been called.
133 _igvn._worklist.push(x);
134 }
135 }
136 }
137 if (x != the_clone && the_clone != NULL)
138 _igvn.remove_dead_node(the_clone);
139 phi->set_req( i, x );
140 }
141 // Too few wins?
142 if (wins <= policy) {
143 _igvn.remove_dead_node(phi);
144 return NULL;
145 }
147 // Record Phi
148 register_new_node( phi, region );
150 for (uint i2 = 1; i2 < phi->req(); i2++) {
151 Node *x = phi->in(i2);
152 // If we commoned up the cloned 'x' with another existing Node,
153 // the existing Node picks up a new use. We need to make the
154 // existing Node occur higher up so it dominates its uses.
155 Node *old_ctrl;
156 IdealLoopTree *old_loop;
158 if (x->is_Con()) {
159 // Constant's control is always root.
160 set_ctrl(x, C->root());
161 continue;
162 }
163 // The occasional new node
164 if (x->_idx >= old_unique) { // Found a new, unplaced node?
165 old_ctrl = NULL;
166 old_loop = NULL; // Not in any prior loop
167 } else {
168 old_ctrl = get_ctrl(x);
169 old_loop = get_loop(old_ctrl); // Get prior loop
170 }
171 // New late point must dominate new use
172 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
173 if (new_ctrl == old_ctrl) // Nothing is changed
174 continue;
176 IdealLoopTree *new_loop = get_loop(new_ctrl);
178 // Don't move x into a loop if its uses are
179 // outside of loop. Otherwise x will be cloned
180 // for each use outside of this loop.
181 IdealLoopTree *use_loop = get_loop(region);
182 if (!new_loop->is_member(use_loop) &&
183 (old_loop == NULL || !new_loop->is_member(old_loop))) {
184 // Take early control, later control will be recalculated
185 // during next iteration of loop optimizations.
186 new_ctrl = get_early_ctrl(x);
187 new_loop = get_loop(new_ctrl);
188 }
189 // Set new location
190 set_ctrl(x, new_ctrl);
191 // If changing loop bodies, see if we need to collect into new body
192 if (old_loop != new_loop) {
193 if (old_loop && !old_loop->_child)
194 old_loop->_body.yank(x);
195 if (!new_loop->_child)
196 new_loop->_body.push(x); // Collect body info
197 }
198 }
200 return phi;
201 }
203 //------------------------------dominated_by------------------------------------
204 // Replace the dominated test with an obvious true or false. Place it on the
205 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
206 // live path up to the dominating control.
207 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) {
208 #ifndef PRODUCT
209 if (VerifyLoopOptimizations && PrintOpto) tty->print_cr("dominating test");
210 #endif
213 // prevdom is the dominating projection of the dominating test.
214 assert( iff->is_If(), "" );
215 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
216 int pop = prevdom->Opcode();
217 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
218 if (flip) {
219 if (pop == Op_IfTrue)
220 pop = Op_IfFalse;
221 else
222 pop = Op_IfTrue;
223 }
224 // 'con' is set to true or false to kill the dominated test.
225 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
226 set_ctrl(con, C->root()); // Constant gets a new use
227 // Hack the dominated test
228 _igvn.replace_input_of(iff, 1, con);
230 // If I dont have a reachable TRUE and FALSE path following the IfNode then
231 // I can assume this path reaches an infinite loop. In this case it's not
232 // important to optimize the data Nodes - either the whole compilation will
233 // be tossed or this path (and all data Nodes) will go dead.
234 if (iff->outcnt() != 2) return;
236 // Make control-dependent data Nodes on the live path (path that will remain
237 // once the dominated IF is removed) become control-dependent on the
238 // dominating projection.
239 Node* dp = iff->as_If()->proj_out(pop == Op_IfTrue);
241 // Loop predicates may have depending checks which should not
242 // be skipped. For example, range check predicate has two checks
243 // for lower and upper bounds.
244 if (dp == NULL)
245 return;
247 ProjNode* dp_proj = dp->as_Proj();
248 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj();
249 if (exclude_loop_predicate &&
250 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) ||
251 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check))) {
252 // If this is a range check (IfNode::is_range_check), do not
253 // reorder because Compile::allow_range_check_smearing might have
254 // changed the check.
255 return; // Let IGVN transformation change control dependence.
256 }
258 IdealLoopTree *old_loop = get_loop(dp);
260 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
261 Node* cd = dp->fast_out(i); // Control-dependent node
262 if (cd->depends_only_on_test()) {
263 assert(cd->in(0) == dp, "");
264 _igvn.replace_input_of(cd, 0, prevdom);
265 set_early_ctrl(cd);
266 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
267 if (old_loop != new_loop) {
268 if (!old_loop->_child) old_loop->_body.yank(cd);
269 if (!new_loop->_child) new_loop->_body.push(cd);
270 }
271 --i;
272 --imax;
273 }
274 }
275 }
277 //------------------------------has_local_phi_input----------------------------
278 // Return TRUE if 'n' has Phi inputs from its local block and no other
279 // block-local inputs (all non-local-phi inputs come from earlier blocks)
280 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
281 Node *n_ctrl = get_ctrl(n);
282 // See if some inputs come from a Phi in this block, or from before
283 // this block.
284 uint i;
285 for( i = 1; i < n->req(); i++ ) {
286 Node *phi = n->in(i);
287 if( phi->is_Phi() && phi->in(0) == n_ctrl )
288 break;
289 }
290 if( i >= n->req() )
291 return NULL; // No Phi inputs; nowhere to clone thru
293 // Check for inputs created between 'n' and the Phi input. These
294 // must split as well; they have already been given the chance
295 // (courtesy of a post-order visit) and since they did not we must
296 // recover the 'cost' of splitting them by being very profitable
297 // when splitting 'n'. Since this is unlikely we simply give up.
298 for( i = 1; i < n->req(); i++ ) {
299 Node *m = n->in(i);
300 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
301 // We allow the special case of AddP's with no local inputs.
302 // This allows us to split-up address expressions.
303 if (m->is_AddP() &&
304 get_ctrl(m->in(2)) != n_ctrl &&
305 get_ctrl(m->in(3)) != n_ctrl) {
306 // Move the AddP up to dominating point
307 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl)));
308 continue;
309 }
310 return NULL;
311 }
312 }
314 return n_ctrl;
315 }
317 //------------------------------remix_address_expressions----------------------
318 // Rework addressing expressions to get the most loop-invariant stuff
319 // moved out. We'd like to do all associative operators, but it's especially
320 // important (common) to do address expressions.
321 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
322 if (!has_ctrl(n)) return NULL;
323 Node *n_ctrl = get_ctrl(n);
324 IdealLoopTree *n_loop = get_loop(n_ctrl);
326 // See if 'n' mixes loop-varying and loop-invariant inputs and
327 // itself is loop-varying.
329 // Only interested in binary ops (and AddP)
330 if( n->req() < 3 || n->req() > 4 ) return NULL;
332 Node *n1_ctrl = get_ctrl(n->in( 1));
333 Node *n2_ctrl = get_ctrl(n->in( 2));
334 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
335 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
336 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
337 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
339 // Does one of my inputs spin in a tighter loop than self?
340 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
341 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
342 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
343 return NULL; // Leave well enough alone
345 // Is at least one of my inputs loop-invariant?
346 if( n1_loop == n_loop &&
347 n2_loop == n_loop &&
348 n3_loop == n_loop )
349 return NULL; // No loop-invariant inputs
352 int n_op = n->Opcode();
354 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
355 if( n_op == Op_LShiftI ) {
356 // Scale is loop invariant
357 Node *scale = n->in(2);
358 Node *scale_ctrl = get_ctrl(scale);
359 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
360 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
361 return NULL;
362 const TypeInt *scale_t = scale->bottom_type()->isa_int();
363 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
364 return NULL; // Dont bother with byte/short masking
365 // Add must vary with loop (else shift would be loop-invariant)
366 Node *add = n->in(1);
367 Node *add_ctrl = get_ctrl(add);
368 IdealLoopTree *add_loop = get_loop(add_ctrl);
369 //assert( n_loop == add_loop, "" );
370 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
372 // Convert I-V into I+ (0-V); same for V-I
373 if( add->Opcode() == Op_SubI &&
374 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
375 Node *zero = _igvn.intcon(0);
376 set_ctrl(zero, C->root());
377 Node *neg = new (C) SubINode( _igvn.intcon(0), add->in(2) );
378 register_new_node( neg, get_ctrl(add->in(2) ) );
379 add = new (C) AddINode( add->in(1), neg );
380 register_new_node( add, add_ctrl );
381 }
382 if( add->Opcode() != Op_AddI ) return NULL;
383 // See if one add input is loop invariant
384 Node *add_var = add->in(1);
385 Node *add_var_ctrl = get_ctrl(add_var);
386 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
387 Node *add_invar = add->in(2);
388 Node *add_invar_ctrl = get_ctrl(add_invar);
389 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
390 if( add_var_loop == n_loop ) {
391 } else if( add_invar_loop == n_loop ) {
392 // Swap to find the invariant part
393 add_invar = add_var;
394 add_invar_ctrl = add_var_ctrl;
395 add_invar_loop = add_var_loop;
396 add_var = add->in(2);
397 Node *add_var_ctrl = get_ctrl(add_var);
398 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
399 } else // Else neither input is loop invariant
400 return NULL;
401 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
402 return NULL; // No invariant part of the add?
404 // Yes! Reshape address expression!
405 Node *inv_scale = new (C) LShiftINode( add_invar, scale );
406 Node *inv_scale_ctrl =
407 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
408 add_invar_ctrl : scale_ctrl;
409 register_new_node( inv_scale, inv_scale_ctrl );
410 Node *var_scale = new (C) LShiftINode( add_var, scale );
411 register_new_node( var_scale, n_ctrl );
412 Node *var_add = new (C) AddINode( var_scale, inv_scale );
413 register_new_node( var_add, n_ctrl );
414 _igvn.replace_node( n, var_add );
415 return var_add;
416 }
418 // Replace (I+V) with (V+I)
419 if( n_op == Op_AddI ||
420 n_op == Op_AddL ||
421 n_op == Op_AddF ||
422 n_op == Op_AddD ||
423 n_op == Op_MulI ||
424 n_op == Op_MulL ||
425 n_op == Op_MulF ||
426 n_op == Op_MulD ) {
427 if( n2_loop == n_loop ) {
428 assert( n1_loop != n_loop, "" );
429 n->swap_edges(1, 2);
430 }
431 }
433 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
434 // but not if I2 is a constant.
435 if( n_op == Op_AddP ) {
436 if( n2_loop == n_loop && n3_loop != n_loop ) {
437 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
438 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
439 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
440 IdealLoopTree *n22loop = get_loop( n22_ctrl );
441 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
442 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
443 n23_loop == n_loop ) {
444 Node *add1 = new (C) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
445 // Stuff new AddP in the loop preheader
446 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
447 Node *add2 = new (C) AddPNode( n->in(1), add1, n->in(2)->in(3) );
448 register_new_node( add2, n_ctrl );
449 _igvn.replace_node( n, add2 );
450 return add2;
451 }
452 }
453 }
455 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
456 if( n2_loop != n_loop && n3_loop == n_loop ) {
457 if( n->in(3)->Opcode() == Op_AddI ) {
458 Node *V = n->in(3)->in(1);
459 Node *I = n->in(3)->in(2);
460 if( is_member(n_loop,get_ctrl(V)) ) {
461 } else {
462 Node *tmp = V; V = I; I = tmp;
463 }
464 if( !is_member(n_loop,get_ctrl(I)) ) {
465 Node *add1 = new (C) AddPNode( n->in(1), n->in(2), I );
466 // Stuff new AddP in the loop preheader
467 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
468 Node *add2 = new (C) AddPNode( n->in(1), add1, V );
469 register_new_node( add2, n_ctrl );
470 _igvn.replace_node( n, add2 );
471 return add2;
472 }
473 }
474 }
475 }
477 return NULL;
478 }
480 //------------------------------conditional_move-------------------------------
481 // Attempt to replace a Phi with a conditional move. We have some pretty
482 // strict profitability requirements. All Phis at the merge point must
483 // be converted, so we can remove the control flow. We need to limit the
484 // number of c-moves to a small handful. All code that was in the side-arms
485 // of the CFG diamond is now speculatively executed. This code has to be
486 // "cheap enough". We are pretty much limited to CFG diamonds that merge
487 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
488 Node *PhaseIdealLoop::conditional_move( Node *region ) {
490 assert(region->is_Region(), "sanity check");
491 if (region->req() != 3) return NULL;
493 // Check for CFG diamond
494 Node *lp = region->in(1);
495 Node *rp = region->in(2);
496 if (!lp || !rp) return NULL;
497 Node *lp_c = lp->in(0);
498 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
499 IfNode *iff = lp_c->as_If();
501 // Check for ops pinned in an arm of the diamond.
502 // Can't remove the control flow in this case
503 if (lp->outcnt() > 1) return NULL;
504 if (rp->outcnt() > 1) return NULL;
506 IdealLoopTree* r_loop = get_loop(region);
507 assert(r_loop == get_loop(iff), "sanity");
508 // Always convert to CMOVE if all results are used only outside this loop.
509 bool used_inside_loop = (r_loop == _ltree_root);
511 // Check profitability
512 int cost = 0;
513 int phis = 0;
514 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
515 Node *out = region->fast_out(i);
516 if (!out->is_Phi()) continue; // Ignore other control edges, etc
517 phis++;
518 PhiNode* phi = out->as_Phi();
519 BasicType bt = phi->type()->basic_type();
520 switch (bt) {
521 case T_FLOAT:
522 case T_DOUBLE: {
523 cost += Matcher::float_cmove_cost(); // Could be very expensive
524 break;
525 }
526 case T_LONG: {
527 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
528 }
529 case T_INT: // These all CMOV fine
530 case T_ADDRESS: { // (RawPtr)
531 cost++;
532 break;
533 }
534 case T_NARROWOOP: // Fall through
535 case T_OBJECT: { // Base oops are OK, but not derived oops
536 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
537 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
538 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
539 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
540 // have a Phi for the base here that we convert to a CMOVE all is well
541 // and good. But if the base is dead, we'll not make a CMOVE. Later
542 // the allocator will have to produce a base by creating a CMOVE of the
543 // relevant bases. This puts the allocator in the business of
544 // manufacturing expensive instructions, generally a bad plan.
545 // Just Say No to Conditionally-Moved Derived Pointers.
546 if (tp && tp->offset() != 0)
547 return NULL;
548 cost++;
549 break;
550 }
551 default:
552 return NULL; // In particular, can't do memory or I/O
553 }
554 // Add in cost any speculative ops
555 for (uint j = 1; j < region->req(); j++) {
556 Node *proj = region->in(j);
557 Node *inp = phi->in(j);
558 if (get_ctrl(inp) == proj) { // Found local op
559 cost++;
560 // Check for a chain of dependent ops; these will all become
561 // speculative in a CMOV.
562 for (uint k = 1; k < inp->req(); k++)
563 if (get_ctrl(inp->in(k)) == proj)
564 cost += ConditionalMoveLimit; // Too much speculative goo
565 }
566 }
567 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
568 // This will likely Split-If, a higher-payoff operation.
569 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
570 Node* use = phi->fast_out(k);
571 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
572 cost += ConditionalMoveLimit;
573 // Is there a use inside the loop?
574 // Note: check only basic types since CMoveP is pinned.
575 if (!used_inside_loop && is_java_primitive(bt)) {
576 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
577 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
578 used_inside_loop = true;
579 }
580 }
581 }
582 }
583 Node* bol = iff->in(1);
584 assert(bol->Opcode() == Op_Bool, "");
585 int cmp_op = bol->in(1)->Opcode();
586 // It is expensive to generate flags from a float compare.
587 // Avoid duplicated float compare.
588 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
590 float infrequent_prob = PROB_UNLIKELY_MAG(3);
591 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
592 if (used_inside_loop) {
593 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
595 // BlockLayoutByFrequency optimization moves infrequent branch
596 // from hot path. No point in CMOV'ing in such case (110 is used
597 // instead of 100 to take into account not exactness of float value).
598 if (BlockLayoutByFrequency) {
599 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
600 }
601 }
602 // Check for highly predictable branch. No point in CMOV'ing if
603 // we are going to predict accurately all the time.
604 if (iff->_prob < infrequent_prob ||
605 iff->_prob > (1.0f - infrequent_prob))
606 return NULL;
608 // --------------
609 // Now replace all Phis with CMOV's
610 Node *cmov_ctrl = iff->in(0);
611 uint flip = (lp->Opcode() == Op_IfTrue);
612 while (1) {
613 PhiNode* phi = NULL;
614 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
615 Node *out = region->fast_out(i);
616 if (out->is_Phi()) {
617 phi = out->as_Phi();
618 break;
619 }
620 }
621 if (phi == NULL) break;
622 #ifndef PRODUCT
623 if (PrintOpto && VerifyLoopOptimizations) tty->print_cr("CMOV");
624 #endif
625 // Move speculative ops
626 for (uint j = 1; j < region->req(); j++) {
627 Node *proj = region->in(j);
628 Node *inp = phi->in(j);
629 if (get_ctrl(inp) == proj) { // Found local op
630 #ifndef PRODUCT
631 if (PrintOpto && VerifyLoopOptimizations) {
632 tty->print(" speculate: ");
633 inp->dump();
634 }
635 #endif
636 set_ctrl(inp, cmov_ctrl);
637 }
638 }
639 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) );
640 register_new_node( cmov, cmov_ctrl );
641 _igvn.replace_node( phi, cmov );
642 #ifndef PRODUCT
643 if (TraceLoopOpts) {
644 tty->print("CMOV ");
645 r_loop->dump_head();
646 if (Verbose) {
647 bol->in(1)->dump(1);
648 cmov->dump(1);
649 }
650 }
651 if (VerifyLoopOptimizations) verify();
652 #endif
653 }
655 // The useless CFG diamond will fold up later; see the optimization in
656 // RegionNode::Ideal.
657 _igvn._worklist.push(region);
659 return iff->in(1);
660 }
662 //------------------------------split_if_with_blocks_pre-----------------------
663 // Do the real work in a non-recursive function. Data nodes want to be
664 // cloned in the pre-order so they can feed each other nicely.
665 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
666 // Cloning these guys is unlikely to win
667 int n_op = n->Opcode();
668 if( n_op == Op_MergeMem ) return n;
669 if( n->is_Proj() ) return n;
670 // Do not clone-up CmpFXXX variations, as these are always
671 // followed by a CmpI
672 if( n->is_Cmp() ) return n;
673 // Attempt to use a conditional move instead of a phi/branch
674 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
675 Node *cmov = conditional_move( n );
676 if( cmov ) return cmov;
677 }
678 if( n->is_CFG() || n->is_LoadStore() )
679 return n;
680 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
681 n_op == Op_Opaque2 ) {
682 if( !C->major_progress() ) // If chance of no more loop opts...
683 _igvn._worklist.push(n); // maybe we'll remove them
684 return n;
685 }
687 if( n->is_Con() ) return n; // No cloning for Con nodes
689 Node *n_ctrl = get_ctrl(n);
690 if( !n_ctrl ) return n; // Dead node
692 // Attempt to remix address expressions for loop invariants
693 Node *m = remix_address_expressions( n );
694 if( m ) return m;
696 // Determine if the Node has inputs from some local Phi.
697 // Returns the block to clone thru.
698 Node *n_blk = has_local_phi_input( n );
699 if( !n_blk ) return n;
700 // Do not clone the trip counter through on a CountedLoop
701 // (messes up the canonical shape).
702 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
704 // Check for having no control input; not pinned. Allow
705 // dominating control.
706 if( n->in(0) ) {
707 Node *dom = idom(n_blk);
708 if( dom_lca( n->in(0), dom ) != n->in(0) )
709 return n;
710 }
711 // Policy: when is it profitable. You must get more wins than
712 // policy before it is considered profitable. Policy is usually 0,
713 // so 1 win is considered profitable. Big merges will require big
714 // cloning, so get a larger policy.
715 int policy = n_blk->req() >> 2;
717 // If the loop is a candidate for range check elimination,
718 // delay splitting through it's phi until a later loop optimization
719 if (n_blk->is_CountedLoop()) {
720 IdealLoopTree *lp = get_loop(n_blk);
721 if (lp && lp->_rce_candidate) {
722 return n;
723 }
724 }
726 // Use same limit as split_if_with_blocks_post
727 if( C->unique() > 35000 ) return n; // Method too big
729 // Split 'n' through the merge point if it is profitable
730 Node *phi = split_thru_phi( n, n_blk, policy );
731 if (!phi) return n;
733 // Found a Phi to split thru!
734 // Replace 'n' with the new phi
735 _igvn.replace_node( n, phi );
736 // Moved a load around the loop, 'en-registering' something.
737 if (n_blk->is_Loop() && n->is_Load() &&
738 !phi->in(LoopNode::LoopBackControl)->is_Load())
739 C->set_major_progress();
741 return phi;
742 }
744 static bool merge_point_too_heavy(Compile* C, Node* region) {
745 // Bail out if the region and its phis have too many users.
746 int weight = 0;
747 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
748 weight += region->fast_out(i)->outcnt();
749 }
750 int nodes_left = C->max_node_limit() - C->live_nodes();
751 if (weight * 8 > nodes_left) {
752 #ifndef PRODUCT
753 if (PrintOpto)
754 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
755 #endif
756 return true;
757 } else {
758 return false;
759 }
760 }
762 static bool merge_point_safe(Node* region) {
763 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
764 // having a PhiNode input. This sidesteps the dangerous case where the split
765 // ConvI2LNode may become TOP if the input Value() does not
766 // overlap the ConvI2L range, leaving a node which may not dominate its
767 // uses.
768 // A better fix for this problem can be found in the BugTraq entry, but
769 // expediency for Mantis demands this hack.
770 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop
771 // split_if_with_blocks from splitting a block because we could not move around
772 // the FastLockNode.
773 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
774 Node* n = region->fast_out(i);
775 if (n->is_Phi()) {
776 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
777 Node* m = n->fast_out(j);
778 if (m->is_FastLock())
779 return false;
780 #ifdef _LP64
781 if (m->Opcode() == Op_ConvI2L)
782 return false;
783 if (m->is_CastII() && m->isa_CastII()->has_range_check()) {
784 return false;
785 }
786 #endif
787 }
788 }
789 }
790 return true;
791 }
794 //------------------------------place_near_use---------------------------------
795 // Place some computation next to use but not inside inner loops.
796 // For inner loop uses move it to the preheader area.
797 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const {
798 IdealLoopTree *u_loop = get_loop( useblock );
799 return (u_loop->_irreducible || u_loop->_child)
800 ? useblock
801 : u_loop->_head->in(LoopNode::EntryControl);
802 }
805 //------------------------------split_if_with_blocks_post----------------------
806 // Do the real work in a non-recursive function. CFG hackery wants to be
807 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
808 // info.
809 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) {
811 // Cloning Cmp through Phi's involves the split-if transform.
812 // FastLock is not used by an If
813 if( n->is_Cmp() && !n->is_FastLock() ) {
814 if( C->unique() > 35000 ) return; // Method too big
816 // Do not do 'split-if' if irreducible loops are present.
817 if( _has_irreducible_loops )
818 return;
820 Node *n_ctrl = get_ctrl(n);
821 // Determine if the Node has inputs from some local Phi.
822 // Returns the block to clone thru.
823 Node *n_blk = has_local_phi_input( n );
824 if( n_blk != n_ctrl ) return;
826 if( merge_point_too_heavy(C, n_ctrl) )
827 return;
829 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare?
830 Node *bol = n->unique_out();
831 assert( bol->is_Bool(), "expect a bool here" );
832 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare?
833 Node *iff = bol->unique_out();
835 // Check some safety conditions
836 if( iff->is_If() ) { // Classic split-if?
837 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if
838 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
839 // Can't split CMove with different control edge.
840 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) return;
841 if( get_ctrl(iff->in(2)) == n_ctrl ||
842 get_ctrl(iff->in(3)) == n_ctrl )
843 return; // Inputs not yet split-up
844 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) {
845 return; // Loop-invar test gates loop-varying CMOVE
846 }
847 } else {
848 return; // some other kind of node, such as an Allocate
849 }
851 // Do not do 'split-if' if some paths are dead. First do dead code
852 // elimination and then see if its still profitable.
853 for( uint i = 1; i < n_ctrl->req(); i++ )
854 if( n_ctrl->in(i) == C->top() )
855 return;
857 // When is split-if profitable? Every 'win' on means some control flow
858 // goes dead, so it's almost always a win.
859 int policy = 0;
860 // If trying to do a 'Split-If' at the loop head, it is only
861 // profitable if the cmp folds up on BOTH paths. Otherwise we
862 // risk peeling a loop forever.
864 // CNC - Disabled for now. Requires careful handling of loop
865 // body selection for the cloned code. Also, make sure we check
866 // for any input path not being in the same loop as n_ctrl. For
867 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
868 // because the alternative loop entry points won't be converted
869 // into LoopNodes.
870 IdealLoopTree *n_loop = get_loop(n_ctrl);
871 for( uint j = 1; j < n_ctrl->req(); j++ )
872 if( get_loop(n_ctrl->in(j)) != n_loop )
873 return;
875 // Check for safety of the merge point.
876 if( !merge_point_safe(n_ctrl) ) {
877 return;
878 }
880 // Split compare 'n' through the merge point if it is profitable
881 Node *phi = split_thru_phi( n, n_ctrl, policy );
882 if( !phi ) return;
884 // Found a Phi to split thru!
885 // Replace 'n' with the new phi
886 _igvn.replace_node( n, phi );
888 // Now split the bool up thru the phi
889 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 );
890 guarantee(bolphi != NULL, "null boolean phi node");
892 _igvn.replace_node( bol, bolphi );
893 assert( iff->in(1) == bolphi, "" );
895 if( bolphi->Value(&_igvn)->singleton() )
896 return;
898 // Conditional-move? Must split up now
899 if( !iff->is_If() ) {
900 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 );
901 _igvn.replace_node( iff, cmovphi );
902 return;
903 }
905 // Now split the IF
906 do_split_if( iff );
907 return;
908 }
910 // Check for an IF ready to split; one that has its
911 // condition codes input coming from a Phi at the block start.
912 int n_op = n->Opcode();
914 // Check for an IF being dominated by another IF same test
915 if (n_op == Op_If) {
916 Node *bol = n->in(1);
917 uint max = bol->outcnt();
918 // Check for same test used more than once?
919 if (max > 1 && bol->is_Bool()) {
920 // Search up IDOMs to see if this IF is dominated.
921 Node *cutoff = get_ctrl(bol);
923 // Now search up IDOMs till cutoff, looking for a dominating test
924 Node *prevdom = n;
925 Node *dom = idom(prevdom);
926 while (dom != cutoff) {
927 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) {
928 // Replace the dominated test with an obvious true or false.
929 // Place it on the IGVN worklist for later cleanup.
930 C->set_major_progress();
931 dominated_by(prevdom, n, false, true);
932 #ifndef PRODUCT
933 if( VerifyLoopOptimizations ) verify();
934 #endif
935 return;
936 }
937 prevdom = dom;
938 dom = idom(prevdom);
939 }
940 }
941 }
943 // See if a shared loop-varying computation has no loop-varying uses.
944 // Happens if something is only used for JVM state in uncommon trap exits,
945 // like various versions of induction variable+offset. Clone the
946 // computation per usage to allow it to sink out of the loop.
947 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
948 Node *n_ctrl = get_ctrl(n);
949 IdealLoopTree *n_loop = get_loop(n_ctrl);
950 if( n_loop != _ltree_root ) {
951 DUIterator_Fast imax, i = n->fast_outs(imax);
952 for (; i < imax; i++) {
953 Node* u = n->fast_out(i);
954 if( !has_ctrl(u) ) break; // Found control user
955 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
956 if( u_loop == n_loop ) break; // Found loop-varying use
957 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
958 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
959 }
960 bool did_break = (i < imax); // Did we break out of the previous loop?
961 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
962 Node *late_load_ctrl = NULL;
963 if (n->is_Load()) {
964 // If n is a load, get and save the result from get_late_ctrl(),
965 // to be later used in calculating the control for n's clones.
966 clear_dom_lca_tags();
967 late_load_ctrl = get_late_ctrl(n, n_ctrl);
968 }
969 // If n is a load, and the late control is the same as the current
970 // control, then the cloning of n is a pointless exercise, because
971 // GVN will ensure that we end up where we started.
972 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
973 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
974 Node *u = n->last_out(j); // Clone private computation per use
975 _igvn.rehash_node_delayed(u);
976 Node *x = n->clone(); // Clone computation
977 Node *x_ctrl = NULL;
978 if( u->is_Phi() ) {
979 // Replace all uses of normal nodes. Replace Phi uses
980 // individually, so the separate Nodes can sink down
981 // different paths.
982 uint k = 1;
983 while( u->in(k) != n ) k++;
984 u->set_req( k, x );
985 // x goes next to Phi input path
986 x_ctrl = u->in(0)->in(k);
987 --j;
988 } else { // Normal use
989 // Replace all uses
990 for( uint k = 0; k < u->req(); k++ ) {
991 if( u->in(k) == n ) {
992 u->set_req( k, x );
993 --j;
994 }
995 }
996 x_ctrl = get_ctrl(u);
997 }
999 // Find control for 'x' next to use but not inside inner loops.
1000 // For inner loop uses get the preheader area.
1001 x_ctrl = place_near_use(x_ctrl);
1003 if (n->is_Load()) {
1004 // For loads, add a control edge to a CFG node outside of the loop
1005 // to force them to not combine and return back inside the loop
1006 // during GVN optimization (4641526).
1007 //
1008 // Because we are setting the actual control input, factor in
1009 // the result from get_late_ctrl() so we respect any
1010 // anti-dependences. (6233005).
1011 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
1013 // Don't allow the control input to be a CFG splitting node.
1014 // Such nodes should only have ProjNodes as outs, e.g. IfNode
1015 // should only have IfTrueNode and IfFalseNode (4985384).
1016 x_ctrl = find_non_split_ctrl(x_ctrl);
1017 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1019 x->set_req(0, x_ctrl);
1020 }
1021 register_new_node(x, x_ctrl);
1023 // Some institutional knowledge is needed here: 'x' is
1024 // yanked because if the optimizer runs GVN on it all the
1025 // cloned x's will common up and undo this optimization and
1026 // be forced back in the loop. This is annoying because it
1027 // makes +VerifyOpto report false-positives on progress. I
1028 // tried setting control edges on the x's to force them to
1029 // not combine, but the matching gets worried when it tries
1030 // to fold a StoreP and an AddP together (as part of an
1031 // address expression) and the AddP and StoreP have
1032 // different controls.
1033 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x);
1034 }
1035 _igvn.remove_dead_node(n);
1036 }
1037 }
1038 }
1039 }
1041 // Check for Opaque2's who's loop has disappeared - who's input is in the
1042 // same loop nest as their output. Remove 'em, they are no longer useful.
1043 if( n_op == Op_Opaque2 &&
1044 n->in(1) != NULL &&
1045 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1046 _igvn.replace_node( n, n->in(1) );
1047 }
1048 }
1050 //------------------------------split_if_with_blocks---------------------------
1051 // Check for aggressive application of 'split-if' optimization,
1052 // using basic block level info.
1053 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) {
1054 Node *n = C->root();
1055 visited.set(n->_idx); // first, mark node as visited
1056 // Do pre-visit work for root
1057 n = split_if_with_blocks_pre( n );
1058 uint cnt = n->outcnt();
1059 uint i = 0;
1060 while (true) {
1061 // Visit all children
1062 if (i < cnt) {
1063 Node* use = n->raw_out(i);
1064 ++i;
1065 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1066 // Now do pre-visit work for this use
1067 use = split_if_with_blocks_pre( use );
1068 nstack.push(n, i); // Save parent and next use's index.
1069 n = use; // Process all children of current use.
1070 cnt = use->outcnt();
1071 i = 0;
1072 }
1073 }
1074 else {
1075 // All of n's children have been processed, complete post-processing.
1076 if (cnt != 0 && !n->is_Con()) {
1077 assert(has_node(n), "no dead nodes");
1078 split_if_with_blocks_post( n );
1079 }
1080 if (nstack.is_empty()) {
1081 // Finished all nodes on stack.
1082 break;
1083 }
1084 // Get saved parent node and next use's index. Visit the rest of uses.
1085 n = nstack.node();
1086 cnt = n->outcnt();
1087 i = nstack.index();
1088 nstack.pop();
1089 }
1090 }
1091 }
1094 //=============================================================================
1095 //
1096 // C L O N E A L O O P B O D Y
1097 //
1099 //------------------------------clone_iff--------------------------------------
1100 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1101 // "Nearly" because all Nodes have been cloned from the original in the loop,
1102 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1103 // through the Phi recursively, and return a Bool.
1104 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) {
1106 // Convert this Phi into a Phi merging Bools
1107 uint i;
1108 for( i = 1; i < phi->req(); i++ ) {
1109 Node *b = phi->in(i);
1110 if( b->is_Phi() ) {
1111 _igvn.replace_input_of(phi, i, clone_iff( b->as_Phi(), loop ));
1112 } else {
1113 assert( b->is_Bool(), "" );
1114 }
1115 }
1117 Node *sample_bool = phi->in(1);
1118 Node *sample_cmp = sample_bool->in(1);
1120 // Make Phis to merge the Cmp's inputs.
1121 PhiNode *phi1 = new (C) PhiNode( phi->in(0), Type::TOP );
1122 PhiNode *phi2 = new (C) PhiNode( phi->in(0), Type::TOP );
1123 for( i = 1; i < phi->req(); i++ ) {
1124 Node *n1 = phi->in(i)->in(1)->in(1);
1125 Node *n2 = phi->in(i)->in(1)->in(2);
1126 phi1->set_req( i, n1 );
1127 phi2->set_req( i, n2 );
1128 phi1->set_type( phi1->type()->meet_speculative(n1->bottom_type()));
1129 phi2->set_type( phi2->type()->meet_speculative(n2->bottom_type()));
1130 }
1131 // See if these Phis have been made before.
1132 // Register with optimizer
1133 Node *hit1 = _igvn.hash_find_insert(phi1);
1134 if( hit1 ) { // Hit, toss just made Phi
1135 _igvn.remove_dead_node(phi1); // Remove new phi
1136 assert( hit1->is_Phi(), "" );
1137 phi1 = (PhiNode*)hit1; // Use existing phi
1138 } else { // Miss
1139 _igvn.register_new_node_with_optimizer(phi1);
1140 }
1141 Node *hit2 = _igvn.hash_find_insert(phi2);
1142 if( hit2 ) { // Hit, toss just made Phi
1143 _igvn.remove_dead_node(phi2); // Remove new phi
1144 assert( hit2->is_Phi(), "" );
1145 phi2 = (PhiNode*)hit2; // Use existing phi
1146 } else { // Miss
1147 _igvn.register_new_node_with_optimizer(phi2);
1148 }
1149 // Register Phis with loop/block info
1150 set_ctrl(phi1, phi->in(0));
1151 set_ctrl(phi2, phi->in(0));
1152 // Make a new Cmp
1153 Node *cmp = sample_cmp->clone();
1154 cmp->set_req( 1, phi1 );
1155 cmp->set_req( 2, phi2 );
1156 _igvn.register_new_node_with_optimizer(cmp);
1157 set_ctrl(cmp, phi->in(0));
1159 // Make a new Bool
1160 Node *b = sample_bool->clone();
1161 b->set_req(1,cmp);
1162 _igvn.register_new_node_with_optimizer(b);
1163 set_ctrl(b, phi->in(0));
1165 assert( b->is_Bool(), "" );
1166 return (BoolNode*)b;
1167 }
1169 //------------------------------clone_bool-------------------------------------
1170 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1171 // "Nearly" because all Nodes have been cloned from the original in the loop,
1172 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1173 // through the Phi recursively, and return a Bool.
1174 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1175 uint i;
1176 // Convert this Phi into a Phi merging Bools
1177 for( i = 1; i < phi->req(); i++ ) {
1178 Node *b = phi->in(i);
1179 if( b->is_Phi() ) {
1180 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1181 } else {
1182 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1183 }
1184 }
1186 Node *sample_cmp = phi->in(1);
1188 // Make Phis to merge the Cmp's inputs.
1189 PhiNode *phi1 = new (C) PhiNode( phi->in(0), Type::TOP );
1190 PhiNode *phi2 = new (C) PhiNode( phi->in(0), Type::TOP );
1191 for( uint j = 1; j < phi->req(); j++ ) {
1192 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1193 Node *n1, *n2;
1194 if( cmp_top->is_Cmp() ) {
1195 n1 = cmp_top->in(1);
1196 n2 = cmp_top->in(2);
1197 } else {
1198 n1 = n2 = cmp_top;
1199 }
1200 phi1->set_req( j, n1 );
1201 phi2->set_req( j, n2 );
1202 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1203 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1204 }
1206 // See if these Phis have been made before.
1207 // Register with optimizer
1208 Node *hit1 = _igvn.hash_find_insert(phi1);
1209 if( hit1 ) { // Hit, toss just made Phi
1210 _igvn.remove_dead_node(phi1); // Remove new phi
1211 assert( hit1->is_Phi(), "" );
1212 phi1 = (PhiNode*)hit1; // Use existing phi
1213 } else { // Miss
1214 _igvn.register_new_node_with_optimizer(phi1);
1215 }
1216 Node *hit2 = _igvn.hash_find_insert(phi2);
1217 if( hit2 ) { // Hit, toss just made Phi
1218 _igvn.remove_dead_node(phi2); // Remove new phi
1219 assert( hit2->is_Phi(), "" );
1220 phi2 = (PhiNode*)hit2; // Use existing phi
1221 } else { // Miss
1222 _igvn.register_new_node_with_optimizer(phi2);
1223 }
1224 // Register Phis with loop/block info
1225 set_ctrl(phi1, phi->in(0));
1226 set_ctrl(phi2, phi->in(0));
1227 // Make a new Cmp
1228 Node *cmp = sample_cmp->clone();
1229 cmp->set_req( 1, phi1 );
1230 cmp->set_req( 2, phi2 );
1231 _igvn.register_new_node_with_optimizer(cmp);
1232 set_ctrl(cmp, phi->in(0));
1234 assert( cmp->is_Cmp(), "" );
1235 return (CmpNode*)cmp;
1236 }
1238 //------------------------------sink_use---------------------------------------
1239 // If 'use' was in the loop-exit block, it now needs to be sunk
1240 // below the post-loop merge point.
1241 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1242 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1243 set_ctrl(use, post_loop);
1244 for (DUIterator j = use->outs(); use->has_out(j); j++)
1245 sink_use(use->out(j), post_loop);
1246 }
1247 }
1249 //------------------------------clone_loop-------------------------------------
1250 //
1251 // C L O N E A L O O P B O D Y
1252 //
1253 // This is the basic building block of the loop optimizations. It clones an
1254 // entire loop body. It makes an old_new loop body mapping; with this mapping
1255 // you can find the new-loop equivalent to an old-loop node. All new-loop
1256 // nodes are exactly equal to their old-loop counterparts, all edges are the
1257 // same. All exits from the old-loop now have a RegionNode that merges the
1258 // equivalent new-loop path. This is true even for the normal "loop-exit"
1259 // condition. All uses of loop-invariant old-loop values now come from (one
1260 // or more) Phis that merge their new-loop equivalents.
1261 //
1262 // This operation leaves the graph in an illegal state: there are two valid
1263 // control edges coming from the loop pre-header to both loop bodies. I'll
1264 // definitely have to hack the graph after running this transform.
1265 //
1266 // From this building block I will further edit edges to perform loop peeling
1267 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1268 //
1269 // Parameter side_by_size_idom:
1270 // When side_by_size_idom is NULL, the dominator tree is constructed for
1271 // the clone loop to dominate the original. Used in construction of
1272 // pre-main-post loop sequence.
1273 // When nonnull, the clone and original are side-by-side, both are
1274 // dominated by the side_by_side_idom node. Used in construction of
1275 // unswitched loops.
1276 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1277 Node* side_by_side_idom) {
1279 // Step 1: Clone the loop body. Make the old->new mapping.
1280 uint i;
1281 for( i = 0; i < loop->_body.size(); i++ ) {
1282 Node *old = loop->_body.at(i);
1283 Node *nnn = old->clone();
1284 old_new.map( old->_idx, nnn );
1285 _igvn.register_new_node_with_optimizer(nnn);
1286 }
1289 // Step 2: Fix the edges in the new body. If the old input is outside the
1290 // loop use it. If the old input is INside the loop, use the corresponding
1291 // new node instead.
1292 for( i = 0; i < loop->_body.size(); i++ ) {
1293 Node *old = loop->_body.at(i);
1294 Node *nnn = old_new[old->_idx];
1295 // Fix CFG/Loop controlling the new node
1296 if (has_ctrl(old)) {
1297 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
1298 } else {
1299 set_loop(nnn, loop->_parent);
1300 if (old->outcnt() > 0) {
1301 set_idom( nnn, old_new[idom(old)->_idx], dd );
1302 }
1303 }
1304 // Correct edges to the new node
1305 for( uint j = 0; j < nnn->req(); j++ ) {
1306 Node *n = nnn->in(j);
1307 if( n ) {
1308 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
1309 if( loop->is_member( old_in_loop ) )
1310 nnn->set_req(j, old_new[n->_idx]);
1311 }
1312 }
1313 _igvn.hash_find_insert(nnn);
1314 }
1315 Node *newhead = old_new[loop->_head->_idx];
1316 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1319 // Step 3: Now fix control uses. Loop varying control uses have already
1320 // been fixed up (as part of all input edges in Step 2). Loop invariant
1321 // control uses must be either an IfFalse or an IfTrue. Make a merge
1322 // point to merge the old and new IfFalse/IfTrue nodes; make the use
1323 // refer to this.
1324 ResourceArea *area = Thread::current()->resource_area();
1325 Node_List worklist(area);
1326 uint new_counter = C->unique();
1327 for( i = 0; i < loop->_body.size(); i++ ) {
1328 Node* old = loop->_body.at(i);
1329 if( !old->is_CFG() ) continue;
1330 Node* nnn = old_new[old->_idx];
1332 // Copy uses to a worklist, so I can munge the def-use info
1333 // with impunity.
1334 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1335 worklist.push(old->fast_out(j));
1337 while( worklist.size() ) { // Visit all uses
1338 Node *use = worklist.pop();
1339 if (!has_node(use)) continue; // Ignore dead nodes
1340 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1341 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
1342 // Both OLD and USE are CFG nodes here.
1343 assert( use->is_Proj(), "" );
1345 // Clone the loop exit control projection
1346 Node *newuse = use->clone();
1347 newuse->set_req(0,nnn);
1348 _igvn.register_new_node_with_optimizer(newuse);
1349 set_loop(newuse, use_loop);
1350 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
1352 // We need a Region to merge the exit from the peeled body and the
1353 // exit from the old loop body.
1354 RegionNode *r = new (C) RegionNode(3);
1355 // Map the old use to the new merge point
1356 old_new.map( use->_idx, r );
1357 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
1358 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
1360 // The original user of 'use' uses 'r' instead.
1361 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
1362 Node* useuse = use->last_out(l);
1363 _igvn.rehash_node_delayed(useuse);
1364 uint uses_found = 0;
1365 if( useuse->in(0) == use ) {
1366 useuse->set_req(0, r);
1367 uses_found++;
1368 if( useuse->is_CFG() ) {
1369 assert( dom_depth(useuse) > dd_r, "" );
1370 set_idom(useuse, r, dom_depth(useuse));
1371 }
1372 }
1373 for( uint k = 1; k < useuse->req(); k++ ) {
1374 if( useuse->in(k) == use ) {
1375 useuse->set_req(k, r);
1376 uses_found++;
1377 }
1378 }
1379 l -= uses_found; // we deleted 1 or more copies of this edge
1380 }
1382 // Now finish up 'r'
1383 r->set_req( 1, newuse );
1384 r->set_req( 2, use );
1385 _igvn.register_new_node_with_optimizer(r);
1386 set_loop(r, use_loop);
1387 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
1388 } // End of if a loop-exit test
1389 }
1390 }
1392 // Step 4: If loop-invariant use is not control, it must be dominated by a
1393 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
1394 // there if needed. Make a Phi there merging old and new used values.
1395 Node_List *split_if_set = NULL;
1396 Node_List *split_bool_set = NULL;
1397 Node_List *split_cex_set = NULL;
1398 for( i = 0; i < loop->_body.size(); i++ ) {
1399 Node* old = loop->_body.at(i);
1400 Node* nnn = old_new[old->_idx];
1401 // Copy uses to a worklist, so I can munge the def-use info
1402 // with impunity.
1403 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1404 worklist.push(old->fast_out(j));
1406 while( worklist.size() ) {
1407 Node *use = worklist.pop();
1408 if (!has_node(use)) continue; // Ignore dead nodes
1409 if (use->in(0) == C->top()) continue;
1410 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1411 // Check for data-use outside of loop - at least one of OLD or USE
1412 // must not be a CFG node.
1413 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) {
1415 // If the Data use is an IF, that means we have an IF outside of the
1416 // loop that is switching on a condition that is set inside of the
1417 // loop. Happens if people set a loop-exit flag; then test the flag
1418 // in the loop to break the loop, then test is again outside of the
1419 // loop to determine which way the loop exited.
1420 // Loop predicate If node connects to Bool node through Opaque1 node.
1421 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use)) {
1422 // Since this code is highly unlikely, we lazily build the worklist
1423 // of such Nodes to go split.
1424 if( !split_if_set )
1425 split_if_set = new Node_List(area);
1426 split_if_set->push(use);
1427 }
1428 if( use->is_Bool() ) {
1429 if( !split_bool_set )
1430 split_bool_set = new Node_List(area);
1431 split_bool_set->push(use);
1432 }
1433 if( use->Opcode() == Op_CreateEx ) {
1434 if( !split_cex_set )
1435 split_cex_set = new Node_List(area);
1436 split_cex_set->push(use);
1437 }
1440 // Get "block" use is in
1441 uint idx = 0;
1442 while( use->in(idx) != old ) idx++;
1443 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1444 assert( !loop->is_member( get_loop( prev ) ), "" );
1445 Node *cfg = prev->_idx >= new_counter
1446 ? prev->in(2)
1447 : idom(prev);
1448 if( use->is_Phi() ) // Phi use is in prior block
1449 cfg = prev->in(idx); // NOT in block of Phi itself
1450 if (cfg->is_top()) { // Use is dead?
1451 _igvn.replace_input_of(use, idx, C->top());
1452 continue;
1453 }
1455 while( !loop->is_member( get_loop( cfg ) ) ) {
1456 prev = cfg;
1457 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1458 }
1459 // If the use occurs after merging several exits from the loop, then
1460 // old value must have dominated all those exits. Since the same old
1461 // value was used on all those exits we did not need a Phi at this
1462 // merge point. NOW we do need a Phi here. Each loop exit value
1463 // is now merged with the peeled body exit; each exit gets its own
1464 // private Phi and those Phis need to be merged here.
1465 Node *phi;
1466 if( prev->is_Region() ) {
1467 if( idx == 0 ) { // Updating control edge?
1468 phi = prev; // Just use existing control
1469 } else { // Else need a new Phi
1470 phi = PhiNode::make( prev, old );
1471 // Now recursively fix up the new uses of old!
1472 for( uint i = 1; i < prev->req(); i++ ) {
1473 worklist.push(phi); // Onto worklist once for each 'old' input
1474 }
1475 }
1476 } else {
1477 // Get new RegionNode merging old and new loop exits
1478 prev = old_new[prev->_idx];
1479 assert( prev, "just made this in step 7" );
1480 if( idx == 0 ) { // Updating control edge?
1481 phi = prev; // Just use existing control
1482 } else { // Else need a new Phi
1483 // Make a new Phi merging data values properly
1484 phi = PhiNode::make( prev, old );
1485 phi->set_req( 1, nnn );
1486 }
1487 }
1488 // If inserting a new Phi, check for prior hits
1489 if( idx != 0 ) {
1490 Node *hit = _igvn.hash_find_insert(phi);
1491 if( hit == NULL ) {
1492 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1493 } else { // or
1494 // Remove the new phi from the graph and use the hit
1495 _igvn.remove_dead_node(phi);
1496 phi = hit; // Use existing phi
1497 }
1498 set_ctrl(phi, prev);
1499 }
1500 // Make 'use' use the Phi instead of the old loop body exit value
1501 _igvn.replace_input_of(use, idx, phi);
1502 if( use->_idx >= new_counter ) { // If updating new phis
1503 // Not needed for correctness, but prevents a weak assert
1504 // in AddPNode from tripping (when we end up with different
1505 // base & derived Phis that will become the same after
1506 // IGVN does CSE).
1507 Node *hit = _igvn.hash_find_insert(use);
1508 if( hit ) // Go ahead and re-hash for hits.
1509 _igvn.replace_node( use, hit );
1510 }
1512 // If 'use' was in the loop-exit block, it now needs to be sunk
1513 // below the post-loop merge point.
1514 sink_use( use, prev );
1515 }
1516 }
1517 }
1519 // Check for IFs that need splitting/cloning. Happens if an IF outside of
1520 // the loop uses a condition set in the loop. The original IF probably
1521 // takes control from one or more OLD Regions (which in turn get from NEW
1522 // Regions). In any case, there will be a set of Phis for each merge point
1523 // from the IF up to where the original BOOL def exists the loop.
1524 if( split_if_set ) {
1525 while( split_if_set->size() ) {
1526 Node *iff = split_if_set->pop();
1527 if( iff->in(1)->is_Phi() ) {
1528 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop );
1529 _igvn.replace_input_of(iff, 1, b);
1530 }
1531 }
1532 }
1533 if( split_bool_set ) {
1534 while( split_bool_set->size() ) {
1535 Node *b = split_bool_set->pop();
1536 Node *phi = b->in(1);
1537 assert( phi->is_Phi(), "" );
1538 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop );
1539 _igvn.replace_input_of(b, 1, cmp);
1540 }
1541 }
1542 if( split_cex_set ) {
1543 while( split_cex_set->size() ) {
1544 Node *b = split_cex_set->pop();
1545 assert( b->in(0)->is_Region(), "" );
1546 assert( b->in(1)->is_Phi(), "" );
1547 assert( b->in(0)->in(0) == b->in(1)->in(0), "" );
1548 split_up( b, b->in(0), NULL );
1549 }
1550 }
1552 }
1555 //---------------------- stride_of_possible_iv -------------------------------------
1556 // Looks for an iff/bool/comp with one operand of the compare
1557 // being a cycle involving an add and a phi,
1558 // with an optional truncation (left-shift followed by a right-shift)
1559 // of the add. Returns zero if not an iv.
1560 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
1561 Node* trunc1 = NULL;
1562 Node* trunc2 = NULL;
1563 const TypeInt* ttype = NULL;
1564 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
1565 return 0;
1566 }
1567 BoolNode* bl = iff->in(1)->as_Bool();
1568 Node* cmp = bl->in(1);
1569 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) {
1570 return 0;
1571 }
1572 // Must have an invariant operand
1573 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
1574 return 0;
1575 }
1576 Node* add2 = NULL;
1577 Node* cmp1 = cmp->in(1);
1578 if (cmp1->is_Phi()) {
1579 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
1580 Node* phi = cmp1;
1581 for (uint i = 1; i < phi->req(); i++) {
1582 Node* in = phi->in(i);
1583 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
1584 &trunc1, &trunc2, &ttype);
1585 if (add && add->in(1) == phi) {
1586 add2 = add->in(2);
1587 break;
1588 }
1589 }
1590 } else {
1591 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
1592 Node* addtrunc = cmp1;
1593 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
1594 &trunc1, &trunc2, &ttype);
1595 if (add && add->in(1)->is_Phi()) {
1596 Node* phi = add->in(1);
1597 for (uint i = 1; i < phi->req(); i++) {
1598 if (phi->in(i) == addtrunc) {
1599 add2 = add->in(2);
1600 break;
1601 }
1602 }
1603 }
1604 }
1605 if (add2 != NULL) {
1606 const TypeInt* add2t = _igvn.type(add2)->is_int();
1607 if (add2t->is_con()) {
1608 return add2t->get_con();
1609 }
1610 }
1611 return 0;
1612 }
1615 //---------------------- stay_in_loop -------------------------------------
1616 // Return the (unique) control output node that's in the loop (if it exists.)
1617 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
1618 Node* unique = NULL;
1619 if (!n) return NULL;
1620 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1621 Node* use = n->fast_out(i);
1622 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
1623 if (unique != NULL) {
1624 return NULL;
1625 }
1626 unique = use;
1627 }
1628 }
1629 return unique;
1630 }
1632 //------------------------------ register_node -------------------------------------
1633 // Utility to register node "n" with PhaseIdealLoop
1634 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
1635 _igvn.register_new_node_with_optimizer(n);
1636 loop->_body.push(n);
1637 if (n->is_CFG()) {
1638 set_loop(n, loop);
1639 set_idom(n, pred, ddepth);
1640 } else {
1641 set_ctrl(n, pred);
1642 }
1643 }
1645 //------------------------------ proj_clone -------------------------------------
1646 // Utility to create an if-projection
1647 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
1648 ProjNode* c = p->clone()->as_Proj();
1649 c->set_req(0, iff);
1650 return c;
1651 }
1653 //------------------------------ short_circuit_if -------------------------------------
1654 // Force the iff control output to be the live_proj
1655 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
1656 guarantee(live_proj != NULL, "null projection");
1657 int proj_con = live_proj->_con;
1658 assert(proj_con == 0 || proj_con == 1, "false or true projection");
1659 Node *con = _igvn.intcon(proj_con);
1660 set_ctrl(con, C->root());
1661 if (iff) {
1662 iff->set_req(1, con);
1663 }
1664 return con;
1665 }
1667 //------------------------------ insert_if_before_proj -------------------------------------
1668 // Insert a new if before an if projection (* - new node)
1669 //
1670 // before
1671 // if(test)
1672 // / \
1673 // v v
1674 // other-proj proj (arg)
1675 //
1676 // after
1677 // if(test)
1678 // / \
1679 // / v
1680 // | * proj-clone
1681 // v |
1682 // other-proj v
1683 // * new_if(relop(cmp[IU](left,right)))
1684 // / \
1685 // v v
1686 // * new-proj proj
1687 // (returned)
1688 //
1689 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
1690 IfNode* iff = proj->in(0)->as_If();
1691 IdealLoopTree *loop = get_loop(proj);
1692 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1693 int ddepth = dom_depth(proj);
1695 _igvn.rehash_node_delayed(iff);
1696 _igvn.rehash_node_delayed(proj);
1698 proj->set_req(0, NULL); // temporary disconnect
1699 ProjNode* proj2 = proj_clone(proj, iff);
1700 register_node(proj2, loop, iff, ddepth);
1702 Node* cmp = Signed ? (Node*) new (C)CmpINode(left, right) : (Node*) new (C)CmpUNode(left, right);
1703 register_node(cmp, loop, proj2, ddepth);
1705 BoolNode* bol = new (C)BoolNode(cmp, relop);
1706 register_node(bol, loop, proj2, ddepth);
1708 IfNode* new_if = new (C)IfNode(proj2, bol, iff->_prob, iff->_fcnt);
1709 register_node(new_if, loop, proj2, ddepth);
1711 proj->set_req(0, new_if); // reattach
1712 set_idom(proj, new_if, ddepth);
1714 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
1715 guarantee(new_exit != NULL, "null exit node");
1716 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
1718 return new_exit;
1719 }
1721 //------------------------------ insert_region_before_proj -------------------------------------
1722 // Insert a region before an if projection (* - new node)
1723 //
1724 // before
1725 // if(test)
1726 // / |
1727 // v |
1728 // proj v
1729 // other-proj
1730 //
1731 // after
1732 // if(test)
1733 // / |
1734 // v |
1735 // * proj-clone v
1736 // | other-proj
1737 // v
1738 // * new-region
1739 // |
1740 // v
1741 // * dum_if
1742 // / \
1743 // v \
1744 // * dum-proj v
1745 // proj
1746 //
1747 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
1748 IfNode* iff = proj->in(0)->as_If();
1749 IdealLoopTree *loop = get_loop(proj);
1750 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1751 int ddepth = dom_depth(proj);
1753 _igvn.rehash_node_delayed(iff);
1754 _igvn.rehash_node_delayed(proj);
1756 proj->set_req(0, NULL); // temporary disconnect
1757 ProjNode* proj2 = proj_clone(proj, iff);
1758 register_node(proj2, loop, iff, ddepth);
1760 RegionNode* reg = new (C)RegionNode(2);
1761 reg->set_req(1, proj2);
1762 register_node(reg, loop, iff, ddepth);
1764 IfNode* dum_if = new (C)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
1765 register_node(dum_if, loop, reg, ddepth);
1767 proj->set_req(0, dum_if); // reattach
1768 set_idom(proj, dum_if, ddepth);
1770 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
1771 register_node(dum_proj, loop, dum_if, ddepth);
1773 return reg;
1774 }
1776 //------------------------------ insert_cmpi_loop_exit -------------------------------------
1777 // Clone a signed compare loop exit from an unsigned compare and
1778 // insert it before the unsigned cmp on the stay-in-loop path.
1779 // All new nodes inserted in the dominator tree between the original
1780 // if and it's projections. The original if test is replaced with
1781 // a constant to force the stay-in-loop path.
1782 //
1783 // This is done to make sure that the original if and it's projections
1784 // still dominate the same set of control nodes, that the ctrl() relation
1785 // from data nodes to them is preserved, and that their loop nesting is
1786 // preserved.
1787 //
1788 // before
1789 // if(i <u limit) unsigned compare loop exit
1790 // / |
1791 // v v
1792 // exit-proj stay-in-loop-proj
1793 //
1794 // after
1795 // if(stay-in-loop-const) original if
1796 // / |
1797 // / v
1798 // / if(i < limit) new signed test
1799 // / / |
1800 // / / v
1801 // / / if(i <u limit) new cloned unsigned test
1802 // / / / |
1803 // v v v |
1804 // region |
1805 // | |
1806 // dum-if |
1807 // / | |
1808 // ether | |
1809 // v v
1810 // exit-proj stay-in-loop-proj
1811 //
1812 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
1813 const bool Signed = true;
1814 const bool Unsigned = false;
1816 BoolNode* bol = if_cmpu->in(1)->as_Bool();
1817 if (bol->_test._test != BoolTest::lt) return NULL;
1818 CmpNode* cmpu = bol->in(1)->as_Cmp();
1819 if (cmpu->Opcode() != Op_CmpU) return NULL;
1820 int stride = stride_of_possible_iv(if_cmpu);
1821 if (stride == 0) return NULL;
1823 Node* lp_proj = stay_in_loop(if_cmpu, loop);
1824 guarantee(lp_proj != NULL, "null loop node");
1826 ProjNode* lp_continue = lp_proj->as_Proj();
1827 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
1829 Node* limit = NULL;
1830 if (stride > 0) {
1831 limit = cmpu->in(2);
1832 } else {
1833 limit = _igvn.makecon(TypeInt::ZERO);
1834 set_ctrl(limit, C->root());
1835 }
1836 // Create a new region on the exit path
1837 RegionNode* reg = insert_region_before_proj(lp_exit);
1838 guarantee(reg != NULL, "null region node");
1840 // Clone the if-cmpu-true-false using a signed compare
1841 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
1842 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
1843 reg->add_req(cmpi_exit);
1845 // Clone the if-cmpu-true-false
1846 BoolTest::mask rel_u = bol->_test._test;
1847 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
1848 reg->add_req(cmpu_exit);
1850 // Force original if to stay in loop.
1851 short_circuit_if(if_cmpu, lp_continue);
1853 return cmpi_exit->in(0)->as_If();
1854 }
1856 //------------------------------ remove_cmpi_loop_exit -------------------------------------
1857 // Remove a previously inserted signed compare loop exit.
1858 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
1859 Node* lp_proj = stay_in_loop(if_cmp, loop);
1860 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
1861 stay_in_loop(lp_proj, loop)->is_If() &&
1862 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
1863 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
1864 set_ctrl(con, C->root());
1865 if_cmp->set_req(1, con);
1866 }
1868 //------------------------------ scheduled_nodelist -------------------------------------
1869 // Create a post order schedule of nodes that are in the
1870 // "member" set. The list is returned in "sched".
1871 // The first node in "sched" is the loop head, followed by
1872 // nodes which have no inputs in the "member" set, and then
1873 // followed by the nodes that have an immediate input dependence
1874 // on a node in "sched".
1875 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
1877 assert(member.test(loop->_head->_idx), "loop head must be in member set");
1878 Arena *a = Thread::current()->resource_area();
1879 VectorSet visited(a);
1880 Node_Stack nstack(a, loop->_body.size());
1882 Node* n = loop->_head; // top of stack is cached in "n"
1883 uint idx = 0;
1884 visited.set(n->_idx);
1886 // Initially push all with no inputs from within member set
1887 for(uint i = 0; i < loop->_body.size(); i++ ) {
1888 Node *elt = loop->_body.at(i);
1889 if (member.test(elt->_idx)) {
1890 bool found = false;
1891 for (uint j = 0; j < elt->req(); j++) {
1892 Node* def = elt->in(j);
1893 if (def && member.test(def->_idx) && def != elt) {
1894 found = true;
1895 break;
1896 }
1897 }
1898 if (!found && elt != loop->_head) {
1899 nstack.push(n, idx);
1900 n = elt;
1901 assert(!visited.test(n->_idx), "not seen yet");
1902 visited.set(n->_idx);
1903 }
1904 }
1905 }
1907 // traverse out's that are in the member set
1908 while (true) {
1909 if (idx < n->outcnt()) {
1910 Node* use = n->raw_out(idx);
1911 idx++;
1912 if (!visited.test_set(use->_idx)) {
1913 if (member.test(use->_idx)) {
1914 nstack.push(n, idx);
1915 n = use;
1916 idx = 0;
1917 }
1918 }
1919 } else {
1920 // All outputs processed
1921 sched.push(n);
1922 if (nstack.is_empty()) break;
1923 n = nstack.node();
1924 idx = nstack.index();
1925 nstack.pop();
1926 }
1927 }
1928 }
1931 //------------------------------ has_use_in_set -------------------------------------
1932 // Has a use in the vector set
1933 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
1934 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1935 Node* use = n->fast_out(j);
1936 if (vset.test(use->_idx)) {
1937 return true;
1938 }
1939 }
1940 return false;
1941 }
1944 //------------------------------ has_use_internal_to_set -------------------------------------
1945 // Has use internal to the vector set (ie. not in a phi at the loop head)
1946 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
1947 Node* head = loop->_head;
1948 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1949 Node* use = n->fast_out(j);
1950 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
1951 return true;
1952 }
1953 }
1954 return false;
1955 }
1958 //------------------------------ clone_for_use_outside_loop -------------------------------------
1959 // clone "n" for uses that are outside of loop
1960 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
1961 int cloned = 0;
1962 assert(worklist.size() == 0, "should be empty");
1963 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1964 Node* use = n->fast_out(j);
1965 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
1966 worklist.push(use);
1967 }
1968 }
1969 while( worklist.size() ) {
1970 Node *use = worklist.pop();
1971 if (!has_node(use) || use->in(0) == C->top()) continue;
1972 uint j;
1973 for (j = 0; j < use->req(); j++) {
1974 if (use->in(j) == n) break;
1975 }
1976 assert(j < use->req(), "must be there");
1978 // clone "n" and insert it between the inputs of "n" and the use outside the loop
1979 Node* n_clone = n->clone();
1980 _igvn.replace_input_of(use, j, n_clone);
1981 cloned++;
1982 Node* use_c;
1983 if (!use->is_Phi()) {
1984 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
1985 } else {
1986 // Use in a phi is considered a use in the associated predecessor block
1987 use_c = use->in(0)->in(j);
1988 }
1989 set_ctrl(n_clone, use_c);
1990 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
1991 get_loop(use_c)->_body.push(n_clone);
1992 _igvn.register_new_node_with_optimizer(n_clone);
1993 #if !defined(PRODUCT)
1994 if (TracePartialPeeling) {
1995 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
1996 }
1997 #endif
1998 }
1999 return cloned;
2000 }
2003 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
2004 // clone "n" for special uses that are in the not_peeled region.
2005 // If these def-uses occur in separate blocks, the code generator
2006 // marks the method as not compilable. For example, if a "BoolNode"
2007 // is in a different basic block than the "IfNode" that uses it, then
2008 // the compilation is aborted in the code generator.
2009 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
2010 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
2011 if (n->is_Phi() || n->is_Load()) {
2012 return;
2013 }
2014 assert(worklist.size() == 0, "should be empty");
2015 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2016 Node* use = n->fast_out(j);
2017 if ( not_peel.test(use->_idx) &&
2018 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
2019 use->in(1) == n) {
2020 worklist.push(use);
2021 }
2022 }
2023 if (worklist.size() > 0) {
2024 // clone "n" and insert it between inputs of "n" and the use
2025 Node* n_clone = n->clone();
2026 loop->_body.push(n_clone);
2027 _igvn.register_new_node_with_optimizer(n_clone);
2028 set_ctrl(n_clone, get_ctrl(n));
2029 sink_list.push(n_clone);
2030 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
2031 #if !defined(PRODUCT)
2032 if (TracePartialPeeling) {
2033 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2034 }
2035 #endif
2036 while( worklist.size() ) {
2037 Node *use = worklist.pop();
2038 _igvn.rehash_node_delayed(use);
2039 for (uint j = 1; j < use->req(); j++) {
2040 if (use->in(j) == n) {
2041 use->set_req(j, n_clone);
2042 }
2043 }
2044 }
2045 }
2046 }
2049 //------------------------------ insert_phi_for_loop -------------------------------------
2050 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2051 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2052 Node *phi = PhiNode::make(lp, back_edge_val);
2053 phi->set_req(LoopNode::EntryControl, lp_entry_val);
2054 // Use existing phi if it already exists
2055 Node *hit = _igvn.hash_find_insert(phi);
2056 if( hit == NULL ) {
2057 _igvn.register_new_node_with_optimizer(phi);
2058 set_ctrl(phi, lp);
2059 } else {
2060 // Remove the new phi from the graph and use the hit
2061 _igvn.remove_dead_node(phi);
2062 phi = hit;
2063 }
2064 _igvn.replace_input_of(use, idx, phi);
2065 }
2067 #ifdef ASSERT
2068 //------------------------------ is_valid_loop_partition -------------------------------------
2069 // Validate the loop partition sets: peel and not_peel
2070 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2071 VectorSet& not_peel ) {
2072 uint i;
2073 // Check that peel_list entries are in the peel set
2074 for (i = 0; i < peel_list.size(); i++) {
2075 if (!peel.test(peel_list.at(i)->_idx)) {
2076 return false;
2077 }
2078 }
2079 // Check at loop members are in one of peel set or not_peel set
2080 for (i = 0; i < loop->_body.size(); i++ ) {
2081 Node *def = loop->_body.at(i);
2082 uint di = def->_idx;
2083 // Check that peel set elements are in peel_list
2084 if (peel.test(di)) {
2085 if (not_peel.test(di)) {
2086 return false;
2087 }
2088 // Must be in peel_list also
2089 bool found = false;
2090 for (uint j = 0; j < peel_list.size(); j++) {
2091 if (peel_list.at(j)->_idx == di) {
2092 found = true;
2093 break;
2094 }
2095 }
2096 if (!found) {
2097 return false;
2098 }
2099 } else if (not_peel.test(di)) {
2100 if (peel.test(di)) {
2101 return false;
2102 }
2103 } else {
2104 return false;
2105 }
2106 }
2107 return true;
2108 }
2110 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2111 // Ensure a use outside of loop is of the right form
2112 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2113 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2114 return (use->is_Phi() &&
2115 use_c->is_Region() && use_c->req() == 3 &&
2116 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2117 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2118 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2119 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2120 }
2122 //------------------------------ is_valid_clone_loop_form -------------------------------------
2123 // Ensure that all uses outside of loop are of the right form
2124 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2125 uint orig_exit_idx, uint clone_exit_idx) {
2126 uint len = peel_list.size();
2127 for (uint i = 0; i < len; i++) {
2128 Node *def = peel_list.at(i);
2130 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2131 Node *use = def->fast_out(j);
2132 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2133 if (!loop->is_member(get_loop(use_c))) {
2134 // use is not in the loop, check for correct structure
2135 if (use->in(0) == def) {
2136 // Okay
2137 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2138 return false;
2139 }
2140 }
2141 }
2142 }
2143 return true;
2144 }
2145 #endif
2147 //------------------------------ partial_peel -------------------------------------
2148 // Partially peel (aka loop rotation) the top portion of a loop (called
2149 // the peel section below) by cloning it and placing one copy just before
2150 // the new loop head and the other copy at the bottom of the new loop.
2151 //
2152 // before after where it came from
2153 //
2154 // stmt1 stmt1
2155 // loop: stmt2 clone
2156 // stmt2 if condA goto exitA clone
2157 // if condA goto exitA new_loop: new
2158 // stmt3 stmt3 clone
2159 // if !condB goto loop if condB goto exitB clone
2160 // exitB: stmt2 orig
2161 // stmt4 if !condA goto new_loop orig
2162 // exitA: goto exitA
2163 // exitB:
2164 // stmt4
2165 // exitA:
2166 //
2167 // Step 1: find the cut point: an exit test on probable
2168 // induction variable.
2169 // Step 2: schedule (with cloning) operations in the peel
2170 // section that can be executed after the cut into
2171 // the section that is not peeled. This may need
2172 // to clone operations into exit blocks. For
2173 // instance, a reference to A[i] in the not-peel
2174 // section and a reference to B[i] in an exit block
2175 // may cause a left-shift of i by 2 to be placed
2176 // in the peel block. This step will clone the left
2177 // shift into the exit block and sink the left shift
2178 // from the peel to the not-peel section.
2179 // Step 3: clone the loop, retarget the control, and insert
2180 // phis for values that are live across the new loop
2181 // head. This is very dependent on the graph structure
2182 // from clone_loop. It creates region nodes for
2183 // exit control and associated phi nodes for values
2184 // flow out of the loop through that exit. The region
2185 // node is dominated by the clone's control projection.
2186 // So the clone's peel section is placed before the
2187 // new loop head, and the clone's not-peel section is
2188 // forms the top part of the new loop. The original
2189 // peel section forms the tail of the new loop.
2190 // Step 4: update the dominator tree and recompute the
2191 // dominator depth.
2192 //
2193 // orig
2194 //
2195 // stmt1
2196 // |
2197 // v
2198 // loop predicate
2199 // |
2200 // v
2201 // loop<----+
2202 // | |
2203 // stmt2 |
2204 // | |
2205 // v |
2206 // ifA |
2207 // / | |
2208 // v v |
2209 // false true ^ <-- last_peel
2210 // / | |
2211 // / ===|==cut |
2212 // / stmt3 | <-- first_not_peel
2213 // / | |
2214 // | v |
2215 // v ifB |
2216 // exitA: / \ |
2217 // / \ |
2218 // v v |
2219 // false true |
2220 // / \ |
2221 // / ----+
2222 // |
2223 // v
2224 // exitB:
2225 // stmt4
2226 //
2227 //
2228 // after clone loop
2229 //
2230 // stmt1
2231 // |
2232 // v
2233 // loop predicate
2234 // / \
2235 // clone / \ orig
2236 // / \
2237 // / \
2238 // v v
2239 // +---->loop loop<----+
2240 // | | | |
2241 // | stmt2 stmt2 |
2242 // | | | |
2243 // | v v |
2244 // | ifA ifA |
2245 // | | \ / | |
2246 // | v v v v |
2247 // ^ true false false true ^ <-- last_peel
2248 // | | ^ \ / | |
2249 // | cut==|== \ \ / ===|==cut |
2250 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2251 // | | dom | | | |
2252 // | v \ 1v v2 v |
2253 // | ifB regionA ifB |
2254 // | / \ | / \ |
2255 // | / \ v / \ |
2256 // | v v exitA: v v |
2257 // | true false false true |
2258 // | / ^ \ / \ |
2259 // +---- \ \ / ----+
2260 // dom \ /
2261 // \ 1v v2
2262 // regionB
2263 // |
2264 // v
2265 // exitB:
2266 // stmt4
2267 //
2268 //
2269 // after partial peel
2270 //
2271 // stmt1
2272 // |
2273 // v
2274 // loop predicate
2275 // /
2276 // clone / orig
2277 // / TOP
2278 // / \
2279 // v v
2280 // TOP->loop loop----+
2281 // | | |
2282 // stmt2 stmt2 |
2283 // | | |
2284 // v v |
2285 // ifA ifA |
2286 // | \ / | |
2287 // v v v v |
2288 // true false false true | <-- last_peel
2289 // | ^ \ / +------|---+
2290 // +->newloop \ \ / === ==cut | |
2291 // | stmt3 \ \ / TOP | |
2292 // | | dom | | stmt3 | | <-- first_not_peel
2293 // | v \ 1v v2 v | |
2294 // | ifB regionA ifB ^ v
2295 // | / \ | / \ | |
2296 // | / \ v / \ | |
2297 // | v v exitA: v v | |
2298 // | true false false true | |
2299 // | / ^ \ / \ | |
2300 // | | \ \ / v | |
2301 // | | dom \ / TOP | |
2302 // | | \ 1v v2 | |
2303 // ^ v regionB | |
2304 // | | | | |
2305 // | | v ^ v
2306 // | | exitB: | |
2307 // | | stmt4 | |
2308 // | +------------>-----------------+ |
2309 // | |
2310 // +-----------------<---------------------+
2311 //
2312 //
2313 // final graph
2314 //
2315 // stmt1
2316 // |
2317 // v
2318 // loop predicate
2319 // |
2320 // v
2321 // stmt2 clone
2322 // |
2323 // v
2324 // ........> ifA clone
2325 // : / |
2326 // dom / |
2327 // : v v
2328 // : false true
2329 // : | |
2330 // : | v
2331 // : | newloop<-----+
2332 // : | | |
2333 // : | stmt3 clone |
2334 // : | | |
2335 // : | v |
2336 // : | ifB |
2337 // : | / \ |
2338 // : | v v |
2339 // : | false true |
2340 // : | | | |
2341 // : | v stmt2 |
2342 // : | exitB: | |
2343 // : | stmt4 v |
2344 // : | ifA orig |
2345 // : | / \ |
2346 // : | / \ |
2347 // : | v v |
2348 // : | false true |
2349 // : | / \ |
2350 // : v v -----+
2351 // RegionA
2352 // |
2353 // v
2354 // exitA
2355 //
2356 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
2358 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
2359 if (!loop->_head->is_Loop()) {
2360 return false; }
2362 LoopNode *head = loop->_head->as_Loop();
2364 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
2365 return false;
2366 }
2368 // Check for complex exit control
2369 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
2370 Node *n = loop->_body.at(ii);
2371 int opc = n->Opcode();
2372 if (n->is_Call() ||
2373 opc == Op_Catch ||
2374 opc == Op_CatchProj ||
2375 opc == Op_Jump ||
2376 opc == Op_JumpProj) {
2377 #if !defined(PRODUCT)
2378 if (TracePartialPeeling) {
2379 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
2380 }
2381 #endif
2382 return false;
2383 }
2384 }
2386 int dd = dom_depth(head);
2388 // Step 1: find cut point
2390 // Walk up dominators to loop head looking for first loop exit
2391 // which is executed on every path thru loop.
2392 IfNode *peel_if = NULL;
2393 IfNode *peel_if_cmpu = NULL;
2395 Node *iff = loop->tail();
2396 while( iff != head ) {
2397 if( iff->is_If() ) {
2398 Node *ctrl = get_ctrl(iff->in(1));
2399 if (ctrl->is_top()) return false; // Dead test on live IF.
2400 // If loop-varying exit-test, check for induction variable
2401 if( loop->is_member(get_loop(ctrl)) &&
2402 loop->is_loop_exit(iff) &&
2403 is_possible_iv_test(iff)) {
2404 Node* cmp = iff->in(1)->in(1);
2405 if (cmp->Opcode() == Op_CmpI) {
2406 peel_if = iff->as_If();
2407 } else {
2408 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
2409 peel_if_cmpu = iff->as_If();
2410 }
2411 }
2412 }
2413 iff = idom(iff);
2414 }
2415 // Prefer signed compare over unsigned compare.
2416 IfNode* new_peel_if = NULL;
2417 if (peel_if == NULL) {
2418 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
2419 return false; // No peel point found
2420 }
2421 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
2422 if (new_peel_if == NULL) {
2423 return false; // No peel point found
2424 }
2425 peel_if = new_peel_if;
2426 }
2427 Node* last_peel = stay_in_loop(peel_if, loop);
2428 Node* first_not_peeled = stay_in_loop(last_peel, loop);
2429 if (first_not_peeled == NULL || first_not_peeled == head) {
2430 return false;
2431 }
2433 #if !defined(PRODUCT)
2434 if (TraceLoopOpts) {
2435 tty->print("PartialPeel ");
2436 loop->dump_head();
2437 }
2439 if (TracePartialPeeling) {
2440 tty->print_cr("before partial peel one iteration");
2441 Node_List wl;
2442 Node* t = head->in(2);
2443 while (true) {
2444 wl.push(t);
2445 if (t == head) break;
2446 t = idom(t);
2447 }
2448 while (wl.size() > 0) {
2449 Node* tt = wl.pop();
2450 tt->dump();
2451 if (tt == last_peel) tty->print_cr("-- cut --");
2452 }
2453 }
2454 #endif
2455 ResourceArea *area = Thread::current()->resource_area();
2456 VectorSet peel(area);
2457 VectorSet not_peel(area);
2458 Node_List peel_list(area);
2459 Node_List worklist(area);
2460 Node_List sink_list(area);
2462 // Set of cfg nodes to peel are those that are executable from
2463 // the head through last_peel.
2464 assert(worklist.size() == 0, "should be empty");
2465 worklist.push(head);
2466 peel.set(head->_idx);
2467 while (worklist.size() > 0) {
2468 Node *n = worklist.pop();
2469 if (n != last_peel) {
2470 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2471 Node* use = n->fast_out(j);
2472 if (use->is_CFG() &&
2473 loop->is_member(get_loop(use)) &&
2474 !peel.test_set(use->_idx)) {
2475 worklist.push(use);
2476 }
2477 }
2478 }
2479 }
2481 // Set of non-cfg nodes to peel are those that are control
2482 // dependent on the cfg nodes.
2483 uint i;
2484 for(i = 0; i < loop->_body.size(); i++ ) {
2485 Node *n = loop->_body.at(i);
2486 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
2487 if (peel.test(n_c->_idx)) {
2488 peel.set(n->_idx);
2489 } else {
2490 not_peel.set(n->_idx);
2491 }
2492 }
2494 // Step 2: move operations from the peeled section down into the
2495 // not-peeled section
2497 // Get a post order schedule of nodes in the peel region
2498 // Result in right-most operand.
2499 scheduled_nodelist(loop, peel, peel_list );
2501 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2503 // For future check for too many new phis
2504 uint old_phi_cnt = 0;
2505 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2506 Node* use = head->fast_out(j);
2507 if (use->is_Phi()) old_phi_cnt++;
2508 }
2510 #if !defined(PRODUCT)
2511 if (TracePartialPeeling) {
2512 tty->print_cr("\npeeled list");
2513 }
2514 #endif
2516 // Evacuate nodes in peel region into the not_peeled region if possible
2517 uint new_phi_cnt = 0;
2518 uint cloned_for_outside_use = 0;
2519 for (i = 0; i < peel_list.size();) {
2520 Node* n = peel_list.at(i);
2521 #if !defined(PRODUCT)
2522 if (TracePartialPeeling) n->dump();
2523 #endif
2524 bool incr = true;
2525 if ( !n->is_CFG() ) {
2527 if ( has_use_in_set(n, not_peel) ) {
2529 // If not used internal to the peeled region,
2530 // move "n" from peeled to not_peeled region.
2532 if ( !has_use_internal_to_set(n, peel, loop) ) {
2534 // if not pinned and not a load (which maybe anti-dependent on a store)
2535 // and not a CMove (Matcher expects only bool->cmove).
2536 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
2537 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist );
2538 sink_list.push(n);
2539 peel >>= n->_idx; // delete n from peel set.
2540 not_peel <<= n->_idx; // add n to not_peel set.
2541 peel_list.remove(i);
2542 incr = false;
2543 #if !defined(PRODUCT)
2544 if (TracePartialPeeling) {
2545 tty->print_cr("sink to not_peeled region: %d newbb: %d",
2546 n->_idx, get_ctrl(n)->_idx);
2547 }
2548 #endif
2549 }
2550 } else {
2551 // Otherwise check for special def-use cases that span
2552 // the peel/not_peel boundary such as bool->if
2553 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
2554 new_phi_cnt++;
2555 }
2556 }
2557 }
2558 if (incr) i++;
2559 }
2561 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
2562 #if !defined(PRODUCT)
2563 if (TracePartialPeeling) {
2564 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
2565 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
2566 }
2567 #endif
2568 if (new_peel_if != NULL) {
2569 remove_cmpi_loop_exit(new_peel_if, loop);
2570 }
2571 // Inhibit more partial peeling on this loop
2572 assert(!head->is_partial_peel_loop(), "not partial peeled");
2573 head->mark_partial_peel_failed();
2574 if (cloned_for_outside_use > 0) {
2575 // Terminate this round of loop opts because
2576 // the graph outside this loop was changed.
2577 C->set_major_progress();
2578 return true;
2579 }
2580 return false;
2581 }
2583 // Step 3: clone loop, retarget control, and insert new phis
2585 // Create new loop head for new phis and to hang
2586 // the nodes being moved (sinked) from the peel region.
2587 LoopNode* new_head = new (C) LoopNode(last_peel, last_peel);
2588 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
2589 _igvn.register_new_node_with_optimizer(new_head);
2590 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
2591 first_not_peeled->set_req(0, new_head);
2592 set_loop(new_head, loop);
2593 loop->_body.push(new_head);
2594 not_peel.set(new_head->_idx);
2595 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
2596 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
2598 while (sink_list.size() > 0) {
2599 Node* n = sink_list.pop();
2600 set_ctrl(n, new_head);
2601 }
2603 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2605 clone_loop( loop, old_new, dd );
2607 const uint clone_exit_idx = 1;
2608 const uint orig_exit_idx = 2;
2609 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
2611 Node* head_clone = old_new[head->_idx];
2612 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
2613 Node* orig_tail_clone = head_clone->in(2);
2615 // Add phi if "def" node is in peel set and "use" is not
2617 for(i = 0; i < peel_list.size(); i++ ) {
2618 Node *def = peel_list.at(i);
2619 if (!def->is_CFG()) {
2620 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2621 Node *use = def->fast_out(j);
2622 if (has_node(use) && use->in(0) != C->top() &&
2623 (!peel.test(use->_idx) ||
2624 (use->is_Phi() && use->in(0) == head)) ) {
2625 worklist.push(use);
2626 }
2627 }
2628 while( worklist.size() ) {
2629 Node *use = worklist.pop();
2630 for (uint j = 1; j < use->req(); j++) {
2631 Node* n = use->in(j);
2632 if (n == def) {
2634 // "def" is in peel set, "use" is not in peel set
2635 // or "use" is in the entry boundary (a phi) of the peel set
2637 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
2639 if ( loop->is_member(get_loop( use_c )) ) {
2640 // use is in loop
2641 if (old_new[use->_idx] != NULL) { // null for dead code
2642 Node* use_clone = old_new[use->_idx];
2643 _igvn.replace_input_of(use, j, C->top());
2644 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
2645 }
2646 } else {
2647 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
2648 // use is not in the loop, check if the live range includes the cut
2649 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
2650 if (not_peel.test(lp_if->_idx)) {
2651 assert(j == orig_exit_idx, "use from original loop");
2652 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
2653 }
2654 }
2655 }
2656 }
2657 }
2658 }
2659 }
2661 // Step 3b: retarget control
2663 // Redirect control to the new loop head if a cloned node in
2664 // the not_peeled region has control that points into the peeled region.
2665 // This necessary because the cloned peeled region will be outside
2666 // the loop.
2667 // from to
2668 // cloned-peeled <---+
2669 // new_head_clone: | <--+
2670 // cloned-not_peeled in(0) in(0)
2671 // orig-peeled
2673 for(i = 0; i < loop->_body.size(); i++ ) {
2674 Node *n = loop->_body.at(i);
2675 if (!n->is_CFG() && n->in(0) != NULL &&
2676 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
2677 Node* n_clone = old_new[n->_idx];
2678 _igvn.replace_input_of(n_clone, 0, new_head_clone);
2679 }
2680 }
2682 // Backedge of the surviving new_head (the clone) is original last_peel
2683 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
2685 // Cut first node in original not_peel set
2686 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
2687 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
2688 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
2690 // Copy head_clone back-branch info to original head
2691 // and remove original head's loop entry and
2692 // clone head's back-branch
2693 _igvn.rehash_node_delayed(head); // Multiple edge updates
2694 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
2695 head->set_req(LoopNode::LoopBackControl, C->top());
2696 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
2698 // Similarly modify the phis
2699 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
2700 Node* use = head->fast_out(k);
2701 if (use->is_Phi() && use->outcnt() > 0) {
2702 Node* use_clone = old_new[use->_idx];
2703 _igvn.rehash_node_delayed(use); // Multiple edge updates
2704 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
2705 use->set_req(LoopNode::LoopBackControl, C->top());
2706 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
2707 }
2708 }
2710 // Step 4: update dominator tree and dominator depth
2712 set_idom(head, orig_tail_clone, dd);
2713 recompute_dom_depth();
2715 // Inhibit more partial peeling on this loop
2716 new_head_clone->set_partial_peel_loop();
2717 C->set_major_progress();
2718 loop->record_for_igvn();
2720 #if !defined(PRODUCT)
2721 if (TracePartialPeeling) {
2722 tty->print_cr("\nafter partial peel one iteration");
2723 Node_List wl(area);
2724 Node* t = last_peel;
2725 while (true) {
2726 wl.push(t);
2727 if (t == head_clone) break;
2728 t = idom(t);
2729 }
2730 while (wl.size() > 0) {
2731 Node* tt = wl.pop();
2732 if (tt == head) tty->print_cr("orig head");
2733 else if (tt == new_head_clone) tty->print_cr("new head");
2734 else if (tt == head_clone) tty->print_cr("clone head");
2735 tt->dump();
2736 }
2737 }
2738 #endif
2739 return true;
2740 }
2742 //------------------------------reorg_offsets----------------------------------
2743 // Reorganize offset computations to lower register pressure. Mostly
2744 // prevent loop-fallout uses of the pre-incremented trip counter (which are
2745 // then alive with the post-incremented trip counter forcing an extra
2746 // register move)
2747 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
2748 // Perform it only for canonical counted loops.
2749 // Loop's shape could be messed up by iteration_split_impl.
2750 if (!loop->_head->is_CountedLoop())
2751 return;
2752 if (!loop->_head->as_Loop()->is_valid_counted_loop())
2753 return;
2755 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2756 CountedLoopEndNode *cle = cl->loopexit();
2757 Node *exit = cle->proj_out(false);
2758 Node *phi = cl->phi();
2760 // Check for the special case of folks using the pre-incremented
2761 // trip-counter on the fall-out path (forces the pre-incremented
2762 // and post-incremented trip counter to be live at the same time).
2763 // Fix this by adjusting to use the post-increment trip counter.
2765 bool progress = true;
2766 while (progress) {
2767 progress = false;
2768 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
2769 Node* use = phi->fast_out(i); // User of trip-counter
2770 if (!has_ctrl(use)) continue;
2771 Node *u_ctrl = get_ctrl(use);
2772 if (use->is_Phi()) {
2773 u_ctrl = NULL;
2774 for (uint j = 1; j < use->req(); j++)
2775 if (use->in(j) == phi)
2776 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
2777 }
2778 IdealLoopTree *u_loop = get_loop(u_ctrl);
2779 // Look for loop-invariant use
2780 if (u_loop == loop) continue;
2781 if (loop->is_member(u_loop)) continue;
2782 // Check that use is live out the bottom. Assuming the trip-counter
2783 // update is right at the bottom, uses of of the loop middle are ok.
2784 if (dom_lca(exit, u_ctrl) != exit) continue;
2785 // Hit! Refactor use to use the post-incremented tripcounter.
2786 // Compute a post-increment tripcounter.
2787 Node *opaq = new (C) Opaque2Node( C, cle->incr() );
2788 register_new_node(opaq, exit);
2789 Node *neg_stride = _igvn.intcon(-cle->stride_con());
2790 set_ctrl(neg_stride, C->root());
2791 Node *post = new (C) AddINode( opaq, neg_stride);
2792 register_new_node(post, exit);
2793 _igvn.rehash_node_delayed(use);
2794 for (uint j = 1; j < use->req(); j++) {
2795 if (use->in(j) == phi)
2796 use->set_req(j, post);
2797 }
2798 // Since DU info changed, rerun loop
2799 progress = true;
2800 break;
2801 }
2802 }
2804 }