1 /*
2 * Copyright (c) 1999, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/shared/barrierSet.hpp"
27 #include "gc/shared/c2/barrierSetC2.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "memory/resourceArea.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/connode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/divnode.hpp"
36 #include "opto/loopnode.hpp"
37 #include "opto/matcher.hpp"
38 #include "opto/mulnode.hpp"
39 #include "opto/movenode.hpp"
40 #include "opto/opaquenode.hpp"
41 #include "opto/rootnode.hpp"
42 #include "opto/subnode.hpp"
43 #include "opto/subtypenode.hpp"
44 #include "utilities/macros.hpp"
45
46 //=============================================================================
47 //------------------------------split_thru_phi---------------------------------
48 // Split Node 'n' through merge point if there is enough win.
split_thru_phi(Node * n,Node * region,int policy)49 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
50 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
51 // ConvI2L may have type information on it which is unsafe to push up
52 // so disable this for now
53 return NULL;
54 }
55
56 // Splitting range check CastIIs through a loop induction Phi can
57 // cause new Phis to be created that are left unrelated to the loop
58 // induction Phi and prevent optimizations (vectorization)
59 if (n->Opcode() == Op_CastII && region->is_CountedLoop() &&
60 n->in(1) == region->as_CountedLoop()->phi()) {
61 return NULL;
62 }
63
64 // Bail out if 'n' is a Div or Mod node whose zero check was removed earlier (i.e. control is NULL) and its divisor is an induction variable
65 // phi p of a trip-counted (integer) loop whose inputs could be zero (include zero in their type range). p could have a more precise type
66 // range that does not necessarily include all values of its inputs. Since each of these inputs will be a divisor of the newly cloned nodes
67 // of 'n', we need to bail out of one of these divisors could be zero (zero in its type range).
68 if ((n->Opcode() == Op_DivI || n->Opcode() == Op_ModI) && n->in(0) == NULL
69 && region->is_CountedLoop() && n->in(2) == region->as_CountedLoop()->phi()) {
70 Node* phi = region->as_CountedLoop()->phi();
71 for (uint i = 1; i < phi->req(); i++) {
72 if (_igvn.type(phi->in(i))->filter_speculative(TypeInt::ZERO) != Type::TOP) {
73 // Zero could be a possible value but we already removed the zero check. Bail out to avoid a possible division by zero at a later point.
74 return NULL;
75 }
76 }
77 }
78
79 int wins = 0;
80 assert(!n->is_CFG(), "");
81 assert(region->is_Region(), "");
82
83 const Type* type = n->bottom_type();
84 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr();
85 Node* phi;
86 if (t_oop != NULL && t_oop->is_known_instance_field()) {
87 int iid = t_oop->instance_id();
88 int index = C->get_alias_index(t_oop);
89 int offset = t_oop->offset();
90 phi = new PhiNode(region, type, NULL, iid, index, offset);
91 } else {
92 phi = PhiNode::make_blank(region, n);
93 }
94 uint old_unique = C->unique();
95 for (uint i = 1; i < region->req(); i++) {
96 Node* x;
97 Node* the_clone = NULL;
98 if (region->in(i) == C->top()) {
99 x = C->top(); // Dead path? Use a dead data op
100 } else {
101 x = n->clone(); // Else clone up the data op
102 the_clone = x; // Remember for possible deletion.
103 // Alter data node to use pre-phi inputs
104 if (n->in(0) == region)
105 x->set_req( 0, region->in(i) );
106 for (uint j = 1; j < n->req(); j++) {
107 Node* in = n->in(j);
108 if (in->is_Phi() && in->in(0) == region)
109 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
110 }
111 }
112 // Check for a 'win' on some paths
113 const Type* t = x->Value(&_igvn);
114
115 bool singleton = t->singleton();
116
117 // A TOP singleton indicates that there are no possible values incoming
118 // along a particular edge. In most cases, this is OK, and the Phi will
119 // be eliminated later in an Ideal call. However, we can't allow this to
120 // happen if the singleton occurs on loop entry, as the elimination of
121 // the PhiNode may cause the resulting node to migrate back to a previous
122 // loop iteration.
123 if (singleton && t == Type::TOP) {
124 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
125 // irreducible loop may not be indicated by an affirmative is_Loop());
126 // therefore, the only top we can split thru a phi is on a backedge of
127 // a loop.
128 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
129 }
130
131 if (singleton) {
132 wins++;
133 x = ((PhaseGVN&)_igvn).makecon(t);
134 } else {
135 // We now call Identity to try to simplify the cloned node.
136 // Note that some Identity methods call phase->type(this).
137 // Make sure that the type array is big enough for
138 // our new node, even though we may throw the node away.
139 // (Note: This tweaking with igvn only works because x is a new node.)
140 _igvn.set_type(x, t);
141 // If x is a TypeNode, capture any more-precise type permanently into Node
142 // otherwise it will be not updated during igvn->transform since
143 // igvn->type(x) is set to x->Value() already.
144 x->raise_bottom_type(t);
145 Node* y = x->Identity(&_igvn);
146 if (y != x) {
147 wins++;
148 x = y;
149 } else {
150 y = _igvn.hash_find(x);
151 if (y) {
152 wins++;
153 x = y;
154 } else {
155 // Else x is a new node we are keeping
156 // We do not need register_new_node_with_optimizer
157 // because set_type has already been called.
158 _igvn._worklist.push(x);
159 }
160 }
161 }
162 if (x != the_clone && the_clone != NULL)
163 _igvn.remove_dead_node(the_clone);
164 phi->set_req( i, x );
165 }
166 // Too few wins?
167 if (wins <= policy) {
168 _igvn.remove_dead_node(phi);
169 return NULL;
170 }
171
172 // Record Phi
173 register_new_node( phi, region );
174
175 for (uint i2 = 1; i2 < phi->req(); i2++) {
176 Node *x = phi->in(i2);
177 // If we commoned up the cloned 'x' with another existing Node,
178 // the existing Node picks up a new use. We need to make the
179 // existing Node occur higher up so it dominates its uses.
180 Node *old_ctrl;
181 IdealLoopTree *old_loop;
182
183 if (x->is_Con()) {
184 // Constant's control is always root.
185 set_ctrl(x, C->root());
186 continue;
187 }
188 // The occasional new node
189 if (x->_idx >= old_unique) { // Found a new, unplaced node?
190 old_ctrl = NULL;
191 old_loop = NULL; // Not in any prior loop
192 } else {
193 old_ctrl = get_ctrl(x);
194 old_loop = get_loop(old_ctrl); // Get prior loop
195 }
196 // New late point must dominate new use
197 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
198 if (new_ctrl == old_ctrl) // Nothing is changed
199 continue;
200
201 IdealLoopTree *new_loop = get_loop(new_ctrl);
202
203 // Don't move x into a loop if its uses are
204 // outside of loop. Otherwise x will be cloned
205 // for each use outside of this loop.
206 IdealLoopTree *use_loop = get_loop(region);
207 if (!new_loop->is_member(use_loop) &&
208 (old_loop == NULL || !new_loop->is_member(old_loop))) {
209 // Take early control, later control will be recalculated
210 // during next iteration of loop optimizations.
211 new_ctrl = get_early_ctrl(x);
212 new_loop = get_loop(new_ctrl);
213 }
214 // Set new location
215 set_ctrl(x, new_ctrl);
216 // If changing loop bodies, see if we need to collect into new body
217 if (old_loop != new_loop) {
218 if (old_loop && !old_loop->_child)
219 old_loop->_body.yank(x);
220 if (!new_loop->_child)
221 new_loop->_body.push(x); // Collect body info
222 }
223 }
224
225 return phi;
226 }
227
228 //------------------------------dominated_by------------------------------------
229 // Replace the dominated test with an obvious true or false. Place it on the
230 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
231 // live path up to the dominating control.
dominated_by(Node * prevdom,Node * iff,bool flip,bool exclude_loop_predicate)232 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) {
233 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
234
235 // prevdom is the dominating projection of the dominating test.
236 assert(iff->is_If(), "must be");
237 assert(iff->Opcode() == Op_If ||
238 iff->Opcode() == Op_CountedLoopEnd ||
239 iff->Opcode() == Op_LongCountedLoopEnd ||
240 iff->Opcode() == Op_RangeCheck,
241 "Check this code when new subtype is added");
242
243 int pop = prevdom->Opcode();
244 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
245 if (flip) {
246 if (pop == Op_IfTrue)
247 pop = Op_IfFalse;
248 else
249 pop = Op_IfTrue;
250 }
251 // 'con' is set to true or false to kill the dominated test.
252 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
253 set_ctrl(con, C->root()); // Constant gets a new use
254 // Hack the dominated test
255 _igvn.replace_input_of(iff, 1, con);
256
257 // If I dont have a reachable TRUE and FALSE path following the IfNode then
258 // I can assume this path reaches an infinite loop. In this case it's not
259 // important to optimize the data Nodes - either the whole compilation will
260 // be tossed or this path (and all data Nodes) will go dead.
261 if (iff->outcnt() != 2) return;
262
263 // Make control-dependent data Nodes on the live path (path that will remain
264 // once the dominated IF is removed) become control-dependent on the
265 // dominating projection.
266 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue);
267
268 // Loop predicates may have depending checks which should not
269 // be skipped. For example, range check predicate has two checks
270 // for lower and upper bounds.
271 if (dp == NULL)
272 return;
273
274 ProjNode* dp_proj = dp->as_Proj();
275 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj();
276 if (exclude_loop_predicate &&
277 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL ||
278 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL ||
279 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) {
280 // If this is a range check (IfNode::is_range_check), do not
281 // reorder because Compile::allow_range_check_smearing might have
282 // changed the check.
283 return; // Let IGVN transformation change control dependence.
284 }
285
286 IdealLoopTree* old_loop = get_loop(dp);
287
288 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
289 Node* cd = dp->fast_out(i); // Control-dependent node
290 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check.
291 if (cd->depends_only_on_test() && _igvn.no_dependent_zero_check(cd)) {
292 assert(cd->in(0) == dp, "");
293 _igvn.replace_input_of(cd, 0, prevdom);
294 set_early_ctrl(cd, false);
295 IdealLoopTree* new_loop = get_loop(get_ctrl(cd));
296 if (old_loop != new_loop) {
297 if (!old_loop->_child) {
298 old_loop->_body.yank(cd);
299 }
300 if (!new_loop->_child) {
301 new_loop->_body.push(cd);
302 }
303 }
304 --i;
305 --imax;
306 }
307 }
308 }
309
310 //------------------------------has_local_phi_input----------------------------
311 // Return TRUE if 'n' has Phi inputs from its local block and no other
312 // block-local inputs (all non-local-phi inputs come from earlier blocks)
has_local_phi_input(Node * n)313 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
314 Node *n_ctrl = get_ctrl(n);
315 // See if some inputs come from a Phi in this block, or from before
316 // this block.
317 uint i;
318 for( i = 1; i < n->req(); i++ ) {
319 Node *phi = n->in(i);
320 if( phi->is_Phi() && phi->in(0) == n_ctrl )
321 break;
322 }
323 if( i >= n->req() )
324 return NULL; // No Phi inputs; nowhere to clone thru
325
326 // Check for inputs created between 'n' and the Phi input. These
327 // must split as well; they have already been given the chance
328 // (courtesy of a post-order visit) and since they did not we must
329 // recover the 'cost' of splitting them by being very profitable
330 // when splitting 'n'. Since this is unlikely we simply give up.
331 for( i = 1; i < n->req(); i++ ) {
332 Node *m = n->in(i);
333 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
334 // We allow the special case of AddP's with no local inputs.
335 // This allows us to split-up address expressions.
336 if (m->is_AddP() &&
337 get_ctrl(m->in(2)) != n_ctrl &&
338 get_ctrl(m->in(3)) != n_ctrl) {
339 // Move the AddP up to dominating point
340 Node* c = find_non_split_ctrl(idom(n_ctrl));
341 if (c->is_OuterStripMinedLoop()) {
342 c->as_Loop()->verify_strip_mined(1);
343 c = c->in(LoopNode::EntryControl);
344 }
345 set_ctrl_and_loop(m, c);
346 continue;
347 }
348 return NULL;
349 }
350 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
351 }
352
353 return n_ctrl;
354 }
355
356 //------------------------------remix_address_expressions----------------------
357 // Rework addressing expressions to get the most loop-invariant stuff
358 // moved out. We'd like to do all associative operators, but it's especially
359 // important (common) to do address expressions.
remix_address_expressions(Node * n)360 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
361 if (!has_ctrl(n)) return NULL;
362 Node *n_ctrl = get_ctrl(n);
363 IdealLoopTree *n_loop = get_loop(n_ctrl);
364
365 // See if 'n' mixes loop-varying and loop-invariant inputs and
366 // itself is loop-varying.
367
368 // Only interested in binary ops (and AddP)
369 if( n->req() < 3 || n->req() > 4 ) return NULL;
370
371 Node *n1_ctrl = get_ctrl(n->in( 1));
372 Node *n2_ctrl = get_ctrl(n->in( 2));
373 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
374 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
375 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
376 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
377
378 // Does one of my inputs spin in a tighter loop than self?
379 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
380 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
381 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
382 return NULL; // Leave well enough alone
383
384 // Is at least one of my inputs loop-invariant?
385 if( n1_loop == n_loop &&
386 n2_loop == n_loop &&
387 n3_loop == n_loop )
388 return NULL; // No loop-invariant inputs
389
390
391 int n_op = n->Opcode();
392
393 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
394 if( n_op == Op_LShiftI ) {
395 // Scale is loop invariant
396 Node *scale = n->in(2);
397 Node *scale_ctrl = get_ctrl(scale);
398 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
399 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
400 return NULL;
401 const TypeInt *scale_t = scale->bottom_type()->isa_int();
402 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
403 return NULL; // Dont bother with byte/short masking
404 // Add must vary with loop (else shift would be loop-invariant)
405 Node *add = n->in(1);
406 Node *add_ctrl = get_ctrl(add);
407 IdealLoopTree *add_loop = get_loop(add_ctrl);
408 //assert( n_loop == add_loop, "" );
409 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
410
411 // Convert I-V into I+ (0-V); same for V-I
412 if( add->Opcode() == Op_SubI &&
413 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
414 Node *zero = _igvn.intcon(0);
415 set_ctrl(zero, C->root());
416 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) );
417 register_new_node( neg, get_ctrl(add->in(2) ) );
418 add = new AddINode( add->in(1), neg );
419 register_new_node( add, add_ctrl );
420 }
421 if( add->Opcode() != Op_AddI ) return NULL;
422 // See if one add input is loop invariant
423 Node *add_var = add->in(1);
424 Node *add_var_ctrl = get_ctrl(add_var);
425 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
426 Node *add_invar = add->in(2);
427 Node *add_invar_ctrl = get_ctrl(add_invar);
428 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
429 if( add_var_loop == n_loop ) {
430 } else if( add_invar_loop == n_loop ) {
431 // Swap to find the invariant part
432 add_invar = add_var;
433 add_invar_ctrl = add_var_ctrl;
434 add_invar_loop = add_var_loop;
435 add_var = add->in(2);
436 Node *add_var_ctrl = get_ctrl(add_var);
437 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
438 } else // Else neither input is loop invariant
439 return NULL;
440 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
441 return NULL; // No invariant part of the add?
442
443 // Yes! Reshape address expression!
444 Node *inv_scale = new LShiftINode( add_invar, scale );
445 Node *inv_scale_ctrl =
446 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
447 add_invar_ctrl : scale_ctrl;
448 register_new_node( inv_scale, inv_scale_ctrl );
449 Node *var_scale = new LShiftINode( add_var, scale );
450 register_new_node( var_scale, n_ctrl );
451 Node *var_add = new AddINode( var_scale, inv_scale );
452 register_new_node( var_add, n_ctrl );
453 _igvn.replace_node( n, var_add );
454 return var_add;
455 }
456
457 // Replace (I+V) with (V+I)
458 if( n_op == Op_AddI ||
459 n_op == Op_AddL ||
460 n_op == Op_AddF ||
461 n_op == Op_AddD ||
462 n_op == Op_MulI ||
463 n_op == Op_MulL ||
464 n_op == Op_MulF ||
465 n_op == Op_MulD ) {
466 if( n2_loop == n_loop ) {
467 assert( n1_loop != n_loop, "" );
468 n->swap_edges(1, 2);
469 }
470 }
471
472 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
473 // but not if I2 is a constant.
474 if( n_op == Op_AddP ) {
475 if( n2_loop == n_loop && n3_loop != n_loop ) {
476 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
477 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
478 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
479 IdealLoopTree *n22loop = get_loop( n22_ctrl );
480 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
481 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
482 n23_loop == n_loop ) {
483 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
484 // Stuff new AddP in the loop preheader
485 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
486 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) );
487 register_new_node( add2, n_ctrl );
488 _igvn.replace_node( n, add2 );
489 return add2;
490 }
491 }
492 }
493
494 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
495 if (n2_loop != n_loop && n3_loop == n_loop) {
496 if (n->in(3)->Opcode() == Op_AddX) {
497 Node *V = n->in(3)->in(1);
498 Node *I = n->in(3)->in(2);
499 if (is_member(n_loop,get_ctrl(V))) {
500 } else {
501 Node *tmp = V; V = I; I = tmp;
502 }
503 if (!is_member(n_loop,get_ctrl(I))) {
504 Node *add1 = new AddPNode(n->in(1), n->in(2), I);
505 // Stuff new AddP in the loop preheader
506 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl));
507 Node *add2 = new AddPNode(n->in(1), add1, V);
508 register_new_node(add2, n_ctrl);
509 _igvn.replace_node(n, add2);
510 return add2;
511 }
512 }
513 }
514 }
515
516 return NULL;
517 }
518
519 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
convert_add_to_muladd(Node * n)520 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
521 assert(n->Opcode() == Op_AddI, "sanity");
522 Node * nn = NULL;
523 Node * in1 = n->in(1);
524 Node * in2 = n->in(2);
525 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
526 IdealLoopTree* loop_n = get_loop(get_ctrl(n));
527 if (loop_n->is_counted() &&
528 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) &&
529 Matcher::match_rule_supported(Op_MulAddVS2VI) &&
530 Matcher::match_rule_supported(Op_MulAddS2I)) {
531 Node* mul_in1 = in1->in(1);
532 Node* mul_in2 = in1->in(2);
533 Node* mul_in3 = in2->in(1);
534 Node* mul_in4 = in2->in(2);
535 if (mul_in1->Opcode() == Op_LoadS &&
536 mul_in2->Opcode() == Op_LoadS &&
537 mul_in3->Opcode() == Op_LoadS &&
538 mul_in4->Opcode() == Op_LoadS) {
539 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
540 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
541 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
542 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
543 IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
544 IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
545 // All nodes should be in the same counted loop.
546 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
547 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
548 Node* adr1 = mul_in1->in(MemNode::Address);
549 Node* adr2 = mul_in2->in(MemNode::Address);
550 Node* adr3 = mul_in3->in(MemNode::Address);
551 Node* adr4 = mul_in4->in(MemNode::Address);
552 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
553 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
554 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
555 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
556 register_new_node(nn, get_ctrl(n));
557 _igvn.replace_node(n, nn);
558 return nn;
559 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
560 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
561 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
562 register_new_node(nn, get_ctrl(n));
563 _igvn.replace_node(n, nn);
564 return nn;
565 }
566 }
567 }
568 }
569 }
570 }
571 return nn;
572 }
573
574 //------------------------------conditional_move-------------------------------
575 // Attempt to replace a Phi with a conditional move. We have some pretty
576 // strict profitability requirements. All Phis at the merge point must
577 // be converted, so we can remove the control flow. We need to limit the
578 // number of c-moves to a small handful. All code that was in the side-arms
579 // of the CFG diamond is now speculatively executed. This code has to be
580 // "cheap enough". We are pretty much limited to CFG diamonds that merge
581 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
conditional_move(Node * region)582 Node *PhaseIdealLoop::conditional_move( Node *region ) {
583
584 assert(region->is_Region(), "sanity check");
585 if (region->req() != 3) return NULL;
586
587 // Check for CFG diamond
588 Node *lp = region->in(1);
589 Node *rp = region->in(2);
590 if (!lp || !rp) return NULL;
591 Node *lp_c = lp->in(0);
592 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
593 IfNode *iff = lp_c->as_If();
594
595 // Check for ops pinned in an arm of the diamond.
596 // Can't remove the control flow in this case
597 if (lp->outcnt() > 1) return NULL;
598 if (rp->outcnt() > 1) return NULL;
599
600 IdealLoopTree* r_loop = get_loop(region);
601 assert(r_loop == get_loop(iff), "sanity");
602 // Always convert to CMOVE if all results are used only outside this loop.
603 bool used_inside_loop = (r_loop == _ltree_root);
604
605 // Check profitability
606 int cost = 0;
607 int phis = 0;
608 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
609 Node *out = region->fast_out(i);
610 if (!out->is_Phi()) continue; // Ignore other control edges, etc
611 phis++;
612 PhiNode* phi = out->as_Phi();
613 BasicType bt = phi->type()->basic_type();
614 switch (bt) {
615 case T_DOUBLE:
616 case T_FLOAT:
617 if (C->use_cmove()) {
618 continue; //TODO: maybe we want to add some cost
619 }
620 cost += Matcher::float_cmove_cost(); // Could be very expensive
621 break;
622 case T_LONG: {
623 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
624 }
625 case T_INT: // These all CMOV fine
626 case T_ADDRESS: { // (RawPtr)
627 cost++;
628 break;
629 }
630 case T_NARROWOOP: // Fall through
631 case T_OBJECT: { // Base oops are OK, but not derived oops
632 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
633 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
634 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
635 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
636 // have a Phi for the base here that we convert to a CMOVE all is well
637 // and good. But if the base is dead, we'll not make a CMOVE. Later
638 // the allocator will have to produce a base by creating a CMOVE of the
639 // relevant bases. This puts the allocator in the business of
640 // manufacturing expensive instructions, generally a bad plan.
641 // Just Say No to Conditionally-Moved Derived Pointers.
642 if (tp && tp->offset() != 0)
643 return NULL;
644 cost++;
645 break;
646 }
647 default:
648 return NULL; // In particular, can't do memory or I/O
649 }
650 // Add in cost any speculative ops
651 for (uint j = 1; j < region->req(); j++) {
652 Node *proj = region->in(j);
653 Node *inp = phi->in(j);
654 if (get_ctrl(inp) == proj) { // Found local op
655 cost++;
656 // Check for a chain of dependent ops; these will all become
657 // speculative in a CMOV.
658 for (uint k = 1; k < inp->req(); k++)
659 if (get_ctrl(inp->in(k)) == proj)
660 cost += ConditionalMoveLimit; // Too much speculative goo
661 }
662 }
663 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
664 // This will likely Split-If, a higher-payoff operation.
665 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
666 Node* use = phi->fast_out(k);
667 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
668 cost += ConditionalMoveLimit;
669 // Is there a use inside the loop?
670 // Note: check only basic types since CMoveP is pinned.
671 if (!used_inside_loop && is_java_primitive(bt)) {
672 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
673 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
674 used_inside_loop = true;
675 }
676 }
677 }
678 }//for
679 Node* bol = iff->in(1);
680 if (bol->Opcode() == Op_Opaque4) {
681 return NULL; // Ignore loop predicate checks (the Opaque4 ensures they will go away)
682 }
683 assert(bol->Opcode() == Op_Bool, "Unexpected node");
684 int cmp_op = bol->in(1)->Opcode();
685 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
686 return NULL;
687 }
688 // It is expensive to generate flags from a float compare.
689 // Avoid duplicated float compare.
690 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
691
692 float infrequent_prob = PROB_UNLIKELY_MAG(3);
693 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
694 if (used_inside_loop) {
695 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
696
697 // BlockLayoutByFrequency optimization moves infrequent branch
698 // from hot path. No point in CMOV'ing in such case (110 is used
699 // instead of 100 to take into account not exactness of float value).
700 if (BlockLayoutByFrequency) {
701 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
702 }
703 }
704 // Check for highly predictable branch. No point in CMOV'ing if
705 // we are going to predict accurately all the time.
706 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
707 //keep going
708 } else if (iff->_prob < infrequent_prob ||
709 iff->_prob > (1.0f - infrequent_prob))
710 return NULL;
711
712 // --------------
713 // Now replace all Phis with CMOV's
714 Node *cmov_ctrl = iff->in(0);
715 uint flip = (lp->Opcode() == Op_IfTrue);
716 Node_List wq;
717 while (1) {
718 PhiNode* phi = NULL;
719 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
720 Node *out = region->fast_out(i);
721 if (out->is_Phi()) {
722 phi = out->as_Phi();
723 break;
724 }
725 }
726 if (phi == NULL || _igvn.type(phi) == Type::TOP) {
727 break;
728 }
729 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
730 // Move speculative ops
731 wq.push(phi);
732 while (wq.size() > 0) {
733 Node *n = wq.pop();
734 for (uint j = 1; j < n->req(); j++) {
735 Node* m = n->in(j);
736 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) {
737 #ifndef PRODUCT
738 if (PrintOpto && VerifyLoopOptimizations) {
739 tty->print(" speculate: ");
740 m->dump();
741 }
742 #endif
743 set_ctrl(m, cmov_ctrl);
744 wq.push(m);
745 }
746 }
747 }
748 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
749 register_new_node( cmov, cmov_ctrl );
750 _igvn.replace_node( phi, cmov );
751 #ifndef PRODUCT
752 if (TraceLoopOpts) {
753 tty->print("CMOV ");
754 r_loop->dump_head();
755 if (Verbose) {
756 bol->in(1)->dump(1);
757 cmov->dump(1);
758 }
759 }
760 if (VerifyLoopOptimizations) verify();
761 #endif
762 }
763
764 // The useless CFG diamond will fold up later; see the optimization in
765 // RegionNode::Ideal.
766 _igvn._worklist.push(region);
767
768 return iff->in(1);
769 }
770
enqueue_cfg_uses(Node * m,Unique_Node_List & wq)771 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
772 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
773 Node* u = m->fast_out(i);
774 if (u->is_CFG()) {
775 if (u->Opcode() == Op_NeverBranch) {
776 u = ((NeverBranchNode*)u)->proj_out(0);
777 enqueue_cfg_uses(u, wq);
778 } else {
779 wq.push(u);
780 }
781 }
782 }
783 }
784
785 // Try moving a store out of a loop, right before the loop
try_move_store_before_loop(Node * n,Node * n_ctrl)786 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
787 // Store has to be first in the loop body
788 IdealLoopTree *n_loop = get_loop(n_ctrl);
789 if (n->is_Store() && n_loop != _ltree_root &&
790 n_loop->is_loop() && n_loop->_head->is_Loop() &&
791 n->in(0) != NULL) {
792 Node* address = n->in(MemNode::Address);
793 Node* value = n->in(MemNode::ValueIn);
794 Node* mem = n->in(MemNode::Memory);
795 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
796 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
797
798 // - address and value must be loop invariant
799 // - memory must be a memory Phi for the loop
800 // - Store must be the only store on this memory slice in the
801 // loop: if there's another store following this one then value
802 // written at iteration i by the second store could be overwritten
803 // at iteration i+n by the first store: it's not safe to move the
804 // first store out of the loop
805 // - nothing must observe the memory Phi: it guarantees no read
806 // before the store, we are also guaranteed the store post
807 // dominates the loop head (ignoring a possible early
808 // exit). Otherwise there would be extra Phi involved between the
809 // loop's Phi and the store.
810 // - there must be no early exit from the loop before the Store
811 // (such an exit most of the time would be an extra use of the
812 // memory Phi but sometimes is a bottom memory Phi that takes the
813 // store as input).
814
815 if (!n_loop->is_member(address_loop) &&
816 !n_loop->is_member(value_loop) &&
817 mem->is_Phi() && mem->in(0) == n_loop->_head &&
818 mem->outcnt() == 1 &&
819 mem->in(LoopNode::LoopBackControl) == n) {
820
821 assert(n_loop->_tail != NULL, "need a tail");
822 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
823
824 // Verify that there's no early exit of the loop before the store.
825 bool ctrl_ok = false;
826 {
827 // Follow control from loop head until n, we exit the loop or
828 // we reach the tail
829 ResourceMark rm;
830 Unique_Node_List wq;
831 wq.push(n_loop->_head);
832
833 for (uint next = 0; next < wq.size(); ++next) {
834 Node *m = wq.at(next);
835 if (m == n->in(0)) {
836 ctrl_ok = true;
837 continue;
838 }
839 assert(!has_ctrl(m), "should be CFG");
840 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
841 ctrl_ok = false;
842 break;
843 }
844 enqueue_cfg_uses(m, wq);
845 if (wq.size() > 10) {
846 ctrl_ok = false;
847 break;
848 }
849 }
850 }
851 if (ctrl_ok) {
852 // move the Store
853 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
854 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
855 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
856 // Disconnect the phi now. An empty phi can confuse other
857 // optimizations in this pass of loop opts.
858 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
859 n_loop->_body.yank(mem);
860
861 set_ctrl_and_loop(n, n->in(0));
862
863 return n;
864 }
865 }
866 }
867 return NULL;
868 }
869
870 // Try moving a store out of a loop, right after the loop
try_move_store_after_loop(Node * n)871 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
872 if (n->is_Store() && n->in(0) != NULL) {
873 Node *n_ctrl = get_ctrl(n);
874 IdealLoopTree *n_loop = get_loop(n_ctrl);
875 // Store must be in a loop
876 if (n_loop != _ltree_root && !n_loop->_irreducible) {
877 Node* address = n->in(MemNode::Address);
878 Node* value = n->in(MemNode::ValueIn);
879 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
880 // address must be loop invariant
881 if (!n_loop->is_member(address_loop)) {
882 // Store must be last on this memory slice in the loop and
883 // nothing in the loop must observe it
884 Node* phi = NULL;
885 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
886 Node* u = n->fast_out(i);
887 if (has_ctrl(u)) { // control use?
888 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
889 if (!n_loop->is_member(u_loop)) {
890 continue;
891 }
892 if (u->is_Phi() && u->in(0) == n_loop->_head) {
893 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
894 // multiple phis on the same slice are possible
895 if (phi != NULL) {
896 return;
897 }
898 phi = u;
899 continue;
900 }
901 }
902 return;
903 }
904 if (phi != NULL) {
905 // Nothing in the loop before the store (next iteration)
906 // must observe the stored value
907 bool mem_ok = true;
908 {
909 ResourceMark rm;
910 Unique_Node_List wq;
911 wq.push(phi);
912 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
913 Node *m = wq.at(next);
914 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
915 Node* u = m->fast_out(i);
916 if (u->is_Store() || u->is_Phi()) {
917 if (u != n) {
918 wq.push(u);
919 mem_ok = (wq.size() <= 10);
920 }
921 } else {
922 mem_ok = false;
923 break;
924 }
925 }
926 }
927 }
928 if (mem_ok) {
929 // Move the store out of the loop if the LCA of all
930 // users (except for the phi) is outside the loop.
931 Node* hook = new Node(1);
932 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead
933 _igvn.rehash_node_delayed(phi);
934 int count = phi->replace_edge(n, hook, &_igvn);
935 assert(count > 0, "inconsistent phi");
936
937 // Compute latest point this store can go
938 Node* lca = get_late_ctrl(n, get_ctrl(n));
939 if (lca->is_OuterStripMinedLoop()) {
940 lca = lca->in(LoopNode::EntryControl);
941 }
942 if (n_loop->is_member(get_loop(lca))) {
943 // LCA is in the loop - bail out
944 _igvn.replace_node(hook, n);
945 return;
946 }
947 #ifdef ASSERT
948 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
949 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
950 n_loop->_head->as_Loop()->verify_strip_mined(1);
951 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
952 IdealLoopTree* outer_loop = get_loop(outer);
953 assert(n_loop->_parent == outer_loop, "broken loop tree");
954 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
955 }
956 #endif
957 lca = place_outside_loop(lca, n_loop);
958 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop");
959 assert(get_loop(lca)->_nest < n_loop->_nest || lca->in(0)->Opcode() == Op_NeverBranch, "must not be moved into inner loop");
960
961 // Move store out of the loop
962 _igvn.replace_node(hook, n->in(MemNode::Memory));
963 _igvn.replace_input_of(n, 0, lca);
964 set_ctrl_and_loop(n, lca);
965
966 // Disconnect the phi now. An empty phi can confuse other
967 // optimizations in this pass of loop opts..
968 if (phi->in(LoopNode::LoopBackControl) == phi) {
969 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
970 n_loop->_body.yank(phi);
971 }
972 }
973 }
974 }
975 }
976 }
977 }
978
979 //------------------------------split_if_with_blocks_pre-----------------------
980 // Do the real work in a non-recursive function. Data nodes want to be
981 // cloned in the pre-order so they can feed each other nicely.
split_if_with_blocks_pre(Node * n)982 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
983 // Cloning these guys is unlikely to win
984 int n_op = n->Opcode();
985 if (n_op == Op_MergeMem) {
986 return n;
987 }
988 if (n->is_Proj()) {
989 return n;
990 }
991 // Do not clone-up CmpFXXX variations, as these are always
992 // followed by a CmpI
993 if (n->is_Cmp()) {
994 return n;
995 }
996 // Attempt to use a conditional move instead of a phi/branch
997 if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
998 Node *cmov = conditional_move( n );
999 if (cmov) {
1000 return cmov;
1001 }
1002 }
1003 if (n->is_CFG() || n->is_LoadStore()) {
1004 return n;
1005 }
1006 if (n->is_Opaque1() || // Opaque nodes cannot be mod'd
1007 n_op == Op_Opaque2) {
1008 if (!C->major_progress()) { // If chance of no more loop opts...
1009 _igvn._worklist.push(n); // maybe we'll remove them
1010 }
1011 return n;
1012 }
1013
1014 if (n->is_Con()) {
1015 return n; // No cloning for Con nodes
1016 }
1017
1018 Node *n_ctrl = get_ctrl(n);
1019 if (!n_ctrl) {
1020 return n; // Dead node
1021 }
1022
1023 Node* res = try_move_store_before_loop(n, n_ctrl);
1024 if (res != NULL) {
1025 return n;
1026 }
1027
1028 // Attempt to remix address expressions for loop invariants
1029 Node *m = remix_address_expressions( n );
1030 if( m ) return m;
1031
1032 if (n_op == Op_AddI) {
1033 Node *nn = convert_add_to_muladd( n );
1034 if ( nn ) return nn;
1035 }
1036
1037 if (n->is_ConstraintCast()) {
1038 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
1039 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
1040 // Node control inputs don't necessarily agree with loop control info (due to
1041 // transformations happened in between), thus additional dominance check is needed
1042 // to keep loop info valid.
1043 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
1044 _igvn.replace_node(n, dom_cast);
1045 return dom_cast;
1046 }
1047 }
1048
1049 // Determine if the Node has inputs from some local Phi.
1050 // Returns the block to clone thru.
1051 Node *n_blk = has_local_phi_input( n );
1052 if( !n_blk ) return n;
1053
1054 // Do not clone the trip counter through on a CountedLoop
1055 // (messes up the canonical shape).
1056 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_transformed_long_inner_loop())) && n->Opcode() == Op_AddI) ||
1057 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) {
1058 return n;
1059 }
1060
1061 // Check for having no control input; not pinned. Allow
1062 // dominating control.
1063 if (n->in(0)) {
1064 Node *dom = idom(n_blk);
1065 if (dom_lca(n->in(0), dom) != n->in(0)) {
1066 return n;
1067 }
1068 }
1069 // Policy: when is it profitable. You must get more wins than
1070 // policy before it is considered profitable. Policy is usually 0,
1071 // so 1 win is considered profitable. Big merges will require big
1072 // cloning, so get a larger policy.
1073 int policy = n_blk->req() >> 2;
1074
1075 // If the loop is a candidate for range check elimination,
1076 // delay splitting through it's phi until a later loop optimization
1077 if (n_blk->is_CountedLoop()) {
1078 IdealLoopTree *lp = get_loop(n_blk);
1079 if (lp && lp->_rce_candidate) {
1080 return n;
1081 }
1082 }
1083
1084 if (must_throttle_split_if()) return n;
1085
1086 // Split 'n' through the merge point if it is profitable
1087 Node *phi = split_thru_phi( n, n_blk, policy );
1088 if (!phi) return n;
1089
1090 // Found a Phi to split thru!
1091 // Replace 'n' with the new phi
1092 _igvn.replace_node( n, phi );
1093 // Moved a load around the loop, 'en-registering' something.
1094 if (n_blk->is_Loop() && n->is_Load() &&
1095 !phi->in(LoopNode::LoopBackControl)->is_Load())
1096 C->set_major_progress();
1097
1098 return phi;
1099 }
1100
merge_point_too_heavy(Compile * C,Node * region)1101 static bool merge_point_too_heavy(Compile* C, Node* region) {
1102 // Bail out if the region and its phis have too many users.
1103 int weight = 0;
1104 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1105 weight += region->fast_out(i)->outcnt();
1106 }
1107 int nodes_left = C->max_node_limit() - C->live_nodes();
1108 if (weight * 8 > nodes_left) {
1109 if (PrintOpto) {
1110 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1111 }
1112 return true;
1113 } else {
1114 return false;
1115 }
1116 }
1117
merge_point_safe(Node * region)1118 static bool merge_point_safe(Node* region) {
1119 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1120 // having a PhiNode input. This sidesteps the dangerous case where the split
1121 // ConvI2LNode may become TOP if the input Value() does not
1122 // overlap the ConvI2L range, leaving a node which may not dominate its
1123 // uses.
1124 // A better fix for this problem can be found in the BugTraq entry, but
1125 // expediency for Mantis demands this hack.
1126 #ifdef _LP64
1127 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1128 Node* n = region->fast_out(i);
1129 if (n->is_Phi()) {
1130 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1131 Node* m = n->fast_out(j);
1132 if (m->Opcode() == Op_ConvI2L)
1133 return false;
1134 if (m->is_CastII()) {
1135 return false;
1136 }
1137 }
1138 }
1139 }
1140 #endif
1141 return true;
1142 }
1143
1144
1145 //------------------------------place_outside_loop---------------------------------
1146 // Place some computation outside of this loop on the path to the use passed as argument
place_outside_loop(Node * useblock,IdealLoopTree * loop) const1147 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const {
1148 Node* head = loop->_head;
1149 assert(!loop->is_member(get_loop(useblock)), "must be outside loop");
1150 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) {
1151 loop = loop->_parent;
1152 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop");
1153 }
1154
1155 // Pick control right outside the loop
1156 for (;;) {
1157 Node* dom = idom(useblock);
1158 if (loop->is_member(get_loop(dom)) ||
1159 // NeverBranch nodes are not assigned to the loop when constructed
1160 (dom->Opcode() == Op_NeverBranch && loop->is_member(get_loop(dom->in(0))))) {
1161 break;
1162 }
1163 useblock = dom;
1164 }
1165 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control");
1166 return useblock;
1167 }
1168
1169
identical_backtoback_ifs(Node * n)1170 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1171 if (!n->is_If() || n->is_BaseCountedLoopEnd()) {
1172 return false;
1173 }
1174 if (!n->in(0)->is_Region()) {
1175 return false;
1176 }
1177 Node* region = n->in(0);
1178 Node* dom = idom(region);
1179 if (!dom->is_If() || dom->in(1) != n->in(1)) {
1180 return false;
1181 }
1182 IfNode* dom_if = dom->as_If();
1183 Node* proj_true = dom_if->proj_out(1);
1184 Node* proj_false = dom_if->proj_out(0);
1185
1186 for (uint i = 1; i < region->req(); i++) {
1187 if (is_dominator(proj_true, region->in(i))) {
1188 continue;
1189 }
1190 if (is_dominator(proj_false, region->in(i))) {
1191 continue;
1192 }
1193 return false;
1194 }
1195
1196 return true;
1197 }
1198
1199
can_split_if(Node * n_ctrl)1200 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1201 if (must_throttle_split_if()) {
1202 return false;
1203 }
1204
1205 // Do not do 'split-if' if irreducible loops are present.
1206 if (_has_irreducible_loops) {
1207 return false;
1208 }
1209
1210 if (merge_point_too_heavy(C, n_ctrl)) {
1211 return false;
1212 }
1213
1214 // Do not do 'split-if' if some paths are dead. First do dead code
1215 // elimination and then see if its still profitable.
1216 for (uint i = 1; i < n_ctrl->req(); i++) {
1217 if (n_ctrl->in(i) == C->top()) {
1218 return false;
1219 }
1220 }
1221
1222 // If trying to do a 'Split-If' at the loop head, it is only
1223 // profitable if the cmp folds up on BOTH paths. Otherwise we
1224 // risk peeling a loop forever.
1225
1226 // CNC - Disabled for now. Requires careful handling of loop
1227 // body selection for the cloned code. Also, make sure we check
1228 // for any input path not being in the same loop as n_ctrl. For
1229 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1230 // because the alternative loop entry points won't be converted
1231 // into LoopNodes.
1232 IdealLoopTree *n_loop = get_loop(n_ctrl);
1233 for (uint j = 1; j < n_ctrl->req(); j++) {
1234 if (get_loop(n_ctrl->in(j)) != n_loop) {
1235 return false;
1236 }
1237 }
1238
1239 // Check for safety of the merge point.
1240 if (!merge_point_safe(n_ctrl)) {
1241 return false;
1242 }
1243
1244 return true;
1245 }
1246
1247 // Detect if the node is the inner strip-mined loop
1248 // Return: NULL if it's not the case, or the exit of outer strip-mined loop
is_inner_of_stripmined_loop(const Node * out)1249 static Node* is_inner_of_stripmined_loop(const Node* out) {
1250 Node* out_le = NULL;
1251
1252 if (out->is_CountedLoopEnd()) {
1253 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1254
1255 if (loop != NULL && loop->is_strip_mined()) {
1256 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1257 }
1258 }
1259
1260 return out_le;
1261 }
1262
1263 //------------------------------split_if_with_blocks_post----------------------
1264 // Do the real work in a non-recursive function. CFG hackery wants to be
1265 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1266 // info.
split_if_with_blocks_post(Node * n)1267 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1268
1269 // Cloning Cmp through Phi's involves the split-if transform.
1270 // FastLock is not used by an If
1271 if (n->is_Cmp() && !n->is_FastLock()) {
1272 Node *n_ctrl = get_ctrl(n);
1273 // Determine if the Node has inputs from some local Phi.
1274 // Returns the block to clone thru.
1275 Node *n_blk = has_local_phi_input(n);
1276 if (n_blk != n_ctrl) {
1277 return;
1278 }
1279
1280 if (!can_split_if(n_ctrl)) {
1281 return;
1282 }
1283
1284 if (n->outcnt() != 1) {
1285 return; // Multiple bool's from 1 compare?
1286 }
1287 Node *bol = n->unique_out();
1288 assert(bol->is_Bool(), "expect a bool here");
1289 if (bol->outcnt() != 1) {
1290 return;// Multiple branches from 1 compare?
1291 }
1292 Node *iff = bol->unique_out();
1293
1294 // Check some safety conditions
1295 if (iff->is_If()) { // Classic split-if?
1296 if (iff->in(0) != n_ctrl) {
1297 return; // Compare must be in same blk as if
1298 }
1299 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1300 // Can't split CMove with different control edge.
1301 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) {
1302 return;
1303 }
1304 if (get_ctrl(iff->in(2)) == n_ctrl ||
1305 get_ctrl(iff->in(3)) == n_ctrl) {
1306 return; // Inputs not yet split-up
1307 }
1308 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1309 return; // Loop-invar test gates loop-varying CMOVE
1310 }
1311 } else {
1312 return; // some other kind of node, such as an Allocate
1313 }
1314
1315 // When is split-if profitable? Every 'win' on means some control flow
1316 // goes dead, so it's almost always a win.
1317 int policy = 0;
1318 // Split compare 'n' through the merge point if it is profitable
1319 Node *phi = split_thru_phi( n, n_ctrl, policy);
1320 if (!phi) {
1321 return;
1322 }
1323
1324 // Found a Phi to split thru!
1325 // Replace 'n' with the new phi
1326 _igvn.replace_node(n, phi);
1327
1328 // Now split the bool up thru the phi
1329 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1330 guarantee(bolphi != NULL, "null boolean phi node");
1331
1332 _igvn.replace_node(bol, bolphi);
1333 assert(iff->in(1) == bolphi, "");
1334
1335 if (bolphi->Value(&_igvn)->singleton()) {
1336 return;
1337 }
1338
1339 // Conditional-move? Must split up now
1340 if (!iff->is_If()) {
1341 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1342 _igvn.replace_node(iff, cmovphi);
1343 return;
1344 }
1345
1346 // Now split the IF
1347 do_split_if(iff);
1348 return;
1349 }
1350
1351 // Two identical ifs back to back can be merged
1352 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1353 Node *n_ctrl = n->in(0);
1354 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1355 IfNode* dom_if = idom(n_ctrl)->as_If();
1356 Node* proj_true = dom_if->proj_out(1);
1357 Node* proj_false = dom_if->proj_out(0);
1358 Node* con_true = _igvn.makecon(TypeInt::ONE);
1359 Node* con_false = _igvn.makecon(TypeInt::ZERO);
1360
1361 for (uint i = 1; i < n_ctrl->req(); i++) {
1362 if (is_dominator(proj_true, n_ctrl->in(i))) {
1363 bolphi->init_req(i, con_true);
1364 } else {
1365 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1366 bolphi->init_req(i, con_false);
1367 }
1368 }
1369 register_new_node(bolphi, n_ctrl);
1370 _igvn.replace_input_of(n, 1, bolphi);
1371
1372 // Now split the IF
1373 do_split_if(n);
1374 return;
1375 }
1376
1377 // Check for an IF ready to split; one that has its
1378 // condition codes input coming from a Phi at the block start.
1379 int n_op = n->Opcode();
1380
1381 // Check for an IF being dominated by another IF same test
1382 if (n_op == Op_If ||
1383 n_op == Op_RangeCheck) {
1384 Node *bol = n->in(1);
1385 uint max = bol->outcnt();
1386 // Check for same test used more than once?
1387 if (max > 1 && bol->is_Bool()) {
1388 // Search up IDOMs to see if this IF is dominated.
1389 Node *cutoff = get_ctrl(bol);
1390
1391 // Now search up IDOMs till cutoff, looking for a dominating test
1392 Node *prevdom = n;
1393 Node *dom = idom(prevdom);
1394 while (dom != cutoff) {
1395 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom &&
1396 safe_for_if_replacement(dom)) {
1397 // It's invalid to move control dependent data nodes in the inner
1398 // strip-mined loop, because:
1399 // 1) break validation of LoopNode::verify_strip_mined()
1400 // 2) move code with side-effect in strip-mined loop
1401 // Move to the exit of outer strip-mined loop in that case.
1402 Node* out_le = is_inner_of_stripmined_loop(dom);
1403 if (out_le != NULL) {
1404 prevdom = out_le;
1405 }
1406 // Replace the dominated test with an obvious true or false.
1407 // Place it on the IGVN worklist for later cleanup.
1408 C->set_major_progress();
1409 dominated_by(prevdom, n, false, true);
1410 #ifndef PRODUCT
1411 if( VerifyLoopOptimizations ) verify();
1412 #endif
1413 return;
1414 }
1415 prevdom = dom;
1416 dom = idom(prevdom);
1417 }
1418 }
1419 }
1420
1421 try_sink_out_of_loop(n);
1422
1423 try_move_store_after_loop(n);
1424
1425 // Check for Opaque2's who's loop has disappeared - who's input is in the
1426 // same loop nest as their output. Remove 'em, they are no longer useful.
1427 if( n_op == Op_Opaque2 &&
1428 n->in(1) != NULL &&
1429 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1430 _igvn.replace_node( n, n->in(1) );
1431 }
1432 }
1433
safe_for_if_replacement(const Node * dom) const1434 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
1435 if (!dom->is_CountedLoopEnd()) {
1436 return true;
1437 }
1438 CountedLoopEndNode* le = dom->as_CountedLoopEnd();
1439 CountedLoopNode* cl = le->loopnode();
1440 if (cl == NULL) {
1441 return true;
1442 }
1443 if (!cl->is_main_loop()) {
1444 return true;
1445 }
1446 if (cl->is_canonical_loop_entry() == NULL) {
1447 return true;
1448 }
1449 // Further unrolling is possible so loop exit condition might change
1450 return false;
1451 }
1452
1453 // See if a shared loop-varying computation has no loop-varying uses.
1454 // Happens if something is only used for JVM state in uncommon trap exits,
1455 // like various versions of induction variable+offset. Clone the
1456 // computation per usage to allow it to sink out of the loop.
try_sink_out_of_loop(Node * n)1457 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
1458 bool is_raw_to_oop_cast = n->is_ConstraintCast() &&
1459 n->in(1)->bottom_type()->isa_rawptr() &&
1460 !n->bottom_type()->isa_rawptr();
1461 if (has_ctrl(n) &&
1462 !n->is_Phi() &&
1463 !n->is_Bool() &&
1464 !n->is_Proj() &&
1465 !n->is_MergeMem() &&
1466 !n->is_CMove() &&
1467 !is_raw_to_oop_cast && // don't extend live ranges of raw oops
1468 n->Opcode() != Op_Opaque4 &&
1469 !n->is_Type()) {
1470 Node *n_ctrl = get_ctrl(n);
1471 IdealLoopTree *n_loop = get_loop(n_ctrl);
1472
1473 if (n->in(0) != NULL) {
1474 IdealLoopTree* loop_ctrl = get_loop(n->in(0));
1475 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) {
1476 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example,
1477 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop).
1478 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not.
1479 _igvn.replace_input_of(n, 0, place_outside_loop(n_ctrl, loop_ctrl));
1480 }
1481 }
1482 if (n_loop != _ltree_root && n->outcnt() > 1) {
1483 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of
1484 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us.
1485 Node* early_ctrl = compute_early_ctrl(n, n_ctrl);
1486 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now
1487 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops!
1488 assert(!n->is_Store() && !n->is_LoadStore(), "no node with a side effect");
1489 Node* outer_loop_clone = NULL;
1490 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) {
1491 Node* u = n->last_out(j); // Clone private computation per use
1492 _igvn.rehash_node_delayed(u);
1493 Node* x = n->clone(); // Clone computation
1494 Node* x_ctrl = NULL;
1495 if (u->is_Phi()) {
1496 // Replace all uses of normal nodes. Replace Phi uses
1497 // individually, so the separate Nodes can sink down
1498 // different paths.
1499 uint k = 1;
1500 while (u->in(k) != n) k++;
1501 u->set_req(k, x);
1502 // x goes next to Phi input path
1503 x_ctrl = u->in(0)->in(k);
1504 // Find control for 'x' next to use but not inside inner loops.
1505 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1506 --j;
1507 } else { // Normal use
1508 if (has_ctrl(u)) {
1509 x_ctrl = get_ctrl(u);
1510 } else {
1511 x_ctrl = u->in(0);
1512 }
1513 // Find control for 'x' next to use but not inside inner loops.
1514 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1515 // Replace all uses
1516 if (u->is_ConstraintCast() && u->bottom_type()->higher_equal(_igvn.type(n)) && u->in(0) == x_ctrl) {
1517 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary
1518 // anymore now that we're going to pin n as well
1519 _igvn.replace_node(u, x);
1520 --j;
1521 } else {
1522 int nb = u->replace_edge(n, x, &_igvn);
1523 j -= nb;
1524 }
1525 }
1526
1527 if (n->is_Load()) {
1528 // For loads, add a control edge to a CFG node outside of the loop
1529 // to force them to not combine and return back inside the loop
1530 // during GVN optimization (4641526).
1531 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked");
1532
1533 IdealLoopTree* x_loop = get_loop(x_ctrl);
1534 Node* x_head = x_loop->_head;
1535 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) {
1536 // Do not add duplicate LoadNodes to the outer strip mined loop
1537 if (outer_loop_clone != NULL) {
1538 _igvn.replace_node(x, outer_loop_clone);
1539 continue;
1540 }
1541 outer_loop_clone = x;
1542 }
1543 x->set_req(0, x_ctrl);
1544 } else if (n->in(0) != NULL){
1545 x->set_req(0, x_ctrl);
1546 }
1547 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1548 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop");
1549 register_new_node(x, x_ctrl);
1550
1551 // Chain of AddP: (AddP base (AddP base )) must keep the same base after sinking so:
1552 // 1- We don't add a CastPP here when the first one is sunk so if the second one is not, their bases remain
1553 // the same.
1554 // (see 2- below)
1555 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() ||
1556 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) ||
1557 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape");
1558 if (x->in(0) == NULL && !x->is_DecodeNarrowPtr() &&
1559 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) {
1560 assert(!x->is_Load(), "load should be pinned");
1561 // Use a cast node to pin clone out of loop
1562 Node* cast = NULL;
1563 for (uint k = 0; k < x->req(); k++) {
1564 Node* in = x->in(k);
1565 if (in != NULL && n_loop->is_member(get_loop(get_ctrl(in)))) {
1566 const Type* in_t = _igvn.type(in);
1567 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t, ConstraintCastNode::UnconditionalDependency);
1568 }
1569 if (cast != NULL) {
1570 register_new_node(cast, x_ctrl);
1571 x->replace_edge(in, cast);
1572 // Chain of AddP:
1573 // 2- A CastPP of the base is only added now that both AddP nodes are sunk
1574 if (x->is_AddP() && k == AddPNode::Base) {
1575 for (DUIterator_Fast imax, i = x->fast_outs(imax); i < imax; i++) {
1576 Node* u = x->fast_out(i);
1577 if (u->is_AddP() && u->in(AddPNode::Base) == n->in(AddPNode::Base)) {
1578 _igvn.replace_input_of(u, AddPNode::Base, cast);
1579 assert(u->find_out_with(Op_AddP) == NULL, "more than 2 chained AddP nodes?");
1580 }
1581 }
1582 }
1583 break;
1584 }
1585 }
1586 assert(cast != NULL, "must have added a cast to pin the node");
1587 }
1588 }
1589 _igvn.remove_dead_node(n);
1590 }
1591 _dom_lca_tags_round = 0;
1592 }
1593 }
1594 }
1595
compute_early_ctrl(Node * n,Node * n_ctrl)1596 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) {
1597 Node* early_ctrl = NULL;
1598 ResourceMark rm;
1599 Unique_Node_List wq;
1600 wq.push(n);
1601 for (uint i = 0; i < wq.size(); i++) {
1602 Node* m = wq.at(i);
1603 Node* c = NULL;
1604 if (m->is_CFG()) {
1605 c = m;
1606 } else if (m->pinned()) {
1607 c = m->in(0);
1608 } else {
1609 for (uint j = 0; j < m->req(); j++) {
1610 Node* in = m->in(j);
1611 if (in == NULL) {
1612 continue;
1613 }
1614 wq.push(in);
1615 }
1616 }
1617 if (c != NULL) {
1618 assert(is_dominator(c, n_ctrl), "");
1619 if (early_ctrl == NULL) {
1620 early_ctrl = c;
1621 } else if (is_dominator(early_ctrl, c)) {
1622 early_ctrl = c;
1623 }
1624 }
1625 }
1626 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control");
1627 return early_ctrl;
1628 }
1629
ctrl_of_all_uses_out_of_loop(const Node * n,Node * n_ctrl,IdealLoopTree * n_loop)1630 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) {
1631 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1632 Node* u = n->fast_out(i);
1633 if (u->Opcode() == Op_Opaque1) {
1634 return false; // Found loop limit, bugfix for 4677003
1635 }
1636 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to
1637 // get_late_ctrl_with_anti_dep() use their own tag
1638 _dom_lca_tags_round++;
1639 assert(_dom_lca_tags_round != 0, "shouldn't wrap around");
1640
1641 if (u->is_Phi()) {
1642 for (uint j = 1; j < u->req(); ++j) {
1643 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) {
1644 return false;
1645 }
1646 }
1647 } else {
1648 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0);
1649 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) {
1650 return false;
1651 }
1652 }
1653 }
1654 return true;
1655 }
1656
ctrl_of_use_out_of_loop(const Node * n,Node * n_ctrl,IdealLoopTree * n_loop,Node * ctrl)1657 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) {
1658 if (n->is_Load()) {
1659 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl);
1660 }
1661 IdealLoopTree *u_loop = get_loop(ctrl);
1662 if (u_loop == n_loop) {
1663 return false; // Found loop-varying use
1664 }
1665 if (n_loop->is_member(u_loop)) {
1666 return false; // Found use in inner loop
1667 }
1668 return true;
1669 }
1670
1671 //------------------------------split_if_with_blocks---------------------------
1672 // Check for aggressive application of 'split-if' optimization,
1673 // using basic block level info.
split_if_with_blocks(VectorSet & visited,Node_Stack & nstack)1674 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
1675 Node* root = C->root();
1676 visited.set(root->_idx); // first, mark root as visited
1677 // Do pre-visit work for root
1678 Node* n = split_if_with_blocks_pre(root);
1679 uint cnt = n->outcnt();
1680 uint i = 0;
1681
1682 while (true) {
1683 // Visit all children
1684 if (i < cnt) {
1685 Node* use = n->raw_out(i);
1686 ++i;
1687 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1688 // Now do pre-visit work for this use
1689 use = split_if_with_blocks_pre(use);
1690 nstack.push(n, i); // Save parent and next use's index.
1691 n = use; // Process all children of current use.
1692 cnt = use->outcnt();
1693 i = 0;
1694 }
1695 }
1696 else {
1697 // All of n's children have been processed, complete post-processing.
1698 if (cnt != 0 && !n->is_Con()) {
1699 assert(has_node(n), "no dead nodes");
1700 split_if_with_blocks_post(n);
1701 }
1702 if (must_throttle_split_if()) {
1703 nstack.clear();
1704 }
1705 if (nstack.is_empty()) {
1706 // Finished all nodes on stack.
1707 break;
1708 }
1709 // Get saved parent node and next use's index. Visit the rest of uses.
1710 n = nstack.node();
1711 cnt = n->outcnt();
1712 i = nstack.index();
1713 nstack.pop();
1714 }
1715 }
1716 }
1717
1718
1719 //=============================================================================
1720 //
1721 // C L O N E A L O O P B O D Y
1722 //
1723
1724 //------------------------------clone_iff--------------------------------------
1725 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1726 // "Nearly" because all Nodes have been cloned from the original in the loop,
1727 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1728 // through the Phi recursively, and return a Bool.
clone_iff(PhiNode * phi,IdealLoopTree * loop)1729 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) {
1730
1731 // Convert this Phi into a Phi merging Bools
1732 uint i;
1733 for (i = 1; i < phi->req(); i++) {
1734 Node *b = phi->in(i);
1735 if (b->is_Phi()) {
1736 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop));
1737 } else {
1738 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
1739 }
1740 }
1741
1742 Node* n = phi->in(1);
1743 Node* sample_opaque = NULL;
1744 Node *sample_bool = NULL;
1745 if (n->Opcode() == Op_Opaque4) {
1746 sample_opaque = n;
1747 sample_bool = n->in(1);
1748 assert(sample_bool->is_Bool(), "wrong type");
1749 } else {
1750 sample_bool = n;
1751 }
1752 Node *sample_cmp = sample_bool->in(1);
1753
1754 // Make Phis to merge the Cmp's inputs.
1755 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
1756 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
1757 for (i = 1; i < phi->req(); i++) {
1758 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
1759 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
1760 phi1->set_req(i, n1);
1761 phi2->set_req(i, n2);
1762 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1763 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1764 }
1765 // See if these Phis have been made before.
1766 // Register with optimizer
1767 Node *hit1 = _igvn.hash_find_insert(phi1);
1768 if (hit1) { // Hit, toss just made Phi
1769 _igvn.remove_dead_node(phi1); // Remove new phi
1770 assert(hit1->is_Phi(), "" );
1771 phi1 = (PhiNode*)hit1; // Use existing phi
1772 } else { // Miss
1773 _igvn.register_new_node_with_optimizer(phi1);
1774 }
1775 Node *hit2 = _igvn.hash_find_insert(phi2);
1776 if (hit2) { // Hit, toss just made Phi
1777 _igvn.remove_dead_node(phi2); // Remove new phi
1778 assert(hit2->is_Phi(), "" );
1779 phi2 = (PhiNode*)hit2; // Use existing phi
1780 } else { // Miss
1781 _igvn.register_new_node_with_optimizer(phi2);
1782 }
1783 // Register Phis with loop/block info
1784 set_ctrl(phi1, phi->in(0));
1785 set_ctrl(phi2, phi->in(0));
1786 // Make a new Cmp
1787 Node *cmp = sample_cmp->clone();
1788 cmp->set_req(1, phi1);
1789 cmp->set_req(2, phi2);
1790 _igvn.register_new_node_with_optimizer(cmp);
1791 set_ctrl(cmp, phi->in(0));
1792
1793 // Make a new Bool
1794 Node *b = sample_bool->clone();
1795 b->set_req(1,cmp);
1796 _igvn.register_new_node_with_optimizer(b);
1797 set_ctrl(b, phi->in(0));
1798
1799 if (sample_opaque != NULL) {
1800 Node* opaque = sample_opaque->clone();
1801 opaque->set_req(1, b);
1802 _igvn.register_new_node_with_optimizer(opaque);
1803 set_ctrl(opaque, phi->in(0));
1804 return opaque;
1805 }
1806
1807 assert(b->is_Bool(), "");
1808 return b;
1809 }
1810
1811 //------------------------------clone_bool-------------------------------------
1812 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1813 // "Nearly" because all Nodes have been cloned from the original in the loop,
1814 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1815 // through the Phi recursively, and return a Bool.
clone_bool(PhiNode * phi,IdealLoopTree * loop)1816 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1817 uint i;
1818 // Convert this Phi into a Phi merging Bools
1819 for( i = 1; i < phi->req(); i++ ) {
1820 Node *b = phi->in(i);
1821 if( b->is_Phi() ) {
1822 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1823 } else {
1824 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1825 }
1826 }
1827
1828 Node *sample_cmp = phi->in(1);
1829
1830 // Make Phis to merge the Cmp's inputs.
1831 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1832 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1833 for( uint j = 1; j < phi->req(); j++ ) {
1834 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1835 Node *n1, *n2;
1836 if( cmp_top->is_Cmp() ) {
1837 n1 = cmp_top->in(1);
1838 n2 = cmp_top->in(2);
1839 } else {
1840 n1 = n2 = cmp_top;
1841 }
1842 phi1->set_req( j, n1 );
1843 phi2->set_req( j, n2 );
1844 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1845 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1846 }
1847
1848 // See if these Phis have been made before.
1849 // Register with optimizer
1850 Node *hit1 = _igvn.hash_find_insert(phi1);
1851 if( hit1 ) { // Hit, toss just made Phi
1852 _igvn.remove_dead_node(phi1); // Remove new phi
1853 assert( hit1->is_Phi(), "" );
1854 phi1 = (PhiNode*)hit1; // Use existing phi
1855 } else { // Miss
1856 _igvn.register_new_node_with_optimizer(phi1);
1857 }
1858 Node *hit2 = _igvn.hash_find_insert(phi2);
1859 if( hit2 ) { // Hit, toss just made Phi
1860 _igvn.remove_dead_node(phi2); // Remove new phi
1861 assert( hit2->is_Phi(), "" );
1862 phi2 = (PhiNode*)hit2; // Use existing phi
1863 } else { // Miss
1864 _igvn.register_new_node_with_optimizer(phi2);
1865 }
1866 // Register Phis with loop/block info
1867 set_ctrl(phi1, phi->in(0));
1868 set_ctrl(phi2, phi->in(0));
1869 // Make a new Cmp
1870 Node *cmp = sample_cmp->clone();
1871 cmp->set_req( 1, phi1 );
1872 cmp->set_req( 2, phi2 );
1873 _igvn.register_new_node_with_optimizer(cmp);
1874 set_ctrl(cmp, phi->in(0));
1875
1876 assert( cmp->is_Cmp(), "" );
1877 return (CmpNode*)cmp;
1878 }
1879
1880 //------------------------------sink_use---------------------------------------
1881 // If 'use' was in the loop-exit block, it now needs to be sunk
1882 // below the post-loop merge point.
sink_use(Node * use,Node * post_loop)1883 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1884 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1885 set_ctrl(use, post_loop);
1886 for (DUIterator j = use->outs(); use->has_out(j); j++)
1887 sink_use(use->out(j), post_loop);
1888 }
1889 }
1890
clone_loop_handle_data_uses(Node * old,Node_List & old_new,IdealLoopTree * loop,IdealLoopTree * outer_loop,Node_List * & split_if_set,Node_List * & split_bool_set,Node_List * & split_cex_set,Node_List & worklist,uint new_counter,CloneLoopMode mode)1891 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1892 IdealLoopTree* loop, IdealLoopTree* outer_loop,
1893 Node_List*& split_if_set, Node_List*& split_bool_set,
1894 Node_List*& split_cex_set, Node_List& worklist,
1895 uint new_counter, CloneLoopMode mode) {
1896 Node* nnn = old_new[old->_idx];
1897 // Copy uses to a worklist, so I can munge the def-use info
1898 // with impunity.
1899 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1900 worklist.push(old->fast_out(j));
1901
1902 while( worklist.size() ) {
1903 Node *use = worklist.pop();
1904 if (!has_node(use)) continue; // Ignore dead nodes
1905 if (use->in(0) == C->top()) continue;
1906 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1907 // Check for data-use outside of loop - at least one of OLD or USE
1908 // must not be a CFG node.
1909 #ifdef ASSERT
1910 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) {
1911 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
1912 assert(mode != IgnoreStripMined, "incorrect cloning mode");
1913 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
1914 }
1915 #endif
1916 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
1917
1918 // If the Data use is an IF, that means we have an IF outside of the
1919 // loop that is switching on a condition that is set inside of the
1920 // loop. Happens if people set a loop-exit flag; then test the flag
1921 // in the loop to break the loop, then test is again outside of the
1922 // loop to determine which way the loop exited.
1923 // Loop predicate If node connects to Bool node through Opaque1 node.
1924 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) {
1925 // Since this code is highly unlikely, we lazily build the worklist
1926 // of such Nodes to go split.
1927 if (!split_if_set) {
1928 split_if_set = new Node_List();
1929 }
1930 split_if_set->push(use);
1931 }
1932 if (use->is_Bool()) {
1933 if (!split_bool_set) {
1934 split_bool_set = new Node_List();
1935 }
1936 split_bool_set->push(use);
1937 }
1938 if (use->Opcode() == Op_CreateEx) {
1939 if (!split_cex_set) {
1940 split_cex_set = new Node_List();
1941 }
1942 split_cex_set->push(use);
1943 }
1944
1945
1946 // Get "block" use is in
1947 uint idx = 0;
1948 while( use->in(idx) != old ) idx++;
1949 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1950 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
1951 Node *cfg = prev->_idx >= new_counter
1952 ? prev->in(2)
1953 : idom(prev);
1954 if( use->is_Phi() ) // Phi use is in prior block
1955 cfg = prev->in(idx); // NOT in block of Phi itself
1956 if (cfg->is_top()) { // Use is dead?
1957 _igvn.replace_input_of(use, idx, C->top());
1958 continue;
1959 }
1960
1961 // If use is referenced through control edge... (idx == 0)
1962 if (mode == IgnoreStripMined && idx == 0) {
1963 LoopNode *head = loop->_head->as_Loop();
1964 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
1965 // That node is outside the inner loop, leave it outside the
1966 // outer loop as well to not confuse verification code.
1967 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
1968 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
1969 continue;
1970 }
1971 }
1972
1973 while(!outer_loop->is_member(get_loop(cfg))) {
1974 prev = cfg;
1975 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1976 }
1977 // If the use occurs after merging several exits from the loop, then
1978 // old value must have dominated all those exits. Since the same old
1979 // value was used on all those exits we did not need a Phi at this
1980 // merge point. NOW we do need a Phi here. Each loop exit value
1981 // is now merged with the peeled body exit; each exit gets its own
1982 // private Phi and those Phis need to be merged here.
1983 Node *phi;
1984 if( prev->is_Region() ) {
1985 if( idx == 0 ) { // Updating control edge?
1986 phi = prev; // Just use existing control
1987 } else { // Else need a new Phi
1988 phi = PhiNode::make( prev, old );
1989 // Now recursively fix up the new uses of old!
1990 for( uint i = 1; i < prev->req(); i++ ) {
1991 worklist.push(phi); // Onto worklist once for each 'old' input
1992 }
1993 }
1994 } else {
1995 // Get new RegionNode merging old and new loop exits
1996 prev = old_new[prev->_idx];
1997 assert( prev, "just made this in step 7" );
1998 if( idx == 0) { // Updating control edge?
1999 phi = prev; // Just use existing control
2000 } else { // Else need a new Phi
2001 // Make a new Phi merging data values properly
2002 phi = PhiNode::make( prev, old );
2003 phi->set_req( 1, nnn );
2004 }
2005 }
2006 // If inserting a new Phi, check for prior hits
2007 if( idx != 0 ) {
2008 Node *hit = _igvn.hash_find_insert(phi);
2009 if( hit == NULL ) {
2010 _igvn.register_new_node_with_optimizer(phi); // Register new phi
2011 } else { // or
2012 // Remove the new phi from the graph and use the hit
2013 _igvn.remove_dead_node(phi);
2014 phi = hit; // Use existing phi
2015 }
2016 set_ctrl(phi, prev);
2017 }
2018 // Make 'use' use the Phi instead of the old loop body exit value
2019 _igvn.replace_input_of(use, idx, phi);
2020 if( use->_idx >= new_counter ) { // If updating new phis
2021 // Not needed for correctness, but prevents a weak assert
2022 // in AddPNode from tripping (when we end up with different
2023 // base & derived Phis that will become the same after
2024 // IGVN does CSE).
2025 Node *hit = _igvn.hash_find_insert(use);
2026 if( hit ) // Go ahead and re-hash for hits.
2027 _igvn.replace_node( use, hit );
2028 }
2029
2030 // If 'use' was in the loop-exit block, it now needs to be sunk
2031 // below the post-loop merge point.
2032 sink_use( use, prev );
2033 }
2034 }
2035 }
2036
clone_outer_loop_helper(Node * n,const IdealLoopTree * loop,const IdealLoopTree * outer_loop,const Node_List & old_new,Unique_Node_List & wq,PhaseIdealLoop * phase,bool check_old_new)2037 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
2038 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
2039 bool check_old_new) {
2040 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2041 Node* u = n->fast_out(j);
2042 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned");
2043 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) {
2044 Node* c = phase->get_ctrl(u);
2045 IdealLoopTree* u_loop = phase->get_loop(c);
2046 assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only");
2047 if (outer_loop->is_member(u_loop) ||
2048 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
2049 // the outer loop too
2050 (u->in(0) != NULL && outer_loop->is_member(phase->get_loop(u->in(0))))) {
2051 wq.push(u);
2052 }
2053 }
2054 }
2055 }
2056
clone_outer_loop(LoopNode * head,CloneLoopMode mode,IdealLoopTree * loop,IdealLoopTree * outer_loop,int dd,Node_List & old_new,Node_List & extra_data_nodes)2057 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2058 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2059 Node_List& extra_data_nodes) {
2060 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2061 CountedLoopNode* cl = head->as_CountedLoop();
2062 Node* l = cl->outer_loop();
2063 Node* tail = cl->outer_loop_tail();
2064 IfNode* le = cl->outer_loop_end();
2065 Node* sfpt = cl->outer_safepoint();
2066 CountedLoopEndNode* cle = cl->loopexit();
2067 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2068 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2069 Node* cle_out = cle->proj_out(false);
2070
2071 Node* new_sfpt = NULL;
2072 Node* new_cle_out = cle_out->clone();
2073 old_new.map(cle_out->_idx, new_cle_out);
2074 if (mode == CloneIncludesStripMined) {
2075 // clone outer loop body
2076 Node* new_l = l->clone();
2077 Node* new_tail = tail->clone();
2078 IfNode* new_le = le->clone()->as_If();
2079 new_sfpt = sfpt->clone();
2080
2081 set_loop(new_l, outer_loop->_parent);
2082 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2083 set_loop(new_cle_out, outer_loop->_parent);
2084 set_idom(new_cle_out, new_cle, dd);
2085 set_loop(new_sfpt, outer_loop->_parent);
2086 set_idom(new_sfpt, new_cle_out, dd);
2087 set_loop(new_le, outer_loop->_parent);
2088 set_idom(new_le, new_sfpt, dd);
2089 set_loop(new_tail, outer_loop->_parent);
2090 set_idom(new_tail, new_le, dd);
2091 set_idom(new_cl, new_l, dd);
2092
2093 old_new.map(l->_idx, new_l);
2094 old_new.map(tail->_idx, new_tail);
2095 old_new.map(le->_idx, new_le);
2096 old_new.map(sfpt->_idx, new_sfpt);
2097
2098 new_l->set_req(LoopNode::LoopBackControl, new_tail);
2099 new_l->set_req(0, new_l);
2100 new_tail->set_req(0, new_le);
2101 new_le->set_req(0, new_sfpt);
2102 new_sfpt->set_req(0, new_cle_out);
2103 new_cle_out->set_req(0, new_cle);
2104 new_cl->set_req(LoopNode::EntryControl, new_l);
2105
2106 _igvn.register_new_node_with_optimizer(new_l);
2107 _igvn.register_new_node_with_optimizer(new_tail);
2108 _igvn.register_new_node_with_optimizer(new_le);
2109 } else {
2110 Node *newhead = old_new[loop->_head->_idx];
2111 newhead->as_Loop()->clear_strip_mined();
2112 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2113 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2114 }
2115 // Look at data node that were assigned a control in the outer
2116 // loop: they are kept in the outer loop by the safepoint so start
2117 // from the safepoint node's inputs.
2118 IdealLoopTree* outer_loop = get_loop(l);
2119 Node_Stack stack(2);
2120 stack.push(sfpt, 1);
2121 uint new_counter = C->unique();
2122 while (stack.size() > 0) {
2123 Node* n = stack.node();
2124 uint i = stack.index();
2125 while (i < n->req() &&
2126 (n->in(i) == NULL ||
2127 !has_ctrl(n->in(i)) ||
2128 get_loop(get_ctrl(n->in(i))) != outer_loop ||
2129 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2130 i++;
2131 }
2132 if (i < n->req()) {
2133 stack.set_index(i+1);
2134 stack.push(n->in(i), 0);
2135 } else {
2136 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2137 Node* m = n == sfpt ? new_sfpt : n->clone();
2138 if (m != NULL) {
2139 for (uint i = 0; i < n->req(); i++) {
2140 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) {
2141 m->set_req(i, old_new[m->in(i)->_idx]);
2142 }
2143 }
2144 } else {
2145 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2146 }
2147 if (n != sfpt) {
2148 extra_data_nodes.push(n);
2149 _igvn.register_new_node_with_optimizer(m);
2150 assert(get_ctrl(n) == cle_out, "what other control?");
2151 set_ctrl(m, new_cle_out);
2152 old_new.map(n->_idx, m);
2153 }
2154 stack.pop();
2155 }
2156 }
2157 if (mode == CloneIncludesStripMined) {
2158 _igvn.register_new_node_with_optimizer(new_sfpt);
2159 _igvn.register_new_node_with_optimizer(new_cle_out);
2160 }
2161 // Some other transformation may have pessimistically assign some
2162 // data nodes to the outer loop. Set their control so they are out
2163 // of the outer loop.
2164 ResourceMark rm;
2165 Unique_Node_List wq;
2166 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2167 Node* old = extra_data_nodes.at(i);
2168 clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true);
2169 }
2170
2171 Node* inner_out = sfpt->in(0);
2172 if (inner_out->outcnt() > 1) {
2173 clone_outer_loop_helper(inner_out, loop, outer_loop, old_new, wq, this, true);
2174 }
2175
2176 Node* new_ctrl = cl->outer_loop_exit();
2177 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2178 for (uint i = 0; i < wq.size(); i++) {
2179 Node* n = wq.at(i);
2180 set_ctrl(n, new_ctrl);
2181 if (n->in(0) != NULL) {
2182 _igvn.replace_input_of(n, 0, new_ctrl);
2183 }
2184 clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false);
2185 }
2186 } else {
2187 Node *newhead = old_new[loop->_head->_idx];
2188 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2189 }
2190 }
2191
2192 //------------------------------clone_loop-------------------------------------
2193 //
2194 // C L O N E A L O O P B O D Y
2195 //
2196 // This is the basic building block of the loop optimizations. It clones an
2197 // entire loop body. It makes an old_new loop body mapping; with this mapping
2198 // you can find the new-loop equivalent to an old-loop node. All new-loop
2199 // nodes are exactly equal to their old-loop counterparts, all edges are the
2200 // same. All exits from the old-loop now have a RegionNode that merges the
2201 // equivalent new-loop path. This is true even for the normal "loop-exit"
2202 // condition. All uses of loop-invariant old-loop values now come from (one
2203 // or more) Phis that merge their new-loop equivalents.
2204 //
2205 // This operation leaves the graph in an illegal state: there are two valid
2206 // control edges coming from the loop pre-header to both loop bodies. I'll
2207 // definitely have to hack the graph after running this transform.
2208 //
2209 // From this building block I will further edit edges to perform loop peeling
2210 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2211 //
2212 // Parameter side_by_size_idom:
2213 // When side_by_size_idom is NULL, the dominator tree is constructed for
2214 // the clone loop to dominate the original. Used in construction of
2215 // pre-main-post loop sequence.
2216 // When nonnull, the clone and original are side-by-side, both are
2217 // dominated by the side_by_side_idom node. Used in construction of
2218 // unswitched loops.
clone_loop(IdealLoopTree * loop,Node_List & old_new,int dd,CloneLoopMode mode,Node * side_by_side_idom)2219 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2220 CloneLoopMode mode, Node* side_by_side_idom) {
2221
2222 LoopNode* head = loop->_head->as_Loop();
2223 head->verify_strip_mined(1);
2224
2225 if (C->do_vector_loop() && PrintOpto) {
2226 const char* mname = C->method()->name()->as_quoted_ascii();
2227 if (mname != NULL) {
2228 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2229 }
2230 }
2231
2232 CloneMap& cm = C->clone_map();
2233 Dict* dict = cm.dict();
2234 if (C->do_vector_loop()) {
2235 cm.set_clone_idx(cm.max_gen()+1);
2236 #ifndef PRODUCT
2237 if (PrintOpto) {
2238 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2239 loop->dump_head();
2240 }
2241 #endif
2242 }
2243
2244 // Step 1: Clone the loop body. Make the old->new mapping.
2245 uint i;
2246 for (i = 0; i < loop->_body.size(); i++) {
2247 Node* old = loop->_body.at(i);
2248 Node* nnn = old->clone();
2249 old_new.map(old->_idx, nnn);
2250 if (old->is_reduction()) {
2251 // Reduction flag is not copied by default. Copy it here when cloning the entire loop body.
2252 nnn->add_flag(Node::Flag_is_reduction);
2253 }
2254 if (C->do_vector_loop()) {
2255 cm.verify_insert_and_clone(old, nnn, cm.clone_idx());
2256 }
2257 _igvn.register_new_node_with_optimizer(nnn);
2258 }
2259
2260 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2261
2262 // Step 2: Fix the edges in the new body. If the old input is outside the
2263 // loop use it. If the old input is INside the loop, use the corresponding
2264 // new node instead.
2265 for( i = 0; i < loop->_body.size(); i++ ) {
2266 Node *old = loop->_body.at(i);
2267 Node *nnn = old_new[old->_idx];
2268 // Fix CFG/Loop controlling the new node
2269 if (has_ctrl(old)) {
2270 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2271 } else {
2272 set_loop(nnn, outer_loop->_parent);
2273 if (old->outcnt() > 0) {
2274 set_idom( nnn, old_new[idom(old)->_idx], dd );
2275 }
2276 }
2277 // Correct edges to the new node
2278 for( uint j = 0; j < nnn->req(); j++ ) {
2279 Node *n = nnn->in(j);
2280 if( n ) {
2281 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
2282 if( loop->is_member( old_in_loop ) )
2283 nnn->set_req(j, old_new[n->_idx]);
2284 }
2285 }
2286 _igvn.hash_find_insert(nnn);
2287 }
2288
2289 Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2290 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2291
2292 // Step 3: Now fix control uses. Loop varying control uses have already
2293 // been fixed up (as part of all input edges in Step 2). Loop invariant
2294 // control uses must be either an IfFalse or an IfTrue. Make a merge
2295 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2296 // refer to this.
2297 Node_List worklist;
2298 uint new_counter = C->unique();
2299 for( i = 0; i < loop->_body.size(); i++ ) {
2300 Node* old = loop->_body.at(i);
2301 if( !old->is_CFG() ) continue;
2302
2303 // Copy uses to a worklist, so I can munge the def-use info
2304 // with impunity.
2305 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2306 worklist.push(old->fast_out(j));
2307
2308 while( worklist.size() ) { // Visit all uses
2309 Node *use = worklist.pop();
2310 if (!has_node(use)) continue; // Ignore dead nodes
2311 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2312 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
2313 // Both OLD and USE are CFG nodes here.
2314 assert( use->is_Proj(), "" );
2315 Node* nnn = old_new[old->_idx];
2316
2317 Node* newuse = NULL;
2318 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2319 CountedLoopNode* cl = head->as_CountedLoop();
2320 CountedLoopEndNode* cle = cl->loopexit();
2321 Node* cle_out = cle->proj_out_or_null(false);
2322 if (use == cle_out) {
2323 IfNode* le = cl->outer_loop_end();
2324 use = le->proj_out(false);
2325 use_loop = get_loop(use);
2326 if (mode == CloneIncludesStripMined) {
2327 nnn = old_new[le->_idx];
2328 } else {
2329 newuse = old_new[cle_out->_idx];
2330 }
2331 }
2332 }
2333 if (newuse == NULL) {
2334 newuse = use->clone();
2335 }
2336
2337 // Clone the loop exit control projection
2338 if (C->do_vector_loop()) {
2339 cm.verify_insert_and_clone(use, newuse, cm.clone_idx());
2340 }
2341 newuse->set_req(0,nnn);
2342 _igvn.register_new_node_with_optimizer(newuse);
2343 set_loop(newuse, use_loop);
2344 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2345
2346 // We need a Region to merge the exit from the peeled body and the
2347 // exit from the old loop body.
2348 RegionNode *r = new RegionNode(3);
2349 // Map the old use to the new merge point
2350 old_new.map( use->_idx, r );
2351 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
2352 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
2353
2354 // The original user of 'use' uses 'r' instead.
2355 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2356 Node* useuse = use->last_out(l);
2357 _igvn.rehash_node_delayed(useuse);
2358 uint uses_found = 0;
2359 if( useuse->in(0) == use ) {
2360 useuse->set_req(0, r);
2361 uses_found++;
2362 if( useuse->is_CFG() ) {
2363 // This is not a dom_depth > dd_r because when new
2364 // control flow is constructed by a loop opt, a node and
2365 // its dominator can end up at the same dom_depth
2366 assert(dom_depth(useuse) >= dd_r, "");
2367 set_idom(useuse, r, dom_depth(useuse));
2368 }
2369 }
2370 for( uint k = 1; k < useuse->req(); k++ ) {
2371 if( useuse->in(k) == use ) {
2372 useuse->set_req(k, r);
2373 uses_found++;
2374 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2375 // This is not a dom_depth > dd_r because when new
2376 // control flow is constructed by a loop opt, a node
2377 // and its dominator can end up at the same dom_depth
2378 assert(dom_depth(useuse) >= dd_r , "");
2379 set_idom(useuse, r, dom_depth(useuse));
2380 }
2381 }
2382 }
2383 l -= uses_found; // we deleted 1 or more copies of this edge
2384 }
2385
2386 // Now finish up 'r'
2387 r->set_req( 1, newuse );
2388 r->set_req( 2, use );
2389 _igvn.register_new_node_with_optimizer(r);
2390 set_loop(r, use_loop);
2391 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
2392 } // End of if a loop-exit test
2393 }
2394 }
2395
2396 // Step 4: If loop-invariant use is not control, it must be dominated by a
2397 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2398 // there if needed. Make a Phi there merging old and new used values.
2399 Node_List *split_if_set = NULL;
2400 Node_List *split_bool_set = NULL;
2401 Node_List *split_cex_set = NULL;
2402 for( i = 0; i < loop->_body.size(); i++ ) {
2403 Node* old = loop->_body.at(i);
2404 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2405 split_bool_set, split_cex_set, worklist, new_counter,
2406 mode);
2407 }
2408
2409 for (i = 0; i < extra_data_nodes.size(); i++) {
2410 Node* old = extra_data_nodes.at(i);
2411 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2412 split_bool_set, split_cex_set, worklist, new_counter,
2413 mode);
2414 }
2415
2416 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2417 // the loop uses a condition set in the loop. The original IF probably
2418 // takes control from one or more OLD Regions (which in turn get from NEW
2419 // Regions). In any case, there will be a set of Phis for each merge point
2420 // from the IF up to where the original BOOL def exists the loop.
2421 if (split_if_set) {
2422 while (split_if_set->size()) {
2423 Node *iff = split_if_set->pop();
2424 if (iff->in(1)->is_Phi()) {
2425 Node *b = clone_iff(iff->in(1)->as_Phi(), loop);
2426 _igvn.replace_input_of(iff, 1, b);
2427 }
2428 }
2429 }
2430 if (split_bool_set) {
2431 while (split_bool_set->size()) {
2432 Node *b = split_bool_set->pop();
2433 Node *phi = b->in(1);
2434 assert(phi->is_Phi(), "");
2435 CmpNode *cmp = clone_bool((PhiNode*)phi, loop);
2436 _igvn.replace_input_of(b, 1, cmp);
2437 }
2438 }
2439 if (split_cex_set) {
2440 while (split_cex_set->size()) {
2441 Node *b = split_cex_set->pop();
2442 assert(b->in(0)->is_Region(), "");
2443 assert(b->in(1)->is_Phi(), "");
2444 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2445 split_up(b, b->in(0), NULL);
2446 }
2447 }
2448
2449 }
2450
2451
2452 //---------------------- stride_of_possible_iv -------------------------------------
2453 // Looks for an iff/bool/comp with one operand of the compare
2454 // being a cycle involving an add and a phi,
2455 // with an optional truncation (left-shift followed by a right-shift)
2456 // of the add. Returns zero if not an iv.
stride_of_possible_iv(Node * iff)2457 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2458 Node* trunc1 = NULL;
2459 Node* trunc2 = NULL;
2460 const TypeInteger* ttype = NULL;
2461 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
2462 return 0;
2463 }
2464 BoolNode* bl = iff->in(1)->as_Bool();
2465 Node* cmp = bl->in(1);
2466 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2467 return 0;
2468 }
2469 // Must have an invariant operand
2470 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2471 return 0;
2472 }
2473 Node* add2 = NULL;
2474 Node* cmp1 = cmp->in(1);
2475 if (cmp1->is_Phi()) {
2476 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2477 Node* phi = cmp1;
2478 for (uint i = 1; i < phi->req(); i++) {
2479 Node* in = phi->in(i);
2480 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2481 &trunc1, &trunc2, &ttype, T_INT);
2482 if (add && add->in(1) == phi) {
2483 add2 = add->in(2);
2484 break;
2485 }
2486 }
2487 } else {
2488 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2489 Node* addtrunc = cmp1;
2490 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2491 &trunc1, &trunc2, &ttype, T_INT);
2492 if (add && add->in(1)->is_Phi()) {
2493 Node* phi = add->in(1);
2494 for (uint i = 1; i < phi->req(); i++) {
2495 if (phi->in(i) == addtrunc) {
2496 add2 = add->in(2);
2497 break;
2498 }
2499 }
2500 }
2501 }
2502 if (add2 != NULL) {
2503 const TypeInt* add2t = _igvn.type(add2)->is_int();
2504 if (add2t->is_con()) {
2505 return add2t->get_con();
2506 }
2507 }
2508 return 0;
2509 }
2510
2511
2512 //---------------------- stay_in_loop -------------------------------------
2513 // Return the (unique) control output node that's in the loop (if it exists.)
stay_in_loop(Node * n,IdealLoopTree * loop)2514 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2515 Node* unique = NULL;
2516 if (!n) return NULL;
2517 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2518 Node* use = n->fast_out(i);
2519 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2520 if (unique != NULL) {
2521 return NULL;
2522 }
2523 unique = use;
2524 }
2525 }
2526 return unique;
2527 }
2528
2529 //------------------------------ register_node -------------------------------------
2530 // Utility to register node "n" with PhaseIdealLoop
register_node(Node * n,IdealLoopTree * loop,Node * pred,int ddepth)2531 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
2532 _igvn.register_new_node_with_optimizer(n);
2533 loop->_body.push(n);
2534 if (n->is_CFG()) {
2535 set_loop(n, loop);
2536 set_idom(n, pred, ddepth);
2537 } else {
2538 set_ctrl(n, pred);
2539 }
2540 }
2541
2542 //------------------------------ proj_clone -------------------------------------
2543 // Utility to create an if-projection
proj_clone(ProjNode * p,IfNode * iff)2544 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2545 ProjNode* c = p->clone()->as_Proj();
2546 c->set_req(0, iff);
2547 return c;
2548 }
2549
2550 //------------------------------ short_circuit_if -------------------------------------
2551 // Force the iff control output to be the live_proj
short_circuit_if(IfNode * iff,ProjNode * live_proj)2552 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2553 guarantee(live_proj != NULL, "null projection");
2554 int proj_con = live_proj->_con;
2555 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2556 Node *con = _igvn.intcon(proj_con);
2557 set_ctrl(con, C->root());
2558 if (iff) {
2559 iff->set_req(1, con);
2560 }
2561 return con;
2562 }
2563
2564 //------------------------------ insert_if_before_proj -------------------------------------
2565 // Insert a new if before an if projection (* - new node)
2566 //
2567 // before
2568 // if(test)
2569 // / \
2570 // v v
2571 // other-proj proj (arg)
2572 //
2573 // after
2574 // if(test)
2575 // / \
2576 // / v
2577 // | * proj-clone
2578 // v |
2579 // other-proj v
2580 // * new_if(relop(cmp[IU](left,right)))
2581 // / \
2582 // v v
2583 // * new-proj proj
2584 // (returned)
2585 //
insert_if_before_proj(Node * left,bool Signed,BoolTest::mask relop,Node * right,ProjNode * proj)2586 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2587 IfNode* iff = proj->in(0)->as_If();
2588 IdealLoopTree *loop = get_loop(proj);
2589 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2590 int ddepth = dom_depth(proj);
2591
2592 _igvn.rehash_node_delayed(iff);
2593 _igvn.rehash_node_delayed(proj);
2594
2595 proj->set_req(0, NULL); // temporary disconnect
2596 ProjNode* proj2 = proj_clone(proj, iff);
2597 register_node(proj2, loop, iff, ddepth);
2598
2599 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2600 register_node(cmp, loop, proj2, ddepth);
2601
2602 BoolNode* bol = new BoolNode(cmp, relop);
2603 register_node(bol, loop, proj2, ddepth);
2604
2605 int opcode = iff->Opcode();
2606 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2607 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
2608 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
2609 register_node(new_if, loop, proj2, ddepth);
2610
2611 proj->set_req(0, new_if); // reattach
2612 set_idom(proj, new_if, ddepth);
2613
2614 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2615 guarantee(new_exit != NULL, "null exit node");
2616 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2617
2618 return new_exit;
2619 }
2620
2621 //------------------------------ insert_region_before_proj -------------------------------------
2622 // Insert a region before an if projection (* - new node)
2623 //
2624 // before
2625 // if(test)
2626 // / |
2627 // v |
2628 // proj v
2629 // other-proj
2630 //
2631 // after
2632 // if(test)
2633 // / |
2634 // v |
2635 // * proj-clone v
2636 // | other-proj
2637 // v
2638 // * new-region
2639 // |
2640 // v
2641 // * dum_if
2642 // / \
2643 // v \
2644 // * dum-proj v
2645 // proj
2646 //
insert_region_before_proj(ProjNode * proj)2647 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2648 IfNode* iff = proj->in(0)->as_If();
2649 IdealLoopTree *loop = get_loop(proj);
2650 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2651 int ddepth = dom_depth(proj);
2652
2653 _igvn.rehash_node_delayed(iff);
2654 _igvn.rehash_node_delayed(proj);
2655
2656 proj->set_req(0, NULL); // temporary disconnect
2657 ProjNode* proj2 = proj_clone(proj, iff);
2658 register_node(proj2, loop, iff, ddepth);
2659
2660 RegionNode* reg = new RegionNode(2);
2661 reg->set_req(1, proj2);
2662 register_node(reg, loop, iff, ddepth);
2663
2664 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
2665 register_node(dum_if, loop, reg, ddepth);
2666
2667 proj->set_req(0, dum_if); // reattach
2668 set_idom(proj, dum_if, ddepth);
2669
2670 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
2671 register_node(dum_proj, loop, dum_if, ddepth);
2672
2673 return reg;
2674 }
2675
2676 //------------------------------ insert_cmpi_loop_exit -------------------------------------
2677 // Clone a signed compare loop exit from an unsigned compare and
2678 // insert it before the unsigned cmp on the stay-in-loop path.
2679 // All new nodes inserted in the dominator tree between the original
2680 // if and it's projections. The original if test is replaced with
2681 // a constant to force the stay-in-loop path.
2682 //
2683 // This is done to make sure that the original if and it's projections
2684 // still dominate the same set of control nodes, that the ctrl() relation
2685 // from data nodes to them is preserved, and that their loop nesting is
2686 // preserved.
2687 //
2688 // before
2689 // if(i <u limit) unsigned compare loop exit
2690 // / |
2691 // v v
2692 // exit-proj stay-in-loop-proj
2693 //
2694 // after
2695 // if(stay-in-loop-const) original if
2696 // / |
2697 // / v
2698 // / if(i < limit) new signed test
2699 // / / |
2700 // / / v
2701 // / / if(i <u limit) new cloned unsigned test
2702 // / / / |
2703 // v v v |
2704 // region |
2705 // | |
2706 // dum-if |
2707 // / | |
2708 // ether | |
2709 // v v
2710 // exit-proj stay-in-loop-proj
2711 //
insert_cmpi_loop_exit(IfNode * if_cmpu,IdealLoopTree * loop)2712 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
2713 const bool Signed = true;
2714 const bool Unsigned = false;
2715
2716 BoolNode* bol = if_cmpu->in(1)->as_Bool();
2717 if (bol->_test._test != BoolTest::lt) return NULL;
2718 CmpNode* cmpu = bol->in(1)->as_Cmp();
2719 if (cmpu->Opcode() != Op_CmpU) return NULL;
2720 int stride = stride_of_possible_iv(if_cmpu);
2721 if (stride == 0) return NULL;
2722
2723 Node* lp_proj = stay_in_loop(if_cmpu, loop);
2724 guarantee(lp_proj != NULL, "null loop node");
2725
2726 ProjNode* lp_continue = lp_proj->as_Proj();
2727 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
2728 if (!lp_exit->is_IfFalse()) {
2729 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
2730 // We therefore can't add a single exit condition.
2731 return NULL;
2732 }
2733 // The loop exit condition is !(i <u limit) ==> (i < 0 || i >= limit).
2734 // Split out the exit condition (i < 0) for stride < 0 or (i >= limit) for stride > 0.
2735 Node* limit = NULL;
2736 if (stride > 0) {
2737 limit = cmpu->in(2);
2738 } else {
2739 limit = _igvn.makecon(TypeInt::ZERO);
2740 set_ctrl(limit, C->root());
2741 }
2742 // Create a new region on the exit path
2743 RegionNode* reg = insert_region_before_proj(lp_exit);
2744 guarantee(reg != NULL, "null region node");
2745
2746 // Clone the if-cmpu-true-false using a signed compare
2747 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
2748 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
2749 reg->add_req(cmpi_exit);
2750
2751 // Clone the if-cmpu-true-false
2752 BoolTest::mask rel_u = bol->_test._test;
2753 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
2754 reg->add_req(cmpu_exit);
2755
2756 // Force original if to stay in loop.
2757 short_circuit_if(if_cmpu, lp_continue);
2758
2759 return cmpi_exit->in(0)->as_If();
2760 }
2761
2762 //------------------------------ remove_cmpi_loop_exit -------------------------------------
2763 // Remove a previously inserted signed compare loop exit.
remove_cmpi_loop_exit(IfNode * if_cmp,IdealLoopTree * loop)2764 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
2765 Node* lp_proj = stay_in_loop(if_cmp, loop);
2766 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
2767 stay_in_loop(lp_proj, loop)->is_If() &&
2768 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
2769 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
2770 set_ctrl(con, C->root());
2771 if_cmp->set_req(1, con);
2772 }
2773
2774 //------------------------------ scheduled_nodelist -------------------------------------
2775 // Create a post order schedule of nodes that are in the
2776 // "member" set. The list is returned in "sched".
2777 // The first node in "sched" is the loop head, followed by
2778 // nodes which have no inputs in the "member" set, and then
2779 // followed by the nodes that have an immediate input dependence
2780 // on a node in "sched".
scheduled_nodelist(IdealLoopTree * loop,VectorSet & member,Node_List & sched)2781 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
2782
2783 assert(member.test(loop->_head->_idx), "loop head must be in member set");
2784 VectorSet visited;
2785 Node_Stack nstack(loop->_body.size());
2786
2787 Node* n = loop->_head; // top of stack is cached in "n"
2788 uint idx = 0;
2789 visited.set(n->_idx);
2790
2791 // Initially push all with no inputs from within member set
2792 for(uint i = 0; i < loop->_body.size(); i++ ) {
2793 Node *elt = loop->_body.at(i);
2794 if (member.test(elt->_idx)) {
2795 bool found = false;
2796 for (uint j = 0; j < elt->req(); j++) {
2797 Node* def = elt->in(j);
2798 if (def && member.test(def->_idx) && def != elt) {
2799 found = true;
2800 break;
2801 }
2802 }
2803 if (!found && elt != loop->_head) {
2804 nstack.push(n, idx);
2805 n = elt;
2806 assert(!visited.test(n->_idx), "not seen yet");
2807 visited.set(n->_idx);
2808 }
2809 }
2810 }
2811
2812 // traverse out's that are in the member set
2813 while (true) {
2814 if (idx < n->outcnt()) {
2815 Node* use = n->raw_out(idx);
2816 idx++;
2817 if (!visited.test_set(use->_idx)) {
2818 if (member.test(use->_idx)) {
2819 nstack.push(n, idx);
2820 n = use;
2821 idx = 0;
2822 }
2823 }
2824 } else {
2825 // All outputs processed
2826 sched.push(n);
2827 if (nstack.is_empty()) break;
2828 n = nstack.node();
2829 idx = nstack.index();
2830 nstack.pop();
2831 }
2832 }
2833 }
2834
2835
2836 //------------------------------ has_use_in_set -------------------------------------
2837 // Has a use in the vector set
has_use_in_set(Node * n,VectorSet & vset)2838 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
2839 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2840 Node* use = n->fast_out(j);
2841 if (vset.test(use->_idx)) {
2842 return true;
2843 }
2844 }
2845 return false;
2846 }
2847
2848
2849 //------------------------------ has_use_internal_to_set -------------------------------------
2850 // Has use internal to the vector set (ie. not in a phi at the loop head)
has_use_internal_to_set(Node * n,VectorSet & vset,IdealLoopTree * loop)2851 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
2852 Node* head = loop->_head;
2853 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2854 Node* use = n->fast_out(j);
2855 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
2856 return true;
2857 }
2858 }
2859 return false;
2860 }
2861
2862
2863 //------------------------------ clone_for_use_outside_loop -------------------------------------
2864 // clone "n" for uses that are outside of loop
clone_for_use_outside_loop(IdealLoopTree * loop,Node * n,Node_List & worklist)2865 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
2866 int cloned = 0;
2867 assert(worklist.size() == 0, "should be empty");
2868 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2869 Node* use = n->fast_out(j);
2870 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
2871 worklist.push(use);
2872 }
2873 }
2874
2875 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
2876 "Too many clones required in clone_for_use_outside_loop in partial peeling")) {
2877 return -1;
2878 }
2879
2880 while( worklist.size() ) {
2881 Node *use = worklist.pop();
2882 if (!has_node(use) || use->in(0) == C->top()) continue;
2883 uint j;
2884 for (j = 0; j < use->req(); j++) {
2885 if (use->in(j) == n) break;
2886 }
2887 assert(j < use->req(), "must be there");
2888
2889 // clone "n" and insert it between the inputs of "n" and the use outside the loop
2890 Node* n_clone = n->clone();
2891 _igvn.replace_input_of(use, j, n_clone);
2892 cloned++;
2893 Node* use_c;
2894 if (!use->is_Phi()) {
2895 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
2896 } else {
2897 // Use in a phi is considered a use in the associated predecessor block
2898 use_c = use->in(0)->in(j);
2899 }
2900 set_ctrl(n_clone, use_c);
2901 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
2902 get_loop(use_c)->_body.push(n_clone);
2903 _igvn.register_new_node_with_optimizer(n_clone);
2904 #ifndef PRODUCT
2905 if (TracePartialPeeling) {
2906 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
2907 }
2908 #endif
2909 }
2910 return cloned;
2911 }
2912
2913
2914 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
2915 // clone "n" for special uses that are in the not_peeled region.
2916 // If these def-uses occur in separate blocks, the code generator
2917 // marks the method as not compilable. For example, if a "BoolNode"
2918 // is in a different basic block than the "IfNode" that uses it, then
2919 // the compilation is aborted in the code generator.
clone_for_special_use_inside_loop(IdealLoopTree * loop,Node * n,VectorSet & not_peel,Node_List & sink_list,Node_List & worklist)2920 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
2921 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
2922 if (n->is_Phi() || n->is_Load()) {
2923 return;
2924 }
2925 assert(worklist.size() == 0, "should be empty");
2926 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2927 Node* use = n->fast_out(j);
2928 if ( not_peel.test(use->_idx) &&
2929 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
2930 use->in(1) == n) {
2931 worklist.push(use);
2932 }
2933 }
2934 if (worklist.size() > 0) {
2935 // clone "n" and insert it between inputs of "n" and the use
2936 Node* n_clone = n->clone();
2937 loop->_body.push(n_clone);
2938 _igvn.register_new_node_with_optimizer(n_clone);
2939 set_ctrl(n_clone, get_ctrl(n));
2940 sink_list.push(n_clone);
2941 not_peel.set(n_clone->_idx);
2942 #ifndef PRODUCT
2943 if (TracePartialPeeling) {
2944 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2945 }
2946 #endif
2947 while( worklist.size() ) {
2948 Node *use = worklist.pop();
2949 _igvn.rehash_node_delayed(use);
2950 for (uint j = 1; j < use->req(); j++) {
2951 if (use->in(j) == n) {
2952 use->set_req(j, n_clone);
2953 }
2954 }
2955 }
2956 }
2957 }
2958
2959
2960 //------------------------------ insert_phi_for_loop -------------------------------------
2961 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
insert_phi_for_loop(Node * use,uint idx,Node * lp_entry_val,Node * back_edge_val,LoopNode * lp)2962 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2963 Node *phi = PhiNode::make(lp, back_edge_val);
2964 phi->set_req(LoopNode::EntryControl, lp_entry_val);
2965 // Use existing phi if it already exists
2966 Node *hit = _igvn.hash_find_insert(phi);
2967 if( hit == NULL ) {
2968 _igvn.register_new_node_with_optimizer(phi);
2969 set_ctrl(phi, lp);
2970 } else {
2971 // Remove the new phi from the graph and use the hit
2972 _igvn.remove_dead_node(phi);
2973 phi = hit;
2974 }
2975 _igvn.replace_input_of(use, idx, phi);
2976 }
2977
2978 #ifdef ASSERT
2979 //------------------------------ is_valid_loop_partition -------------------------------------
2980 // Validate the loop partition sets: peel and not_peel
is_valid_loop_partition(IdealLoopTree * loop,VectorSet & peel,Node_List & peel_list,VectorSet & not_peel)2981 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2982 VectorSet& not_peel ) {
2983 uint i;
2984 // Check that peel_list entries are in the peel set
2985 for (i = 0; i < peel_list.size(); i++) {
2986 if (!peel.test(peel_list.at(i)->_idx)) {
2987 return false;
2988 }
2989 }
2990 // Check at loop members are in one of peel set or not_peel set
2991 for (i = 0; i < loop->_body.size(); i++ ) {
2992 Node *def = loop->_body.at(i);
2993 uint di = def->_idx;
2994 // Check that peel set elements are in peel_list
2995 if (peel.test(di)) {
2996 if (not_peel.test(di)) {
2997 return false;
2998 }
2999 // Must be in peel_list also
3000 bool found = false;
3001 for (uint j = 0; j < peel_list.size(); j++) {
3002 if (peel_list.at(j)->_idx == di) {
3003 found = true;
3004 break;
3005 }
3006 }
3007 if (!found) {
3008 return false;
3009 }
3010 } else if (not_peel.test(di)) {
3011 if (peel.test(di)) {
3012 return false;
3013 }
3014 } else {
3015 return false;
3016 }
3017 }
3018 return true;
3019 }
3020
3021 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
3022 // Ensure a use outside of loop is of the right form
is_valid_clone_loop_exit_use(IdealLoopTree * loop,Node * use,uint exit_idx)3023 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
3024 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3025 return (use->is_Phi() &&
3026 use_c->is_Region() && use_c->req() == 3 &&
3027 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
3028 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
3029 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
3030 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
3031 }
3032
3033 //------------------------------ is_valid_clone_loop_form -------------------------------------
3034 // Ensure that all uses outside of loop are of the right form
is_valid_clone_loop_form(IdealLoopTree * loop,Node_List & peel_list,uint orig_exit_idx,uint clone_exit_idx)3035 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
3036 uint orig_exit_idx, uint clone_exit_idx) {
3037 uint len = peel_list.size();
3038 for (uint i = 0; i < len; i++) {
3039 Node *def = peel_list.at(i);
3040
3041 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3042 Node *use = def->fast_out(j);
3043 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3044 if (!loop->is_member(get_loop(use_c))) {
3045 // use is not in the loop, check for correct structure
3046 if (use->in(0) == def) {
3047 // Okay
3048 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
3049 return false;
3050 }
3051 }
3052 }
3053 }
3054 return true;
3055 }
3056 #endif
3057
3058 //------------------------------ partial_peel -------------------------------------
3059 // Partially peel (aka loop rotation) the top portion of a loop (called
3060 // the peel section below) by cloning it and placing one copy just before
3061 // the new loop head and the other copy at the bottom of the new loop.
3062 //
3063 // before after where it came from
3064 //
3065 // stmt1 stmt1
3066 // loop: stmt2 clone
3067 // stmt2 if condA goto exitA clone
3068 // if condA goto exitA new_loop: new
3069 // stmt3 stmt3 clone
3070 // if !condB goto loop if condB goto exitB clone
3071 // exitB: stmt2 orig
3072 // stmt4 if !condA goto new_loop orig
3073 // exitA: goto exitA
3074 // exitB:
3075 // stmt4
3076 // exitA:
3077 //
3078 // Step 1: find the cut point: an exit test on probable
3079 // induction variable.
3080 // Step 2: schedule (with cloning) operations in the peel
3081 // section that can be executed after the cut into
3082 // the section that is not peeled. This may need
3083 // to clone operations into exit blocks. For
3084 // instance, a reference to A[i] in the not-peel
3085 // section and a reference to B[i] in an exit block
3086 // may cause a left-shift of i by 2 to be placed
3087 // in the peel block. This step will clone the left
3088 // shift into the exit block and sink the left shift
3089 // from the peel to the not-peel section.
3090 // Step 3: clone the loop, retarget the control, and insert
3091 // phis for values that are live across the new loop
3092 // head. This is very dependent on the graph structure
3093 // from clone_loop. It creates region nodes for
3094 // exit control and associated phi nodes for values
3095 // flow out of the loop through that exit. The region
3096 // node is dominated by the clone's control projection.
3097 // So the clone's peel section is placed before the
3098 // new loop head, and the clone's not-peel section is
3099 // forms the top part of the new loop. The original
3100 // peel section forms the tail of the new loop.
3101 // Step 4: update the dominator tree and recompute the
3102 // dominator depth.
3103 //
3104 // orig
3105 //
3106 // stmt1
3107 // |
3108 // v
3109 // loop predicate
3110 // |
3111 // v
3112 // loop<----+
3113 // | |
3114 // stmt2 |
3115 // | |
3116 // v |
3117 // ifA |
3118 // / | |
3119 // v v |
3120 // false true ^ <-- last_peel
3121 // / | |
3122 // / ===|==cut |
3123 // / stmt3 | <-- first_not_peel
3124 // / | |
3125 // | v |
3126 // v ifB |
3127 // exitA: / \ |
3128 // / \ |
3129 // v v |
3130 // false true |
3131 // / \ |
3132 // / ----+
3133 // |
3134 // v
3135 // exitB:
3136 // stmt4
3137 //
3138 //
3139 // after clone loop
3140 //
3141 // stmt1
3142 // |
3143 // v
3144 // loop predicate
3145 // / \
3146 // clone / \ orig
3147 // / \
3148 // / \
3149 // v v
3150 // +---->loop loop<----+
3151 // | | | |
3152 // | stmt2 stmt2 |
3153 // | | | |
3154 // | v v |
3155 // | ifA ifA |
3156 // | | \ / | |
3157 // | v v v v |
3158 // ^ true false false true ^ <-- last_peel
3159 // | | ^ \ / | |
3160 // | cut==|== \ \ / ===|==cut |
3161 // | stmt3 \ \ / stmt3 | <-- first_not_peel
3162 // | | dom | | | |
3163 // | v \ 1v v2 v |
3164 // | ifB regionA ifB |
3165 // | / \ | / \ |
3166 // | / \ v / \ |
3167 // | v v exitA: v v |
3168 // | true false false true |
3169 // | / ^ \ / \ |
3170 // +---- \ \ / ----+
3171 // dom \ /
3172 // \ 1v v2
3173 // regionB
3174 // |
3175 // v
3176 // exitB:
3177 // stmt4
3178 //
3179 //
3180 // after partial peel
3181 //
3182 // stmt1
3183 // |
3184 // v
3185 // loop predicate
3186 // /
3187 // clone / orig
3188 // / TOP
3189 // / \
3190 // v v
3191 // TOP->loop loop----+
3192 // | | |
3193 // stmt2 stmt2 |
3194 // | | |
3195 // v v |
3196 // ifA ifA |
3197 // | \ / | |
3198 // v v v v |
3199 // true false false true | <-- last_peel
3200 // | ^ \ / +------|---+
3201 // +->newloop \ \ / === ==cut | |
3202 // | stmt3 \ \ / TOP | |
3203 // | | dom | | stmt3 | | <-- first_not_peel
3204 // | v \ 1v v2 v | |
3205 // | ifB regionA ifB ^ v
3206 // | / \ | / \ | |
3207 // | / \ v / \ | |
3208 // | v v exitA: v v | |
3209 // | true false false true | |
3210 // | / ^ \ / \ | |
3211 // | | \ \ / v | |
3212 // | | dom \ / TOP | |
3213 // | | \ 1v v2 | |
3214 // ^ v regionB | |
3215 // | | | | |
3216 // | | v ^ v
3217 // | | exitB: | |
3218 // | | stmt4 | |
3219 // | +------------>-----------------+ |
3220 // | |
3221 // +-----------------<---------------------+
3222 //
3223 //
3224 // final graph
3225 //
3226 // stmt1
3227 // |
3228 // v
3229 // loop predicate
3230 // |
3231 // v
3232 // stmt2 clone
3233 // |
3234 // v
3235 // ........> ifA clone
3236 // : / |
3237 // dom / |
3238 // : v v
3239 // : false true
3240 // : | |
3241 // : | v
3242 // : | newloop<-----+
3243 // : | | |
3244 // : | stmt3 clone |
3245 // : | | |
3246 // : | v |
3247 // : | ifB |
3248 // : | / \ |
3249 // : | v v |
3250 // : | false true |
3251 // : | | | |
3252 // : | v stmt2 |
3253 // : | exitB: | |
3254 // : | stmt4 v |
3255 // : | ifA orig |
3256 // : | / \ |
3257 // : | / \ |
3258 // : | v v |
3259 // : | false true |
3260 // : | / \ |
3261 // : v v -----+
3262 // RegionA
3263 // |
3264 // v
3265 // exitA
3266 //
partial_peel(IdealLoopTree * loop,Node_List & old_new)3267 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3268
3269 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3270 if (!loop->_head->is_Loop()) {
3271 return false;
3272 }
3273 LoopNode *head = loop->_head->as_Loop();
3274
3275 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3276 return false;
3277 }
3278
3279 // Check for complex exit control
3280 for (uint ii = 0; ii < loop->_body.size(); ii++) {
3281 Node *n = loop->_body.at(ii);
3282 int opc = n->Opcode();
3283 if (n->is_Call() ||
3284 opc == Op_Catch ||
3285 opc == Op_CatchProj ||
3286 opc == Op_Jump ||
3287 opc == Op_JumpProj) {
3288 #ifndef PRODUCT
3289 if (TracePartialPeeling) {
3290 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3291 }
3292 #endif
3293 return false;
3294 }
3295 }
3296
3297 int dd = dom_depth(head);
3298
3299 // Step 1: find cut point
3300
3301 // Walk up dominators to loop head looking for first loop exit
3302 // which is executed on every path thru loop.
3303 IfNode *peel_if = NULL;
3304 IfNode *peel_if_cmpu = NULL;
3305
3306 Node *iff = loop->tail();
3307 while (iff != head) {
3308 if (iff->is_If()) {
3309 Node *ctrl = get_ctrl(iff->in(1));
3310 if (ctrl->is_top()) return false; // Dead test on live IF.
3311 // If loop-varying exit-test, check for induction variable
3312 if (loop->is_member(get_loop(ctrl)) &&
3313 loop->is_loop_exit(iff) &&
3314 is_possible_iv_test(iff)) {
3315 Node* cmp = iff->in(1)->in(1);
3316 if (cmp->Opcode() == Op_CmpI) {
3317 peel_if = iff->as_If();
3318 } else {
3319 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3320 peel_if_cmpu = iff->as_If();
3321 }
3322 }
3323 }
3324 iff = idom(iff);
3325 }
3326
3327 // Prefer signed compare over unsigned compare.
3328 IfNode* new_peel_if = NULL;
3329 if (peel_if == NULL) {
3330 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
3331 return false; // No peel point found
3332 }
3333 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3334 if (new_peel_if == NULL) {
3335 return false; // No peel point found
3336 }
3337 peel_if = new_peel_if;
3338 }
3339 Node* last_peel = stay_in_loop(peel_if, loop);
3340 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3341 if (first_not_peeled == NULL || first_not_peeled == head) {
3342 return false;
3343 }
3344
3345 #ifndef PRODUCT
3346 if (TraceLoopOpts) {
3347 tty->print("PartialPeel ");
3348 loop->dump_head();
3349 }
3350
3351 if (TracePartialPeeling) {
3352 tty->print_cr("before partial peel one iteration");
3353 Node_List wl;
3354 Node* t = head->in(2);
3355 while (true) {
3356 wl.push(t);
3357 if (t == head) break;
3358 t = idom(t);
3359 }
3360 while (wl.size() > 0) {
3361 Node* tt = wl.pop();
3362 tt->dump();
3363 if (tt == last_peel) tty->print_cr("-- cut --");
3364 }
3365 }
3366 #endif
3367 VectorSet peel;
3368 VectorSet not_peel;
3369 Node_List peel_list;
3370 Node_List worklist;
3371 Node_List sink_list;
3372
3373 uint estimate = loop->est_loop_clone_sz(1);
3374 if (exceeding_node_budget(estimate)) {
3375 return false;
3376 }
3377
3378 // Set of cfg nodes to peel are those that are executable from
3379 // the head through last_peel.
3380 assert(worklist.size() == 0, "should be empty");
3381 worklist.push(head);
3382 peel.set(head->_idx);
3383 while (worklist.size() > 0) {
3384 Node *n = worklist.pop();
3385 if (n != last_peel) {
3386 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3387 Node* use = n->fast_out(j);
3388 if (use->is_CFG() &&
3389 loop->is_member(get_loop(use)) &&
3390 !peel.test_set(use->_idx)) {
3391 worklist.push(use);
3392 }
3393 }
3394 }
3395 }
3396
3397 // Set of non-cfg nodes to peel are those that are control
3398 // dependent on the cfg nodes.
3399 for (uint i = 0; i < loop->_body.size(); i++) {
3400 Node *n = loop->_body.at(i);
3401 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3402 if (peel.test(n_c->_idx)) {
3403 peel.set(n->_idx);
3404 } else {
3405 not_peel.set(n->_idx);
3406 }
3407 }
3408
3409 // Step 2: move operations from the peeled section down into the
3410 // not-peeled section
3411
3412 // Get a post order schedule of nodes in the peel region
3413 // Result in right-most operand.
3414 scheduled_nodelist(loop, peel, peel_list);
3415
3416 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3417
3418 // For future check for too many new phis
3419 uint old_phi_cnt = 0;
3420 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3421 Node* use = head->fast_out(j);
3422 if (use->is_Phi()) old_phi_cnt++;
3423 }
3424
3425 #ifndef PRODUCT
3426 if (TracePartialPeeling) {
3427 tty->print_cr("\npeeled list");
3428 }
3429 #endif
3430
3431 // Evacuate nodes in peel region into the not_peeled region if possible
3432 bool too_many_clones = false;
3433 uint new_phi_cnt = 0;
3434 uint cloned_for_outside_use = 0;
3435 for (uint i = 0; i < peel_list.size();) {
3436 Node* n = peel_list.at(i);
3437 #ifndef PRODUCT
3438 if (TracePartialPeeling) n->dump();
3439 #endif
3440 bool incr = true;
3441 if (!n->is_CFG()) {
3442 if (has_use_in_set(n, not_peel)) {
3443 // If not used internal to the peeled region,
3444 // move "n" from peeled to not_peeled region.
3445 if (!has_use_internal_to_set(n, peel, loop)) {
3446 // if not pinned and not a load (which maybe anti-dependent on a store)
3447 // and not a CMove (Matcher expects only bool->cmove).
3448 if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove()) {
3449 int new_clones = clone_for_use_outside_loop(loop, n, worklist);
3450 if (new_clones == -1) {
3451 too_many_clones = true;
3452 break;
3453 }
3454 cloned_for_outside_use += new_clones;
3455 sink_list.push(n);
3456 peel.remove(n->_idx);
3457 not_peel.set(n->_idx);
3458 peel_list.remove(i);
3459 incr = false;
3460 #ifndef PRODUCT
3461 if (TracePartialPeeling) {
3462 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3463 n->_idx, get_ctrl(n)->_idx);
3464 }
3465 #endif
3466 }
3467 } else {
3468 // Otherwise check for special def-use cases that span
3469 // the peel/not_peel boundary such as bool->if
3470 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
3471 new_phi_cnt++;
3472 }
3473 }
3474 }
3475 if (incr) i++;
3476 }
3477
3478 estimate += cloned_for_outside_use + new_phi_cnt;
3479 bool exceed_node_budget = !may_require_nodes(estimate);
3480 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
3481
3482 if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
3483 #ifndef PRODUCT
3484 if (TracePartialPeeling && exceed_phi_limit) {
3485 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3486 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
3487 }
3488 #endif
3489 if (new_peel_if != NULL) {
3490 remove_cmpi_loop_exit(new_peel_if, loop);
3491 }
3492 // Inhibit more partial peeling on this loop
3493 assert(!head->is_partial_peel_loop(), "not partial peeled");
3494 head->mark_partial_peel_failed();
3495 if (cloned_for_outside_use > 0) {
3496 // Terminate this round of loop opts because
3497 // the graph outside this loop was changed.
3498 C->set_major_progress();
3499 return true;
3500 }
3501 return false;
3502 }
3503
3504 // Step 3: clone loop, retarget control, and insert new phis
3505
3506 // Create new loop head for new phis and to hang
3507 // the nodes being moved (sinked) from the peel region.
3508 LoopNode* new_head = new LoopNode(last_peel, last_peel);
3509 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3510 _igvn.register_new_node_with_optimizer(new_head);
3511 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3512 _igvn.replace_input_of(first_not_peeled, 0, new_head);
3513 set_loop(new_head, loop);
3514 loop->_body.push(new_head);
3515 not_peel.set(new_head->_idx);
3516 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3517 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3518
3519 while (sink_list.size() > 0) {
3520 Node* n = sink_list.pop();
3521 set_ctrl(n, new_head);
3522 }
3523
3524 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3525
3526 clone_loop(loop, old_new, dd, IgnoreStripMined);
3527
3528 const uint clone_exit_idx = 1;
3529 const uint orig_exit_idx = 2;
3530 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
3531
3532 Node* head_clone = old_new[head->_idx];
3533 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
3534 Node* orig_tail_clone = head_clone->in(2);
3535
3536 // Add phi if "def" node is in peel set and "use" is not
3537
3538 for (uint i = 0; i < peel_list.size(); i++) {
3539 Node *def = peel_list.at(i);
3540 if (!def->is_CFG()) {
3541 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3542 Node *use = def->fast_out(j);
3543 if (has_node(use) && use->in(0) != C->top() &&
3544 (!peel.test(use->_idx) ||
3545 (use->is_Phi() && use->in(0) == head)) ) {
3546 worklist.push(use);
3547 }
3548 }
3549 while( worklist.size() ) {
3550 Node *use = worklist.pop();
3551 for (uint j = 1; j < use->req(); j++) {
3552 Node* n = use->in(j);
3553 if (n == def) {
3554
3555 // "def" is in peel set, "use" is not in peel set
3556 // or "use" is in the entry boundary (a phi) of the peel set
3557
3558 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
3559
3560 if ( loop->is_member(get_loop( use_c )) ) {
3561 // use is in loop
3562 if (old_new[use->_idx] != NULL) { // null for dead code
3563 Node* use_clone = old_new[use->_idx];
3564 _igvn.replace_input_of(use, j, C->top());
3565 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
3566 }
3567 } else {
3568 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
3569 // use is not in the loop, check if the live range includes the cut
3570 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
3571 if (not_peel.test(lp_if->_idx)) {
3572 assert(j == orig_exit_idx, "use from original loop");
3573 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
3574 }
3575 }
3576 }
3577 }
3578 }
3579 }
3580 }
3581
3582 // Step 3b: retarget control
3583
3584 // Redirect control to the new loop head if a cloned node in
3585 // the not_peeled region has control that points into the peeled region.
3586 // This necessary because the cloned peeled region will be outside
3587 // the loop.
3588 // from to
3589 // cloned-peeled <---+
3590 // new_head_clone: | <--+
3591 // cloned-not_peeled in(0) in(0)
3592 // orig-peeled
3593
3594 for (uint i = 0; i < loop->_body.size(); i++) {
3595 Node *n = loop->_body.at(i);
3596 if (!n->is_CFG() && n->in(0) != NULL &&
3597 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
3598 Node* n_clone = old_new[n->_idx];
3599 _igvn.replace_input_of(n_clone, 0, new_head_clone);
3600 }
3601 }
3602
3603 // Backedge of the surviving new_head (the clone) is original last_peel
3604 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
3605
3606 // Cut first node in original not_peel set
3607 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
3608 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
3609 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
3610
3611 // Copy head_clone back-branch info to original head
3612 // and remove original head's loop entry and
3613 // clone head's back-branch
3614 _igvn.rehash_node_delayed(head); // Multiple edge updates
3615 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
3616 head->set_req(LoopNode::LoopBackControl, C->top());
3617 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
3618
3619 // Similarly modify the phis
3620 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
3621 Node* use = head->fast_out(k);
3622 if (use->is_Phi() && use->outcnt() > 0) {
3623 Node* use_clone = old_new[use->_idx];
3624 _igvn.rehash_node_delayed(use); // Multiple edge updates
3625 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
3626 use->set_req(LoopNode::LoopBackControl, C->top());
3627 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
3628 }
3629 }
3630
3631 // Step 4: update dominator tree and dominator depth
3632
3633 set_idom(head, orig_tail_clone, dd);
3634 recompute_dom_depth();
3635
3636 // Inhibit more partial peeling on this loop
3637 new_head_clone->set_partial_peel_loop();
3638 C->set_major_progress();
3639 loop->record_for_igvn();
3640
3641 #ifndef PRODUCT
3642 if (TracePartialPeeling) {
3643 tty->print_cr("\nafter partial peel one iteration");
3644 Node_List wl;
3645 Node* t = last_peel;
3646 while (true) {
3647 wl.push(t);
3648 if (t == head_clone) break;
3649 t = idom(t);
3650 }
3651 while (wl.size() > 0) {
3652 Node* tt = wl.pop();
3653 if (tt == head) tty->print_cr("orig head");
3654 else if (tt == new_head_clone) tty->print_cr("new head");
3655 else if (tt == head_clone) tty->print_cr("clone head");
3656 tt->dump();
3657 }
3658 }
3659 #endif
3660 return true;
3661 }
3662
3663 //------------------------------reorg_offsets----------------------------------
3664 // Reorganize offset computations to lower register pressure. Mostly
3665 // prevent loop-fallout uses of the pre-incremented trip counter (which are
3666 // then alive with the post-incremented trip counter forcing an extra
3667 // register move)
reorg_offsets(IdealLoopTree * loop)3668 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
3669 // Perform it only for canonical counted loops.
3670 // Loop's shape could be messed up by iteration_split_impl.
3671 if (!loop->_head->is_CountedLoop())
3672 return;
3673 if (!loop->_head->as_Loop()->is_valid_counted_loop(T_INT))
3674 return;
3675
3676 CountedLoopNode *cl = loop->_head->as_CountedLoop();
3677 CountedLoopEndNode *cle = cl->loopexit();
3678 Node *exit = cle->proj_out(false);
3679 Node *phi = cl->phi();
3680
3681 // Check for the special case when using the pre-incremented trip-counter on
3682 // the fall-out path (forces the pre-incremented and post-incremented trip
3683 // counter to be live at the same time). Fix this by adjusting to use the
3684 // post-increment trip counter.
3685
3686 bool progress = true;
3687 while (progress) {
3688 progress = false;
3689 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3690 Node* use = phi->fast_out(i); // User of trip-counter
3691 if (!has_ctrl(use)) continue;
3692 Node *u_ctrl = get_ctrl(use);
3693 if (use->is_Phi()) {
3694 u_ctrl = NULL;
3695 for (uint j = 1; j < use->req(); j++)
3696 if (use->in(j) == phi)
3697 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
3698 }
3699 IdealLoopTree *u_loop = get_loop(u_ctrl);
3700 // Look for loop-invariant use
3701 if (u_loop == loop) continue;
3702 if (loop->is_member(u_loop)) continue;
3703 // Check that use is live out the bottom. Assuming the trip-counter
3704 // update is right at the bottom, uses of of the loop middle are ok.
3705 if (dom_lca(exit, u_ctrl) != exit) continue;
3706 // Hit! Refactor use to use the post-incremented tripcounter.
3707 // Compute a post-increment tripcounter.
3708 Node* c = exit;
3709 if (cl->is_strip_mined()) {
3710 IdealLoopTree* outer_loop = get_loop(cl->outer_loop());
3711 if (!outer_loop->is_member(u_loop)) {
3712 c = cl->outer_loop_exit();
3713 }
3714 }
3715 Node *opaq = new Opaque2Node(C, cle->incr());
3716 register_new_node(opaq, c);
3717 Node *neg_stride = _igvn.intcon(-cle->stride_con());
3718 set_ctrl(neg_stride, C->root());
3719 Node *post = new AddINode(opaq, neg_stride);
3720 register_new_node(post, c);
3721 _igvn.rehash_node_delayed(use);
3722 for (uint j = 1; j < use->req(); j++) {
3723 if (use->in(j) == phi)
3724 use->set_req(j, post);
3725 }
3726 // Since DU info changed, rerun loop
3727 progress = true;
3728 break;
3729 }
3730 }
3731
3732 }
3733