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