0
|
1 /*
|
|
2 * Copyright 1999-2006 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
|
|
20 * CA 95054 USA or visit www.sun.com if you need additional information or
|
|
21 * have any questions.
|
|
22 *
|
|
23 */
|
|
24
|
|
25 #include "incls/_precompiled.incl"
|
|
26 #include "incls/_loopopts.cpp.incl"
|
|
27
|
|
28 //=============================================================================
|
|
29 //------------------------------split_thru_phi---------------------------------
|
|
30 // Split Node 'n' through merge point if there is enough win.
|
|
31 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
|
|
32 int wins = 0;
|
|
33 assert( !n->is_CFG(), "" );
|
|
34 assert( region->is_Region(), "" );
|
|
35 Node *phi = new (C, region->req()) PhiNode( region, n->bottom_type() );
|
|
36 uint old_unique = C->unique();
|
|
37 for( uint i = 1; i < region->req(); i++ ) {
|
|
38 Node *x;
|
|
39 Node* the_clone = NULL;
|
|
40 if( region->in(i) == C->top() ) {
|
|
41 x = C->top(); // Dead path? Use a dead data op
|
|
42 } else {
|
|
43 x = n->clone(); // Else clone up the data op
|
|
44 the_clone = x; // Remember for possible deletion.
|
|
45 // Alter data node to use pre-phi inputs
|
|
46 if( n->in(0) == region )
|
|
47 x->set_req( 0, region->in(i) );
|
|
48 for( uint j = 1; j < n->req(); j++ ) {
|
|
49 Node *in = n->in(j);
|
|
50 if( in->is_Phi() && in->in(0) == region )
|
|
51 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
|
|
52 }
|
|
53 }
|
|
54 // Check for a 'win' on some paths
|
|
55 const Type *t = x->Value(&_igvn);
|
|
56
|
|
57 bool singleton = t->singleton();
|
|
58
|
|
59 // A TOP singleton indicates that there are no possible values incoming
|
|
60 // along a particular edge. In most cases, this is OK, and the Phi will
|
|
61 // be eliminated later in an Ideal call. However, we can't allow this to
|
|
62 // happen if the singleton occurs on loop entry, as the elimination of
|
|
63 // the PhiNode may cause the resulting node to migrate back to a previous
|
|
64 // loop iteration.
|
|
65 if( singleton && t == Type::TOP ) {
|
|
66 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
|
|
67 // irreducible loop may not be indicated by an affirmative is_Loop());
|
|
68 // therefore, the only top we can split thru a phi is on a backedge of
|
|
69 // a loop.
|
|
70 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
|
|
71 }
|
|
72
|
|
73 if( singleton ) {
|
|
74 wins++;
|
|
75 x = ((PhaseGVN&)_igvn).makecon(t);
|
|
76 } else {
|
|
77 // We now call Identity to try to simplify the cloned node.
|
|
78 // Note that some Identity methods call phase->type(this).
|
|
79 // Make sure that the type array is big enough for
|
|
80 // our new node, even though we may throw the node away.
|
|
81 // (Note: This tweaking with igvn only works because x is a new node.)
|
|
82 _igvn.set_type(x, t);
|
|
83 Node *y = x->Identity(&_igvn);
|
|
84 if( y != x ) {
|
|
85 wins++;
|
|
86 x = y;
|
|
87 } else {
|
|
88 y = _igvn.hash_find(x);
|
|
89 if( y ) {
|
|
90 wins++;
|
|
91 x = y;
|
|
92 } else {
|
|
93 // Else x is a new node we are keeping
|
|
94 // We do not need register_new_node_with_optimizer
|
|
95 // because set_type has already been called.
|
|
96 _igvn._worklist.push(x);
|
|
97 }
|
|
98 }
|
|
99 }
|
|
100 if (x != the_clone && the_clone != NULL)
|
|
101 _igvn.remove_dead_node(the_clone);
|
|
102 phi->set_req( i, x );
|
|
103 }
|
|
104 // Too few wins?
|
|
105 if( wins <= policy ) {
|
|
106 _igvn.remove_dead_node(phi);
|
|
107 return NULL;
|
|
108 }
|
|
109
|
|
110 // Record Phi
|
|
111 register_new_node( phi, region );
|
|
112
|
|
113 for( uint i2 = 1; i2 < phi->req(); i2++ ) {
|
|
114 Node *x = phi->in(i2);
|
|
115 // If we commoned up the cloned 'x' with another existing Node,
|
|
116 // the existing Node picks up a new use. We need to make the
|
|
117 // existing Node occur higher up so it dominates its uses.
|
|
118 Node *old_ctrl;
|
|
119 IdealLoopTree *old_loop;
|
|
120
|
|
121 // The occasional new node
|
|
122 if( x->_idx >= old_unique ) { // Found a new, unplaced node?
|
|
123 old_ctrl = x->is_Con() ? C->root() : NULL;
|
|
124 old_loop = NULL; // Not in any prior loop
|
|
125 } else {
|
|
126 old_ctrl = x->is_Con() ? C->root() : get_ctrl(x);
|
|
127 old_loop = get_loop(old_ctrl); // Get prior loop
|
|
128 }
|
|
129 // New late point must dominate new use
|
|
130 Node *new_ctrl = dom_lca( old_ctrl, region->in(i2) );
|
|
131 // Set new location
|
|
132 set_ctrl(x, new_ctrl);
|
|
133 IdealLoopTree *new_loop = get_loop( new_ctrl );
|
|
134 // If changing loop bodies, see if we need to collect into new body
|
|
135 if( old_loop != new_loop ) {
|
|
136 if( old_loop && !old_loop->_child )
|
|
137 old_loop->_body.yank(x);
|
|
138 if( !new_loop->_child )
|
|
139 new_loop->_body.push(x); // Collect body info
|
|
140 }
|
|
141 }
|
|
142
|
|
143 return phi;
|
|
144 }
|
|
145
|
|
146 //------------------------------dominated_by------------------------------------
|
|
147 // Replace the dominated test with an obvious true or false. Place it on the
|
|
148 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
|
|
149 // live path up to the dominating control.
|
|
150 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff ) {
|
|
151 #ifndef PRODUCT
|
|
152 if( VerifyLoopOptimizations && PrintOpto ) tty->print_cr("dominating test");
|
|
153 #endif
|
|
154
|
|
155
|
|
156 // prevdom is the dominating projection of the dominating test.
|
|
157 assert( iff->is_If(), "" );
|
|
158 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
|
|
159 int pop = prevdom->Opcode();
|
|
160 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
|
|
161 // 'con' is set to true or false to kill the dominated test.
|
|
162 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
|
|
163 set_ctrl(con, C->root()); // Constant gets a new use
|
|
164 // Hack the dominated test
|
|
165 _igvn.hash_delete(iff);
|
|
166 iff->set_req(1, con);
|
|
167 _igvn._worklist.push(iff);
|
|
168
|
|
169 // If I dont have a reachable TRUE and FALSE path following the IfNode then
|
|
170 // I can assume this path reaches an infinite loop. In this case it's not
|
|
171 // important to optimize the data Nodes - either the whole compilation will
|
|
172 // be tossed or this path (and all data Nodes) will go dead.
|
|
173 if( iff->outcnt() != 2 ) return;
|
|
174
|
|
175 // Make control-dependent data Nodes on the live path (path that will remain
|
|
176 // once the dominated IF is removed) become control-dependent on the
|
|
177 // dominating projection.
|
|
178 Node* dp = ((IfNode*)iff)->proj_out(pop == Op_IfTrue);
|
|
179 IdealLoopTree *old_loop = get_loop(dp);
|
|
180
|
|
181 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
|
|
182 Node* cd = dp->fast_out(i); // Control-dependent node
|
|
183 if( cd->depends_only_on_test() ) {
|
|
184 assert( cd->in(0) == dp, "" );
|
|
185 _igvn.hash_delete( cd );
|
|
186 cd->set_req(0, prevdom);
|
|
187 set_early_ctrl( cd );
|
|
188 _igvn._worklist.push(cd);
|
|
189 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
|
|
190 if( old_loop != new_loop ) {
|
|
191 if( !old_loop->_child ) old_loop->_body.yank(cd);
|
|
192 if( !new_loop->_child ) new_loop->_body.push(cd);
|
|
193 }
|
|
194 --i;
|
|
195 --imax;
|
|
196 }
|
|
197 }
|
|
198 }
|
|
199
|
|
200 //------------------------------has_local_phi_input----------------------------
|
|
201 // Return TRUE if 'n' has Phi inputs from its local block and no other
|
|
202 // block-local inputs (all non-local-phi inputs come from earlier blocks)
|
|
203 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
|
|
204 Node *n_ctrl = get_ctrl(n);
|
|
205 // See if some inputs come from a Phi in this block, or from before
|
|
206 // this block.
|
|
207 uint i;
|
|
208 for( i = 1; i < n->req(); i++ ) {
|
|
209 Node *phi = n->in(i);
|
|
210 if( phi->is_Phi() && phi->in(0) == n_ctrl )
|
|
211 break;
|
|
212 }
|
|
213 if( i >= n->req() )
|
|
214 return NULL; // No Phi inputs; nowhere to clone thru
|
|
215
|
|
216 // Check for inputs created between 'n' and the Phi input. These
|
|
217 // must split as well; they have already been given the chance
|
|
218 // (courtesy of a post-order visit) and since they did not we must
|
|
219 // recover the 'cost' of splitting them by being very profitable
|
|
220 // when splitting 'n'. Since this is unlikely we simply give up.
|
|
221 for( i = 1; i < n->req(); i++ ) {
|
|
222 Node *m = n->in(i);
|
|
223 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
|
|
224 // We allow the special case of AddP's with no local inputs.
|
|
225 // This allows us to split-up address expressions.
|
|
226 if (m->is_AddP() &&
|
|
227 get_ctrl(m->in(2)) != n_ctrl &&
|
|
228 get_ctrl(m->in(3)) != n_ctrl) {
|
|
229 // Move the AddP up to dominating point
|
|
230 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl)));
|
|
231 continue;
|
|
232 }
|
|
233 return NULL;
|
|
234 }
|
|
235 }
|
|
236
|
|
237 return n_ctrl;
|
|
238 }
|
|
239
|
|
240 //------------------------------remix_address_expressions----------------------
|
|
241 // Rework addressing expressions to get the most loop-invariant stuff
|
|
242 // moved out. We'd like to do all associative operators, but it's especially
|
|
243 // important (common) to do address expressions.
|
|
244 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
|
|
245 if (!has_ctrl(n)) return NULL;
|
|
246 Node *n_ctrl = get_ctrl(n);
|
|
247 IdealLoopTree *n_loop = get_loop(n_ctrl);
|
|
248
|
|
249 // See if 'n' mixes loop-varying and loop-invariant inputs and
|
|
250 // itself is loop-varying.
|
|
251
|
|
252 // Only interested in binary ops (and AddP)
|
|
253 if( n->req() < 3 || n->req() > 4 ) return NULL;
|
|
254
|
|
255 Node *n1_ctrl = get_ctrl(n->in( 1));
|
|
256 Node *n2_ctrl = get_ctrl(n->in( 2));
|
|
257 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
|
|
258 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
|
|
259 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
|
|
260 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
|
|
261
|
|
262 // Does one of my inputs spin in a tighter loop than self?
|
|
263 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
|
|
264 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
|
|
265 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
|
|
266 return NULL; // Leave well enough alone
|
|
267
|
|
268 // Is at least one of my inputs loop-invariant?
|
|
269 if( n1_loop == n_loop &&
|
|
270 n2_loop == n_loop &&
|
|
271 n3_loop == n_loop )
|
|
272 return NULL; // No loop-invariant inputs
|
|
273
|
|
274
|
|
275 int n_op = n->Opcode();
|
|
276
|
|
277 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
|
|
278 if( n_op == Op_LShiftI ) {
|
|
279 // Scale is loop invariant
|
|
280 Node *scale = n->in(2);
|
|
281 Node *scale_ctrl = get_ctrl(scale);
|
|
282 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
|
|
283 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
|
|
284 return NULL;
|
|
285 const TypeInt *scale_t = scale->bottom_type()->isa_int();
|
|
286 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
|
|
287 return NULL; // Dont bother with byte/short masking
|
|
288 // Add must vary with loop (else shift would be loop-invariant)
|
|
289 Node *add = n->in(1);
|
|
290 Node *add_ctrl = get_ctrl(add);
|
|
291 IdealLoopTree *add_loop = get_loop(add_ctrl);
|
|
292 //assert( n_loop == add_loop, "" );
|
|
293 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
|
|
294
|
|
295 // Convert I-V into I+ (0-V); same for V-I
|
|
296 if( add->Opcode() == Op_SubI &&
|
|
297 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
|
|
298 Node *zero = _igvn.intcon(0);
|
|
299 set_ctrl(zero, C->root());
|
|
300 Node *neg = new (C, 3) SubINode( _igvn.intcon(0), add->in(2) );
|
|
301 register_new_node( neg, get_ctrl(add->in(2) ) );
|
|
302 add = new (C, 3) AddINode( add->in(1), neg );
|
|
303 register_new_node( add, add_ctrl );
|
|
304 }
|
|
305 if( add->Opcode() != Op_AddI ) return NULL;
|
|
306 // See if one add input is loop invariant
|
|
307 Node *add_var = add->in(1);
|
|
308 Node *add_var_ctrl = get_ctrl(add_var);
|
|
309 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
|
|
310 Node *add_invar = add->in(2);
|
|
311 Node *add_invar_ctrl = get_ctrl(add_invar);
|
|
312 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
|
|
313 if( add_var_loop == n_loop ) {
|
|
314 } else if( add_invar_loop == n_loop ) {
|
|
315 // Swap to find the invariant part
|
|
316 add_invar = add_var;
|
|
317 add_invar_ctrl = add_var_ctrl;
|
|
318 add_invar_loop = add_var_loop;
|
|
319 add_var = add->in(2);
|
|
320 Node *add_var_ctrl = get_ctrl(add_var);
|
|
321 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
|
|
322 } else // Else neither input is loop invariant
|
|
323 return NULL;
|
|
324 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
|
|
325 return NULL; // No invariant part of the add?
|
|
326
|
|
327 // Yes! Reshape address expression!
|
|
328 Node *inv_scale = new (C, 3) LShiftINode( add_invar, scale );
|
|
329 register_new_node( inv_scale, add_invar_ctrl );
|
|
330 Node *var_scale = new (C, 3) LShiftINode( add_var, scale );
|
|
331 register_new_node( var_scale, n_ctrl );
|
|
332 Node *var_add = new (C, 3) AddINode( var_scale, inv_scale );
|
|
333 register_new_node( var_add, n_ctrl );
|
|
334 _igvn.hash_delete( n );
|
|
335 _igvn.subsume_node( n, var_add );
|
|
336 return var_add;
|
|
337 }
|
|
338
|
|
339 // Replace (I+V) with (V+I)
|
|
340 if( n_op == Op_AddI ||
|
|
341 n_op == Op_AddL ||
|
|
342 n_op == Op_AddF ||
|
|
343 n_op == Op_AddD ||
|
|
344 n_op == Op_MulI ||
|
|
345 n_op == Op_MulL ||
|
|
346 n_op == Op_MulF ||
|
|
347 n_op == Op_MulD ) {
|
|
348 if( n2_loop == n_loop ) {
|
|
349 assert( n1_loop != n_loop, "" );
|
|
350 n->swap_edges(1, 2);
|
|
351 }
|
|
352 }
|
|
353
|
|
354 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
|
|
355 // but not if I2 is a constant.
|
|
356 if( n_op == Op_AddP ) {
|
|
357 if( n2_loop == n_loop && n3_loop != n_loop ) {
|
|
358 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
|
|
359 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
|
|
360 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
|
|
361 IdealLoopTree *n22loop = get_loop( n22_ctrl );
|
|
362 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
|
|
363 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
|
|
364 n23_loop == n_loop ) {
|
|
365 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
|
|
366 // Stuff new AddP in the loop preheader
|
|
367 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
|
|
368 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) );
|
|
369 register_new_node( add2, n_ctrl );
|
|
370 _igvn.hash_delete( n );
|
|
371 _igvn.subsume_node( n, add2 );
|
|
372 return add2;
|
|
373 }
|
|
374 }
|
|
375 }
|
|
376
|
|
377 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
|
|
378 if( n2_loop != n_loop && n3_loop == n_loop ) {
|
|
379 if( n->in(3)->Opcode() == Op_AddI ) {
|
|
380 Node *V = n->in(3)->in(1);
|
|
381 Node *I = n->in(3)->in(2);
|
|
382 if( is_member(n_loop,get_ctrl(V)) ) {
|
|
383 } else {
|
|
384 Node *tmp = V; V = I; I = tmp;
|
|
385 }
|
|
386 if( !is_member(n_loop,get_ctrl(I)) ) {
|
|
387 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I );
|
|
388 // Stuff new AddP in the loop preheader
|
|
389 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
|
|
390 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V );
|
|
391 register_new_node( add2, n_ctrl );
|
|
392 _igvn.hash_delete( n );
|
|
393 _igvn.subsume_node( n, add2 );
|
|
394 return add2;
|
|
395 }
|
|
396 }
|
|
397 }
|
|
398 }
|
|
399
|
|
400 return NULL;
|
|
401 }
|
|
402
|
|
403 //------------------------------conditional_move-------------------------------
|
|
404 // Attempt to replace a Phi with a conditional move. We have some pretty
|
|
405 // strict profitability requirements. All Phis at the merge point must
|
|
406 // be converted, so we can remove the control flow. We need to limit the
|
|
407 // number of c-moves to a small handful. All code that was in the side-arms
|
|
408 // of the CFG diamond is now speculatively executed. This code has to be
|
|
409 // "cheap enough". We are pretty much limited to CFG diamonds that merge
|
|
410 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
|
|
411 Node *PhaseIdealLoop::conditional_move( Node *region ) {
|
|
412
|
|
413 assert( region->is_Region(), "sanity check" );
|
|
414 if( region->req() != 3 ) return NULL;
|
|
415
|
|
416 // Check for CFG diamond
|
|
417 Node *lp = region->in(1);
|
|
418 Node *rp = region->in(2);
|
|
419 if( !lp || !rp ) return NULL;
|
|
420 Node *lp_c = lp->in(0);
|
|
421 if( lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If() ) return NULL;
|
|
422 IfNode *iff = lp_c->as_If();
|
|
423
|
|
424 // Check for highly predictable branch. No point in CMOV'ing if
|
|
425 // we are going to predict accurately all the time.
|
|
426 // %%% This hides patterns produced by utility methods like Math.min.
|
|
427 if( iff->_prob < PROB_UNLIKELY_MAG(3) ||
|
|
428 iff->_prob > PROB_LIKELY_MAG(3) )
|
|
429 return NULL;
|
|
430
|
|
431 // Check for ops pinned in an arm of the diamond.
|
|
432 // Can't remove the control flow in this case
|
|
433 if( lp->outcnt() > 1 ) return NULL;
|
|
434 if( rp->outcnt() > 1 ) return NULL;
|
|
435
|
|
436 // Check profitability
|
|
437 int cost = 0;
|
|
438 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
|
|
439 Node *out = region->fast_out(i);
|
|
440 if( !out->is_Phi() ) continue; // Ignore other control edges, etc
|
|
441 PhiNode* phi = out->as_Phi();
|
|
442 switch (phi->type()->basic_type()) {
|
|
443 case T_LONG:
|
|
444 cost++; // Probably encodes as 2 CMOV's
|
|
445 case T_INT: // These all CMOV fine
|
|
446 case T_FLOAT:
|
|
447 case T_DOUBLE:
|
|
448 case T_ADDRESS: // (RawPtr)
|
|
449 cost++;
|
|
450 break;
|
|
451 case T_OBJECT: { // Base oops are OK, but not derived oops
|
|
452 const TypeOopPtr *tp = phi->type()->isa_oopptr();
|
|
453 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
|
|
454 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
|
|
455 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
|
|
456 // have a Phi for the base here that we convert to a CMOVE all is well
|
|
457 // and good. But if the base is dead, we'll not make a CMOVE. Later
|
|
458 // the allocator will have to produce a base by creating a CMOVE of the
|
|
459 // relevant bases. This puts the allocator in the business of
|
|
460 // manufacturing expensive instructions, generally a bad plan.
|
|
461 // Just Say No to Conditionally-Moved Derived Pointers.
|
|
462 if( tp && tp->offset() != 0 )
|
|
463 return NULL;
|
|
464 cost++;
|
|
465 break;
|
|
466 }
|
|
467 default:
|
|
468 return NULL; // In particular, can't do memory or I/O
|
|
469 }
|
|
470 // Add in cost any speculative ops
|
|
471 for( uint j = 1; j < region->req(); j++ ) {
|
|
472 Node *proj = region->in(j);
|
|
473 Node *inp = phi->in(j);
|
|
474 if (get_ctrl(inp) == proj) { // Found local op
|
|
475 cost++;
|
|
476 // Check for a chain of dependent ops; these will all become
|
|
477 // speculative in a CMOV.
|
|
478 for( uint k = 1; k < inp->req(); k++ )
|
|
479 if (get_ctrl(inp->in(k)) == proj)
|
|
480 return NULL; // Too much speculative goo
|
|
481 }
|
|
482 }
|
|
483 // See if the Phi is used by a Cmp. This will likely Split-If, a
|
|
484 // higher-payoff operation.
|
|
485 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
|
|
486 Node* use = phi->fast_out(k);
|
|
487 if( use->is_Cmp() )
|
|
488 return NULL;
|
|
489 }
|
|
490 }
|
|
491 if( cost >= ConditionalMoveLimit ) return NULL; // Too much goo
|
|
492
|
|
493 // --------------
|
|
494 // Now replace all Phis with CMOV's
|
|
495 Node *cmov_ctrl = iff->in(0);
|
|
496 uint flip = (lp->Opcode() == Op_IfTrue);
|
|
497 while( 1 ) {
|
|
498 PhiNode* phi = NULL;
|
|
499 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
|
|
500 Node *out = region->fast_out(i);
|
|
501 if (out->is_Phi()) {
|
|
502 phi = out->as_Phi();
|
|
503 break;
|
|
504 }
|
|
505 }
|
|
506 if (phi == NULL) break;
|
|
507 #ifndef PRODUCT
|
|
508 if( PrintOpto && VerifyLoopOptimizations ) tty->print_cr("CMOV");
|
|
509 #endif
|
|
510 // Move speculative ops
|
|
511 for( uint j = 1; j < region->req(); j++ ) {
|
|
512 Node *proj = region->in(j);
|
|
513 Node *inp = phi->in(j);
|
|
514 if (get_ctrl(inp) == proj) { // Found local op
|
|
515 #ifndef PRODUCT
|
|
516 if( PrintOpto && VerifyLoopOptimizations ) {
|
|
517 tty->print(" speculate: ");
|
|
518 inp->dump();
|
|
519 }
|
|
520 #endif
|
|
521 set_ctrl(inp, cmov_ctrl);
|
|
522 }
|
|
523 }
|
|
524 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) );
|
|
525 register_new_node( cmov, cmov_ctrl );
|
|
526 _igvn.hash_delete(phi);
|
|
527 _igvn.subsume_node( phi, cmov );
|
|
528 #ifndef PRODUCT
|
|
529 if( VerifyLoopOptimizations ) verify();
|
|
530 #endif
|
|
531 }
|
|
532
|
|
533 // The useless CFG diamond will fold up later; see the optimization in
|
|
534 // RegionNode::Ideal.
|
|
535 _igvn._worklist.push(region);
|
|
536
|
|
537 return iff->in(1);
|
|
538 }
|
|
539
|
|
540 //------------------------------split_if_with_blocks_pre-----------------------
|
|
541 // Do the real work in a non-recursive function. Data nodes want to be
|
|
542 // cloned in the pre-order so they can feed each other nicely.
|
|
543 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
|
|
544 // Cloning these guys is unlikely to win
|
|
545 int n_op = n->Opcode();
|
|
546 if( n_op == Op_MergeMem ) return n;
|
|
547 if( n->is_Proj() ) return n;
|
|
548 // Do not clone-up CmpFXXX variations, as these are always
|
|
549 // followed by a CmpI
|
|
550 if( n->is_Cmp() ) return n;
|
|
551 // Attempt to use a conditional move instead of a phi/branch
|
|
552 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
|
|
553 Node *cmov = conditional_move( n );
|
|
554 if( cmov ) return cmov;
|
|
555 }
|
|
556 if( n->is_CFG() || n_op == Op_StorePConditional || n_op == Op_StoreLConditional || n_op == Op_CompareAndSwapI || n_op == Op_CompareAndSwapL ||n_op == Op_CompareAndSwapP) return n;
|
|
557 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
|
|
558 n_op == Op_Opaque2 ) {
|
|
559 if( !C->major_progress() ) // If chance of no more loop opts...
|
|
560 _igvn._worklist.push(n); // maybe we'll remove them
|
|
561 return n;
|
|
562 }
|
|
563
|
|
564 if( n->is_Con() ) return n; // No cloning for Con nodes
|
|
565
|
|
566 Node *n_ctrl = get_ctrl(n);
|
|
567 if( !n_ctrl ) return n; // Dead node
|
|
568
|
|
569 // Attempt to remix address expressions for loop invariants
|
|
570 Node *m = remix_address_expressions( n );
|
|
571 if( m ) return m;
|
|
572
|
|
573 // Determine if the Node has inputs from some local Phi.
|
|
574 // Returns the block to clone thru.
|
|
575 Node *n_blk = has_local_phi_input( n );
|
|
576 if( !n_blk ) return n;
|
|
577 // Do not clone the trip counter through on a CountedLoop
|
|
578 // (messes up the canonical shape).
|
|
579 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
|
|
580
|
|
581 // Check for having no control input; not pinned. Allow
|
|
582 // dominating control.
|
|
583 if( n->in(0) ) {
|
|
584 Node *dom = idom(n_blk);
|
|
585 if( dom_lca( n->in(0), dom ) != n->in(0) )
|
|
586 return n;
|
|
587 }
|
|
588 // Policy: when is it profitable. You must get more wins than
|
|
589 // policy before it is considered profitable. Policy is usually 0,
|
|
590 // so 1 win is considered profitable. Big merges will require big
|
|
591 // cloning, so get a larger policy.
|
|
592 int policy = n_blk->req() >> 2;
|
|
593
|
|
594 // If the loop is a candidate for range check elimination,
|
|
595 // delay splitting through it's phi until a later loop optimization
|
|
596 if (n_blk->is_CountedLoop()) {
|
|
597 IdealLoopTree *lp = get_loop(n_blk);
|
|
598 if (lp && lp->_rce_candidate) {
|
|
599 return n;
|
|
600 }
|
|
601 }
|
|
602
|
|
603 // Use same limit as split_if_with_blocks_post
|
|
604 if( C->unique() > 35000 ) return n; // Method too big
|
|
605
|
|
606 // Split 'n' through the merge point if it is profitable
|
|
607 Node *phi = split_thru_phi( n, n_blk, policy );
|
|
608 if( !phi ) return n;
|
|
609
|
|
610 // Found a Phi to split thru!
|
|
611 // Replace 'n' with the new phi
|
|
612 _igvn.hash_delete(n);
|
|
613 _igvn.subsume_node( n, phi );
|
|
614 // Moved a load around the loop, 'en-registering' something.
|
|
615 if( n_blk->Opcode() == Op_Loop && n->is_Load() &&
|
|
616 !phi->in(LoopNode::LoopBackControl)->is_Load() )
|
|
617 C->set_major_progress();
|
|
618
|
|
619 return phi;
|
|
620 }
|
|
621
|
|
622 static bool merge_point_too_heavy(Compile* C, Node* region) {
|
|
623 // Bail out if the region and its phis have too many users.
|
|
624 int weight = 0;
|
|
625 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
|
|
626 weight += region->fast_out(i)->outcnt();
|
|
627 }
|
|
628 int nodes_left = MaxNodeLimit - C->unique();
|
|
629 if (weight * 8 > nodes_left) {
|
|
630 #ifndef PRODUCT
|
|
631 if (PrintOpto)
|
|
632 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
|
|
633 #endif
|
|
634 return true;
|
|
635 } else {
|
|
636 return false;
|
|
637 }
|
|
638 }
|
|
639
|
|
640 #ifdef _LP64
|
|
641 static bool merge_point_safe(Node* region) {
|
|
642 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
|
|
643 // having a PhiNode input. This sidesteps the dangerous case where the split
|
|
644 // ConvI2LNode may become TOP if the input Value() does not
|
|
645 // overlap the ConvI2L range, leaving a node which may not dominate its
|
|
646 // uses.
|
|
647 // A better fix for this problem can be found in the BugTraq entry, but
|
|
648 // expediency for Mantis demands this hack.
|
|
649 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
|
|
650 Node* n = region->fast_out(i);
|
|
651 if (n->is_Phi()) {
|
|
652 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
|
|
653 Node* m = n->fast_out(j);
|
|
654 if (m->Opcode() == Op_ConvI2L) {
|
|
655 return false;
|
|
656 }
|
|
657 }
|
|
658 }
|
|
659 }
|
|
660 return true;
|
|
661 }
|
|
662 #endif
|
|
663
|
|
664
|
|
665 //------------------------------place_near_use---------------------------------
|
|
666 // Place some computation next to use but not inside inner loops.
|
|
667 // For inner loop uses move it to the preheader area.
|
|
668 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const {
|
|
669 IdealLoopTree *u_loop = get_loop( useblock );
|
|
670 return (u_loop->_irreducible || u_loop->_child)
|
|
671 ? useblock
|
|
672 : u_loop->_head->in(LoopNode::EntryControl);
|
|
673 }
|
|
674
|
|
675
|
|
676 //------------------------------split_if_with_blocks_post----------------------
|
|
677 // Do the real work in a non-recursive function. CFG hackery wants to be
|
|
678 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
|
|
679 // info.
|
|
680 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) {
|
|
681
|
|
682 // Cloning Cmp through Phi's involves the split-if transform.
|
|
683 // FastLock is not used by an If
|
|
684 if( n->is_Cmp() && !n->is_FastLock() ) {
|
|
685 if( C->unique() > 35000 ) return; // Method too big
|
|
686
|
|
687 // Do not do 'split-if' if irreducible loops are present.
|
|
688 if( _has_irreducible_loops )
|
|
689 return;
|
|
690
|
|
691 Node *n_ctrl = get_ctrl(n);
|
|
692 // Determine if the Node has inputs from some local Phi.
|
|
693 // Returns the block to clone thru.
|
|
694 Node *n_blk = has_local_phi_input( n );
|
|
695 if( n_blk != n_ctrl ) return;
|
|
696
|
|
697 if( merge_point_too_heavy(C, n_ctrl) )
|
|
698 return;
|
|
699
|
|
700 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare?
|
|
701 Node *bol = n->unique_out();
|
|
702 assert( bol->is_Bool(), "expect a bool here" );
|
|
703 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare?
|
|
704 Node *iff = bol->unique_out();
|
|
705
|
|
706 // Check some safety conditions
|
|
707 if( iff->is_If() ) { // Classic split-if?
|
|
708 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if
|
|
709 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
|
|
710 if( get_ctrl(iff->in(2)) == n_ctrl ||
|
|
711 get_ctrl(iff->in(3)) == n_ctrl )
|
|
712 return; // Inputs not yet split-up
|
|
713 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) {
|
|
714 return; // Loop-invar test gates loop-varying CMOVE
|
|
715 }
|
|
716 } else {
|
|
717 return; // some other kind of node, such as an Allocate
|
|
718 }
|
|
719
|
|
720 // Do not do 'split-if' if some paths are dead. First do dead code
|
|
721 // elimination and then see if its still profitable.
|
|
722 for( uint i = 1; i < n_ctrl->req(); i++ )
|
|
723 if( n_ctrl->in(i) == C->top() )
|
|
724 return;
|
|
725
|
|
726 // When is split-if profitable? Every 'win' on means some control flow
|
|
727 // goes dead, so it's almost always a win.
|
|
728 int policy = 0;
|
|
729 // If trying to do a 'Split-If' at the loop head, it is only
|
|
730 // profitable if the cmp folds up on BOTH paths. Otherwise we
|
|
731 // risk peeling a loop forever.
|
|
732
|
|
733 // CNC - Disabled for now. Requires careful handling of loop
|
|
734 // body selection for the cloned code. Also, make sure we check
|
|
735 // for any input path not being in the same loop as n_ctrl. For
|
|
736 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
|
|
737 // because the alternative loop entry points won't be converted
|
|
738 // into LoopNodes.
|
|
739 IdealLoopTree *n_loop = get_loop(n_ctrl);
|
|
740 for( uint j = 1; j < n_ctrl->req(); j++ )
|
|
741 if( get_loop(n_ctrl->in(j)) != n_loop )
|
|
742 return;
|
|
743
|
|
744 #ifdef _LP64
|
|
745 // Check for safety of the merge point.
|
|
746 if( !merge_point_safe(n_ctrl) ) {
|
|
747 return;
|
|
748 }
|
|
749 #endif
|
|
750
|
|
751 // Split compare 'n' through the merge point if it is profitable
|
|
752 Node *phi = split_thru_phi( n, n_ctrl, policy );
|
|
753 if( !phi ) return;
|
|
754
|
|
755 // Found a Phi to split thru!
|
|
756 // Replace 'n' with the new phi
|
|
757 _igvn.hash_delete(n);
|
|
758 _igvn.subsume_node( n, phi );
|
|
759
|
|
760 // Now split the bool up thru the phi
|
|
761 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 );
|
|
762 _igvn.hash_delete(bol);
|
|
763 _igvn.subsume_node( bol, bolphi );
|
|
764 assert( iff->in(1) == bolphi, "" );
|
|
765 if( bolphi->Value(&_igvn)->singleton() )
|
|
766 return;
|
|
767
|
|
768 // Conditional-move? Must split up now
|
|
769 if( !iff->is_If() ) {
|
|
770 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 );
|
|
771 _igvn.hash_delete(iff);
|
|
772 _igvn.subsume_node( iff, cmovphi );
|
|
773 return;
|
|
774 }
|
|
775
|
|
776 // Now split the IF
|
|
777 do_split_if( iff );
|
|
778 return;
|
|
779 }
|
|
780
|
|
781 // Check for an IF ready to split; one that has its
|
|
782 // condition codes input coming from a Phi at the block start.
|
|
783 int n_op = n->Opcode();
|
|
784
|
|
785 // Check for an IF being dominated by another IF same test
|
|
786 if( n_op == Op_If ) {
|
|
787 Node *bol = n->in(1);
|
|
788 uint max = bol->outcnt();
|
|
789 // Check for same test used more than once?
|
|
790 if( n_op == Op_If && max > 1 && bol->is_Bool() ) {
|
|
791 // Search up IDOMs to see if this IF is dominated.
|
|
792 Node *cutoff = get_ctrl(bol);
|
|
793
|
|
794 // Now search up IDOMs till cutoff, looking for a dominating test
|
|
795 Node *prevdom = n;
|
|
796 Node *dom = idom(prevdom);
|
|
797 while( dom != cutoff ) {
|
|
798 if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) {
|
|
799 // Replace the dominated test with an obvious true or false.
|
|
800 // Place it on the IGVN worklist for later cleanup.
|
|
801 C->set_major_progress();
|
|
802 dominated_by( prevdom, n );
|
|
803 #ifndef PRODUCT
|
|
804 if( VerifyLoopOptimizations ) verify();
|
|
805 #endif
|
|
806 return;
|
|
807 }
|
|
808 prevdom = dom;
|
|
809 dom = idom(prevdom);
|
|
810 }
|
|
811 }
|
|
812 }
|
|
813
|
|
814 // See if a shared loop-varying computation has no loop-varying uses.
|
|
815 // Happens if something is only used for JVM state in uncommon trap exits,
|
|
816 // like various versions of induction variable+offset. Clone the
|
|
817 // computation per usage to allow it to sink out of the loop.
|
|
818 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
|
|
819 Node *n_ctrl = get_ctrl(n);
|
|
820 IdealLoopTree *n_loop = get_loop(n_ctrl);
|
|
821 if( n_loop != _ltree_root ) {
|
|
822 DUIterator_Fast imax, i = n->fast_outs(imax);
|
|
823 for (; i < imax; i++) {
|
|
824 Node* u = n->fast_out(i);
|
|
825 if( !has_ctrl(u) ) break; // Found control user
|
|
826 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
|
|
827 if( u_loop == n_loop ) break; // Found loop-varying use
|
|
828 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
|
|
829 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
|
|
830 }
|
|
831 bool did_break = (i < imax); // Did we break out of the previous loop?
|
|
832 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
|
|
833 Node *late_load_ctrl;
|
|
834 if (n->is_Load()) {
|
|
835 // If n is a load, get and save the result from get_late_ctrl(),
|
|
836 // to be later used in calculating the control for n's clones.
|
|
837 clear_dom_lca_tags();
|
|
838 late_load_ctrl = get_late_ctrl(n, n_ctrl);
|
|
839 }
|
|
840 // If n is a load, and the late control is the same as the current
|
|
841 // control, then the cloning of n is a pointless exercise, because
|
|
842 // GVN will ensure that we end up where we started.
|
|
843 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
|
|
844 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
|
|
845 Node *u = n->last_out(j); // Clone private computation per use
|
|
846 _igvn.hash_delete(u);
|
|
847 _igvn._worklist.push(u);
|
|
848 Node *x = n->clone(); // Clone computation
|
|
849 Node *x_ctrl = NULL;
|
|
850 if( u->is_Phi() ) {
|
|
851 // Replace all uses of normal nodes. Replace Phi uses
|
|
852 // individually, so the seperate Nodes can sink down
|
|
853 // different paths.
|
|
854 uint k = 1;
|
|
855 while( u->in(k) != n ) k++;
|
|
856 u->set_req( k, x );
|
|
857 // x goes next to Phi input path
|
|
858 x_ctrl = u->in(0)->in(k);
|
|
859 --j;
|
|
860 } else { // Normal use
|
|
861 // Replace all uses
|
|
862 for( uint k = 0; k < u->req(); k++ ) {
|
|
863 if( u->in(k) == n ) {
|
|
864 u->set_req( k, x );
|
|
865 --j;
|
|
866 }
|
|
867 }
|
|
868 x_ctrl = get_ctrl(u);
|
|
869 }
|
|
870
|
|
871 // Find control for 'x' next to use but not inside inner loops.
|
|
872 // For inner loop uses get the preheader area.
|
|
873 x_ctrl = place_near_use(x_ctrl);
|
|
874
|
|
875 if (n->is_Load()) {
|
|
876 // For loads, add a control edge to a CFG node outside of the loop
|
|
877 // to force them to not combine and return back inside the loop
|
|
878 // during GVN optimization (4641526).
|
|
879 //
|
|
880 // Because we are setting the actual control input, factor in
|
|
881 // the result from get_late_ctrl() so we respect any
|
|
882 // anti-dependences. (6233005).
|
|
883 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
|
|
884
|
|
885 // Don't allow the control input to be a CFG splitting node.
|
|
886 // Such nodes should only have ProjNodes as outs, e.g. IfNode
|
|
887 // should only have IfTrueNode and IfFalseNode (4985384).
|
|
888 x_ctrl = find_non_split_ctrl(x_ctrl);
|
|
889 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
|
|
890
|
|
891 x->set_req(0, x_ctrl);
|
|
892 }
|
|
893 register_new_node(x, x_ctrl);
|
|
894
|
|
895 // Some institutional knowledge is needed here: 'x' is
|
|
896 // yanked because if the optimizer runs GVN on it all the
|
|
897 // cloned x's will common up and undo this optimization and
|
|
898 // be forced back in the loop. This is annoying because it
|
|
899 // makes +VerifyOpto report false-positives on progress. I
|
|
900 // tried setting control edges on the x's to force them to
|
|
901 // not combine, but the matching gets worried when it tries
|
|
902 // to fold a StoreP and an AddP together (as part of an
|
|
903 // address expression) and the AddP and StoreP have
|
|
904 // different controls.
|
|
905 if( !x->is_Load() ) _igvn._worklist.yank(x);
|
|
906 }
|
|
907 _igvn.remove_dead_node(n);
|
|
908 }
|
|
909 }
|
|
910 }
|
|
911 }
|
|
912
|
|
913 // Check for Opaque2's who's loop has disappeared - who's input is in the
|
|
914 // same loop nest as their output. Remove 'em, they are no longer useful.
|
|
915 if( n_op == Op_Opaque2 &&
|
|
916 n->in(1) != NULL &&
|
|
917 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
|
|
918 _igvn.add_users_to_worklist(n);
|
|
919 _igvn.hash_delete(n);
|
|
920 _igvn.subsume_node( n, n->in(1) );
|
|
921 }
|
|
922 }
|
|
923
|
|
924 //------------------------------split_if_with_blocks---------------------------
|
|
925 // Check for aggressive application of 'split-if' optimization,
|
|
926 // using basic block level info.
|
|
927 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) {
|
|
928 Node *n = C->root();
|
|
929 visited.set(n->_idx); // first, mark node as visited
|
|
930 // Do pre-visit work for root
|
|
931 n = split_if_with_blocks_pre( n );
|
|
932 uint cnt = n->outcnt();
|
|
933 uint i = 0;
|
|
934 while (true) {
|
|
935 // Visit all children
|
|
936 if (i < cnt) {
|
|
937 Node* use = n->raw_out(i);
|
|
938 ++i;
|
|
939 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
|
|
940 // Now do pre-visit work for this use
|
|
941 use = split_if_with_blocks_pre( use );
|
|
942 nstack.push(n, i); // Save parent and next use's index.
|
|
943 n = use; // Process all children of current use.
|
|
944 cnt = use->outcnt();
|
|
945 i = 0;
|
|
946 }
|
|
947 }
|
|
948 else {
|
|
949 // All of n's children have been processed, complete post-processing.
|
|
950 if (cnt != 0 && !n->is_Con()) {
|
|
951 assert(has_node(n), "no dead nodes");
|
|
952 split_if_with_blocks_post( n );
|
|
953 }
|
|
954 if (nstack.is_empty()) {
|
|
955 // Finished all nodes on stack.
|
|
956 break;
|
|
957 }
|
|
958 // Get saved parent node and next use's index. Visit the rest of uses.
|
|
959 n = nstack.node();
|
|
960 cnt = n->outcnt();
|
|
961 i = nstack.index();
|
|
962 nstack.pop();
|
|
963 }
|
|
964 }
|
|
965 }
|
|
966
|
|
967
|
|
968 //=============================================================================
|
|
969 //
|
|
970 // C L O N E A L O O P B O D Y
|
|
971 //
|
|
972
|
|
973 //------------------------------clone_iff--------------------------------------
|
|
974 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
|
|
975 // "Nearly" because all Nodes have been cloned from the original in the loop,
|
|
976 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
|
|
977 // through the Phi recursively, and return a Bool.
|
|
978 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) {
|
|
979
|
|
980 // Convert this Phi into a Phi merging Bools
|
|
981 uint i;
|
|
982 for( i = 1; i < phi->req(); i++ ) {
|
|
983 Node *b = phi->in(i);
|
|
984 if( b->is_Phi() ) {
|
|
985 _igvn.hash_delete(phi);
|
|
986 _igvn._worklist.push(phi);
|
|
987 phi->set_req(i, clone_iff( b->as_Phi(), loop ));
|
|
988 } else {
|
|
989 assert( b->is_Bool(), "" );
|
|
990 }
|
|
991 }
|
|
992
|
|
993 Node *sample_bool = phi->in(1);
|
|
994 Node *sample_cmp = sample_bool->in(1);
|
|
995
|
|
996 // Make Phis to merge the Cmp's inputs.
|
|
997 int size = phi->in(0)->req();
|
|
998 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
|
|
999 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
|
|
1000 for( i = 1; i < phi->req(); i++ ) {
|
|
1001 Node *n1 = phi->in(i)->in(1)->in(1);
|
|
1002 Node *n2 = phi->in(i)->in(1)->in(2);
|
|
1003 phi1->set_req( i, n1 );
|
|
1004 phi2->set_req( i, n2 );
|
|
1005 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
|
|
1006 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
|
|
1007 }
|
|
1008 // See if these Phis have been made before.
|
|
1009 // Register with optimizer
|
|
1010 Node *hit1 = _igvn.hash_find_insert(phi1);
|
|
1011 if( hit1 ) { // Hit, toss just made Phi
|
|
1012 _igvn.remove_dead_node(phi1); // Remove new phi
|
|
1013 assert( hit1->is_Phi(), "" );
|
|
1014 phi1 = (PhiNode*)hit1; // Use existing phi
|
|
1015 } else { // Miss
|
|
1016 _igvn.register_new_node_with_optimizer(phi1);
|
|
1017 }
|
|
1018 Node *hit2 = _igvn.hash_find_insert(phi2);
|
|
1019 if( hit2 ) { // Hit, toss just made Phi
|
|
1020 _igvn.remove_dead_node(phi2); // Remove new phi
|
|
1021 assert( hit2->is_Phi(), "" );
|
|
1022 phi2 = (PhiNode*)hit2; // Use existing phi
|
|
1023 } else { // Miss
|
|
1024 _igvn.register_new_node_with_optimizer(phi2);
|
|
1025 }
|
|
1026 // Register Phis with loop/block info
|
|
1027 set_ctrl(phi1, phi->in(0));
|
|
1028 set_ctrl(phi2, phi->in(0));
|
|
1029 // Make a new Cmp
|
|
1030 Node *cmp = sample_cmp->clone();
|
|
1031 cmp->set_req( 1, phi1 );
|
|
1032 cmp->set_req( 2, phi2 );
|
|
1033 _igvn.register_new_node_with_optimizer(cmp);
|
|
1034 set_ctrl(cmp, phi->in(0));
|
|
1035
|
|
1036 // Make a new Bool
|
|
1037 Node *b = sample_bool->clone();
|
|
1038 b->set_req(1,cmp);
|
|
1039 _igvn.register_new_node_with_optimizer(b);
|
|
1040 set_ctrl(b, phi->in(0));
|
|
1041
|
|
1042 assert( b->is_Bool(), "" );
|
|
1043 return (BoolNode*)b;
|
|
1044 }
|
|
1045
|
|
1046 //------------------------------clone_bool-------------------------------------
|
|
1047 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
|
|
1048 // "Nearly" because all Nodes have been cloned from the original in the loop,
|
|
1049 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
|
|
1050 // through the Phi recursively, and return a Bool.
|
|
1051 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
|
|
1052 uint i;
|
|
1053 // Convert this Phi into a Phi merging Bools
|
|
1054 for( i = 1; i < phi->req(); i++ ) {
|
|
1055 Node *b = phi->in(i);
|
|
1056 if( b->is_Phi() ) {
|
|
1057 _igvn.hash_delete(phi);
|
|
1058 _igvn._worklist.push(phi);
|
|
1059 phi->set_req(i, clone_bool( b->as_Phi(), loop ));
|
|
1060 } else {
|
|
1061 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
|
|
1062 }
|
|
1063 }
|
|
1064
|
|
1065 Node *sample_cmp = phi->in(1);
|
|
1066
|
|
1067 // Make Phis to merge the Cmp's inputs.
|
|
1068 int size = phi->in(0)->req();
|
|
1069 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
|
|
1070 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
|
|
1071 for( uint j = 1; j < phi->req(); j++ ) {
|
|
1072 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
|
|
1073 Node *n1, *n2;
|
|
1074 if( cmp_top->is_Cmp() ) {
|
|
1075 n1 = cmp_top->in(1);
|
|
1076 n2 = cmp_top->in(2);
|
|
1077 } else {
|
|
1078 n1 = n2 = cmp_top;
|
|
1079 }
|
|
1080 phi1->set_req( j, n1 );
|
|
1081 phi2->set_req( j, n2 );
|
|
1082 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
|
|
1083 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
|
|
1084 }
|
|
1085
|
|
1086 // See if these Phis have been made before.
|
|
1087 // Register with optimizer
|
|
1088 Node *hit1 = _igvn.hash_find_insert(phi1);
|
|
1089 if( hit1 ) { // Hit, toss just made Phi
|
|
1090 _igvn.remove_dead_node(phi1); // Remove new phi
|
|
1091 assert( hit1->is_Phi(), "" );
|
|
1092 phi1 = (PhiNode*)hit1; // Use existing phi
|
|
1093 } else { // Miss
|
|
1094 _igvn.register_new_node_with_optimizer(phi1);
|
|
1095 }
|
|
1096 Node *hit2 = _igvn.hash_find_insert(phi2);
|
|
1097 if( hit2 ) { // Hit, toss just made Phi
|
|
1098 _igvn.remove_dead_node(phi2); // Remove new phi
|
|
1099 assert( hit2->is_Phi(), "" );
|
|
1100 phi2 = (PhiNode*)hit2; // Use existing phi
|
|
1101 } else { // Miss
|
|
1102 _igvn.register_new_node_with_optimizer(phi2);
|
|
1103 }
|
|
1104 // Register Phis with loop/block info
|
|
1105 set_ctrl(phi1, phi->in(0));
|
|
1106 set_ctrl(phi2, phi->in(0));
|
|
1107 // Make a new Cmp
|
|
1108 Node *cmp = sample_cmp->clone();
|
|
1109 cmp->set_req( 1, phi1 );
|
|
1110 cmp->set_req( 2, phi2 );
|
|
1111 _igvn.register_new_node_with_optimizer(cmp);
|
|
1112 set_ctrl(cmp, phi->in(0));
|
|
1113
|
|
1114 assert( cmp->is_Cmp(), "" );
|
|
1115 return (CmpNode*)cmp;
|
|
1116 }
|
|
1117
|
|
1118 //------------------------------sink_use---------------------------------------
|
|
1119 // If 'use' was in the loop-exit block, it now needs to be sunk
|
|
1120 // below the post-loop merge point.
|
|
1121 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
|
|
1122 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
|
|
1123 set_ctrl(use, post_loop);
|
|
1124 for (DUIterator j = use->outs(); use->has_out(j); j++)
|
|
1125 sink_use(use->out(j), post_loop);
|
|
1126 }
|
|
1127 }
|
|
1128
|
|
1129 //------------------------------clone_loop-------------------------------------
|
|
1130 //
|
|
1131 // C L O N E A L O O P B O D Y
|
|
1132 //
|
|
1133 // This is the basic building block of the loop optimizations. It clones an
|
|
1134 // entire loop body. It makes an old_new loop body mapping; with this mapping
|
|
1135 // you can find the new-loop equivalent to an old-loop node. All new-loop
|
|
1136 // nodes are exactly equal to their old-loop counterparts, all edges are the
|
|
1137 // same. All exits from the old-loop now have a RegionNode that merges the
|
|
1138 // equivalent new-loop path. This is true even for the normal "loop-exit"
|
|
1139 // condition. All uses of loop-invariant old-loop values now come from (one
|
|
1140 // or more) Phis that merge their new-loop equivalents.
|
|
1141 //
|
|
1142 // This operation leaves the graph in an illegal state: there are two valid
|
|
1143 // control edges coming from the loop pre-header to both loop bodies. I'll
|
|
1144 // definitely have to hack the graph after running this transform.
|
|
1145 //
|
|
1146 // From this building block I will further edit edges to perform loop peeling
|
|
1147 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
|
|
1148 //
|
|
1149 // Parameter side_by_size_idom:
|
|
1150 // When side_by_size_idom is NULL, the dominator tree is constructed for
|
|
1151 // the clone loop to dominate the original. Used in construction of
|
|
1152 // pre-main-post loop sequence.
|
|
1153 // When nonnull, the clone and original are side-by-side, both are
|
|
1154 // dominated by the side_by_side_idom node. Used in construction of
|
|
1155 // unswitched loops.
|
|
1156 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
|
|
1157 Node* side_by_side_idom) {
|
|
1158
|
|
1159 // Step 1: Clone the loop body. Make the old->new mapping.
|
|
1160 uint i;
|
|
1161 for( i = 0; i < loop->_body.size(); i++ ) {
|
|
1162 Node *old = loop->_body.at(i);
|
|
1163 Node *nnn = old->clone();
|
|
1164 old_new.map( old->_idx, nnn );
|
|
1165 _igvn.register_new_node_with_optimizer(nnn);
|
|
1166 }
|
|
1167
|
|
1168
|
|
1169 // Step 2: Fix the edges in the new body. If the old input is outside the
|
|
1170 // loop use it. If the old input is INside the loop, use the corresponding
|
|
1171 // new node instead.
|
|
1172 for( i = 0; i < loop->_body.size(); i++ ) {
|
|
1173 Node *old = loop->_body.at(i);
|
|
1174 Node *nnn = old_new[old->_idx];
|
|
1175 // Fix CFG/Loop controlling the new node
|
|
1176 if (has_ctrl(old)) {
|
|
1177 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
|
|
1178 } else {
|
|
1179 set_loop(nnn, loop->_parent);
|
|
1180 if (old->outcnt() > 0) {
|
|
1181 set_idom( nnn, old_new[idom(old)->_idx], dd );
|
|
1182 }
|
|
1183 }
|
|
1184 // Correct edges to the new node
|
|
1185 for( uint j = 0; j < nnn->req(); j++ ) {
|
|
1186 Node *n = nnn->in(j);
|
|
1187 if( n ) {
|
|
1188 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
|
|
1189 if( loop->is_member( old_in_loop ) )
|
|
1190 nnn->set_req(j, old_new[n->_idx]);
|
|
1191 }
|
|
1192 }
|
|
1193 _igvn.hash_find_insert(nnn);
|
|
1194 }
|
|
1195 Node *newhead = old_new[loop->_head->_idx];
|
|
1196 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
|
|
1197
|
|
1198
|
|
1199 // Step 3: Now fix control uses. Loop varying control uses have already
|
|
1200 // been fixed up (as part of all input edges in Step 2). Loop invariant
|
|
1201 // control uses must be either an IfFalse or an IfTrue. Make a merge
|
|
1202 // point to merge the old and new IfFalse/IfTrue nodes; make the use
|
|
1203 // refer to this.
|
|
1204 ResourceArea *area = Thread::current()->resource_area();
|
|
1205 Node_List worklist(area);
|
|
1206 uint new_counter = C->unique();
|
|
1207 for( i = 0; i < loop->_body.size(); i++ ) {
|
|
1208 Node* old = loop->_body.at(i);
|
|
1209 if( !old->is_CFG() ) continue;
|
|
1210 Node* nnn = old_new[old->_idx];
|
|
1211
|
|
1212 // Copy uses to a worklist, so I can munge the def-use info
|
|
1213 // with impunity.
|
|
1214 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
|
|
1215 worklist.push(old->fast_out(j));
|
|
1216
|
|
1217 while( worklist.size() ) { // Visit all uses
|
|
1218 Node *use = worklist.pop();
|
|
1219 if (!has_node(use)) continue; // Ignore dead nodes
|
|
1220 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
|
|
1221 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
|
|
1222 // Both OLD and USE are CFG nodes here.
|
|
1223 assert( use->is_Proj(), "" );
|
|
1224
|
|
1225 // Clone the loop exit control projection
|
|
1226 Node *newuse = use->clone();
|
|
1227 newuse->set_req(0,nnn);
|
|
1228 _igvn.register_new_node_with_optimizer(newuse);
|
|
1229 set_loop(newuse, use_loop);
|
|
1230 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
|
|
1231
|
|
1232 // We need a Region to merge the exit from the peeled body and the
|
|
1233 // exit from the old loop body.
|
|
1234 RegionNode *r = new (C, 3) RegionNode(3);
|
|
1235 // Map the old use to the new merge point
|
|
1236 old_new.map( use->_idx, r );
|
|
1237 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
|
|
1238 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
|
|
1239
|
|
1240 // The original user of 'use' uses 'r' instead.
|
|
1241 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
|
|
1242 Node* useuse = use->last_out(l);
|
|
1243 _igvn.hash_delete(useuse);
|
|
1244 _igvn._worklist.push(useuse);
|
|
1245 uint uses_found = 0;
|
|
1246 if( useuse->in(0) == use ) {
|
|
1247 useuse->set_req(0, r);
|
|
1248 uses_found++;
|
|
1249 if( useuse->is_CFG() ) {
|
|
1250 assert( dom_depth(useuse) > dd_r, "" );
|
|
1251 set_idom(useuse, r, dom_depth(useuse));
|
|
1252 }
|
|
1253 }
|
|
1254 for( uint k = 1; k < useuse->req(); k++ ) {
|
|
1255 if( useuse->in(k) == use ) {
|
|
1256 useuse->set_req(k, r);
|
|
1257 uses_found++;
|
|
1258 }
|
|
1259 }
|
|
1260 l -= uses_found; // we deleted 1 or more copies of this edge
|
|
1261 }
|
|
1262
|
|
1263 // Now finish up 'r'
|
|
1264 r->set_req( 1, newuse );
|
|
1265 r->set_req( 2, use );
|
|
1266 _igvn.register_new_node_with_optimizer(r);
|
|
1267 set_loop(r, use_loop);
|
|
1268 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
|
|
1269 } // End of if a loop-exit test
|
|
1270 }
|
|
1271 }
|
|
1272
|
|
1273 // Step 4: If loop-invariant use is not control, it must be dominated by a
|
|
1274 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
|
|
1275 // there if needed. Make a Phi there merging old and new used values.
|
|
1276 Node_List *split_if_set = NULL;
|
|
1277 Node_List *split_bool_set = NULL;
|
|
1278 Node_List *split_cex_set = NULL;
|
|
1279 for( i = 0; i < loop->_body.size(); i++ ) {
|
|
1280 Node* old = loop->_body.at(i);
|
|
1281 Node* nnn = old_new[old->_idx];
|
|
1282 // Copy uses to a worklist, so I can munge the def-use info
|
|
1283 // with impunity.
|
|
1284 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
|
|
1285 worklist.push(old->fast_out(j));
|
|
1286
|
|
1287 while( worklist.size() ) {
|
|
1288 Node *use = worklist.pop();
|
|
1289 if (!has_node(use)) continue; // Ignore dead nodes
|
|
1290 if (use->in(0) == C->top()) continue;
|
|
1291 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
|
|
1292 // Check for data-use outside of loop - at least one of OLD or USE
|
|
1293 // must not be a CFG node.
|
|
1294 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) {
|
|
1295
|
|
1296 // If the Data use is an IF, that means we have an IF outside of the
|
|
1297 // loop that is switching on a condition that is set inside of the
|
|
1298 // loop. Happens if people set a loop-exit flag; then test the flag
|
|
1299 // in the loop to break the loop, then test is again outside of the
|
|
1300 // loop to determine which way the loop exited.
|
|
1301 if( use->is_If() || use->is_CMove() ) {
|
|
1302 // Since this code is highly unlikely, we lazily build the worklist
|
|
1303 // of such Nodes to go split.
|
|
1304 if( !split_if_set )
|
|
1305 split_if_set = new Node_List(area);
|
|
1306 split_if_set->push(use);
|
|
1307 }
|
|
1308 if( use->is_Bool() ) {
|
|
1309 if( !split_bool_set )
|
|
1310 split_bool_set = new Node_List(area);
|
|
1311 split_bool_set->push(use);
|
|
1312 }
|
|
1313 if( use->Opcode() == Op_CreateEx ) {
|
|
1314 if( !split_cex_set )
|
|
1315 split_cex_set = new Node_List(area);
|
|
1316 split_cex_set->push(use);
|
|
1317 }
|
|
1318
|
|
1319
|
|
1320 // Get "block" use is in
|
|
1321 uint idx = 0;
|
|
1322 while( use->in(idx) != old ) idx++;
|
|
1323 Node *prev = use->is_CFG() ? use : get_ctrl(use);
|
|
1324 assert( !loop->is_member( get_loop( prev ) ), "" );
|
|
1325 Node *cfg = prev->_idx >= new_counter
|
|
1326 ? prev->in(2)
|
|
1327 : idom(prev);
|
|
1328 if( use->is_Phi() ) // Phi use is in prior block
|
|
1329 cfg = prev->in(idx); // NOT in block of Phi itself
|
|
1330 if (cfg->is_top()) { // Use is dead?
|
|
1331 _igvn.hash_delete(use);
|
|
1332 _igvn._worklist.push(use);
|
|
1333 use->set_req(idx, C->top());
|
|
1334 continue;
|
|
1335 }
|
|
1336
|
|
1337 while( !loop->is_member( get_loop( cfg ) ) ) {
|
|
1338 prev = cfg;
|
|
1339 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
|
|
1340 }
|
|
1341 // If the use occurs after merging several exits from the loop, then
|
|
1342 // old value must have dominated all those exits. Since the same old
|
|
1343 // value was used on all those exits we did not need a Phi at this
|
|
1344 // merge point. NOW we do need a Phi here. Each loop exit value
|
|
1345 // is now merged with the peeled body exit; each exit gets its own
|
|
1346 // private Phi and those Phis need to be merged here.
|
|
1347 Node *phi;
|
|
1348 if( prev->is_Region() ) {
|
|
1349 if( idx == 0 ) { // Updating control edge?
|
|
1350 phi = prev; // Just use existing control
|
|
1351 } else { // Else need a new Phi
|
|
1352 phi = PhiNode::make( prev, old );
|
|
1353 // Now recursively fix up the new uses of old!
|
|
1354 for( uint i = 1; i < prev->req(); i++ ) {
|
|
1355 worklist.push(phi); // Onto worklist once for each 'old' input
|
|
1356 }
|
|
1357 }
|
|
1358 } else {
|
|
1359 // Get new RegionNode merging old and new loop exits
|
|
1360 prev = old_new[prev->_idx];
|
|
1361 assert( prev, "just made this in step 7" );
|
|
1362 if( idx == 0 ) { // Updating control edge?
|
|
1363 phi = prev; // Just use existing control
|
|
1364 } else { // Else need a new Phi
|
|
1365 // Make a new Phi merging data values properly
|
|
1366 phi = PhiNode::make( prev, old );
|
|
1367 phi->set_req( 1, nnn );
|
|
1368 }
|
|
1369 }
|
|
1370 // If inserting a new Phi, check for prior hits
|
|
1371 if( idx != 0 ) {
|
|
1372 Node *hit = _igvn.hash_find_insert(phi);
|
|
1373 if( hit == NULL ) {
|
|
1374 _igvn.register_new_node_with_optimizer(phi); // Register new phi
|
|
1375 } else { // or
|
|
1376 // Remove the new phi from the graph and use the hit
|
|
1377 _igvn.remove_dead_node(phi);
|
|
1378 phi = hit; // Use existing phi
|
|
1379 }
|
|
1380 set_ctrl(phi, prev);
|
|
1381 }
|
|
1382 // Make 'use' use the Phi instead of the old loop body exit value
|
|
1383 _igvn.hash_delete(use);
|
|
1384 _igvn._worklist.push(use);
|
|
1385 use->set_req(idx, phi);
|
|
1386 if( use->_idx >= new_counter ) { // If updating new phis
|
|
1387 // Not needed for correctness, but prevents a weak assert
|
|
1388 // in AddPNode from tripping (when we end up with different
|
|
1389 // base & derived Phis that will become the same after
|
|
1390 // IGVN does CSE).
|
|
1391 Node *hit = _igvn.hash_find_insert(use);
|
|
1392 if( hit ) // Go ahead and re-hash for hits.
|
|
1393 _igvn.subsume_node( use, hit );
|
|
1394 }
|
|
1395
|
|
1396 // If 'use' was in the loop-exit block, it now needs to be sunk
|
|
1397 // below the post-loop merge point.
|
|
1398 sink_use( use, prev );
|
|
1399 }
|
|
1400 }
|
|
1401 }
|
|
1402
|
|
1403 // Check for IFs that need splitting/cloning. Happens if an IF outside of
|
|
1404 // the loop uses a condition set in the loop. The original IF probably
|
|
1405 // takes control from one or more OLD Regions (which in turn get from NEW
|
|
1406 // Regions). In any case, there will be a set of Phis for each merge point
|
|
1407 // from the IF up to where the original BOOL def exists the loop.
|
|
1408 if( split_if_set ) {
|
|
1409 while( split_if_set->size() ) {
|
|
1410 Node *iff = split_if_set->pop();
|
|
1411 if( iff->in(1)->is_Phi() ) {
|
|
1412 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop );
|
|
1413 _igvn.hash_delete(iff);
|
|
1414 _igvn._worklist.push(iff);
|
|
1415 iff->set_req(1, b);
|
|
1416 }
|
|
1417 }
|
|
1418 }
|
|
1419 if( split_bool_set ) {
|
|
1420 while( split_bool_set->size() ) {
|
|
1421 Node *b = split_bool_set->pop();
|
|
1422 Node *phi = b->in(1);
|
|
1423 assert( phi->is_Phi(), "" );
|
|
1424 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop );
|
|
1425 _igvn.hash_delete(b);
|
|
1426 _igvn._worklist.push(b);
|
|
1427 b->set_req(1, cmp);
|
|
1428 }
|
|
1429 }
|
|
1430 if( split_cex_set ) {
|
|
1431 while( split_cex_set->size() ) {
|
|
1432 Node *b = split_cex_set->pop();
|
|
1433 assert( b->in(0)->is_Region(), "" );
|
|
1434 assert( b->in(1)->is_Phi(), "" );
|
|
1435 assert( b->in(0)->in(0) == b->in(1)->in(0), "" );
|
|
1436 split_up( b, b->in(0), NULL );
|
|
1437 }
|
|
1438 }
|
|
1439
|
|
1440 }
|
|
1441
|
|
1442
|
|
1443 //---------------------- stride_of_possible_iv -------------------------------------
|
|
1444 // Looks for an iff/bool/comp with one operand of the compare
|
|
1445 // being a cycle involving an add and a phi,
|
|
1446 // with an optional truncation (left-shift followed by a right-shift)
|
|
1447 // of the add. Returns zero if not an iv.
|
|
1448 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
|
|
1449 Node* trunc1 = NULL;
|
|
1450 Node* trunc2 = NULL;
|
|
1451 const TypeInt* ttype = NULL;
|
|
1452 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
|
|
1453 return 0;
|
|
1454 }
|
|
1455 BoolNode* bl = iff->in(1)->as_Bool();
|
|
1456 Node* cmp = bl->in(1);
|
|
1457 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) {
|
|
1458 return 0;
|
|
1459 }
|
|
1460 // Must have an invariant operand
|
|
1461 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
|
|
1462 return 0;
|
|
1463 }
|
|
1464 Node* add2 = NULL;
|
|
1465 Node* cmp1 = cmp->in(1);
|
|
1466 if (cmp1->is_Phi()) {
|
|
1467 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
|
|
1468 Node* phi = cmp1;
|
|
1469 for (uint i = 1; i < phi->req(); i++) {
|
|
1470 Node* in = phi->in(i);
|
|
1471 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
|
|
1472 &trunc1, &trunc2, &ttype);
|
|
1473 if (add && add->in(1) == phi) {
|
|
1474 add2 = add->in(2);
|
|
1475 break;
|
|
1476 }
|
|
1477 }
|
|
1478 } else {
|
|
1479 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
|
|
1480 Node* addtrunc = cmp1;
|
|
1481 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
|
|
1482 &trunc1, &trunc2, &ttype);
|
|
1483 if (add && add->in(1)->is_Phi()) {
|
|
1484 Node* phi = add->in(1);
|
|
1485 for (uint i = 1; i < phi->req(); i++) {
|
|
1486 if (phi->in(i) == addtrunc) {
|
|
1487 add2 = add->in(2);
|
|
1488 break;
|
|
1489 }
|
|
1490 }
|
|
1491 }
|
|
1492 }
|
|
1493 if (add2 != NULL) {
|
|
1494 const TypeInt* add2t = _igvn.type(add2)->is_int();
|
|
1495 if (add2t->is_con()) {
|
|
1496 return add2t->get_con();
|
|
1497 }
|
|
1498 }
|
|
1499 return 0;
|
|
1500 }
|
|
1501
|
|
1502
|
|
1503 //---------------------- stay_in_loop -------------------------------------
|
|
1504 // Return the (unique) control output node that's in the loop (if it exists.)
|
|
1505 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
|
|
1506 Node* unique = NULL;
|
|
1507 if (!n) return NULL;
|
|
1508 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
|
|
1509 Node* use = n->fast_out(i);
|
|
1510 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
|
|
1511 if (unique != NULL) {
|
|
1512 return NULL;
|
|
1513 }
|
|
1514 unique = use;
|
|
1515 }
|
|
1516 }
|
|
1517 return unique;
|
|
1518 }
|
|
1519
|
|
1520 //------------------------------ register_node -------------------------------------
|
|
1521 // Utility to register node "n" with PhaseIdealLoop
|
|
1522 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
|
|
1523 _igvn.register_new_node_with_optimizer(n);
|
|
1524 loop->_body.push(n);
|
|
1525 if (n->is_CFG()) {
|
|
1526 set_loop(n, loop);
|
|
1527 set_idom(n, pred, ddepth);
|
|
1528 } else {
|
|
1529 set_ctrl(n, pred);
|
|
1530 }
|
|
1531 }
|
|
1532
|
|
1533 //------------------------------ proj_clone -------------------------------------
|
|
1534 // Utility to create an if-projection
|
|
1535 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
|
|
1536 ProjNode* c = p->clone()->as_Proj();
|
|
1537 c->set_req(0, iff);
|
|
1538 return c;
|
|
1539 }
|
|
1540
|
|
1541 //------------------------------ short_circuit_if -------------------------------------
|
|
1542 // Force the iff control output to be the live_proj
|
|
1543 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
|
|
1544 int proj_con = live_proj->_con;
|
|
1545 assert(proj_con == 0 || proj_con == 1, "false or true projection");
|
|
1546 Node *con = _igvn.intcon(proj_con);
|
|
1547 set_ctrl(con, C->root());
|
|
1548 if (iff) {
|
|
1549 iff->set_req(1, con);
|
|
1550 }
|
|
1551 return con;
|
|
1552 }
|
|
1553
|
|
1554 //------------------------------ insert_if_before_proj -------------------------------------
|
|
1555 // Insert a new if before an if projection (* - new node)
|
|
1556 //
|
|
1557 // before
|
|
1558 // if(test)
|
|
1559 // / \
|
|
1560 // v v
|
|
1561 // other-proj proj (arg)
|
|
1562 //
|
|
1563 // after
|
|
1564 // if(test)
|
|
1565 // / \
|
|
1566 // / v
|
|
1567 // | * proj-clone
|
|
1568 // v |
|
|
1569 // other-proj v
|
|
1570 // * new_if(relop(cmp[IU](left,right)))
|
|
1571 // / \
|
|
1572 // v v
|
|
1573 // * new-proj proj
|
|
1574 // (returned)
|
|
1575 //
|
|
1576 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
|
|
1577 IfNode* iff = proj->in(0)->as_If();
|
|
1578 IdealLoopTree *loop = get_loop(proj);
|
|
1579 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
|
|
1580 int ddepth = dom_depth(proj);
|
|
1581
|
|
1582 _igvn.hash_delete(iff);
|
|
1583 _igvn._worklist.push(iff);
|
|
1584 _igvn.hash_delete(proj);
|
|
1585 _igvn._worklist.push(proj);
|
|
1586
|
|
1587 proj->set_req(0, NULL); // temporary disconnect
|
|
1588 ProjNode* proj2 = proj_clone(proj, iff);
|
|
1589 register_node(proj2, loop, iff, ddepth);
|
|
1590
|
|
1591 Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right);
|
|
1592 register_node(cmp, loop, proj2, ddepth);
|
|
1593
|
|
1594 BoolNode* bol = new (C,2)BoolNode(cmp, relop);
|
|
1595 register_node(bol, loop, proj2, ddepth);
|
|
1596
|
|
1597 IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt);
|
|
1598 register_node(new_if, loop, proj2, ddepth);
|
|
1599
|
|
1600 proj->set_req(0, new_if); // reattach
|
|
1601 set_idom(proj, new_if, ddepth);
|
|
1602
|
|
1603 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
|
|
1604 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
|
|
1605
|
|
1606 return new_exit;
|
|
1607 }
|
|
1608
|
|
1609 //------------------------------ insert_region_before_proj -------------------------------------
|
|
1610 // Insert a region before an if projection (* - new node)
|
|
1611 //
|
|
1612 // before
|
|
1613 // if(test)
|
|
1614 // / |
|
|
1615 // v |
|
|
1616 // proj v
|
|
1617 // other-proj
|
|
1618 //
|
|
1619 // after
|
|
1620 // if(test)
|
|
1621 // / |
|
|
1622 // v |
|
|
1623 // * proj-clone v
|
|
1624 // | other-proj
|
|
1625 // v
|
|
1626 // * new-region
|
|
1627 // |
|
|
1628 // v
|
|
1629 // * dum_if
|
|
1630 // / \
|
|
1631 // v \
|
|
1632 // * dum-proj v
|
|
1633 // proj
|
|
1634 //
|
|
1635 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
|
|
1636 IfNode* iff = proj->in(0)->as_If();
|
|
1637 IdealLoopTree *loop = get_loop(proj);
|
|
1638 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
|
|
1639 int ddepth = dom_depth(proj);
|
|
1640
|
|
1641 _igvn.hash_delete(iff);
|
|
1642 _igvn._worklist.push(iff);
|
|
1643 _igvn.hash_delete(proj);
|
|
1644 _igvn._worklist.push(proj);
|
|
1645
|
|
1646 proj->set_req(0, NULL); // temporary disconnect
|
|
1647 ProjNode* proj2 = proj_clone(proj, iff);
|
|
1648 register_node(proj2, loop, iff, ddepth);
|
|
1649
|
|
1650 RegionNode* reg = new (C,2)RegionNode(2);
|
|
1651 reg->set_req(1, proj2);
|
|
1652 register_node(reg, loop, iff, ddepth);
|
|
1653
|
|
1654 IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
|
|
1655 register_node(dum_if, loop, reg, ddepth);
|
|
1656
|
|
1657 proj->set_req(0, dum_if); // reattach
|
|
1658 set_idom(proj, dum_if, ddepth);
|
|
1659
|
|
1660 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
|
|
1661 register_node(dum_proj, loop, dum_if, ddepth);
|
|
1662
|
|
1663 return reg;
|
|
1664 }
|
|
1665
|
|
1666 //------------------------------ insert_cmpi_loop_exit -------------------------------------
|
|
1667 // Clone a signed compare loop exit from an unsigned compare and
|
|
1668 // insert it before the unsigned cmp on the stay-in-loop path.
|
|
1669 // All new nodes inserted in the dominator tree between the original
|
|
1670 // if and it's projections. The original if test is replaced with
|
|
1671 // a constant to force the stay-in-loop path.
|
|
1672 //
|
|
1673 // This is done to make sure that the original if and it's projections
|
|
1674 // still dominate the same set of control nodes, that the ctrl() relation
|
|
1675 // from data nodes to them is preserved, and that their loop nesting is
|
|
1676 // preserved.
|
|
1677 //
|
|
1678 // before
|
|
1679 // if(i <u limit) unsigned compare loop exit
|
|
1680 // / |
|
|
1681 // v v
|
|
1682 // exit-proj stay-in-loop-proj
|
|
1683 //
|
|
1684 // after
|
|
1685 // if(stay-in-loop-const) original if
|
|
1686 // / |
|
|
1687 // / v
|
|
1688 // / if(i < limit) new signed test
|
|
1689 // / / |
|
|
1690 // / / v
|
|
1691 // / / if(i <u limit) new cloned unsigned test
|
|
1692 // / / / |
|
|
1693 // v v v |
|
|
1694 // region |
|
|
1695 // | |
|
|
1696 // dum-if |
|
|
1697 // / | |
|
|
1698 // ether | |
|
|
1699 // v v
|
|
1700 // exit-proj stay-in-loop-proj
|
|
1701 //
|
|
1702 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
|
|
1703 const bool Signed = true;
|
|
1704 const bool Unsigned = false;
|
|
1705
|
|
1706 BoolNode* bol = if_cmpu->in(1)->as_Bool();
|
|
1707 if (bol->_test._test != BoolTest::lt) return NULL;
|
|
1708 CmpNode* cmpu = bol->in(1)->as_Cmp();
|
|
1709 if (cmpu->Opcode() != Op_CmpU) return NULL;
|
|
1710 int stride = stride_of_possible_iv(if_cmpu);
|
|
1711 if (stride == 0) return NULL;
|
|
1712
|
|
1713 ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj();
|
|
1714 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
|
|
1715
|
|
1716 Node* limit = NULL;
|
|
1717 if (stride > 0) {
|
|
1718 limit = cmpu->in(2);
|
|
1719 } else {
|
|
1720 limit = _igvn.makecon(TypeInt::ZERO);
|
|
1721 set_ctrl(limit, C->root());
|
|
1722 }
|
|
1723 // Create a new region on the exit path
|
|
1724 RegionNode* reg = insert_region_before_proj(lp_exit);
|
|
1725
|
|
1726 // Clone the if-cmpu-true-false using a signed compare
|
|
1727 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
|
|
1728 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
|
|
1729 reg->add_req(cmpi_exit);
|
|
1730
|
|
1731 // Clone the if-cmpu-true-false
|
|
1732 BoolTest::mask rel_u = bol->_test._test;
|
|
1733 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
|
|
1734 reg->add_req(cmpu_exit);
|
|
1735
|
|
1736 // Force original if to stay in loop.
|
|
1737 short_circuit_if(if_cmpu, lp_continue);
|
|
1738
|
|
1739 return cmpi_exit->in(0)->as_If();
|
|
1740 }
|
|
1741
|
|
1742 //------------------------------ remove_cmpi_loop_exit -------------------------------------
|
|
1743 // Remove a previously inserted signed compare loop exit.
|
|
1744 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
|
|
1745 Node* lp_proj = stay_in_loop(if_cmp, loop);
|
|
1746 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
|
|
1747 stay_in_loop(lp_proj, loop)->is_If() &&
|
|
1748 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
|
|
1749 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
|
|
1750 set_ctrl(con, C->root());
|
|
1751 if_cmp->set_req(1, con);
|
|
1752 }
|
|
1753
|
|
1754 //------------------------------ scheduled_nodelist -------------------------------------
|
|
1755 // Create a post order schedule of nodes that are in the
|
|
1756 // "member" set. The list is returned in "sched".
|
|
1757 // The first node in "sched" is the loop head, followed by
|
|
1758 // nodes which have no inputs in the "member" set, and then
|
|
1759 // followed by the nodes that have an immediate input dependence
|
|
1760 // on a node in "sched".
|
|
1761 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
|
|
1762
|
|
1763 assert(member.test(loop->_head->_idx), "loop head must be in member set");
|
|
1764 Arena *a = Thread::current()->resource_area();
|
|
1765 VectorSet visited(a);
|
|
1766 Node_Stack nstack(a, loop->_body.size());
|
|
1767
|
|
1768 Node* n = loop->_head; // top of stack is cached in "n"
|
|
1769 uint idx = 0;
|
|
1770 visited.set(n->_idx);
|
|
1771
|
|
1772 // Initially push all with no inputs from within member set
|
|
1773 for(uint i = 0; i < loop->_body.size(); i++ ) {
|
|
1774 Node *elt = loop->_body.at(i);
|
|
1775 if (member.test(elt->_idx)) {
|
|
1776 bool found = false;
|
|
1777 for (uint j = 0; j < elt->req(); j++) {
|
|
1778 Node* def = elt->in(j);
|
|
1779 if (def && member.test(def->_idx) && def != elt) {
|
|
1780 found = true;
|
|
1781 break;
|
|
1782 }
|
|
1783 }
|
|
1784 if (!found && elt != loop->_head) {
|
|
1785 nstack.push(n, idx);
|
|
1786 n = elt;
|
|
1787 assert(!visited.test(n->_idx), "not seen yet");
|
|
1788 visited.set(n->_idx);
|
|
1789 }
|
|
1790 }
|
|
1791 }
|
|
1792
|
|
1793 // traverse out's that are in the member set
|
|
1794 while (true) {
|
|
1795 if (idx < n->outcnt()) {
|
|
1796 Node* use = n->raw_out(idx);
|
|
1797 idx++;
|
|
1798 if (!visited.test_set(use->_idx)) {
|
|
1799 if (member.test(use->_idx)) {
|
|
1800 nstack.push(n, idx);
|
|
1801 n = use;
|
|
1802 idx = 0;
|
|
1803 }
|
|
1804 }
|
|
1805 } else {
|
|
1806 // All outputs processed
|
|
1807 sched.push(n);
|
|
1808 if (nstack.is_empty()) break;
|
|
1809 n = nstack.node();
|
|
1810 idx = nstack.index();
|
|
1811 nstack.pop();
|
|
1812 }
|
|
1813 }
|
|
1814 }
|
|
1815
|
|
1816
|
|
1817 //------------------------------ has_use_in_set -------------------------------------
|
|
1818 // Has a use in the vector set
|
|
1819 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
|
|
1820 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
|
|
1821 Node* use = n->fast_out(j);
|
|
1822 if (vset.test(use->_idx)) {
|
|
1823 return true;
|
|
1824 }
|
|
1825 }
|
|
1826 return false;
|
|
1827 }
|
|
1828
|
|
1829
|
|
1830 //------------------------------ has_use_internal_to_set -------------------------------------
|
|
1831 // Has use internal to the vector set (ie. not in a phi at the loop head)
|
|
1832 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
|
|
1833 Node* head = loop->_head;
|
|
1834 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
|
|
1835 Node* use = n->fast_out(j);
|
|
1836 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
|
|
1837 return true;
|
|
1838 }
|
|
1839 }
|
|
1840 return false;
|
|
1841 }
|
|
1842
|
|
1843
|
|
1844 //------------------------------ clone_for_use_outside_loop -------------------------------------
|
|
1845 // clone "n" for uses that are outside of loop
|
|
1846 void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
|
|
1847
|
|
1848 assert(worklist.size() == 0, "should be empty");
|
|
1849 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
|
|
1850 Node* use = n->fast_out(j);
|
|
1851 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
|
|
1852 worklist.push(use);
|
|
1853 }
|
|
1854 }
|
|
1855 while( worklist.size() ) {
|
|
1856 Node *use = worklist.pop();
|
|
1857 if (!has_node(use) || use->in(0) == C->top()) continue;
|
|
1858 uint j;
|
|
1859 for (j = 0; j < use->req(); j++) {
|
|
1860 if (use->in(j) == n) break;
|
|
1861 }
|
|
1862 assert(j < use->req(), "must be there");
|
|
1863
|
|
1864 // clone "n" and insert it between the inputs of "n" and the use outside the loop
|
|
1865 Node* n_clone = n->clone();
|
|
1866 _igvn.hash_delete(use);
|
|
1867 use->set_req(j, n_clone);
|
|
1868 _igvn._worklist.push(use);
|
|
1869 if (!use->is_Phi()) {
|
|
1870 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
|
|
1871 set_ctrl(n_clone, use_c);
|
|
1872 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
|
|
1873 get_loop(use_c)->_body.push(n_clone);
|
|
1874 } else {
|
|
1875 // Use in a phi is considered a use in the associated predecessor block
|
|
1876 Node *prevbb = use->in(0)->in(j);
|
|
1877 set_ctrl(n_clone, prevbb);
|
|
1878 assert(!loop->is_member(get_loop(prevbb)), "should be outside loop");
|
|
1879 get_loop(prevbb)->_body.push(n_clone);
|
|
1880 }
|
|
1881 _igvn.register_new_node_with_optimizer(n_clone);
|
|
1882 #if !defined(PRODUCT)
|
|
1883 if (TracePartialPeeling) {
|
|
1884 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
|
|
1885 }
|
|
1886 #endif
|
|
1887 }
|
|
1888 }
|
|
1889
|
|
1890
|
|
1891 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
|
|
1892 // clone "n" for special uses that are in the not_peeled region.
|
|
1893 // If these def-uses occur in separate blocks, the code generator
|
|
1894 // marks the method as not compilable. For example, if a "BoolNode"
|
|
1895 // is in a different basic block than the "IfNode" that uses it, then
|
|
1896 // the compilation is aborted in the code generator.
|
|
1897 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
|
|
1898 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
|
|
1899 if (n->is_Phi() || n->is_Load()) {
|
|
1900 return;
|
|
1901 }
|
|
1902 assert(worklist.size() == 0, "should be empty");
|
|
1903 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
|
|
1904 Node* use = n->fast_out(j);
|
|
1905 if ( not_peel.test(use->_idx) &&
|
|
1906 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
|
|
1907 use->in(1) == n) {
|
|
1908 worklist.push(use);
|
|
1909 }
|
|
1910 }
|
|
1911 if (worklist.size() > 0) {
|
|
1912 // clone "n" and insert it between inputs of "n" and the use
|
|
1913 Node* n_clone = n->clone();
|
|
1914 loop->_body.push(n_clone);
|
|
1915 _igvn.register_new_node_with_optimizer(n_clone);
|
|
1916 set_ctrl(n_clone, get_ctrl(n));
|
|
1917 sink_list.push(n_clone);
|
|
1918 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
|
|
1919 #if !defined(PRODUCT)
|
|
1920 if (TracePartialPeeling) {
|
|
1921 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
|
|
1922 }
|
|
1923 #endif
|
|
1924 while( worklist.size() ) {
|
|
1925 Node *use = worklist.pop();
|
|
1926 _igvn.hash_delete(use);
|
|
1927 _igvn._worklist.push(use);
|
|
1928 for (uint j = 1; j < use->req(); j++) {
|
|
1929 if (use->in(j) == n) {
|
|
1930 use->set_req(j, n_clone);
|
|
1931 }
|
|
1932 }
|
|
1933 }
|
|
1934 }
|
|
1935 }
|
|
1936
|
|
1937
|
|
1938 //------------------------------ insert_phi_for_loop -------------------------------------
|
|
1939 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
|
|
1940 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
|
|
1941 Node *phi = PhiNode::make(lp, back_edge_val);
|
|
1942 phi->set_req(LoopNode::EntryControl, lp_entry_val);
|
|
1943 // Use existing phi if it already exists
|
|
1944 Node *hit = _igvn.hash_find_insert(phi);
|
|
1945 if( hit == NULL ) {
|
|
1946 _igvn.register_new_node_with_optimizer(phi);
|
|
1947 set_ctrl(phi, lp);
|
|
1948 } else {
|
|
1949 // Remove the new phi from the graph and use the hit
|
|
1950 _igvn.remove_dead_node(phi);
|
|
1951 phi = hit;
|
|
1952 }
|
|
1953 _igvn.hash_delete(use);
|
|
1954 _igvn._worklist.push(use);
|
|
1955 use->set_req(idx, phi);
|
|
1956 }
|
|
1957
|
|
1958 #ifdef ASSERT
|
|
1959 //------------------------------ is_valid_loop_partition -------------------------------------
|
|
1960 // Validate the loop partition sets: peel and not_peel
|
|
1961 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
|
|
1962 VectorSet& not_peel ) {
|
|
1963 uint i;
|
|
1964 // Check that peel_list entries are in the peel set
|
|
1965 for (i = 0; i < peel_list.size(); i++) {
|
|
1966 if (!peel.test(peel_list.at(i)->_idx)) {
|
|
1967 return false;
|
|
1968 }
|
|
1969 }
|
|
1970 // Check at loop members are in one of peel set or not_peel set
|
|
1971 for (i = 0; i < loop->_body.size(); i++ ) {
|
|
1972 Node *def = loop->_body.at(i);
|
|
1973 uint di = def->_idx;
|
|
1974 // Check that peel set elements are in peel_list
|
|
1975 if (peel.test(di)) {
|
|
1976 if (not_peel.test(di)) {
|
|
1977 return false;
|
|
1978 }
|
|
1979 // Must be in peel_list also
|
|
1980 bool found = false;
|
|
1981 for (uint j = 0; j < peel_list.size(); j++) {
|
|
1982 if (peel_list.at(j)->_idx == di) {
|
|
1983 found = true;
|
|
1984 break;
|
|
1985 }
|
|
1986 }
|
|
1987 if (!found) {
|
|
1988 return false;
|
|
1989 }
|
|
1990 } else if (not_peel.test(di)) {
|
|
1991 if (peel.test(di)) {
|
|
1992 return false;
|
|
1993 }
|
|
1994 } else {
|
|
1995 return false;
|
|
1996 }
|
|
1997 }
|
|
1998 return true;
|
|
1999 }
|
|
2000
|
|
2001 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
|
|
2002 // Ensure a use outside of loop is of the right form
|
|
2003 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
|
|
2004 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
|
|
2005 return (use->is_Phi() &&
|
|
2006 use_c->is_Region() && use_c->req() == 3 &&
|
|
2007 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
|
|
2008 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
|
|
2009 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
|
|
2010 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
|
|
2011 }
|
|
2012
|
|
2013 //------------------------------ is_valid_clone_loop_form -------------------------------------
|
|
2014 // Ensure that all uses outside of loop are of the right form
|
|
2015 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
|
|
2016 uint orig_exit_idx, uint clone_exit_idx) {
|
|
2017 uint len = peel_list.size();
|
|
2018 for (uint i = 0; i < len; i++) {
|
|
2019 Node *def = peel_list.at(i);
|
|
2020
|
|
2021 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
|
|
2022 Node *use = def->fast_out(j);
|
|
2023 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
|
|
2024 if (!loop->is_member(get_loop(use_c))) {
|
|
2025 // use is not in the loop, check for correct structure
|
|
2026 if (use->in(0) == def) {
|
|
2027 // Okay
|
|
2028 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
|
|
2029 return false;
|
|
2030 }
|
|
2031 }
|
|
2032 }
|
|
2033 }
|
|
2034 return true;
|
|
2035 }
|
|
2036 #endif
|
|
2037
|
|
2038 //------------------------------ partial_peel -------------------------------------
|
|
2039 // Partially peel (aka loop rotation) the top portion of a loop (called
|
|
2040 // the peel section below) by cloning it and placing one copy just before
|
|
2041 // the new loop head and the other copy at the bottom of the new loop.
|
|
2042 //
|
|
2043 // before after where it came from
|
|
2044 //
|
|
2045 // stmt1 stmt1
|
|
2046 // loop: stmt2 clone
|
|
2047 // stmt2 if condA goto exitA clone
|
|
2048 // if condA goto exitA new_loop: new
|
|
2049 // stmt3 stmt3 clone
|
|
2050 // if !condB goto loop if condB goto exitB clone
|
|
2051 // exitB: stmt2 orig
|
|
2052 // stmt4 if !condA goto new_loop orig
|
|
2053 // exitA: goto exitA
|
|
2054 // exitB:
|
|
2055 // stmt4
|
|
2056 // exitA:
|
|
2057 //
|
|
2058 // Step 1: find the cut point: an exit test on probable
|
|
2059 // induction variable.
|
|
2060 // Step 2: schedule (with cloning) operations in the peel
|
|
2061 // section that can be executed after the cut into
|
|
2062 // the section that is not peeled. This may need
|
|
2063 // to clone operations into exit blocks. For
|
|
2064 // instance, a reference to A[i] in the not-peel
|
|
2065 // section and a reference to B[i] in an exit block
|
|
2066 // may cause a left-shift of i by 2 to be placed
|
|
2067 // in the peel block. This step will clone the left
|
|
2068 // shift into the exit block and sink the left shift
|
|
2069 // from the peel to the not-peel section.
|
|
2070 // Step 3: clone the loop, retarget the control, and insert
|
|
2071 // phis for values that are live across the new loop
|
|
2072 // head. This is very dependent on the graph structure
|
|
2073 // from clone_loop. It creates region nodes for
|
|
2074 // exit control and associated phi nodes for values
|
|
2075 // flow out of the loop through that exit. The region
|
|
2076 // node is dominated by the clone's control projection.
|
|
2077 // So the clone's peel section is placed before the
|
|
2078 // new loop head, and the clone's not-peel section is
|
|
2079 // forms the top part of the new loop. The original
|
|
2080 // peel section forms the tail of the new loop.
|
|
2081 // Step 4: update the dominator tree and recompute the
|
|
2082 // dominator depth.
|
|
2083 //
|
|
2084 // orig
|
|
2085 //
|
|
2086 // stmt1
|
|
2087 // |
|
|
2088 // v
|
|
2089 // loop<----+
|
|
2090 // | |
|
|
2091 // stmt2 |
|
|
2092 // | |
|
|
2093 // v |
|
|
2094 // ifA |
|
|
2095 // / | |
|
|
2096 // v v |
|
|
2097 // false true ^ <-- last_peel
|
|
2098 // / | |
|
|
2099 // / ===|==cut |
|
|
2100 // / stmt3 | <-- first_not_peel
|
|
2101 // / | |
|
|
2102 // | v |
|
|
2103 // v ifB |
|
|
2104 // exitA: / \ |
|
|
2105 // / \ |
|
|
2106 // v v |
|
|
2107 // false true |
|
|
2108 // / \ |
|
|
2109 // / ----+
|
|
2110 // |
|
|
2111 // v
|
|
2112 // exitB:
|
|
2113 // stmt4
|
|
2114 //
|
|
2115 //
|
|
2116 // after clone loop
|
|
2117 //
|
|
2118 // stmt1
|
|
2119 // / \
|
|
2120 // clone / \ orig
|
|
2121 // / \
|
|
2122 // / \
|
|
2123 // v v
|
|
2124 // +---->loop loop<----+
|
|
2125 // | | | |
|
|
2126 // | stmt2 stmt2 |
|
|
2127 // | | | |
|
|
2128 // | v v |
|
|
2129 // | ifA ifA |
|
|
2130 // | | \ / | |
|
|
2131 // | v v v v |
|
|
2132 // ^ true false false true ^ <-- last_peel
|
|
2133 // | | ^ \ / | |
|
|
2134 // | cut==|== \ \ / ===|==cut |
|
|
2135 // | stmt3 \ \ / stmt3 | <-- first_not_peel
|
|
2136 // | | dom | | | |
|
|
2137 // | v \ 1v v2 v |
|
|
2138 // | ifB regionA ifB |
|
|
2139 // | / \ | / \ |
|
|
2140 // | / \ v / \ |
|
|
2141 // | v v exitA: v v |
|
|
2142 // | true false false true |
|
|
2143 // | / ^ \ / \ |
|
|
2144 // +---- \ \ / ----+
|
|
2145 // dom \ /
|
|
2146 // \ 1v v2
|
|
2147 // regionB
|
|
2148 // |
|
|
2149 // v
|
|
2150 // exitB:
|
|
2151 // stmt4
|
|
2152 //
|
|
2153 //
|
|
2154 // after partial peel
|
|
2155 //
|
|
2156 // stmt1
|
|
2157 // /
|
|
2158 // clone / orig
|
|
2159 // / TOP
|
|
2160 // / \
|
|
2161 // v v
|
|
2162 // TOP->region region----+
|
|
2163 // | | |
|
|
2164 // stmt2 stmt2 |
|
|
2165 // | | |
|
|
2166 // v v |
|
|
2167 // ifA ifA |
|
|
2168 // | \ / | |
|
|
2169 // v v v v |
|
|
2170 // true false false true | <-- last_peel
|
|
2171 // | ^ \ / +------|---+
|
|
2172 // +->newloop \ \ / === ==cut | |
|
|
2173 // | stmt3 \ \ / TOP | |
|
|
2174 // | | dom | | stmt3 | | <-- first_not_peel
|
|
2175 // | v \ 1v v2 v | |
|
|
2176 // | ifB regionA ifB ^ v
|
|
2177 // | / \ | / \ | |
|
|
2178 // | / \ v / \ | |
|
|
2179 // | v v exitA: v v | |
|
|
2180 // | true false false true | |
|
|
2181 // | / ^ \ / \ | |
|
|
2182 // | | \ \ / v | |
|
|
2183 // | | dom \ / TOP | |
|
|
2184 // | | \ 1v v2 | |
|
|
2185 // ^ v regionB | |
|
|
2186 // | | | | |
|
|
2187 // | | v ^ v
|
|
2188 // | | exitB: | |
|
|
2189 // | | stmt4 | |
|
|
2190 // | +------------>-----------------+ |
|
|
2191 // | |
|
|
2192 // +-----------------<---------------------+
|
|
2193 //
|
|
2194 //
|
|
2195 // final graph
|
|
2196 //
|
|
2197 // stmt1
|
|
2198 // |
|
|
2199 // v
|
|
2200 // ........> ifA clone
|
|
2201 // : / |
|
|
2202 // dom / |
|
|
2203 // : v v
|
|
2204 // : false true
|
|
2205 // : | |
|
|
2206 // : | stmt2 clone
|
|
2207 // : | |
|
|
2208 // : | v
|
|
2209 // : | newloop<-----+
|
|
2210 // : | | |
|
|
2211 // : | stmt3 clone |
|
|
2212 // : | | |
|
|
2213 // : | v |
|
|
2214 // : | ifB |
|
|
2215 // : | / \ |
|
|
2216 // : | v v |
|
|
2217 // : | false true |
|
|
2218 // : | | | |
|
|
2219 // : | v stmt2 |
|
|
2220 // : | exitB: | |
|
|
2221 // : | stmt4 v |
|
|
2222 // : | ifA orig |
|
|
2223 // : | / \ |
|
|
2224 // : | / \ |
|
|
2225 // : | v v |
|
|
2226 // : | false true |
|
|
2227 // : | / \ |
|
|
2228 // : v v -----+
|
|
2229 // RegionA
|
|
2230 // |
|
|
2231 // v
|
|
2232 // exitA
|
|
2233 //
|
|
2234 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
|
|
2235
|
|
2236 LoopNode *head = loop->_head->as_Loop();
|
|
2237
|
|
2238 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
|
|
2239 return false;
|
|
2240 }
|
|
2241
|
|
2242 // Check for complex exit control
|
|
2243 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
|
|
2244 Node *n = loop->_body.at(ii);
|
|
2245 int opc = n->Opcode();
|
|
2246 if (n->is_Call() ||
|
|
2247 opc == Op_Catch ||
|
|
2248 opc == Op_CatchProj ||
|
|
2249 opc == Op_Jump ||
|
|
2250 opc == Op_JumpProj) {
|
|
2251 #if !defined(PRODUCT)
|
|
2252 if (TracePartialPeeling) {
|
|
2253 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
|
|
2254 }
|
|
2255 #endif
|
|
2256 return false;
|
|
2257 }
|
|
2258 }
|
|
2259
|
|
2260 int dd = dom_depth(head);
|
|
2261
|
|
2262 // Step 1: find cut point
|
|
2263
|
|
2264 // Walk up dominators to loop head looking for first loop exit
|
|
2265 // which is executed on every path thru loop.
|
|
2266 IfNode *peel_if = NULL;
|
|
2267 IfNode *peel_if_cmpu = NULL;
|
|
2268
|
|
2269 Node *iff = loop->tail();
|
|
2270 while( iff != head ) {
|
|
2271 if( iff->is_If() ) {
|
|
2272 Node *ctrl = get_ctrl(iff->in(1));
|
|
2273 if (ctrl->is_top()) return false; // Dead test on live IF.
|
|
2274 // If loop-varying exit-test, check for induction variable
|
|
2275 if( loop->is_member(get_loop(ctrl)) &&
|
|
2276 loop->is_loop_exit(iff) &&
|
|
2277 is_possible_iv_test(iff)) {
|
|
2278 Node* cmp = iff->in(1)->in(1);
|
|
2279 if (cmp->Opcode() == Op_CmpI) {
|
|
2280 peel_if = iff->as_If();
|
|
2281 } else {
|
|
2282 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
|
|
2283 peel_if_cmpu = iff->as_If();
|
|
2284 }
|
|
2285 }
|
|
2286 }
|
|
2287 iff = idom(iff);
|
|
2288 }
|
|
2289 // Prefer signed compare over unsigned compare.
|
|
2290 IfNode* new_peel_if = NULL;
|
|
2291 if (peel_if == NULL) {
|
|
2292 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
|
|
2293 return false; // No peel point found
|
|
2294 }
|
|
2295 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
|
|
2296 if (new_peel_if == NULL) {
|
|
2297 return false; // No peel point found
|
|
2298 }
|
|
2299 peel_if = new_peel_if;
|
|
2300 }
|
|
2301 Node* last_peel = stay_in_loop(peel_if, loop);
|
|
2302 Node* first_not_peeled = stay_in_loop(last_peel, loop);
|
|
2303 if (first_not_peeled == NULL || first_not_peeled == head) {
|
|
2304 return false;
|
|
2305 }
|
|
2306
|
|
2307 #if !defined(PRODUCT)
|
|
2308 if (TracePartialPeeling) {
|
|
2309 tty->print_cr("before partial peel one iteration");
|
|
2310 Node_List wl;
|
|
2311 Node* t = head->in(2);
|
|
2312 while (true) {
|
|
2313 wl.push(t);
|
|
2314 if (t == head) break;
|
|
2315 t = idom(t);
|
|
2316 }
|
|
2317 while (wl.size() > 0) {
|
|
2318 Node* tt = wl.pop();
|
|
2319 tt->dump();
|
|
2320 if (tt == last_peel) tty->print_cr("-- cut --");
|
|
2321 }
|
|
2322 }
|
|
2323 #endif
|
|
2324 ResourceArea *area = Thread::current()->resource_area();
|
|
2325 VectorSet peel(area);
|
|
2326 VectorSet not_peel(area);
|
|
2327 Node_List peel_list(area);
|
|
2328 Node_List worklist(area);
|
|
2329 Node_List sink_list(area);
|
|
2330
|
|
2331 // Set of cfg nodes to peel are those that are executable from
|
|
2332 // the head through last_peel.
|
|
2333 assert(worklist.size() == 0, "should be empty");
|
|
2334 worklist.push(head);
|
|
2335 peel.set(head->_idx);
|
|
2336 while (worklist.size() > 0) {
|
|
2337 Node *n = worklist.pop();
|
|
2338 if (n != last_peel) {
|
|
2339 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
|
|
2340 Node* use = n->fast_out(j);
|
|
2341 if (use->is_CFG() &&
|
|
2342 loop->is_member(get_loop(use)) &&
|
|
2343 !peel.test_set(use->_idx)) {
|
|
2344 worklist.push(use);
|
|
2345 }
|
|
2346 }
|
|
2347 }
|
|
2348 }
|
|
2349
|
|
2350 // Set of non-cfg nodes to peel are those that are control
|
|
2351 // dependent on the cfg nodes.
|
|
2352 uint i;
|
|
2353 for(i = 0; i < loop->_body.size(); i++ ) {
|
|
2354 Node *n = loop->_body.at(i);
|
|
2355 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
|
|
2356 if (peel.test(n_c->_idx)) {
|
|
2357 peel.set(n->_idx);
|
|
2358 } else {
|
|
2359 not_peel.set(n->_idx);
|
|
2360 }
|
|
2361 }
|
|
2362
|
|
2363 // Step 2: move operations from the peeled section down into the
|
|
2364 // not-peeled section
|
|
2365
|
|
2366 // Get a post order schedule of nodes in the peel region
|
|
2367 // Result in right-most operand.
|
|
2368 scheduled_nodelist(loop, peel, peel_list );
|
|
2369
|
|
2370 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
|
|
2371
|
|
2372 // For future check for too many new phis
|
|
2373 uint old_phi_cnt = 0;
|
|
2374 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
|
|
2375 Node* use = head->fast_out(j);
|
|
2376 if (use->is_Phi()) old_phi_cnt++;
|
|
2377 }
|
|
2378
|
|
2379 #if !defined(PRODUCT)
|
|
2380 if (TracePartialPeeling) {
|
|
2381 tty->print_cr("\npeeled list");
|
|
2382 }
|
|
2383 #endif
|
|
2384
|
|
2385 // Evacuate nodes in peel region into the not_peeled region if possible
|
|
2386 uint new_phi_cnt = 0;
|
|
2387 for (i = 0; i < peel_list.size();) {
|
|
2388 Node* n = peel_list.at(i);
|
|
2389 #if !defined(PRODUCT)
|
|
2390 if (TracePartialPeeling) n->dump();
|
|
2391 #endif
|
|
2392 bool incr = true;
|
|
2393 if ( !n->is_CFG() ) {
|
|
2394
|
|
2395 if ( has_use_in_set(n, not_peel) ) {
|
|
2396
|
|
2397 // If not used internal to the peeled region,
|
|
2398 // move "n" from peeled to not_peeled region.
|
|
2399
|
|
2400 if ( !has_use_internal_to_set(n, peel, loop) ) {
|
|
2401
|
|
2402 // if not pinned and not a load (which maybe anti-dependent on a store)
|
|
2403 // and not a CMove (Matcher expects only bool->cmove).
|
|
2404 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
|
|
2405 clone_for_use_outside_loop( loop, n, worklist );
|
|
2406
|
|
2407 sink_list.push(n);
|
|
2408 peel >>= n->_idx; // delete n from peel set.
|
|
2409 not_peel <<= n->_idx; // add n to not_peel set.
|
|
2410 peel_list.remove(i);
|
|
2411 incr = false;
|
|
2412 #if !defined(PRODUCT)
|
|
2413 if (TracePartialPeeling) {
|
|
2414 tty->print_cr("sink to not_peeled region: %d newbb: %d",
|
|
2415 n->_idx, get_ctrl(n)->_idx);
|
|
2416 }
|
|
2417 #endif
|
|
2418 }
|
|
2419 } else {
|
|
2420 // Otherwise check for special def-use cases that span
|
|
2421 // the peel/not_peel boundary such as bool->if
|
|
2422 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
|
|
2423 new_phi_cnt++;
|
|
2424 }
|
|
2425 }
|
|
2426 }
|
|
2427 if (incr) i++;
|
|
2428 }
|
|
2429
|
|
2430 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
|
|
2431 #if !defined(PRODUCT)
|
|
2432 if (TracePartialPeeling) {
|
|
2433 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
|
|
2434 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
|
|
2435 }
|
|
2436 #endif
|
|
2437 if (new_peel_if != NULL) {
|
|
2438 remove_cmpi_loop_exit(new_peel_if, loop);
|
|
2439 }
|
|
2440 // Inhibit more partial peeling on this loop
|
|
2441 assert(!head->is_partial_peel_loop(), "not partial peeled");
|
|
2442 head->mark_partial_peel_failed();
|
|
2443 return false;
|
|
2444 }
|
|
2445
|
|
2446 // Step 3: clone loop, retarget control, and insert new phis
|
|
2447
|
|
2448 // Create new loop head for new phis and to hang
|
|
2449 // the nodes being moved (sinked) from the peel region.
|
|
2450 LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel);
|
|
2451 _igvn.register_new_node_with_optimizer(new_head);
|
|
2452 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
|
|
2453 first_not_peeled->set_req(0, new_head);
|
|
2454 set_loop(new_head, loop);
|
|
2455 loop->_body.push(new_head);
|
|
2456 not_peel.set(new_head->_idx);
|
|
2457 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
|
|
2458 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
|
|
2459
|
|
2460 while (sink_list.size() > 0) {
|
|
2461 Node* n = sink_list.pop();
|
|
2462 set_ctrl(n, new_head);
|
|
2463 }
|
|
2464
|
|
2465 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
|
|
2466
|
|
2467 clone_loop( loop, old_new, dd );
|
|
2468
|
|
2469 const uint clone_exit_idx = 1;
|
|
2470 const uint orig_exit_idx = 2;
|
|
2471 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
|
|
2472
|
|
2473 Node* head_clone = old_new[head->_idx];
|
|
2474 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
|
|
2475 Node* orig_tail_clone = head_clone->in(2);
|
|
2476
|
|
2477 // Add phi if "def" node is in peel set and "use" is not
|
|
2478
|
|
2479 for(i = 0; i < peel_list.size(); i++ ) {
|
|
2480 Node *def = peel_list.at(i);
|
|
2481 if (!def->is_CFG()) {
|
|
2482 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
|
|
2483 Node *use = def->fast_out(j);
|
|
2484 if (has_node(use) && use->in(0) != C->top() &&
|
|
2485 (!peel.test(use->_idx) ||
|
|
2486 (use->is_Phi() && use->in(0) == head)) ) {
|
|
2487 worklist.push(use);
|
|
2488 }
|
|
2489 }
|
|
2490 while( worklist.size() ) {
|
|
2491 Node *use = worklist.pop();
|
|
2492 for (uint j = 1; j < use->req(); j++) {
|
|
2493 Node* n = use->in(j);
|
|
2494 if (n == def) {
|
|
2495
|
|
2496 // "def" is in peel set, "use" is not in peel set
|
|
2497 // or "use" is in the entry boundary (a phi) of the peel set
|
|
2498
|
|
2499 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
|
|
2500
|
|
2501 if ( loop->is_member(get_loop( use_c )) ) {
|
|
2502 // use is in loop
|
|
2503 if (old_new[use->_idx] != NULL) { // null for dead code
|
|
2504 Node* use_clone = old_new[use->_idx];
|
|
2505 _igvn.hash_delete(use);
|
|
2506 use->set_req(j, C->top());
|
|
2507 _igvn._worklist.push(use);
|
|
2508 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
|
|
2509 }
|
|
2510 } else {
|
|
2511 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
|
|
2512 // use is not in the loop, check if the live range includes the cut
|
|
2513 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
|
|
2514 if (not_peel.test(lp_if->_idx)) {
|
|
2515 assert(j == orig_exit_idx, "use from original loop");
|
|
2516 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
|
|
2517 }
|
|
2518 }
|
|
2519 }
|
|
2520 }
|
|
2521 }
|
|
2522 }
|
|
2523 }
|
|
2524
|
|
2525 // Step 3b: retarget control
|
|
2526
|
|
2527 // Redirect control to the new loop head if a cloned node in
|
|
2528 // the not_peeled region has control that points into the peeled region.
|
|
2529 // This necessary because the cloned peeled region will be outside
|
|
2530 // the loop.
|
|
2531 // from to
|
|
2532 // cloned-peeled <---+
|
|
2533 // new_head_clone: | <--+
|
|
2534 // cloned-not_peeled in(0) in(0)
|
|
2535 // orig-peeled
|
|
2536
|
|
2537 for(i = 0; i < loop->_body.size(); i++ ) {
|
|
2538 Node *n = loop->_body.at(i);
|
|
2539 if (!n->is_CFG() && n->in(0) != NULL &&
|
|
2540 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
|
|
2541 Node* n_clone = old_new[n->_idx];
|
|
2542 _igvn.hash_delete(n_clone);
|
|
2543 n_clone->set_req(0, new_head_clone);
|
|
2544 _igvn._worklist.push(n_clone);
|
|
2545 }
|
|
2546 }
|
|
2547
|
|
2548 // Backedge of the surviving new_head (the clone) is original last_peel
|
|
2549 _igvn.hash_delete(new_head_clone);
|
|
2550 new_head_clone->set_req(LoopNode::LoopBackControl, last_peel);
|
|
2551 _igvn._worklist.push(new_head_clone);
|
|
2552
|
|
2553 // Cut first node in original not_peel set
|
|
2554 _igvn.hash_delete(new_head);
|
|
2555 new_head->set_req(LoopNode::EntryControl, C->top());
|
|
2556 new_head->set_req(LoopNode::LoopBackControl, C->top());
|
|
2557 _igvn._worklist.push(new_head);
|
|
2558
|
|
2559 // Copy head_clone back-branch info to original head
|
|
2560 // and remove original head's loop entry and
|
|
2561 // clone head's back-branch
|
|
2562 _igvn.hash_delete(head);
|
|
2563 _igvn.hash_delete(head_clone);
|
|
2564 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
|
|
2565 head->set_req(LoopNode::LoopBackControl, C->top());
|
|
2566 head_clone->set_req(LoopNode::LoopBackControl, C->top());
|
|
2567 _igvn._worklist.push(head);
|
|
2568 _igvn._worklist.push(head_clone);
|
|
2569
|
|
2570 // Similarly modify the phis
|
|
2571 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
|
|
2572 Node* use = head->fast_out(k);
|
|
2573 if (use->is_Phi() && use->outcnt() > 0) {
|
|
2574 Node* use_clone = old_new[use->_idx];
|
|
2575 _igvn.hash_delete(use);
|
|
2576 _igvn.hash_delete(use_clone);
|
|
2577 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
|
|
2578 use->set_req(LoopNode::LoopBackControl, C->top());
|
|
2579 use_clone->set_req(LoopNode::LoopBackControl, C->top());
|
|
2580 _igvn._worklist.push(use);
|
|
2581 _igvn._worklist.push(use_clone);
|
|
2582 }
|
|
2583 }
|
|
2584
|
|
2585 // Step 4: update dominator tree and dominator depth
|
|
2586
|
|
2587 set_idom(head, orig_tail_clone, dd);
|
|
2588 recompute_dom_depth();
|
|
2589
|
|
2590 // Inhibit more partial peeling on this loop
|
|
2591 new_head_clone->set_partial_peel_loop();
|
|
2592 C->set_major_progress();
|
|
2593
|
|
2594 #if !defined(PRODUCT)
|
|
2595 if (TracePartialPeeling) {
|
|
2596 tty->print_cr("\nafter partial peel one iteration");
|
|
2597 Node_List wl(area);
|
|
2598 Node* t = last_peel;
|
|
2599 while (true) {
|
|
2600 wl.push(t);
|
|
2601 if (t == head_clone) break;
|
|
2602 t = idom(t);
|
|
2603 }
|
|
2604 while (wl.size() > 0) {
|
|
2605 Node* tt = wl.pop();
|
|
2606 if (tt == head) tty->print_cr("orig head");
|
|
2607 else if (tt == new_head_clone) tty->print_cr("new head");
|
|
2608 else if (tt == head_clone) tty->print_cr("clone head");
|
|
2609 tt->dump();
|
|
2610 }
|
|
2611 }
|
|
2612 #endif
|
|
2613 return true;
|
|
2614 }
|
|
2615
|
|
2616 //------------------------------reorg_offsets----------------------------------
|
|
2617 // Reorganize offset computations to lower register pressure. Mostly
|
|
2618 // prevent loop-fallout uses of the pre-incremented trip counter (which are
|
|
2619 // then alive with the post-incremented trip counter forcing an extra
|
|
2620 // register move)
|
|
2621 void PhaseIdealLoop::reorg_offsets( IdealLoopTree *loop ) {
|
|
2622
|
|
2623 CountedLoopNode *cl = loop->_head->as_CountedLoop();
|
|
2624 CountedLoopEndNode *cle = cl->loopexit();
|
|
2625 if( !cle ) return; // The occasional dead loop
|
|
2626 // Find loop exit control
|
|
2627 Node *exit = cle->proj_out(false);
|
|
2628 assert( exit->Opcode() == Op_IfFalse, "" );
|
|
2629
|
|
2630 // Check for the special case of folks using the pre-incremented
|
|
2631 // trip-counter on the fall-out path (forces the pre-incremented
|
|
2632 // and post-incremented trip counter to be live at the same time).
|
|
2633 // Fix this by adjusting to use the post-increment trip counter.
|
|
2634 Node *phi = cl->phi();
|
|
2635 if( !phi ) return; // Dead infinite loop
|
|
2636 bool progress = true;
|
|
2637 while (progress) {
|
|
2638 progress = false;
|
|
2639 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
|
|
2640 Node* use = phi->fast_out(i); // User of trip-counter
|
|
2641 if (!has_ctrl(use)) continue;
|
|
2642 Node *u_ctrl = get_ctrl(use);
|
|
2643 if( use->is_Phi() ) {
|
|
2644 u_ctrl = NULL;
|
|
2645 for( uint j = 1; j < use->req(); j++ )
|
|
2646 if( use->in(j) == phi )
|
|
2647 u_ctrl = dom_lca( u_ctrl, use->in(0)->in(j) );
|
|
2648 }
|
|
2649 IdealLoopTree *u_loop = get_loop(u_ctrl);
|
|
2650 // Look for loop-invariant use
|
|
2651 if( u_loop == loop ) continue;
|
|
2652 if( loop->is_member( u_loop ) ) continue;
|
|
2653 // Check that use is live out the bottom. Assuming the trip-counter
|
|
2654 // update is right at the bottom, uses of of the loop middle are ok.
|
|
2655 if( dom_lca( exit, u_ctrl ) != exit ) continue;
|
|
2656 // protect against stride not being a constant
|
|
2657 if( !cle->stride_is_con() ) continue;
|
|
2658 // Hit! Refactor use to use the post-incremented tripcounter.
|
|
2659 // Compute a post-increment tripcounter.
|
|
2660 Node *opaq = new (C, 2) Opaque2Node( cle->incr() );
|
|
2661 register_new_node( opaq, u_ctrl );
|
|
2662 Node *neg_stride = _igvn.intcon(-cle->stride_con());
|
|
2663 set_ctrl(neg_stride, C->root());
|
|
2664 Node *post = new (C, 3) AddINode( opaq, neg_stride);
|
|
2665 register_new_node( post, u_ctrl );
|
|
2666 _igvn.hash_delete(use);
|
|
2667 _igvn._worklist.push(use);
|
|
2668 for( uint j = 1; j < use->req(); j++ )
|
|
2669 if( use->in(j) == phi )
|
|
2670 use->set_req(j, post);
|
|
2671 // Since DU info changed, rerun loop
|
|
2672 progress = true;
|
|
2673 break;
|
|
2674 }
|
|
2675 }
|
|
2676
|
|
2677 }
|