Print this page
rev 4533 : 8014811: loopTransform.cpp assert(cmp_end->in(2) == limit) failed
Summary: Stop current iteration of loop opts if partial_peel() failed and it created node clones outside processed loop.
Reviewed-by: roland
Split |
Split |
Close |
Expand all |
Collapse all |
--- old/src/share/vm/opto/loopopts.cpp
+++ new/src/share/vm/opto/loopopts.cpp
1 1 /*
2 2 * Copyright (c) 1999, 2011, Oracle and/or its affiliates. All rights reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 20 * or visit www.oracle.com if you need additional information or have any
21 21 * questions.
22 22 *
23 23 */
24 24
25 25 #include "precompiled.hpp"
26 26 #include "memory/allocation.inline.hpp"
27 27 #include "opto/addnode.hpp"
28 28 #include "opto/connode.hpp"
29 29 #include "opto/divnode.hpp"
30 30 #include "opto/loopnode.hpp"
31 31 #include "opto/matcher.hpp"
32 32 #include "opto/mulnode.hpp"
33 33 #include "opto/rootnode.hpp"
34 34 #include "opto/subnode.hpp"
35 35
36 36 //=============================================================================
37 37 //------------------------------split_thru_phi---------------------------------
38 38 // Split Node 'n' through merge point if there is enough win.
39 39 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
40 40 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
41 41 // ConvI2L may have type information on it which is unsafe to push up
42 42 // so disable this for now
43 43 return NULL;
44 44 }
45 45 int wins = 0;
46 46 assert(!n->is_CFG(), "");
47 47 assert(region->is_Region(), "");
48 48
49 49 const Type* type = n->bottom_type();
50 50 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr();
51 51 Node *phi;
52 52 if (t_oop != NULL && t_oop->is_known_instance_field()) {
53 53 int iid = t_oop->instance_id();
54 54 int index = C->get_alias_index(t_oop);
55 55 int offset = t_oop->offset();
56 56 phi = new (C) PhiNode(region, type, NULL, iid, index, offset);
57 57 } else {
58 58 phi = PhiNode::make_blank(region, n);
59 59 }
60 60 uint old_unique = C->unique();
61 61 for (uint i = 1; i < region->req(); i++) {
62 62 Node *x;
63 63 Node* the_clone = NULL;
64 64 if (region->in(i) == C->top()) {
65 65 x = C->top(); // Dead path? Use a dead data op
66 66 } else {
67 67 x = n->clone(); // Else clone up the data op
68 68 the_clone = x; // Remember for possible deletion.
69 69 // Alter data node to use pre-phi inputs
70 70 if (n->in(0) == region)
71 71 x->set_req( 0, region->in(i) );
72 72 for (uint j = 1; j < n->req(); j++) {
73 73 Node *in = n->in(j);
74 74 if (in->is_Phi() && in->in(0) == region)
75 75 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
76 76 }
77 77 }
78 78 // Check for a 'win' on some paths
79 79 const Type *t = x->Value(&_igvn);
80 80
81 81 bool singleton = t->singleton();
82 82
83 83 // A TOP singleton indicates that there are no possible values incoming
84 84 // along a particular edge. In most cases, this is OK, and the Phi will
85 85 // be eliminated later in an Ideal call. However, we can't allow this to
86 86 // happen if the singleton occurs on loop entry, as the elimination of
87 87 // the PhiNode may cause the resulting node to migrate back to a previous
88 88 // loop iteration.
89 89 if (singleton && t == Type::TOP) {
90 90 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
91 91 // irreducible loop may not be indicated by an affirmative is_Loop());
92 92 // therefore, the only top we can split thru a phi is on a backedge of
93 93 // a loop.
94 94 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
95 95 }
96 96
97 97 if (singleton) {
98 98 wins++;
99 99 x = ((PhaseGVN&)_igvn).makecon(t);
100 100 } else {
101 101 // We now call Identity to try to simplify the cloned node.
102 102 // Note that some Identity methods call phase->type(this).
103 103 // Make sure that the type array is big enough for
104 104 // our new node, even though we may throw the node away.
105 105 // (Note: This tweaking with igvn only works because x is a new node.)
106 106 _igvn.set_type(x, t);
107 107 // If x is a TypeNode, capture any more-precise type permanently into Node
108 108 // otherwise it will be not updated during igvn->transform since
109 109 // igvn->type(x) is set to x->Value() already.
110 110 x->raise_bottom_type(t);
111 111 Node *y = x->Identity(&_igvn);
112 112 if (y != x) {
113 113 wins++;
114 114 x = y;
115 115 } else {
116 116 y = _igvn.hash_find(x);
117 117 if (y) {
118 118 wins++;
119 119 x = y;
120 120 } else {
121 121 // Else x is a new node we are keeping
122 122 // We do not need register_new_node_with_optimizer
123 123 // because set_type has already been called.
124 124 _igvn._worklist.push(x);
125 125 }
126 126 }
127 127 }
128 128 if (x != the_clone && the_clone != NULL)
129 129 _igvn.remove_dead_node(the_clone);
130 130 phi->set_req( i, x );
131 131 }
132 132 // Too few wins?
133 133 if (wins <= policy) {
134 134 _igvn.remove_dead_node(phi);
135 135 return NULL;
136 136 }
137 137
138 138 // Record Phi
139 139 register_new_node( phi, region );
140 140
141 141 for (uint i2 = 1; i2 < phi->req(); i2++) {
142 142 Node *x = phi->in(i2);
143 143 // If we commoned up the cloned 'x' with another existing Node,
144 144 // the existing Node picks up a new use. We need to make the
145 145 // existing Node occur higher up so it dominates its uses.
146 146 Node *old_ctrl;
147 147 IdealLoopTree *old_loop;
148 148
149 149 if (x->is_Con()) {
150 150 // Constant's control is always root.
151 151 set_ctrl(x, C->root());
152 152 continue;
153 153 }
154 154 // The occasional new node
155 155 if (x->_idx >= old_unique) { // Found a new, unplaced node?
156 156 old_ctrl = NULL;
157 157 old_loop = NULL; // Not in any prior loop
158 158 } else {
159 159 old_ctrl = get_ctrl(x);
160 160 old_loop = get_loop(old_ctrl); // Get prior loop
161 161 }
162 162 // New late point must dominate new use
163 163 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
164 164 if (new_ctrl == old_ctrl) // Nothing is changed
165 165 continue;
166 166
167 167 IdealLoopTree *new_loop = get_loop(new_ctrl);
168 168
169 169 // Don't move x into a loop if its uses are
170 170 // outside of loop. Otherwise x will be cloned
171 171 // for each use outside of this loop.
172 172 IdealLoopTree *use_loop = get_loop(region);
173 173 if (!new_loop->is_member(use_loop) &&
174 174 (old_loop == NULL || !new_loop->is_member(old_loop))) {
175 175 // Take early control, later control will be recalculated
176 176 // during next iteration of loop optimizations.
177 177 new_ctrl = get_early_ctrl(x);
178 178 new_loop = get_loop(new_ctrl);
179 179 }
180 180 // Set new location
181 181 set_ctrl(x, new_ctrl);
182 182 // If changing loop bodies, see if we need to collect into new body
183 183 if (old_loop != new_loop) {
184 184 if (old_loop && !old_loop->_child)
185 185 old_loop->_body.yank(x);
186 186 if (!new_loop->_child)
187 187 new_loop->_body.push(x); // Collect body info
188 188 }
189 189 }
190 190
191 191 return phi;
192 192 }
193 193
194 194 //------------------------------dominated_by------------------------------------
195 195 // Replace the dominated test with an obvious true or false. Place it on the
196 196 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
197 197 // live path up to the dominating control.
198 198 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) {
199 199 #ifndef PRODUCT
200 200 if (VerifyLoopOptimizations && PrintOpto) tty->print_cr("dominating test");
201 201 #endif
202 202
203 203
204 204 // prevdom is the dominating projection of the dominating test.
205 205 assert( iff->is_If(), "" );
206 206 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
207 207 int pop = prevdom->Opcode();
208 208 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
209 209 if (flip) {
210 210 if (pop == Op_IfTrue)
211 211 pop = Op_IfFalse;
212 212 else
213 213 pop = Op_IfTrue;
214 214 }
215 215 // 'con' is set to true or false to kill the dominated test.
216 216 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
217 217 set_ctrl(con, C->root()); // Constant gets a new use
218 218 // Hack the dominated test
219 219 _igvn.replace_input_of(iff, 1, con);
220 220
221 221 // If I dont have a reachable TRUE and FALSE path following the IfNode then
222 222 // I can assume this path reaches an infinite loop. In this case it's not
223 223 // important to optimize the data Nodes - either the whole compilation will
224 224 // be tossed or this path (and all data Nodes) will go dead.
225 225 if (iff->outcnt() != 2) return;
226 226
227 227 // Make control-dependent data Nodes on the live path (path that will remain
228 228 // once the dominated IF is removed) become control-dependent on the
229 229 // dominating projection.
230 230 Node* dp = iff->as_If()->proj_out(pop == Op_IfTrue);
231 231
232 232 // Loop predicates may have depending checks which should not
233 233 // be skipped. For example, range check predicate has two checks
234 234 // for lower and upper bounds.
235 235 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp->as_Proj()->_con)->as_Proj();
236 236 if (exclude_loop_predicate &&
237 237 is_uncommon_trap_proj(unc_proj, Deoptimization::Reason_predicate))
238 238 return; // Let IGVN transformation change control dependence.
239 239
240 240 IdealLoopTree *old_loop = get_loop(dp);
241 241
242 242 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
243 243 Node* cd = dp->fast_out(i); // Control-dependent node
244 244 if (cd->depends_only_on_test()) {
245 245 assert(cd->in(0) == dp, "");
246 246 _igvn.replace_input_of(cd, 0, prevdom);
247 247 set_early_ctrl(cd);
248 248 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
249 249 if (old_loop != new_loop) {
250 250 if (!old_loop->_child) old_loop->_body.yank(cd);
251 251 if (!new_loop->_child) new_loop->_body.push(cd);
252 252 }
253 253 --i;
254 254 --imax;
255 255 }
256 256 }
257 257 }
258 258
259 259 //------------------------------has_local_phi_input----------------------------
260 260 // Return TRUE if 'n' has Phi inputs from its local block and no other
261 261 // block-local inputs (all non-local-phi inputs come from earlier blocks)
262 262 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
263 263 Node *n_ctrl = get_ctrl(n);
264 264 // See if some inputs come from a Phi in this block, or from before
265 265 // this block.
266 266 uint i;
267 267 for( i = 1; i < n->req(); i++ ) {
268 268 Node *phi = n->in(i);
269 269 if( phi->is_Phi() && phi->in(0) == n_ctrl )
270 270 break;
271 271 }
272 272 if( i >= n->req() )
273 273 return NULL; // No Phi inputs; nowhere to clone thru
274 274
275 275 // Check for inputs created between 'n' and the Phi input. These
276 276 // must split as well; they have already been given the chance
277 277 // (courtesy of a post-order visit) and since they did not we must
278 278 // recover the 'cost' of splitting them by being very profitable
279 279 // when splitting 'n'. Since this is unlikely we simply give up.
280 280 for( i = 1; i < n->req(); i++ ) {
281 281 Node *m = n->in(i);
282 282 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
283 283 // We allow the special case of AddP's with no local inputs.
284 284 // This allows us to split-up address expressions.
285 285 if (m->is_AddP() &&
286 286 get_ctrl(m->in(2)) != n_ctrl &&
287 287 get_ctrl(m->in(3)) != n_ctrl) {
288 288 // Move the AddP up to dominating point
289 289 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl)));
290 290 continue;
291 291 }
292 292 return NULL;
293 293 }
294 294 }
295 295
296 296 return n_ctrl;
297 297 }
298 298
299 299 //------------------------------remix_address_expressions----------------------
300 300 // Rework addressing expressions to get the most loop-invariant stuff
301 301 // moved out. We'd like to do all associative operators, but it's especially
302 302 // important (common) to do address expressions.
303 303 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
304 304 if (!has_ctrl(n)) return NULL;
305 305 Node *n_ctrl = get_ctrl(n);
306 306 IdealLoopTree *n_loop = get_loop(n_ctrl);
307 307
308 308 // See if 'n' mixes loop-varying and loop-invariant inputs and
309 309 // itself is loop-varying.
310 310
311 311 // Only interested in binary ops (and AddP)
312 312 if( n->req() < 3 || n->req() > 4 ) return NULL;
313 313
314 314 Node *n1_ctrl = get_ctrl(n->in( 1));
315 315 Node *n2_ctrl = get_ctrl(n->in( 2));
316 316 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
317 317 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
318 318 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
319 319 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
320 320
321 321 // Does one of my inputs spin in a tighter loop than self?
322 322 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
323 323 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
324 324 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
325 325 return NULL; // Leave well enough alone
326 326
327 327 // Is at least one of my inputs loop-invariant?
328 328 if( n1_loop == n_loop &&
329 329 n2_loop == n_loop &&
330 330 n3_loop == n_loop )
331 331 return NULL; // No loop-invariant inputs
332 332
333 333
334 334 int n_op = n->Opcode();
335 335
336 336 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
337 337 if( n_op == Op_LShiftI ) {
338 338 // Scale is loop invariant
339 339 Node *scale = n->in(2);
340 340 Node *scale_ctrl = get_ctrl(scale);
341 341 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
342 342 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
343 343 return NULL;
344 344 const TypeInt *scale_t = scale->bottom_type()->isa_int();
345 345 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
346 346 return NULL; // Dont bother with byte/short masking
347 347 // Add must vary with loop (else shift would be loop-invariant)
348 348 Node *add = n->in(1);
349 349 Node *add_ctrl = get_ctrl(add);
350 350 IdealLoopTree *add_loop = get_loop(add_ctrl);
351 351 //assert( n_loop == add_loop, "" );
352 352 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
353 353
354 354 // Convert I-V into I+ (0-V); same for V-I
355 355 if( add->Opcode() == Op_SubI &&
356 356 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
357 357 Node *zero = _igvn.intcon(0);
358 358 set_ctrl(zero, C->root());
359 359 Node *neg = new (C) SubINode( _igvn.intcon(0), add->in(2) );
360 360 register_new_node( neg, get_ctrl(add->in(2) ) );
361 361 add = new (C) AddINode( add->in(1), neg );
362 362 register_new_node( add, add_ctrl );
363 363 }
364 364 if( add->Opcode() != Op_AddI ) return NULL;
365 365 // See if one add input is loop invariant
366 366 Node *add_var = add->in(1);
367 367 Node *add_var_ctrl = get_ctrl(add_var);
368 368 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
369 369 Node *add_invar = add->in(2);
370 370 Node *add_invar_ctrl = get_ctrl(add_invar);
371 371 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
372 372 if( add_var_loop == n_loop ) {
373 373 } else if( add_invar_loop == n_loop ) {
374 374 // Swap to find the invariant part
375 375 add_invar = add_var;
376 376 add_invar_ctrl = add_var_ctrl;
377 377 add_invar_loop = add_var_loop;
378 378 add_var = add->in(2);
379 379 Node *add_var_ctrl = get_ctrl(add_var);
380 380 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
381 381 } else // Else neither input is loop invariant
382 382 return NULL;
383 383 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
384 384 return NULL; // No invariant part of the add?
385 385
386 386 // Yes! Reshape address expression!
387 387 Node *inv_scale = new (C) LShiftINode( add_invar, scale );
388 388 Node *inv_scale_ctrl =
389 389 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
390 390 add_invar_ctrl : scale_ctrl;
391 391 register_new_node( inv_scale, inv_scale_ctrl );
392 392 Node *var_scale = new (C) LShiftINode( add_var, scale );
393 393 register_new_node( var_scale, n_ctrl );
394 394 Node *var_add = new (C) AddINode( var_scale, inv_scale );
395 395 register_new_node( var_add, n_ctrl );
396 396 _igvn.replace_node( n, var_add );
397 397 return var_add;
398 398 }
399 399
400 400 // Replace (I+V) with (V+I)
401 401 if( n_op == Op_AddI ||
402 402 n_op == Op_AddL ||
403 403 n_op == Op_AddF ||
404 404 n_op == Op_AddD ||
405 405 n_op == Op_MulI ||
406 406 n_op == Op_MulL ||
407 407 n_op == Op_MulF ||
408 408 n_op == Op_MulD ) {
409 409 if( n2_loop == n_loop ) {
410 410 assert( n1_loop != n_loop, "" );
411 411 n->swap_edges(1, 2);
412 412 }
413 413 }
414 414
415 415 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
416 416 // but not if I2 is a constant.
417 417 if( n_op == Op_AddP ) {
418 418 if( n2_loop == n_loop && n3_loop != n_loop ) {
419 419 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
420 420 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
421 421 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
422 422 IdealLoopTree *n22loop = get_loop( n22_ctrl );
423 423 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
424 424 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
425 425 n23_loop == n_loop ) {
426 426 Node *add1 = new (C) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
427 427 // Stuff new AddP in the loop preheader
428 428 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
429 429 Node *add2 = new (C) AddPNode( n->in(1), add1, n->in(2)->in(3) );
430 430 register_new_node( add2, n_ctrl );
431 431 _igvn.replace_node( n, add2 );
432 432 return add2;
433 433 }
434 434 }
435 435 }
436 436
437 437 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
438 438 if( n2_loop != n_loop && n3_loop == n_loop ) {
439 439 if( n->in(3)->Opcode() == Op_AddI ) {
440 440 Node *V = n->in(3)->in(1);
441 441 Node *I = n->in(3)->in(2);
442 442 if( is_member(n_loop,get_ctrl(V)) ) {
443 443 } else {
444 444 Node *tmp = V; V = I; I = tmp;
445 445 }
446 446 if( !is_member(n_loop,get_ctrl(I)) ) {
447 447 Node *add1 = new (C) AddPNode( n->in(1), n->in(2), I );
448 448 // Stuff new AddP in the loop preheader
449 449 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
450 450 Node *add2 = new (C) AddPNode( n->in(1), add1, V );
451 451 register_new_node( add2, n_ctrl );
452 452 _igvn.replace_node( n, add2 );
453 453 return add2;
454 454 }
455 455 }
456 456 }
457 457 }
458 458
459 459 return NULL;
460 460 }
461 461
462 462 //------------------------------conditional_move-------------------------------
463 463 // Attempt to replace a Phi with a conditional move. We have some pretty
464 464 // strict profitability requirements. All Phis at the merge point must
465 465 // be converted, so we can remove the control flow. We need to limit the
466 466 // number of c-moves to a small handful. All code that was in the side-arms
467 467 // of the CFG diamond is now speculatively executed. This code has to be
468 468 // "cheap enough". We are pretty much limited to CFG diamonds that merge
469 469 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
470 470 Node *PhaseIdealLoop::conditional_move( Node *region ) {
471 471
472 472 assert(region->is_Region(), "sanity check");
473 473 if (region->req() != 3) return NULL;
474 474
475 475 // Check for CFG diamond
476 476 Node *lp = region->in(1);
477 477 Node *rp = region->in(2);
478 478 if (!lp || !rp) return NULL;
479 479 Node *lp_c = lp->in(0);
480 480 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
481 481 IfNode *iff = lp_c->as_If();
482 482
483 483 // Check for ops pinned in an arm of the diamond.
484 484 // Can't remove the control flow in this case
485 485 if (lp->outcnt() > 1) return NULL;
486 486 if (rp->outcnt() > 1) return NULL;
487 487
488 488 IdealLoopTree* r_loop = get_loop(region);
489 489 assert(r_loop == get_loop(iff), "sanity");
490 490 // Always convert to CMOVE if all results are used only outside this loop.
491 491 bool used_inside_loop = (r_loop == _ltree_root);
492 492
493 493 // Check profitability
494 494 int cost = 0;
495 495 int phis = 0;
496 496 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
497 497 Node *out = region->fast_out(i);
498 498 if (!out->is_Phi()) continue; // Ignore other control edges, etc
499 499 phis++;
500 500 PhiNode* phi = out->as_Phi();
501 501 BasicType bt = phi->type()->basic_type();
502 502 switch (bt) {
503 503 case T_FLOAT:
504 504 case T_DOUBLE: {
505 505 cost += Matcher::float_cmove_cost(); // Could be very expensive
506 506 break;
507 507 }
508 508 case T_LONG: {
509 509 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
510 510 }
511 511 case T_INT: // These all CMOV fine
512 512 case T_ADDRESS: { // (RawPtr)
513 513 cost++;
514 514 break;
515 515 }
516 516 case T_NARROWOOP: // Fall through
517 517 case T_OBJECT: { // Base oops are OK, but not derived oops
518 518 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
519 519 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
520 520 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
521 521 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
522 522 // have a Phi for the base here that we convert to a CMOVE all is well
523 523 // and good. But if the base is dead, we'll not make a CMOVE. Later
524 524 // the allocator will have to produce a base by creating a CMOVE of the
525 525 // relevant bases. This puts the allocator in the business of
526 526 // manufacturing expensive instructions, generally a bad plan.
527 527 // Just Say No to Conditionally-Moved Derived Pointers.
528 528 if (tp && tp->offset() != 0)
529 529 return NULL;
530 530 cost++;
531 531 break;
532 532 }
533 533 default:
534 534 return NULL; // In particular, can't do memory or I/O
535 535 }
536 536 // Add in cost any speculative ops
537 537 for (uint j = 1; j < region->req(); j++) {
538 538 Node *proj = region->in(j);
539 539 Node *inp = phi->in(j);
540 540 if (get_ctrl(inp) == proj) { // Found local op
541 541 cost++;
542 542 // Check for a chain of dependent ops; these will all become
543 543 // speculative in a CMOV.
544 544 for (uint k = 1; k < inp->req(); k++)
545 545 if (get_ctrl(inp->in(k)) == proj)
546 546 cost += ConditionalMoveLimit; // Too much speculative goo
547 547 }
548 548 }
549 549 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
550 550 // This will likely Split-If, a higher-payoff operation.
551 551 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
552 552 Node* use = phi->fast_out(k);
553 553 if (use->is_Cmp() || use->is_DecodeN() || use->is_EncodeP())
554 554 cost += ConditionalMoveLimit;
555 555 // Is there a use inside the loop?
556 556 // Note: check only basic types since CMoveP is pinned.
557 557 if (!used_inside_loop && is_java_primitive(bt)) {
558 558 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
559 559 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
560 560 used_inside_loop = true;
561 561 }
562 562 }
563 563 }
564 564 }
565 565 Node* bol = iff->in(1);
566 566 assert(bol->Opcode() == Op_Bool, "");
567 567 int cmp_op = bol->in(1)->Opcode();
568 568 // It is expensive to generate flags from a float compare.
569 569 // Avoid duplicated float compare.
570 570 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
571 571
572 572 float infrequent_prob = PROB_UNLIKELY_MAG(3);
573 573 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
574 574 if (used_inside_loop) {
575 575 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
576 576
577 577 // BlockLayoutByFrequency optimization moves infrequent branch
578 578 // from hot path. No point in CMOV'ing in such case (110 is used
579 579 // instead of 100 to take into account not exactness of float value).
580 580 if (BlockLayoutByFrequency) {
581 581 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
582 582 }
583 583 }
584 584 // Check for highly predictable branch. No point in CMOV'ing if
585 585 // we are going to predict accurately all the time.
586 586 if (iff->_prob < infrequent_prob ||
587 587 iff->_prob > (1.0f - infrequent_prob))
588 588 return NULL;
589 589
590 590 // --------------
591 591 // Now replace all Phis with CMOV's
592 592 Node *cmov_ctrl = iff->in(0);
593 593 uint flip = (lp->Opcode() == Op_IfTrue);
594 594 while (1) {
595 595 PhiNode* phi = NULL;
596 596 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
597 597 Node *out = region->fast_out(i);
598 598 if (out->is_Phi()) {
599 599 phi = out->as_Phi();
600 600 break;
601 601 }
602 602 }
603 603 if (phi == NULL) break;
604 604 #ifndef PRODUCT
605 605 if (PrintOpto && VerifyLoopOptimizations) tty->print_cr("CMOV");
606 606 #endif
607 607 // Move speculative ops
608 608 for (uint j = 1; j < region->req(); j++) {
609 609 Node *proj = region->in(j);
610 610 Node *inp = phi->in(j);
611 611 if (get_ctrl(inp) == proj) { // Found local op
612 612 #ifndef PRODUCT
613 613 if (PrintOpto && VerifyLoopOptimizations) {
614 614 tty->print(" speculate: ");
615 615 inp->dump();
616 616 }
617 617 #endif
618 618 set_ctrl(inp, cmov_ctrl);
619 619 }
620 620 }
621 621 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) );
622 622 register_new_node( cmov, cmov_ctrl );
623 623 _igvn.replace_node( phi, cmov );
624 624 #ifndef PRODUCT
625 625 if (TraceLoopOpts) {
626 626 tty->print("CMOV ");
627 627 r_loop->dump_head();
628 628 if (Verbose) {
629 629 bol->in(1)->dump(1);
630 630 cmov->dump(1);
631 631 }
632 632 }
633 633 if (VerifyLoopOptimizations) verify();
634 634 #endif
635 635 }
636 636
637 637 // The useless CFG diamond will fold up later; see the optimization in
638 638 // RegionNode::Ideal.
639 639 _igvn._worklist.push(region);
640 640
641 641 return iff->in(1);
642 642 }
643 643
644 644 //------------------------------split_if_with_blocks_pre-----------------------
645 645 // Do the real work in a non-recursive function. Data nodes want to be
646 646 // cloned in the pre-order so they can feed each other nicely.
647 647 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
648 648 // Cloning these guys is unlikely to win
649 649 int n_op = n->Opcode();
650 650 if( n_op == Op_MergeMem ) return n;
651 651 if( n->is_Proj() ) return n;
652 652 // Do not clone-up CmpFXXX variations, as these are always
653 653 // followed by a CmpI
654 654 if( n->is_Cmp() ) return n;
655 655 // Attempt to use a conditional move instead of a phi/branch
656 656 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
657 657 Node *cmov = conditional_move( n );
658 658 if( cmov ) return cmov;
659 659 }
660 660 if( n->is_CFG() || n->is_LoadStore() )
661 661 return n;
662 662 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
663 663 n_op == Op_Opaque2 ) {
664 664 if( !C->major_progress() ) // If chance of no more loop opts...
665 665 _igvn._worklist.push(n); // maybe we'll remove them
666 666 return n;
667 667 }
668 668
669 669 if( n->is_Con() ) return n; // No cloning for Con nodes
670 670
671 671 Node *n_ctrl = get_ctrl(n);
672 672 if( !n_ctrl ) return n; // Dead node
673 673
674 674 // Attempt to remix address expressions for loop invariants
675 675 Node *m = remix_address_expressions( n );
676 676 if( m ) return m;
677 677
678 678 // Determine if the Node has inputs from some local Phi.
679 679 // Returns the block to clone thru.
680 680 Node *n_blk = has_local_phi_input( n );
681 681 if( !n_blk ) return n;
682 682 // Do not clone the trip counter through on a CountedLoop
683 683 // (messes up the canonical shape).
684 684 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
685 685
686 686 // Check for having no control input; not pinned. Allow
687 687 // dominating control.
688 688 if( n->in(0) ) {
689 689 Node *dom = idom(n_blk);
690 690 if( dom_lca( n->in(0), dom ) != n->in(0) )
691 691 return n;
692 692 }
693 693 // Policy: when is it profitable. You must get more wins than
694 694 // policy before it is considered profitable. Policy is usually 0,
695 695 // so 1 win is considered profitable. Big merges will require big
696 696 // cloning, so get a larger policy.
697 697 int policy = n_blk->req() >> 2;
698 698
699 699 // If the loop is a candidate for range check elimination,
700 700 // delay splitting through it's phi until a later loop optimization
701 701 if (n_blk->is_CountedLoop()) {
702 702 IdealLoopTree *lp = get_loop(n_blk);
703 703 if (lp && lp->_rce_candidate) {
704 704 return n;
705 705 }
706 706 }
707 707
708 708 // Use same limit as split_if_with_blocks_post
709 709 if( C->unique() > 35000 ) return n; // Method too big
710 710
711 711 // Split 'n' through the merge point if it is profitable
712 712 Node *phi = split_thru_phi( n, n_blk, policy );
713 713 if (!phi) return n;
714 714
715 715 // Found a Phi to split thru!
716 716 // Replace 'n' with the new phi
717 717 _igvn.replace_node( n, phi );
718 718 // Moved a load around the loop, 'en-registering' something.
719 719 if (n_blk->is_Loop() && n->is_Load() &&
720 720 !phi->in(LoopNode::LoopBackControl)->is_Load())
721 721 C->set_major_progress();
722 722
723 723 return phi;
724 724 }
725 725
726 726 static bool merge_point_too_heavy(Compile* C, Node* region) {
727 727 // Bail out if the region and its phis have too many users.
728 728 int weight = 0;
729 729 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
730 730 weight += region->fast_out(i)->outcnt();
731 731 }
732 732 int nodes_left = MaxNodeLimit - C->live_nodes();
733 733 if (weight * 8 > nodes_left) {
734 734 #ifndef PRODUCT
735 735 if (PrintOpto)
736 736 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
737 737 #endif
738 738 return true;
739 739 } else {
740 740 return false;
741 741 }
742 742 }
743 743
744 744 static bool merge_point_safe(Node* region) {
745 745 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
746 746 // having a PhiNode input. This sidesteps the dangerous case where the split
747 747 // ConvI2LNode may become TOP if the input Value() does not
748 748 // overlap the ConvI2L range, leaving a node which may not dominate its
749 749 // uses.
750 750 // A better fix for this problem can be found in the BugTraq entry, but
751 751 // expediency for Mantis demands this hack.
752 752 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop
753 753 // split_if_with_blocks from splitting a block because we could not move around
754 754 // the FastLockNode.
755 755 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
756 756 Node* n = region->fast_out(i);
757 757 if (n->is_Phi()) {
758 758 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
759 759 Node* m = n->fast_out(j);
760 760 if (m->is_FastLock())
761 761 return false;
762 762 #ifdef _LP64
763 763 if (m->Opcode() == Op_ConvI2L)
764 764 return false;
765 765 #endif
766 766 }
767 767 }
768 768 }
769 769 return true;
770 770 }
771 771
772 772
773 773 //------------------------------place_near_use---------------------------------
774 774 // Place some computation next to use but not inside inner loops.
775 775 // For inner loop uses move it to the preheader area.
776 776 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const {
777 777 IdealLoopTree *u_loop = get_loop( useblock );
778 778 return (u_loop->_irreducible || u_loop->_child)
779 779 ? useblock
780 780 : u_loop->_head->in(LoopNode::EntryControl);
781 781 }
782 782
783 783
784 784 //------------------------------split_if_with_blocks_post----------------------
785 785 // Do the real work in a non-recursive function. CFG hackery wants to be
786 786 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
787 787 // info.
788 788 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) {
789 789
790 790 // Cloning Cmp through Phi's involves the split-if transform.
791 791 // FastLock is not used by an If
792 792 if( n->is_Cmp() && !n->is_FastLock() ) {
793 793 if( C->unique() > 35000 ) return; // Method too big
794 794
795 795 // Do not do 'split-if' if irreducible loops are present.
796 796 if( _has_irreducible_loops )
797 797 return;
798 798
799 799 Node *n_ctrl = get_ctrl(n);
800 800 // Determine if the Node has inputs from some local Phi.
801 801 // Returns the block to clone thru.
802 802 Node *n_blk = has_local_phi_input( n );
803 803 if( n_blk != n_ctrl ) return;
804 804
805 805 if( merge_point_too_heavy(C, n_ctrl) )
806 806 return;
807 807
808 808 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare?
809 809 Node *bol = n->unique_out();
810 810 assert( bol->is_Bool(), "expect a bool here" );
811 811 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare?
812 812 Node *iff = bol->unique_out();
813 813
814 814 // Check some safety conditions
815 815 if( iff->is_If() ) { // Classic split-if?
816 816 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if
817 817 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
818 818 // Can't split CMove with different control edge.
819 819 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) return;
820 820 if( get_ctrl(iff->in(2)) == n_ctrl ||
821 821 get_ctrl(iff->in(3)) == n_ctrl )
822 822 return; // Inputs not yet split-up
823 823 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) {
824 824 return; // Loop-invar test gates loop-varying CMOVE
825 825 }
826 826 } else {
827 827 return; // some other kind of node, such as an Allocate
828 828 }
829 829
830 830 // Do not do 'split-if' if some paths are dead. First do dead code
831 831 // elimination and then see if its still profitable.
832 832 for( uint i = 1; i < n_ctrl->req(); i++ )
833 833 if( n_ctrl->in(i) == C->top() )
834 834 return;
835 835
836 836 // When is split-if profitable? Every 'win' on means some control flow
837 837 // goes dead, so it's almost always a win.
838 838 int policy = 0;
839 839 // If trying to do a 'Split-If' at the loop head, it is only
840 840 // profitable if the cmp folds up on BOTH paths. Otherwise we
841 841 // risk peeling a loop forever.
842 842
843 843 // CNC - Disabled for now. Requires careful handling of loop
844 844 // body selection for the cloned code. Also, make sure we check
845 845 // for any input path not being in the same loop as n_ctrl. For
846 846 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
847 847 // because the alternative loop entry points won't be converted
848 848 // into LoopNodes.
849 849 IdealLoopTree *n_loop = get_loop(n_ctrl);
850 850 for( uint j = 1; j < n_ctrl->req(); j++ )
851 851 if( get_loop(n_ctrl->in(j)) != n_loop )
852 852 return;
853 853
854 854 // Check for safety of the merge point.
855 855 if( !merge_point_safe(n_ctrl) ) {
856 856 return;
857 857 }
858 858
859 859 // Split compare 'n' through the merge point if it is profitable
860 860 Node *phi = split_thru_phi( n, n_ctrl, policy );
861 861 if( !phi ) return;
862 862
863 863 // Found a Phi to split thru!
864 864 // Replace 'n' with the new phi
865 865 _igvn.replace_node( n, phi );
866 866
867 867 // Now split the bool up thru the phi
868 868 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 );
869 869 _igvn.replace_node( bol, bolphi );
870 870 assert( iff->in(1) == bolphi, "" );
871 871 if( bolphi->Value(&_igvn)->singleton() )
872 872 return;
873 873
874 874 // Conditional-move? Must split up now
875 875 if( !iff->is_If() ) {
876 876 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 );
877 877 _igvn.replace_node( iff, cmovphi );
878 878 return;
879 879 }
880 880
881 881 // Now split the IF
882 882 do_split_if( iff );
883 883 return;
884 884 }
885 885
886 886 // Check for an IF ready to split; one that has its
887 887 // condition codes input coming from a Phi at the block start.
888 888 int n_op = n->Opcode();
889 889
890 890 // Check for an IF being dominated by another IF same test
891 891 if( n_op == Op_If ) {
892 892 Node *bol = n->in(1);
893 893 uint max = bol->outcnt();
894 894 // Check for same test used more than once?
895 895 if( n_op == Op_If && max > 1 && bol->is_Bool() ) {
896 896 // Search up IDOMs to see if this IF is dominated.
897 897 Node *cutoff = get_ctrl(bol);
898 898
899 899 // Now search up IDOMs till cutoff, looking for a dominating test
900 900 Node *prevdom = n;
901 901 Node *dom = idom(prevdom);
902 902 while( dom != cutoff ) {
903 903 if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) {
904 904 // Replace the dominated test with an obvious true or false.
905 905 // Place it on the IGVN worklist for later cleanup.
906 906 C->set_major_progress();
907 907 dominated_by( prevdom, n, false, true );
908 908 #ifndef PRODUCT
909 909 if( VerifyLoopOptimizations ) verify();
910 910 #endif
911 911 return;
912 912 }
913 913 prevdom = dom;
914 914 dom = idom(prevdom);
915 915 }
916 916 }
917 917 }
918 918
919 919 // See if a shared loop-varying computation has no loop-varying uses.
920 920 // Happens if something is only used for JVM state in uncommon trap exits,
921 921 // like various versions of induction variable+offset. Clone the
922 922 // computation per usage to allow it to sink out of the loop.
923 923 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
924 924 Node *n_ctrl = get_ctrl(n);
925 925 IdealLoopTree *n_loop = get_loop(n_ctrl);
926 926 if( n_loop != _ltree_root ) {
927 927 DUIterator_Fast imax, i = n->fast_outs(imax);
928 928 for (; i < imax; i++) {
929 929 Node* u = n->fast_out(i);
930 930 if( !has_ctrl(u) ) break; // Found control user
931 931 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
932 932 if( u_loop == n_loop ) break; // Found loop-varying use
933 933 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
934 934 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
935 935 }
936 936 bool did_break = (i < imax); // Did we break out of the previous loop?
937 937 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
938 938 Node *late_load_ctrl = NULL;
939 939 if (n->is_Load()) {
940 940 // If n is a load, get and save the result from get_late_ctrl(),
941 941 // to be later used in calculating the control for n's clones.
942 942 clear_dom_lca_tags();
943 943 late_load_ctrl = get_late_ctrl(n, n_ctrl);
944 944 }
945 945 // If n is a load, and the late control is the same as the current
946 946 // control, then the cloning of n is a pointless exercise, because
947 947 // GVN will ensure that we end up where we started.
948 948 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
949 949 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
950 950 Node *u = n->last_out(j); // Clone private computation per use
951 951 _igvn.rehash_node_delayed(u);
952 952 Node *x = n->clone(); // Clone computation
953 953 Node *x_ctrl = NULL;
954 954 if( u->is_Phi() ) {
955 955 // Replace all uses of normal nodes. Replace Phi uses
956 956 // individually, so the separate Nodes can sink down
957 957 // different paths.
958 958 uint k = 1;
959 959 while( u->in(k) != n ) k++;
960 960 u->set_req( k, x );
961 961 // x goes next to Phi input path
962 962 x_ctrl = u->in(0)->in(k);
963 963 --j;
964 964 } else { // Normal use
965 965 // Replace all uses
966 966 for( uint k = 0; k < u->req(); k++ ) {
967 967 if( u->in(k) == n ) {
968 968 u->set_req( k, x );
969 969 --j;
970 970 }
971 971 }
972 972 x_ctrl = get_ctrl(u);
973 973 }
974 974
975 975 // Find control for 'x' next to use but not inside inner loops.
976 976 // For inner loop uses get the preheader area.
977 977 x_ctrl = place_near_use(x_ctrl);
978 978
979 979 if (n->is_Load()) {
980 980 // For loads, add a control edge to a CFG node outside of the loop
981 981 // to force them to not combine and return back inside the loop
982 982 // during GVN optimization (4641526).
983 983 //
984 984 // Because we are setting the actual control input, factor in
985 985 // the result from get_late_ctrl() so we respect any
986 986 // anti-dependences. (6233005).
987 987 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
988 988
989 989 // Don't allow the control input to be a CFG splitting node.
990 990 // Such nodes should only have ProjNodes as outs, e.g. IfNode
991 991 // should only have IfTrueNode and IfFalseNode (4985384).
992 992 x_ctrl = find_non_split_ctrl(x_ctrl);
993 993 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
994 994
995 995 x->set_req(0, x_ctrl);
996 996 }
997 997 register_new_node(x, x_ctrl);
998 998
999 999 // Some institutional knowledge is needed here: 'x' is
1000 1000 // yanked because if the optimizer runs GVN on it all the
1001 1001 // cloned x's will common up and undo this optimization and
1002 1002 // be forced back in the loop. This is annoying because it
1003 1003 // makes +VerifyOpto report false-positives on progress. I
1004 1004 // tried setting control edges on the x's to force them to
1005 1005 // not combine, but the matching gets worried when it tries
1006 1006 // to fold a StoreP and an AddP together (as part of an
1007 1007 // address expression) and the AddP and StoreP have
1008 1008 // different controls.
1009 1009 if( !x->is_Load() && !x->is_DecodeN() ) _igvn._worklist.yank(x);
1010 1010 }
1011 1011 _igvn.remove_dead_node(n);
1012 1012 }
1013 1013 }
1014 1014 }
1015 1015 }
1016 1016
1017 1017 // Check for Opaque2's who's loop has disappeared - who's input is in the
1018 1018 // same loop nest as their output. Remove 'em, they are no longer useful.
1019 1019 if( n_op == Op_Opaque2 &&
1020 1020 n->in(1) != NULL &&
1021 1021 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1022 1022 _igvn.replace_node( n, n->in(1) );
1023 1023 }
1024 1024 }
1025 1025
1026 1026 //------------------------------split_if_with_blocks---------------------------
1027 1027 // Check for aggressive application of 'split-if' optimization,
1028 1028 // using basic block level info.
1029 1029 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) {
1030 1030 Node *n = C->root();
1031 1031 visited.set(n->_idx); // first, mark node as visited
1032 1032 // Do pre-visit work for root
1033 1033 n = split_if_with_blocks_pre( n );
1034 1034 uint cnt = n->outcnt();
1035 1035 uint i = 0;
1036 1036 while (true) {
1037 1037 // Visit all children
1038 1038 if (i < cnt) {
1039 1039 Node* use = n->raw_out(i);
1040 1040 ++i;
1041 1041 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1042 1042 // Now do pre-visit work for this use
1043 1043 use = split_if_with_blocks_pre( use );
1044 1044 nstack.push(n, i); // Save parent and next use's index.
1045 1045 n = use; // Process all children of current use.
1046 1046 cnt = use->outcnt();
1047 1047 i = 0;
1048 1048 }
1049 1049 }
1050 1050 else {
1051 1051 // All of n's children have been processed, complete post-processing.
1052 1052 if (cnt != 0 && !n->is_Con()) {
1053 1053 assert(has_node(n), "no dead nodes");
1054 1054 split_if_with_blocks_post( n );
1055 1055 }
1056 1056 if (nstack.is_empty()) {
1057 1057 // Finished all nodes on stack.
1058 1058 break;
1059 1059 }
1060 1060 // Get saved parent node and next use's index. Visit the rest of uses.
1061 1061 n = nstack.node();
1062 1062 cnt = n->outcnt();
1063 1063 i = nstack.index();
1064 1064 nstack.pop();
1065 1065 }
1066 1066 }
1067 1067 }
1068 1068
1069 1069
1070 1070 //=============================================================================
1071 1071 //
1072 1072 // C L O N E A L O O P B O D Y
1073 1073 //
1074 1074
1075 1075 //------------------------------clone_iff--------------------------------------
1076 1076 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1077 1077 // "Nearly" because all Nodes have been cloned from the original in the loop,
1078 1078 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1079 1079 // through the Phi recursively, and return a Bool.
1080 1080 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) {
1081 1081
1082 1082 // Convert this Phi into a Phi merging Bools
1083 1083 uint i;
1084 1084 for( i = 1; i < phi->req(); i++ ) {
1085 1085 Node *b = phi->in(i);
1086 1086 if( b->is_Phi() ) {
1087 1087 _igvn.replace_input_of(phi, i, clone_iff( b->as_Phi(), loop ));
1088 1088 } else {
1089 1089 assert( b->is_Bool(), "" );
1090 1090 }
1091 1091 }
1092 1092
1093 1093 Node *sample_bool = phi->in(1);
1094 1094 Node *sample_cmp = sample_bool->in(1);
1095 1095
1096 1096 // Make Phis to merge the Cmp's inputs.
1097 1097 PhiNode *phi1 = new (C) PhiNode( phi->in(0), Type::TOP );
1098 1098 PhiNode *phi2 = new (C) PhiNode( phi->in(0), Type::TOP );
1099 1099 for( i = 1; i < phi->req(); i++ ) {
1100 1100 Node *n1 = phi->in(i)->in(1)->in(1);
1101 1101 Node *n2 = phi->in(i)->in(1)->in(2);
1102 1102 phi1->set_req( i, n1 );
1103 1103 phi2->set_req( i, n2 );
1104 1104 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1105 1105 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1106 1106 }
1107 1107 // See if these Phis have been made before.
1108 1108 // Register with optimizer
1109 1109 Node *hit1 = _igvn.hash_find_insert(phi1);
1110 1110 if( hit1 ) { // Hit, toss just made Phi
1111 1111 _igvn.remove_dead_node(phi1); // Remove new phi
1112 1112 assert( hit1->is_Phi(), "" );
1113 1113 phi1 = (PhiNode*)hit1; // Use existing phi
1114 1114 } else { // Miss
1115 1115 _igvn.register_new_node_with_optimizer(phi1);
1116 1116 }
1117 1117 Node *hit2 = _igvn.hash_find_insert(phi2);
1118 1118 if( hit2 ) { // Hit, toss just made Phi
1119 1119 _igvn.remove_dead_node(phi2); // Remove new phi
1120 1120 assert( hit2->is_Phi(), "" );
1121 1121 phi2 = (PhiNode*)hit2; // Use existing phi
1122 1122 } else { // Miss
1123 1123 _igvn.register_new_node_with_optimizer(phi2);
1124 1124 }
1125 1125 // Register Phis with loop/block info
1126 1126 set_ctrl(phi1, phi->in(0));
1127 1127 set_ctrl(phi2, phi->in(0));
1128 1128 // Make a new Cmp
1129 1129 Node *cmp = sample_cmp->clone();
1130 1130 cmp->set_req( 1, phi1 );
1131 1131 cmp->set_req( 2, phi2 );
1132 1132 _igvn.register_new_node_with_optimizer(cmp);
1133 1133 set_ctrl(cmp, phi->in(0));
1134 1134
1135 1135 // Make a new Bool
1136 1136 Node *b = sample_bool->clone();
1137 1137 b->set_req(1,cmp);
1138 1138 _igvn.register_new_node_with_optimizer(b);
1139 1139 set_ctrl(b, phi->in(0));
1140 1140
1141 1141 assert( b->is_Bool(), "" );
1142 1142 return (BoolNode*)b;
1143 1143 }
1144 1144
1145 1145 //------------------------------clone_bool-------------------------------------
1146 1146 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1147 1147 // "Nearly" because all Nodes have been cloned from the original in the loop,
1148 1148 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1149 1149 // through the Phi recursively, and return a Bool.
1150 1150 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1151 1151 uint i;
1152 1152 // Convert this Phi into a Phi merging Bools
1153 1153 for( i = 1; i < phi->req(); i++ ) {
1154 1154 Node *b = phi->in(i);
1155 1155 if( b->is_Phi() ) {
1156 1156 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1157 1157 } else {
1158 1158 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1159 1159 }
1160 1160 }
1161 1161
1162 1162 Node *sample_cmp = phi->in(1);
1163 1163
1164 1164 // Make Phis to merge the Cmp's inputs.
1165 1165 PhiNode *phi1 = new (C) PhiNode( phi->in(0), Type::TOP );
1166 1166 PhiNode *phi2 = new (C) PhiNode( phi->in(0), Type::TOP );
1167 1167 for( uint j = 1; j < phi->req(); j++ ) {
1168 1168 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1169 1169 Node *n1, *n2;
1170 1170 if( cmp_top->is_Cmp() ) {
1171 1171 n1 = cmp_top->in(1);
1172 1172 n2 = cmp_top->in(2);
1173 1173 } else {
1174 1174 n1 = n2 = cmp_top;
1175 1175 }
1176 1176 phi1->set_req( j, n1 );
1177 1177 phi2->set_req( j, n2 );
1178 1178 phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1179 1179 phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1180 1180 }
1181 1181
1182 1182 // See if these Phis have been made before.
1183 1183 // Register with optimizer
1184 1184 Node *hit1 = _igvn.hash_find_insert(phi1);
1185 1185 if( hit1 ) { // Hit, toss just made Phi
1186 1186 _igvn.remove_dead_node(phi1); // Remove new phi
1187 1187 assert( hit1->is_Phi(), "" );
1188 1188 phi1 = (PhiNode*)hit1; // Use existing phi
1189 1189 } else { // Miss
1190 1190 _igvn.register_new_node_with_optimizer(phi1);
1191 1191 }
1192 1192 Node *hit2 = _igvn.hash_find_insert(phi2);
1193 1193 if( hit2 ) { // Hit, toss just made Phi
1194 1194 _igvn.remove_dead_node(phi2); // Remove new phi
1195 1195 assert( hit2->is_Phi(), "" );
1196 1196 phi2 = (PhiNode*)hit2; // Use existing phi
1197 1197 } else { // Miss
1198 1198 _igvn.register_new_node_with_optimizer(phi2);
1199 1199 }
1200 1200 // Register Phis with loop/block info
1201 1201 set_ctrl(phi1, phi->in(0));
1202 1202 set_ctrl(phi2, phi->in(0));
1203 1203 // Make a new Cmp
1204 1204 Node *cmp = sample_cmp->clone();
1205 1205 cmp->set_req( 1, phi1 );
1206 1206 cmp->set_req( 2, phi2 );
1207 1207 _igvn.register_new_node_with_optimizer(cmp);
1208 1208 set_ctrl(cmp, phi->in(0));
1209 1209
1210 1210 assert( cmp->is_Cmp(), "" );
1211 1211 return (CmpNode*)cmp;
1212 1212 }
1213 1213
1214 1214 //------------------------------sink_use---------------------------------------
1215 1215 // If 'use' was in the loop-exit block, it now needs to be sunk
1216 1216 // below the post-loop merge point.
1217 1217 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1218 1218 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1219 1219 set_ctrl(use, post_loop);
1220 1220 for (DUIterator j = use->outs(); use->has_out(j); j++)
1221 1221 sink_use(use->out(j), post_loop);
1222 1222 }
1223 1223 }
1224 1224
1225 1225 //------------------------------clone_loop-------------------------------------
1226 1226 //
1227 1227 // C L O N E A L O O P B O D Y
1228 1228 //
1229 1229 // This is the basic building block of the loop optimizations. It clones an
1230 1230 // entire loop body. It makes an old_new loop body mapping; with this mapping
1231 1231 // you can find the new-loop equivalent to an old-loop node. All new-loop
1232 1232 // nodes are exactly equal to their old-loop counterparts, all edges are the
1233 1233 // same. All exits from the old-loop now have a RegionNode that merges the
1234 1234 // equivalent new-loop path. This is true even for the normal "loop-exit"
1235 1235 // condition. All uses of loop-invariant old-loop values now come from (one
1236 1236 // or more) Phis that merge their new-loop equivalents.
1237 1237 //
1238 1238 // This operation leaves the graph in an illegal state: there are two valid
1239 1239 // control edges coming from the loop pre-header to both loop bodies. I'll
1240 1240 // definitely have to hack the graph after running this transform.
1241 1241 //
1242 1242 // From this building block I will further edit edges to perform loop peeling
1243 1243 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1244 1244 //
1245 1245 // Parameter side_by_size_idom:
1246 1246 // When side_by_size_idom is NULL, the dominator tree is constructed for
1247 1247 // the clone loop to dominate the original. Used in construction of
1248 1248 // pre-main-post loop sequence.
1249 1249 // When nonnull, the clone and original are side-by-side, both are
1250 1250 // dominated by the side_by_side_idom node. Used in construction of
1251 1251 // unswitched loops.
1252 1252 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1253 1253 Node* side_by_side_idom) {
1254 1254
1255 1255 // Step 1: Clone the loop body. Make the old->new mapping.
1256 1256 uint i;
1257 1257 for( i = 0; i < loop->_body.size(); i++ ) {
1258 1258 Node *old = loop->_body.at(i);
1259 1259 Node *nnn = old->clone();
1260 1260 old_new.map( old->_idx, nnn );
1261 1261 _igvn.register_new_node_with_optimizer(nnn);
1262 1262 }
1263 1263
1264 1264
1265 1265 // Step 2: Fix the edges in the new body. If the old input is outside the
1266 1266 // loop use it. If the old input is INside the loop, use the corresponding
1267 1267 // new node instead.
1268 1268 for( i = 0; i < loop->_body.size(); i++ ) {
1269 1269 Node *old = loop->_body.at(i);
1270 1270 Node *nnn = old_new[old->_idx];
1271 1271 // Fix CFG/Loop controlling the new node
1272 1272 if (has_ctrl(old)) {
1273 1273 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
1274 1274 } else {
1275 1275 set_loop(nnn, loop->_parent);
1276 1276 if (old->outcnt() > 0) {
1277 1277 set_idom( nnn, old_new[idom(old)->_idx], dd );
1278 1278 }
1279 1279 }
1280 1280 // Correct edges to the new node
1281 1281 for( uint j = 0; j < nnn->req(); j++ ) {
1282 1282 Node *n = nnn->in(j);
1283 1283 if( n ) {
1284 1284 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
1285 1285 if( loop->is_member( old_in_loop ) )
1286 1286 nnn->set_req(j, old_new[n->_idx]);
1287 1287 }
1288 1288 }
1289 1289 _igvn.hash_find_insert(nnn);
1290 1290 }
1291 1291 Node *newhead = old_new[loop->_head->_idx];
1292 1292 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1293 1293
1294 1294
1295 1295 // Step 3: Now fix control uses. Loop varying control uses have already
1296 1296 // been fixed up (as part of all input edges in Step 2). Loop invariant
1297 1297 // control uses must be either an IfFalse or an IfTrue. Make a merge
1298 1298 // point to merge the old and new IfFalse/IfTrue nodes; make the use
1299 1299 // refer to this.
1300 1300 ResourceArea *area = Thread::current()->resource_area();
1301 1301 Node_List worklist(area);
1302 1302 uint new_counter = C->unique();
1303 1303 for( i = 0; i < loop->_body.size(); i++ ) {
1304 1304 Node* old = loop->_body.at(i);
1305 1305 if( !old->is_CFG() ) continue;
1306 1306 Node* nnn = old_new[old->_idx];
1307 1307
1308 1308 // Copy uses to a worklist, so I can munge the def-use info
1309 1309 // with impunity.
1310 1310 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1311 1311 worklist.push(old->fast_out(j));
1312 1312
1313 1313 while( worklist.size() ) { // Visit all uses
1314 1314 Node *use = worklist.pop();
1315 1315 if (!has_node(use)) continue; // Ignore dead nodes
1316 1316 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1317 1317 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
1318 1318 // Both OLD and USE are CFG nodes here.
1319 1319 assert( use->is_Proj(), "" );
1320 1320
1321 1321 // Clone the loop exit control projection
1322 1322 Node *newuse = use->clone();
1323 1323 newuse->set_req(0,nnn);
1324 1324 _igvn.register_new_node_with_optimizer(newuse);
1325 1325 set_loop(newuse, use_loop);
1326 1326 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
1327 1327
1328 1328 // We need a Region to merge the exit from the peeled body and the
1329 1329 // exit from the old loop body.
1330 1330 RegionNode *r = new (C) RegionNode(3);
1331 1331 // Map the old use to the new merge point
1332 1332 old_new.map( use->_idx, r );
1333 1333 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
1334 1334 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
1335 1335
1336 1336 // The original user of 'use' uses 'r' instead.
1337 1337 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
1338 1338 Node* useuse = use->last_out(l);
1339 1339 _igvn.rehash_node_delayed(useuse);
1340 1340 uint uses_found = 0;
1341 1341 if( useuse->in(0) == use ) {
1342 1342 useuse->set_req(0, r);
1343 1343 uses_found++;
1344 1344 if( useuse->is_CFG() ) {
1345 1345 assert( dom_depth(useuse) > dd_r, "" );
1346 1346 set_idom(useuse, r, dom_depth(useuse));
1347 1347 }
1348 1348 }
1349 1349 for( uint k = 1; k < useuse->req(); k++ ) {
1350 1350 if( useuse->in(k) == use ) {
1351 1351 useuse->set_req(k, r);
1352 1352 uses_found++;
1353 1353 }
1354 1354 }
1355 1355 l -= uses_found; // we deleted 1 or more copies of this edge
1356 1356 }
1357 1357
1358 1358 // Now finish up 'r'
1359 1359 r->set_req( 1, newuse );
1360 1360 r->set_req( 2, use );
1361 1361 _igvn.register_new_node_with_optimizer(r);
1362 1362 set_loop(r, use_loop);
1363 1363 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
1364 1364 } // End of if a loop-exit test
1365 1365 }
1366 1366 }
1367 1367
1368 1368 // Step 4: If loop-invariant use is not control, it must be dominated by a
1369 1369 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
1370 1370 // there if needed. Make a Phi there merging old and new used values.
1371 1371 Node_List *split_if_set = NULL;
1372 1372 Node_List *split_bool_set = NULL;
1373 1373 Node_List *split_cex_set = NULL;
1374 1374 for( i = 0; i < loop->_body.size(); i++ ) {
1375 1375 Node* old = loop->_body.at(i);
1376 1376 Node* nnn = old_new[old->_idx];
1377 1377 // Copy uses to a worklist, so I can munge the def-use info
1378 1378 // with impunity.
1379 1379 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1380 1380 worklist.push(old->fast_out(j));
1381 1381
1382 1382 while( worklist.size() ) {
1383 1383 Node *use = worklist.pop();
1384 1384 if (!has_node(use)) continue; // Ignore dead nodes
1385 1385 if (use->in(0) == C->top()) continue;
1386 1386 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1387 1387 // Check for data-use outside of loop - at least one of OLD or USE
1388 1388 // must not be a CFG node.
1389 1389 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) {
1390 1390
1391 1391 // If the Data use is an IF, that means we have an IF outside of the
1392 1392 // loop that is switching on a condition that is set inside of the
1393 1393 // loop. Happens if people set a loop-exit flag; then test the flag
1394 1394 // in the loop to break the loop, then test is again outside of the
1395 1395 // loop to determine which way the loop exited.
1396 1396 if( use->is_If() || use->is_CMove() ) {
1397 1397 // Since this code is highly unlikely, we lazily build the worklist
1398 1398 // of such Nodes to go split.
1399 1399 if( !split_if_set )
1400 1400 split_if_set = new Node_List(area);
1401 1401 split_if_set->push(use);
1402 1402 }
1403 1403 if( use->is_Bool() ) {
1404 1404 if( !split_bool_set )
1405 1405 split_bool_set = new Node_List(area);
1406 1406 split_bool_set->push(use);
1407 1407 }
1408 1408 if( use->Opcode() == Op_CreateEx ) {
1409 1409 if( !split_cex_set )
1410 1410 split_cex_set = new Node_List(area);
1411 1411 split_cex_set->push(use);
1412 1412 }
1413 1413
1414 1414
1415 1415 // Get "block" use is in
1416 1416 uint idx = 0;
1417 1417 while( use->in(idx) != old ) idx++;
1418 1418 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1419 1419 assert( !loop->is_member( get_loop( prev ) ), "" );
1420 1420 Node *cfg = prev->_idx >= new_counter
1421 1421 ? prev->in(2)
1422 1422 : idom(prev);
1423 1423 if( use->is_Phi() ) // Phi use is in prior block
1424 1424 cfg = prev->in(idx); // NOT in block of Phi itself
1425 1425 if (cfg->is_top()) { // Use is dead?
1426 1426 _igvn.replace_input_of(use, idx, C->top());
1427 1427 continue;
1428 1428 }
1429 1429
1430 1430 while( !loop->is_member( get_loop( cfg ) ) ) {
1431 1431 prev = cfg;
1432 1432 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1433 1433 }
1434 1434 // If the use occurs after merging several exits from the loop, then
1435 1435 // old value must have dominated all those exits. Since the same old
1436 1436 // value was used on all those exits we did not need a Phi at this
1437 1437 // merge point. NOW we do need a Phi here. Each loop exit value
1438 1438 // is now merged with the peeled body exit; each exit gets its own
1439 1439 // private Phi and those Phis need to be merged here.
1440 1440 Node *phi;
1441 1441 if( prev->is_Region() ) {
1442 1442 if( idx == 0 ) { // Updating control edge?
1443 1443 phi = prev; // Just use existing control
1444 1444 } else { // Else need a new Phi
1445 1445 phi = PhiNode::make( prev, old );
1446 1446 // Now recursively fix up the new uses of old!
1447 1447 for( uint i = 1; i < prev->req(); i++ ) {
1448 1448 worklist.push(phi); // Onto worklist once for each 'old' input
1449 1449 }
1450 1450 }
1451 1451 } else {
1452 1452 // Get new RegionNode merging old and new loop exits
1453 1453 prev = old_new[prev->_idx];
1454 1454 assert( prev, "just made this in step 7" );
1455 1455 if( idx == 0 ) { // Updating control edge?
1456 1456 phi = prev; // Just use existing control
1457 1457 } else { // Else need a new Phi
1458 1458 // Make a new Phi merging data values properly
1459 1459 phi = PhiNode::make( prev, old );
1460 1460 phi->set_req( 1, nnn );
1461 1461 }
1462 1462 }
1463 1463 // If inserting a new Phi, check for prior hits
1464 1464 if( idx != 0 ) {
1465 1465 Node *hit = _igvn.hash_find_insert(phi);
1466 1466 if( hit == NULL ) {
1467 1467 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1468 1468 } else { // or
1469 1469 // Remove the new phi from the graph and use the hit
1470 1470 _igvn.remove_dead_node(phi);
1471 1471 phi = hit; // Use existing phi
1472 1472 }
1473 1473 set_ctrl(phi, prev);
1474 1474 }
1475 1475 // Make 'use' use the Phi instead of the old loop body exit value
1476 1476 _igvn.replace_input_of(use, idx, phi);
1477 1477 if( use->_idx >= new_counter ) { // If updating new phis
1478 1478 // Not needed for correctness, but prevents a weak assert
1479 1479 // in AddPNode from tripping (when we end up with different
1480 1480 // base & derived Phis that will become the same after
1481 1481 // IGVN does CSE).
1482 1482 Node *hit = _igvn.hash_find_insert(use);
1483 1483 if( hit ) // Go ahead and re-hash for hits.
1484 1484 _igvn.replace_node( use, hit );
1485 1485 }
1486 1486
1487 1487 // If 'use' was in the loop-exit block, it now needs to be sunk
1488 1488 // below the post-loop merge point.
1489 1489 sink_use( use, prev );
1490 1490 }
1491 1491 }
1492 1492 }
1493 1493
1494 1494 // Check for IFs that need splitting/cloning. Happens if an IF outside of
1495 1495 // the loop uses a condition set in the loop. The original IF probably
1496 1496 // takes control from one or more OLD Regions (which in turn get from NEW
1497 1497 // Regions). In any case, there will be a set of Phis for each merge point
1498 1498 // from the IF up to where the original BOOL def exists the loop.
1499 1499 if( split_if_set ) {
1500 1500 while( split_if_set->size() ) {
1501 1501 Node *iff = split_if_set->pop();
1502 1502 if( iff->in(1)->is_Phi() ) {
1503 1503 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop );
1504 1504 _igvn.replace_input_of(iff, 1, b);
1505 1505 }
1506 1506 }
1507 1507 }
1508 1508 if( split_bool_set ) {
1509 1509 while( split_bool_set->size() ) {
1510 1510 Node *b = split_bool_set->pop();
1511 1511 Node *phi = b->in(1);
1512 1512 assert( phi->is_Phi(), "" );
1513 1513 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop );
1514 1514 _igvn.replace_input_of(b, 1, cmp);
1515 1515 }
1516 1516 }
1517 1517 if( split_cex_set ) {
1518 1518 while( split_cex_set->size() ) {
1519 1519 Node *b = split_cex_set->pop();
1520 1520 assert( b->in(0)->is_Region(), "" );
1521 1521 assert( b->in(1)->is_Phi(), "" );
1522 1522 assert( b->in(0)->in(0) == b->in(1)->in(0), "" );
1523 1523 split_up( b, b->in(0), NULL );
1524 1524 }
1525 1525 }
1526 1526
1527 1527 }
1528 1528
1529 1529
1530 1530 //---------------------- stride_of_possible_iv -------------------------------------
1531 1531 // Looks for an iff/bool/comp with one operand of the compare
1532 1532 // being a cycle involving an add and a phi,
1533 1533 // with an optional truncation (left-shift followed by a right-shift)
1534 1534 // of the add. Returns zero if not an iv.
1535 1535 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
1536 1536 Node* trunc1 = NULL;
1537 1537 Node* trunc2 = NULL;
1538 1538 const TypeInt* ttype = NULL;
1539 1539 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
1540 1540 return 0;
1541 1541 }
1542 1542 BoolNode* bl = iff->in(1)->as_Bool();
1543 1543 Node* cmp = bl->in(1);
1544 1544 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) {
1545 1545 return 0;
1546 1546 }
1547 1547 // Must have an invariant operand
1548 1548 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
1549 1549 return 0;
1550 1550 }
1551 1551 Node* add2 = NULL;
1552 1552 Node* cmp1 = cmp->in(1);
1553 1553 if (cmp1->is_Phi()) {
1554 1554 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
1555 1555 Node* phi = cmp1;
1556 1556 for (uint i = 1; i < phi->req(); i++) {
1557 1557 Node* in = phi->in(i);
1558 1558 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
1559 1559 &trunc1, &trunc2, &ttype);
1560 1560 if (add && add->in(1) == phi) {
1561 1561 add2 = add->in(2);
1562 1562 break;
1563 1563 }
1564 1564 }
1565 1565 } else {
1566 1566 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
1567 1567 Node* addtrunc = cmp1;
1568 1568 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
1569 1569 &trunc1, &trunc2, &ttype);
1570 1570 if (add && add->in(1)->is_Phi()) {
1571 1571 Node* phi = add->in(1);
1572 1572 for (uint i = 1; i < phi->req(); i++) {
1573 1573 if (phi->in(i) == addtrunc) {
1574 1574 add2 = add->in(2);
1575 1575 break;
1576 1576 }
1577 1577 }
1578 1578 }
1579 1579 }
1580 1580 if (add2 != NULL) {
1581 1581 const TypeInt* add2t = _igvn.type(add2)->is_int();
1582 1582 if (add2t->is_con()) {
1583 1583 return add2t->get_con();
1584 1584 }
1585 1585 }
1586 1586 return 0;
1587 1587 }
1588 1588
1589 1589
1590 1590 //---------------------- stay_in_loop -------------------------------------
1591 1591 // Return the (unique) control output node that's in the loop (if it exists.)
1592 1592 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
1593 1593 Node* unique = NULL;
1594 1594 if (!n) return NULL;
1595 1595 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1596 1596 Node* use = n->fast_out(i);
1597 1597 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
1598 1598 if (unique != NULL) {
1599 1599 return NULL;
1600 1600 }
1601 1601 unique = use;
1602 1602 }
1603 1603 }
1604 1604 return unique;
1605 1605 }
1606 1606
1607 1607 //------------------------------ register_node -------------------------------------
1608 1608 // Utility to register node "n" with PhaseIdealLoop
1609 1609 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
1610 1610 _igvn.register_new_node_with_optimizer(n);
1611 1611 loop->_body.push(n);
1612 1612 if (n->is_CFG()) {
1613 1613 set_loop(n, loop);
1614 1614 set_idom(n, pred, ddepth);
1615 1615 } else {
1616 1616 set_ctrl(n, pred);
1617 1617 }
1618 1618 }
1619 1619
1620 1620 //------------------------------ proj_clone -------------------------------------
1621 1621 // Utility to create an if-projection
1622 1622 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
1623 1623 ProjNode* c = p->clone()->as_Proj();
1624 1624 c->set_req(0, iff);
1625 1625 return c;
1626 1626 }
1627 1627
1628 1628 //------------------------------ short_circuit_if -------------------------------------
1629 1629 // Force the iff control output to be the live_proj
1630 1630 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
1631 1631 int proj_con = live_proj->_con;
1632 1632 assert(proj_con == 0 || proj_con == 1, "false or true projection");
1633 1633 Node *con = _igvn.intcon(proj_con);
1634 1634 set_ctrl(con, C->root());
1635 1635 if (iff) {
1636 1636 iff->set_req(1, con);
1637 1637 }
1638 1638 return con;
1639 1639 }
1640 1640
1641 1641 //------------------------------ insert_if_before_proj -------------------------------------
1642 1642 // Insert a new if before an if projection (* - new node)
1643 1643 //
1644 1644 // before
1645 1645 // if(test)
1646 1646 // / \
1647 1647 // v v
1648 1648 // other-proj proj (arg)
1649 1649 //
1650 1650 // after
1651 1651 // if(test)
1652 1652 // / \
1653 1653 // / v
1654 1654 // | * proj-clone
1655 1655 // v |
1656 1656 // other-proj v
1657 1657 // * new_if(relop(cmp[IU](left,right)))
1658 1658 // / \
1659 1659 // v v
1660 1660 // * new-proj proj
1661 1661 // (returned)
1662 1662 //
1663 1663 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
1664 1664 IfNode* iff = proj->in(0)->as_If();
1665 1665 IdealLoopTree *loop = get_loop(proj);
1666 1666 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1667 1667 int ddepth = dom_depth(proj);
1668 1668
1669 1669 _igvn.rehash_node_delayed(iff);
1670 1670 _igvn.rehash_node_delayed(proj);
1671 1671
1672 1672 proj->set_req(0, NULL); // temporary disconnect
1673 1673 ProjNode* proj2 = proj_clone(proj, iff);
1674 1674 register_node(proj2, loop, iff, ddepth);
1675 1675
1676 1676 Node* cmp = Signed ? (Node*) new (C)CmpINode(left, right) : (Node*) new (C)CmpUNode(left, right);
1677 1677 register_node(cmp, loop, proj2, ddepth);
1678 1678
1679 1679 BoolNode* bol = new (C)BoolNode(cmp, relop);
1680 1680 register_node(bol, loop, proj2, ddepth);
1681 1681
1682 1682 IfNode* new_if = new (C)IfNode(proj2, bol, iff->_prob, iff->_fcnt);
1683 1683 register_node(new_if, loop, proj2, ddepth);
1684 1684
1685 1685 proj->set_req(0, new_if); // reattach
1686 1686 set_idom(proj, new_if, ddepth);
1687 1687
1688 1688 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
1689 1689 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
1690 1690
1691 1691 return new_exit;
1692 1692 }
1693 1693
1694 1694 //------------------------------ insert_region_before_proj -------------------------------------
1695 1695 // Insert a region before an if projection (* - new node)
1696 1696 //
1697 1697 // before
1698 1698 // if(test)
1699 1699 // / |
1700 1700 // v |
1701 1701 // proj v
1702 1702 // other-proj
1703 1703 //
1704 1704 // after
1705 1705 // if(test)
1706 1706 // / |
1707 1707 // v |
1708 1708 // * proj-clone v
1709 1709 // | other-proj
1710 1710 // v
1711 1711 // * new-region
1712 1712 // |
1713 1713 // v
1714 1714 // * dum_if
1715 1715 // / \
1716 1716 // v \
1717 1717 // * dum-proj v
1718 1718 // proj
1719 1719 //
1720 1720 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
1721 1721 IfNode* iff = proj->in(0)->as_If();
1722 1722 IdealLoopTree *loop = get_loop(proj);
1723 1723 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1724 1724 int ddepth = dom_depth(proj);
1725 1725
1726 1726 _igvn.rehash_node_delayed(iff);
1727 1727 _igvn.rehash_node_delayed(proj);
1728 1728
1729 1729 proj->set_req(0, NULL); // temporary disconnect
1730 1730 ProjNode* proj2 = proj_clone(proj, iff);
1731 1731 register_node(proj2, loop, iff, ddepth);
1732 1732
1733 1733 RegionNode* reg = new (C)RegionNode(2);
1734 1734 reg->set_req(1, proj2);
1735 1735 register_node(reg, loop, iff, ddepth);
1736 1736
1737 1737 IfNode* dum_if = new (C)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
1738 1738 register_node(dum_if, loop, reg, ddepth);
1739 1739
1740 1740 proj->set_req(0, dum_if); // reattach
1741 1741 set_idom(proj, dum_if, ddepth);
1742 1742
1743 1743 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
1744 1744 register_node(dum_proj, loop, dum_if, ddepth);
1745 1745
1746 1746 return reg;
1747 1747 }
1748 1748
1749 1749 //------------------------------ insert_cmpi_loop_exit -------------------------------------
1750 1750 // Clone a signed compare loop exit from an unsigned compare and
1751 1751 // insert it before the unsigned cmp on the stay-in-loop path.
1752 1752 // All new nodes inserted in the dominator tree between the original
1753 1753 // if and it's projections. The original if test is replaced with
1754 1754 // a constant to force the stay-in-loop path.
1755 1755 //
1756 1756 // This is done to make sure that the original if and it's projections
1757 1757 // still dominate the same set of control nodes, that the ctrl() relation
1758 1758 // from data nodes to them is preserved, and that their loop nesting is
1759 1759 // preserved.
1760 1760 //
1761 1761 // before
1762 1762 // if(i <u limit) unsigned compare loop exit
1763 1763 // / |
1764 1764 // v v
1765 1765 // exit-proj stay-in-loop-proj
1766 1766 //
1767 1767 // after
1768 1768 // if(stay-in-loop-const) original if
1769 1769 // / |
1770 1770 // / v
1771 1771 // / if(i < limit) new signed test
1772 1772 // / / |
1773 1773 // / / v
1774 1774 // / / if(i <u limit) new cloned unsigned test
1775 1775 // / / / |
1776 1776 // v v v |
1777 1777 // region |
1778 1778 // | |
1779 1779 // dum-if |
1780 1780 // / | |
1781 1781 // ether | |
1782 1782 // v v
1783 1783 // exit-proj stay-in-loop-proj
1784 1784 //
1785 1785 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
1786 1786 const bool Signed = true;
1787 1787 const bool Unsigned = false;
1788 1788
1789 1789 BoolNode* bol = if_cmpu->in(1)->as_Bool();
1790 1790 if (bol->_test._test != BoolTest::lt) return NULL;
1791 1791 CmpNode* cmpu = bol->in(1)->as_Cmp();
1792 1792 if (cmpu->Opcode() != Op_CmpU) return NULL;
1793 1793 int stride = stride_of_possible_iv(if_cmpu);
1794 1794 if (stride == 0) return NULL;
1795 1795
1796 1796 ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj();
1797 1797 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
1798 1798
1799 1799 Node* limit = NULL;
1800 1800 if (stride > 0) {
1801 1801 limit = cmpu->in(2);
1802 1802 } else {
1803 1803 limit = _igvn.makecon(TypeInt::ZERO);
1804 1804 set_ctrl(limit, C->root());
1805 1805 }
1806 1806 // Create a new region on the exit path
1807 1807 RegionNode* reg = insert_region_before_proj(lp_exit);
1808 1808
1809 1809 // Clone the if-cmpu-true-false using a signed compare
1810 1810 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
1811 1811 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
1812 1812 reg->add_req(cmpi_exit);
1813 1813
1814 1814 // Clone the if-cmpu-true-false
1815 1815 BoolTest::mask rel_u = bol->_test._test;
1816 1816 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
1817 1817 reg->add_req(cmpu_exit);
1818 1818
1819 1819 // Force original if to stay in loop.
1820 1820 short_circuit_if(if_cmpu, lp_continue);
1821 1821
1822 1822 return cmpi_exit->in(0)->as_If();
1823 1823 }
1824 1824
1825 1825 //------------------------------ remove_cmpi_loop_exit -------------------------------------
1826 1826 // Remove a previously inserted signed compare loop exit.
1827 1827 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
1828 1828 Node* lp_proj = stay_in_loop(if_cmp, loop);
1829 1829 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
1830 1830 stay_in_loop(lp_proj, loop)->is_If() &&
1831 1831 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
1832 1832 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
1833 1833 set_ctrl(con, C->root());
1834 1834 if_cmp->set_req(1, con);
1835 1835 }
1836 1836
1837 1837 //------------------------------ scheduled_nodelist -------------------------------------
1838 1838 // Create a post order schedule of nodes that are in the
1839 1839 // "member" set. The list is returned in "sched".
1840 1840 // The first node in "sched" is the loop head, followed by
1841 1841 // nodes which have no inputs in the "member" set, and then
1842 1842 // followed by the nodes that have an immediate input dependence
1843 1843 // on a node in "sched".
1844 1844 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
1845 1845
1846 1846 assert(member.test(loop->_head->_idx), "loop head must be in member set");
1847 1847 Arena *a = Thread::current()->resource_area();
1848 1848 VectorSet visited(a);
1849 1849 Node_Stack nstack(a, loop->_body.size());
1850 1850
1851 1851 Node* n = loop->_head; // top of stack is cached in "n"
1852 1852 uint idx = 0;
1853 1853 visited.set(n->_idx);
1854 1854
1855 1855 // Initially push all with no inputs from within member set
1856 1856 for(uint i = 0; i < loop->_body.size(); i++ ) {
1857 1857 Node *elt = loop->_body.at(i);
1858 1858 if (member.test(elt->_idx)) {
1859 1859 bool found = false;
1860 1860 for (uint j = 0; j < elt->req(); j++) {
1861 1861 Node* def = elt->in(j);
1862 1862 if (def && member.test(def->_idx) && def != elt) {
1863 1863 found = true;
1864 1864 break;
1865 1865 }
1866 1866 }
1867 1867 if (!found && elt != loop->_head) {
1868 1868 nstack.push(n, idx);
1869 1869 n = elt;
1870 1870 assert(!visited.test(n->_idx), "not seen yet");
1871 1871 visited.set(n->_idx);
1872 1872 }
1873 1873 }
1874 1874 }
1875 1875
1876 1876 // traverse out's that are in the member set
1877 1877 while (true) {
1878 1878 if (idx < n->outcnt()) {
1879 1879 Node* use = n->raw_out(idx);
1880 1880 idx++;
1881 1881 if (!visited.test_set(use->_idx)) {
1882 1882 if (member.test(use->_idx)) {
1883 1883 nstack.push(n, idx);
1884 1884 n = use;
1885 1885 idx = 0;
1886 1886 }
1887 1887 }
1888 1888 } else {
1889 1889 // All outputs processed
1890 1890 sched.push(n);
1891 1891 if (nstack.is_empty()) break;
1892 1892 n = nstack.node();
1893 1893 idx = nstack.index();
1894 1894 nstack.pop();
1895 1895 }
1896 1896 }
1897 1897 }
1898 1898
1899 1899
1900 1900 //------------------------------ has_use_in_set -------------------------------------
1901 1901 // Has a use in the vector set
1902 1902 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
1903 1903 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1904 1904 Node* use = n->fast_out(j);
1905 1905 if (vset.test(use->_idx)) {
1906 1906 return true;
1907 1907 }
1908 1908 }
1909 1909 return false;
1910 1910 }
1911 1911
1912 1912
1913 1913 //------------------------------ has_use_internal_to_set -------------------------------------
1914 1914 // Has use internal to the vector set (ie. not in a phi at the loop head)
1915 1915 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
1916 1916 Node* head = loop->_head;
1917 1917 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1918 1918 Node* use = n->fast_out(j);
↓ open down ↓ |
1918 lines elided |
↑ open up ↑ |
1919 1919 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
1920 1920 return true;
1921 1921 }
1922 1922 }
1923 1923 return false;
1924 1924 }
1925 1925
1926 1926
1927 1927 //------------------------------ clone_for_use_outside_loop -------------------------------------
1928 1928 // clone "n" for uses that are outside of loop
1929 -void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
1930 -
1929 +int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
1930 + int cloned = 0;
1931 1931 assert(worklist.size() == 0, "should be empty");
1932 1932 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1933 1933 Node* use = n->fast_out(j);
1934 1934 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
1935 1935 worklist.push(use);
1936 1936 }
1937 1937 }
1938 1938 while( worklist.size() ) {
1939 1939 Node *use = worklist.pop();
1940 1940 if (!has_node(use) || use->in(0) == C->top()) continue;
1941 1941 uint j;
1942 1942 for (j = 0; j < use->req(); j++) {
1943 1943 if (use->in(j) == n) break;
1944 1944 }
1945 1945 assert(j < use->req(), "must be there");
1946 1946
1947 1947 // clone "n" and insert it between the inputs of "n" and the use outside the loop
1948 1948 Node* n_clone = n->clone();
1949 1949 _igvn.replace_input_of(use, j, n_clone);
1950 + cloned++;
1950 1951 Node* use_c;
1951 1952 if (!use->is_Phi()) {
1952 1953 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
1953 1954 } else {
1954 1955 // Use in a phi is considered a use in the associated predecessor block
1955 1956 use_c = use->in(0)->in(j);
1956 1957 }
1957 1958 set_ctrl(n_clone, use_c);
1958 1959 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
1959 1960 get_loop(use_c)->_body.push(n_clone);
1960 1961 _igvn.register_new_node_with_optimizer(n_clone);
1961 1962 #if !defined(PRODUCT)
1962 1963 if (TracePartialPeeling) {
1963 1964 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
1964 1965 }
1965 1966 #endif
1966 1967 }
1968 + return cloned;
1967 1969 }
1968 1970
1969 1971
1970 1972 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
1971 1973 // clone "n" for special uses that are in the not_peeled region.
1972 1974 // If these def-uses occur in separate blocks, the code generator
1973 1975 // marks the method as not compilable. For example, if a "BoolNode"
1974 1976 // is in a different basic block than the "IfNode" that uses it, then
1975 1977 // the compilation is aborted in the code generator.
1976 1978 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1977 1979 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
1978 1980 if (n->is_Phi() || n->is_Load()) {
1979 1981 return;
1980 1982 }
1981 1983 assert(worklist.size() == 0, "should be empty");
1982 1984 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1983 1985 Node* use = n->fast_out(j);
1984 1986 if ( not_peel.test(use->_idx) &&
1985 1987 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
1986 1988 use->in(1) == n) {
1987 1989 worklist.push(use);
1988 1990 }
1989 1991 }
1990 1992 if (worklist.size() > 0) {
1991 1993 // clone "n" and insert it between inputs of "n" and the use
1992 1994 Node* n_clone = n->clone();
1993 1995 loop->_body.push(n_clone);
1994 1996 _igvn.register_new_node_with_optimizer(n_clone);
1995 1997 set_ctrl(n_clone, get_ctrl(n));
1996 1998 sink_list.push(n_clone);
1997 1999 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
1998 2000 #if !defined(PRODUCT)
1999 2001 if (TracePartialPeeling) {
2000 2002 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2001 2003 }
2002 2004 #endif
2003 2005 while( worklist.size() ) {
2004 2006 Node *use = worklist.pop();
2005 2007 _igvn.rehash_node_delayed(use);
2006 2008 for (uint j = 1; j < use->req(); j++) {
2007 2009 if (use->in(j) == n) {
2008 2010 use->set_req(j, n_clone);
2009 2011 }
2010 2012 }
2011 2013 }
2012 2014 }
2013 2015 }
2014 2016
2015 2017
2016 2018 //------------------------------ insert_phi_for_loop -------------------------------------
2017 2019 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2018 2020 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2019 2021 Node *phi = PhiNode::make(lp, back_edge_val);
2020 2022 phi->set_req(LoopNode::EntryControl, lp_entry_val);
2021 2023 // Use existing phi if it already exists
2022 2024 Node *hit = _igvn.hash_find_insert(phi);
2023 2025 if( hit == NULL ) {
2024 2026 _igvn.register_new_node_with_optimizer(phi);
2025 2027 set_ctrl(phi, lp);
2026 2028 } else {
2027 2029 // Remove the new phi from the graph and use the hit
2028 2030 _igvn.remove_dead_node(phi);
2029 2031 phi = hit;
2030 2032 }
2031 2033 _igvn.replace_input_of(use, idx, phi);
2032 2034 }
2033 2035
2034 2036 #ifdef ASSERT
2035 2037 //------------------------------ is_valid_loop_partition -------------------------------------
2036 2038 // Validate the loop partition sets: peel and not_peel
2037 2039 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2038 2040 VectorSet& not_peel ) {
2039 2041 uint i;
2040 2042 // Check that peel_list entries are in the peel set
2041 2043 for (i = 0; i < peel_list.size(); i++) {
2042 2044 if (!peel.test(peel_list.at(i)->_idx)) {
2043 2045 return false;
2044 2046 }
2045 2047 }
2046 2048 // Check at loop members are in one of peel set or not_peel set
2047 2049 for (i = 0; i < loop->_body.size(); i++ ) {
2048 2050 Node *def = loop->_body.at(i);
2049 2051 uint di = def->_idx;
2050 2052 // Check that peel set elements are in peel_list
2051 2053 if (peel.test(di)) {
2052 2054 if (not_peel.test(di)) {
2053 2055 return false;
2054 2056 }
2055 2057 // Must be in peel_list also
2056 2058 bool found = false;
2057 2059 for (uint j = 0; j < peel_list.size(); j++) {
2058 2060 if (peel_list.at(j)->_idx == di) {
2059 2061 found = true;
2060 2062 break;
2061 2063 }
2062 2064 }
2063 2065 if (!found) {
2064 2066 return false;
2065 2067 }
2066 2068 } else if (not_peel.test(di)) {
2067 2069 if (peel.test(di)) {
2068 2070 return false;
2069 2071 }
2070 2072 } else {
2071 2073 return false;
2072 2074 }
2073 2075 }
2074 2076 return true;
2075 2077 }
2076 2078
2077 2079 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2078 2080 // Ensure a use outside of loop is of the right form
2079 2081 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2080 2082 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2081 2083 return (use->is_Phi() &&
2082 2084 use_c->is_Region() && use_c->req() == 3 &&
2083 2085 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2084 2086 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2085 2087 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2086 2088 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2087 2089 }
2088 2090
2089 2091 //------------------------------ is_valid_clone_loop_form -------------------------------------
2090 2092 // Ensure that all uses outside of loop are of the right form
2091 2093 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2092 2094 uint orig_exit_idx, uint clone_exit_idx) {
2093 2095 uint len = peel_list.size();
2094 2096 for (uint i = 0; i < len; i++) {
2095 2097 Node *def = peel_list.at(i);
2096 2098
2097 2099 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2098 2100 Node *use = def->fast_out(j);
2099 2101 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2100 2102 if (!loop->is_member(get_loop(use_c))) {
2101 2103 // use is not in the loop, check for correct structure
2102 2104 if (use->in(0) == def) {
2103 2105 // Okay
2104 2106 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2105 2107 return false;
2106 2108 }
2107 2109 }
2108 2110 }
2109 2111 }
2110 2112 return true;
2111 2113 }
2112 2114 #endif
2113 2115
2114 2116 //------------------------------ partial_peel -------------------------------------
2115 2117 // Partially peel (aka loop rotation) the top portion of a loop (called
2116 2118 // the peel section below) by cloning it and placing one copy just before
2117 2119 // the new loop head and the other copy at the bottom of the new loop.
2118 2120 //
2119 2121 // before after where it came from
2120 2122 //
2121 2123 // stmt1 stmt1
2122 2124 // loop: stmt2 clone
2123 2125 // stmt2 if condA goto exitA clone
2124 2126 // if condA goto exitA new_loop: new
2125 2127 // stmt3 stmt3 clone
2126 2128 // if !condB goto loop if condB goto exitB clone
2127 2129 // exitB: stmt2 orig
2128 2130 // stmt4 if !condA goto new_loop orig
2129 2131 // exitA: goto exitA
2130 2132 // exitB:
2131 2133 // stmt4
2132 2134 // exitA:
2133 2135 //
2134 2136 // Step 1: find the cut point: an exit test on probable
2135 2137 // induction variable.
2136 2138 // Step 2: schedule (with cloning) operations in the peel
2137 2139 // section that can be executed after the cut into
2138 2140 // the section that is not peeled. This may need
2139 2141 // to clone operations into exit blocks. For
2140 2142 // instance, a reference to A[i] in the not-peel
2141 2143 // section and a reference to B[i] in an exit block
2142 2144 // may cause a left-shift of i by 2 to be placed
2143 2145 // in the peel block. This step will clone the left
2144 2146 // shift into the exit block and sink the left shift
2145 2147 // from the peel to the not-peel section.
2146 2148 // Step 3: clone the loop, retarget the control, and insert
2147 2149 // phis for values that are live across the new loop
2148 2150 // head. This is very dependent on the graph structure
2149 2151 // from clone_loop. It creates region nodes for
2150 2152 // exit control and associated phi nodes for values
2151 2153 // flow out of the loop through that exit. The region
2152 2154 // node is dominated by the clone's control projection.
2153 2155 // So the clone's peel section is placed before the
2154 2156 // new loop head, and the clone's not-peel section is
2155 2157 // forms the top part of the new loop. The original
2156 2158 // peel section forms the tail of the new loop.
2157 2159 // Step 4: update the dominator tree and recompute the
2158 2160 // dominator depth.
2159 2161 //
2160 2162 // orig
2161 2163 //
2162 2164 // stmt1
2163 2165 // |
2164 2166 // v
2165 2167 // loop predicate
2166 2168 // |
2167 2169 // v
2168 2170 // loop<----+
2169 2171 // | |
2170 2172 // stmt2 |
2171 2173 // | |
2172 2174 // v |
2173 2175 // ifA |
2174 2176 // / | |
2175 2177 // v v |
2176 2178 // false true ^ <-- last_peel
2177 2179 // / | |
2178 2180 // / ===|==cut |
2179 2181 // / stmt3 | <-- first_not_peel
2180 2182 // / | |
2181 2183 // | v |
2182 2184 // v ifB |
2183 2185 // exitA: / \ |
2184 2186 // / \ |
2185 2187 // v v |
2186 2188 // false true |
2187 2189 // / \ |
2188 2190 // / ----+
2189 2191 // |
2190 2192 // v
2191 2193 // exitB:
2192 2194 // stmt4
2193 2195 //
2194 2196 //
2195 2197 // after clone loop
2196 2198 //
2197 2199 // stmt1
2198 2200 // |
2199 2201 // v
2200 2202 // loop predicate
2201 2203 // / \
2202 2204 // clone / \ orig
2203 2205 // / \
2204 2206 // / \
2205 2207 // v v
2206 2208 // +---->loop loop<----+
2207 2209 // | | | |
2208 2210 // | stmt2 stmt2 |
2209 2211 // | | | |
2210 2212 // | v v |
2211 2213 // | ifA ifA |
2212 2214 // | | \ / | |
2213 2215 // | v v v v |
2214 2216 // ^ true false false true ^ <-- last_peel
2215 2217 // | | ^ \ / | |
2216 2218 // | cut==|== \ \ / ===|==cut |
2217 2219 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2218 2220 // | | dom | | | |
2219 2221 // | v \ 1v v2 v |
2220 2222 // | ifB regionA ifB |
2221 2223 // | / \ | / \ |
2222 2224 // | / \ v / \ |
2223 2225 // | v v exitA: v v |
2224 2226 // | true false false true |
2225 2227 // | / ^ \ / \ |
2226 2228 // +---- \ \ / ----+
2227 2229 // dom \ /
2228 2230 // \ 1v v2
2229 2231 // regionB
2230 2232 // |
2231 2233 // v
2232 2234 // exitB:
2233 2235 // stmt4
2234 2236 //
2235 2237 //
2236 2238 // after partial peel
2237 2239 //
2238 2240 // stmt1
2239 2241 // |
2240 2242 // v
2241 2243 // loop predicate
2242 2244 // /
2243 2245 // clone / orig
2244 2246 // / TOP
2245 2247 // / \
2246 2248 // v v
2247 2249 // TOP->loop loop----+
2248 2250 // | | |
2249 2251 // stmt2 stmt2 |
2250 2252 // | | |
2251 2253 // v v |
2252 2254 // ifA ifA |
2253 2255 // | \ / | |
2254 2256 // v v v v |
2255 2257 // true false false true | <-- last_peel
2256 2258 // | ^ \ / +------|---+
2257 2259 // +->newloop \ \ / === ==cut | |
2258 2260 // | stmt3 \ \ / TOP | |
2259 2261 // | | dom | | stmt3 | | <-- first_not_peel
2260 2262 // | v \ 1v v2 v | |
2261 2263 // | ifB regionA ifB ^ v
2262 2264 // | / \ | / \ | |
2263 2265 // | / \ v / \ | |
2264 2266 // | v v exitA: v v | |
2265 2267 // | true false false true | |
2266 2268 // | / ^ \ / \ | |
2267 2269 // | | \ \ / v | |
2268 2270 // | | dom \ / TOP | |
2269 2271 // | | \ 1v v2 | |
2270 2272 // ^ v regionB | |
2271 2273 // | | | | |
2272 2274 // | | v ^ v
2273 2275 // | | exitB: | |
2274 2276 // | | stmt4 | |
2275 2277 // | +------------>-----------------+ |
2276 2278 // | |
2277 2279 // +-----------------<---------------------+
2278 2280 //
2279 2281 //
2280 2282 // final graph
2281 2283 //
2282 2284 // stmt1
2283 2285 // |
2284 2286 // v
2285 2287 // loop predicate
2286 2288 // |
2287 2289 // v
2288 2290 // stmt2 clone
2289 2291 // |
2290 2292 // v
2291 2293 // ........> ifA clone
2292 2294 // : / |
2293 2295 // dom / |
2294 2296 // : v v
2295 2297 // : false true
2296 2298 // : | |
2297 2299 // : | v
2298 2300 // : | newloop<-----+
2299 2301 // : | | |
2300 2302 // : | stmt3 clone |
2301 2303 // : | | |
2302 2304 // : | v |
2303 2305 // : | ifB |
2304 2306 // : | / \ |
2305 2307 // : | v v |
2306 2308 // : | false true |
2307 2309 // : | | | |
2308 2310 // : | v stmt2 |
2309 2311 // : | exitB: | |
2310 2312 // : | stmt4 v |
2311 2313 // : | ifA orig |
2312 2314 // : | / \ |
2313 2315 // : | / \ |
2314 2316 // : | v v |
2315 2317 // : | false true |
2316 2318 // : | / \ |
2317 2319 // : v v -----+
2318 2320 // RegionA
2319 2321 // |
2320 2322 // v
2321 2323 // exitA
2322 2324 //
2323 2325 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
2324 2326
2325 2327 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
2326 2328 if (!loop->_head->is_Loop()) {
2327 2329 return false; }
2328 2330
2329 2331 LoopNode *head = loop->_head->as_Loop();
2330 2332
2331 2333 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
2332 2334 return false;
2333 2335 }
2334 2336
2335 2337 // Check for complex exit control
2336 2338 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
2337 2339 Node *n = loop->_body.at(ii);
2338 2340 int opc = n->Opcode();
2339 2341 if (n->is_Call() ||
2340 2342 opc == Op_Catch ||
2341 2343 opc == Op_CatchProj ||
2342 2344 opc == Op_Jump ||
2343 2345 opc == Op_JumpProj) {
2344 2346 #if !defined(PRODUCT)
2345 2347 if (TracePartialPeeling) {
2346 2348 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
2347 2349 }
2348 2350 #endif
2349 2351 return false;
2350 2352 }
2351 2353 }
2352 2354
2353 2355 int dd = dom_depth(head);
2354 2356
2355 2357 // Step 1: find cut point
2356 2358
2357 2359 // Walk up dominators to loop head looking for first loop exit
2358 2360 // which is executed on every path thru loop.
2359 2361 IfNode *peel_if = NULL;
2360 2362 IfNode *peel_if_cmpu = NULL;
2361 2363
2362 2364 Node *iff = loop->tail();
2363 2365 while( iff != head ) {
2364 2366 if( iff->is_If() ) {
2365 2367 Node *ctrl = get_ctrl(iff->in(1));
2366 2368 if (ctrl->is_top()) return false; // Dead test on live IF.
2367 2369 // If loop-varying exit-test, check for induction variable
2368 2370 if( loop->is_member(get_loop(ctrl)) &&
2369 2371 loop->is_loop_exit(iff) &&
2370 2372 is_possible_iv_test(iff)) {
2371 2373 Node* cmp = iff->in(1)->in(1);
2372 2374 if (cmp->Opcode() == Op_CmpI) {
2373 2375 peel_if = iff->as_If();
2374 2376 } else {
2375 2377 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
2376 2378 peel_if_cmpu = iff->as_If();
2377 2379 }
2378 2380 }
2379 2381 }
2380 2382 iff = idom(iff);
2381 2383 }
2382 2384 // Prefer signed compare over unsigned compare.
2383 2385 IfNode* new_peel_if = NULL;
2384 2386 if (peel_if == NULL) {
2385 2387 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
2386 2388 return false; // No peel point found
2387 2389 }
2388 2390 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
2389 2391 if (new_peel_if == NULL) {
2390 2392 return false; // No peel point found
2391 2393 }
2392 2394 peel_if = new_peel_if;
2393 2395 }
2394 2396 Node* last_peel = stay_in_loop(peel_if, loop);
2395 2397 Node* first_not_peeled = stay_in_loop(last_peel, loop);
2396 2398 if (first_not_peeled == NULL || first_not_peeled == head) {
2397 2399 return false;
2398 2400 }
2399 2401
2400 2402 #if !defined(PRODUCT)
2401 2403 if (TraceLoopOpts) {
2402 2404 tty->print("PartialPeel ");
2403 2405 loop->dump_head();
2404 2406 }
2405 2407
2406 2408 if (TracePartialPeeling) {
2407 2409 tty->print_cr("before partial peel one iteration");
2408 2410 Node_List wl;
2409 2411 Node* t = head->in(2);
2410 2412 while (true) {
2411 2413 wl.push(t);
2412 2414 if (t == head) break;
2413 2415 t = idom(t);
2414 2416 }
2415 2417 while (wl.size() > 0) {
2416 2418 Node* tt = wl.pop();
2417 2419 tt->dump();
2418 2420 if (tt == last_peel) tty->print_cr("-- cut --");
2419 2421 }
2420 2422 }
2421 2423 #endif
2422 2424 ResourceArea *area = Thread::current()->resource_area();
2423 2425 VectorSet peel(area);
2424 2426 VectorSet not_peel(area);
2425 2427 Node_List peel_list(area);
2426 2428 Node_List worklist(area);
2427 2429 Node_List sink_list(area);
2428 2430
2429 2431 // Set of cfg nodes to peel are those that are executable from
2430 2432 // the head through last_peel.
2431 2433 assert(worklist.size() == 0, "should be empty");
2432 2434 worklist.push(head);
2433 2435 peel.set(head->_idx);
2434 2436 while (worklist.size() > 0) {
2435 2437 Node *n = worklist.pop();
2436 2438 if (n != last_peel) {
2437 2439 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2438 2440 Node* use = n->fast_out(j);
2439 2441 if (use->is_CFG() &&
2440 2442 loop->is_member(get_loop(use)) &&
2441 2443 !peel.test_set(use->_idx)) {
2442 2444 worklist.push(use);
2443 2445 }
2444 2446 }
2445 2447 }
2446 2448 }
2447 2449
2448 2450 // Set of non-cfg nodes to peel are those that are control
2449 2451 // dependent on the cfg nodes.
2450 2452 uint i;
2451 2453 for(i = 0; i < loop->_body.size(); i++ ) {
2452 2454 Node *n = loop->_body.at(i);
2453 2455 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
2454 2456 if (peel.test(n_c->_idx)) {
2455 2457 peel.set(n->_idx);
2456 2458 } else {
2457 2459 not_peel.set(n->_idx);
2458 2460 }
2459 2461 }
2460 2462
2461 2463 // Step 2: move operations from the peeled section down into the
2462 2464 // not-peeled section
2463 2465
2464 2466 // Get a post order schedule of nodes in the peel region
2465 2467 // Result in right-most operand.
2466 2468 scheduled_nodelist(loop, peel, peel_list );
2467 2469
2468 2470 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2469 2471
2470 2472 // For future check for too many new phis
2471 2473 uint old_phi_cnt = 0;
2472 2474 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2473 2475 Node* use = head->fast_out(j);
2474 2476 if (use->is_Phi()) old_phi_cnt++;
↓ open down ↓ |
498 lines elided |
↑ open up ↑ |
2475 2477 }
2476 2478
2477 2479 #if !defined(PRODUCT)
2478 2480 if (TracePartialPeeling) {
2479 2481 tty->print_cr("\npeeled list");
2480 2482 }
2481 2483 #endif
2482 2484
2483 2485 // Evacuate nodes in peel region into the not_peeled region if possible
2484 2486 uint new_phi_cnt = 0;
2487 + uint cloned_for_outside_use = 0;
2485 2488 for (i = 0; i < peel_list.size();) {
2486 2489 Node* n = peel_list.at(i);
2487 2490 #if !defined(PRODUCT)
2488 2491 if (TracePartialPeeling) n->dump();
2489 2492 #endif
2490 2493 bool incr = true;
2491 2494 if ( !n->is_CFG() ) {
2492 2495
2493 2496 if ( has_use_in_set(n, not_peel) ) {
2494 2497
2495 2498 // If not used internal to the peeled region,
2496 2499 // move "n" from peeled to not_peeled region.
2497 2500
2498 2501 if ( !has_use_internal_to_set(n, peel, loop) ) {
2499 2502
2500 2503 // if not pinned and not a load (which maybe anti-dependent on a store)
2501 2504 // and not a CMove (Matcher expects only bool->cmove).
2502 2505 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
2503 - clone_for_use_outside_loop( loop, n, worklist );
2504 -
2506 + cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist );
2505 2507 sink_list.push(n);
2506 2508 peel >>= n->_idx; // delete n from peel set.
2507 2509 not_peel <<= n->_idx; // add n to not_peel set.
2508 2510 peel_list.remove(i);
2509 2511 incr = false;
2510 2512 #if !defined(PRODUCT)
2511 2513 if (TracePartialPeeling) {
2512 2514 tty->print_cr("sink to not_peeled region: %d newbb: %d",
2513 2515 n->_idx, get_ctrl(n)->_idx);
2514 2516 }
2515 2517 #endif
2516 2518 }
2517 2519 } else {
2518 2520 // Otherwise check for special def-use cases that span
2519 2521 // the peel/not_peel boundary such as bool->if
2520 2522 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
2521 2523 new_phi_cnt++;
2522 2524 }
2523 2525 }
2524 2526 }
2525 2527 if (incr) i++;
2526 2528 }
2527 2529
2528 2530 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
2529 2531 #if !defined(PRODUCT)
2530 2532 if (TracePartialPeeling) {
↓ open down ↓ |
16 lines elided |
↑ open up ↑ |
2531 2533 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
2532 2534 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
2533 2535 }
2534 2536 #endif
2535 2537 if (new_peel_if != NULL) {
2536 2538 remove_cmpi_loop_exit(new_peel_if, loop);
2537 2539 }
2538 2540 // Inhibit more partial peeling on this loop
2539 2541 assert(!head->is_partial_peel_loop(), "not partial peeled");
2540 2542 head->mark_partial_peel_failed();
2543 + if (cloned_for_outside_use > 0) {
2544 + // Terminate this round of loop opts because
2545 + // the graph outside this loop was changed.
2546 + C->set_major_progress();
2547 + return true;
2548 + }
2541 2549 return false;
2542 2550 }
2543 2551
2544 2552 // Step 3: clone loop, retarget control, and insert new phis
2545 2553
2546 2554 // Create new loop head for new phis and to hang
2547 2555 // the nodes being moved (sinked) from the peel region.
2548 2556 LoopNode* new_head = new (C) LoopNode(last_peel, last_peel);
2549 2557 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
2550 2558 _igvn.register_new_node_with_optimizer(new_head);
2551 2559 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
2552 2560 first_not_peeled->set_req(0, new_head);
2553 2561 set_loop(new_head, loop);
2554 2562 loop->_body.push(new_head);
2555 2563 not_peel.set(new_head->_idx);
2556 2564 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
2557 2565 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
2558 2566
2559 2567 while (sink_list.size() > 0) {
2560 2568 Node* n = sink_list.pop();
2561 2569 set_ctrl(n, new_head);
2562 2570 }
2563 2571
2564 2572 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2565 2573
2566 2574 clone_loop( loop, old_new, dd );
2567 2575
2568 2576 const uint clone_exit_idx = 1;
2569 2577 const uint orig_exit_idx = 2;
2570 2578 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
2571 2579
2572 2580 Node* head_clone = old_new[head->_idx];
2573 2581 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
2574 2582 Node* orig_tail_clone = head_clone->in(2);
2575 2583
2576 2584 // Add phi if "def" node is in peel set and "use" is not
2577 2585
2578 2586 for(i = 0; i < peel_list.size(); i++ ) {
2579 2587 Node *def = peel_list.at(i);
2580 2588 if (!def->is_CFG()) {
2581 2589 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2582 2590 Node *use = def->fast_out(j);
2583 2591 if (has_node(use) && use->in(0) != C->top() &&
2584 2592 (!peel.test(use->_idx) ||
2585 2593 (use->is_Phi() && use->in(0) == head)) ) {
2586 2594 worklist.push(use);
2587 2595 }
2588 2596 }
2589 2597 while( worklist.size() ) {
2590 2598 Node *use = worklist.pop();
2591 2599 for (uint j = 1; j < use->req(); j++) {
2592 2600 Node* n = use->in(j);
2593 2601 if (n == def) {
2594 2602
2595 2603 // "def" is in peel set, "use" is not in peel set
2596 2604 // or "use" is in the entry boundary (a phi) of the peel set
2597 2605
2598 2606 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
2599 2607
2600 2608 if ( loop->is_member(get_loop( use_c )) ) {
2601 2609 // use is in loop
2602 2610 if (old_new[use->_idx] != NULL) { // null for dead code
2603 2611 Node* use_clone = old_new[use->_idx];
2604 2612 _igvn.replace_input_of(use, j, C->top());
2605 2613 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
2606 2614 }
2607 2615 } else {
2608 2616 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
2609 2617 // use is not in the loop, check if the live range includes the cut
2610 2618 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
2611 2619 if (not_peel.test(lp_if->_idx)) {
2612 2620 assert(j == orig_exit_idx, "use from original loop");
2613 2621 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
2614 2622 }
2615 2623 }
2616 2624 }
2617 2625 }
2618 2626 }
2619 2627 }
2620 2628 }
2621 2629
2622 2630 // Step 3b: retarget control
2623 2631
2624 2632 // Redirect control to the new loop head if a cloned node in
2625 2633 // the not_peeled region has control that points into the peeled region.
2626 2634 // This necessary because the cloned peeled region will be outside
2627 2635 // the loop.
2628 2636 // from to
2629 2637 // cloned-peeled <---+
2630 2638 // new_head_clone: | <--+
2631 2639 // cloned-not_peeled in(0) in(0)
2632 2640 // orig-peeled
2633 2641
2634 2642 for(i = 0; i < loop->_body.size(); i++ ) {
2635 2643 Node *n = loop->_body.at(i);
2636 2644 if (!n->is_CFG() && n->in(0) != NULL &&
2637 2645 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
2638 2646 Node* n_clone = old_new[n->_idx];
2639 2647 _igvn.replace_input_of(n_clone, 0, new_head_clone);
2640 2648 }
2641 2649 }
2642 2650
2643 2651 // Backedge of the surviving new_head (the clone) is original last_peel
2644 2652 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
2645 2653
2646 2654 // Cut first node in original not_peel set
2647 2655 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
2648 2656 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
2649 2657 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
2650 2658
2651 2659 // Copy head_clone back-branch info to original head
2652 2660 // and remove original head's loop entry and
2653 2661 // clone head's back-branch
2654 2662 _igvn.rehash_node_delayed(head); // Multiple edge updates
2655 2663 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
2656 2664 head->set_req(LoopNode::LoopBackControl, C->top());
2657 2665 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
2658 2666
2659 2667 // Similarly modify the phis
2660 2668 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
2661 2669 Node* use = head->fast_out(k);
2662 2670 if (use->is_Phi() && use->outcnt() > 0) {
2663 2671 Node* use_clone = old_new[use->_idx];
2664 2672 _igvn.rehash_node_delayed(use); // Multiple edge updates
2665 2673 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
2666 2674 use->set_req(LoopNode::LoopBackControl, C->top());
2667 2675 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
2668 2676 }
2669 2677 }
2670 2678
2671 2679 // Step 4: update dominator tree and dominator depth
2672 2680
2673 2681 set_idom(head, orig_tail_clone, dd);
2674 2682 recompute_dom_depth();
2675 2683
2676 2684 // Inhibit more partial peeling on this loop
2677 2685 new_head_clone->set_partial_peel_loop();
2678 2686 C->set_major_progress();
2679 2687
2680 2688 #if !defined(PRODUCT)
2681 2689 if (TracePartialPeeling) {
2682 2690 tty->print_cr("\nafter partial peel one iteration");
2683 2691 Node_List wl(area);
2684 2692 Node* t = last_peel;
2685 2693 while (true) {
2686 2694 wl.push(t);
2687 2695 if (t == head_clone) break;
2688 2696 t = idom(t);
2689 2697 }
2690 2698 while (wl.size() > 0) {
2691 2699 Node* tt = wl.pop();
2692 2700 if (tt == head) tty->print_cr("orig head");
2693 2701 else if (tt == new_head_clone) tty->print_cr("new head");
2694 2702 else if (tt == head_clone) tty->print_cr("clone head");
2695 2703 tt->dump();
2696 2704 }
2697 2705 }
2698 2706 #endif
2699 2707 return true;
2700 2708 }
2701 2709
2702 2710 //------------------------------reorg_offsets----------------------------------
2703 2711 // Reorganize offset computations to lower register pressure. Mostly
2704 2712 // prevent loop-fallout uses of the pre-incremented trip counter (which are
2705 2713 // then alive with the post-incremented trip counter forcing an extra
2706 2714 // register move)
2707 2715 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
2708 2716 // Perform it only for canonical counted loops.
2709 2717 // Loop's shape could be messed up by iteration_split_impl.
2710 2718 if (!loop->_head->is_CountedLoop())
2711 2719 return;
2712 2720 if (!loop->_head->as_Loop()->is_valid_counted_loop())
2713 2721 return;
2714 2722
2715 2723 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2716 2724 CountedLoopEndNode *cle = cl->loopexit();
2717 2725 Node *exit = cle->proj_out(false);
2718 2726 Node *phi = cl->phi();
2719 2727
2720 2728 // Check for the special case of folks using the pre-incremented
2721 2729 // trip-counter on the fall-out path (forces the pre-incremented
2722 2730 // and post-incremented trip counter to be live at the same time).
2723 2731 // Fix this by adjusting to use the post-increment trip counter.
2724 2732
2725 2733 bool progress = true;
2726 2734 while (progress) {
2727 2735 progress = false;
2728 2736 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
2729 2737 Node* use = phi->fast_out(i); // User of trip-counter
2730 2738 if (!has_ctrl(use)) continue;
2731 2739 Node *u_ctrl = get_ctrl(use);
2732 2740 if (use->is_Phi()) {
2733 2741 u_ctrl = NULL;
2734 2742 for (uint j = 1; j < use->req(); j++)
2735 2743 if (use->in(j) == phi)
2736 2744 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
2737 2745 }
2738 2746 IdealLoopTree *u_loop = get_loop(u_ctrl);
2739 2747 // Look for loop-invariant use
2740 2748 if (u_loop == loop) continue;
2741 2749 if (loop->is_member(u_loop)) continue;
2742 2750 // Check that use is live out the bottom. Assuming the trip-counter
2743 2751 // update is right at the bottom, uses of of the loop middle are ok.
2744 2752 if (dom_lca(exit, u_ctrl) != exit) continue;
2745 2753 // Hit! Refactor use to use the post-incremented tripcounter.
2746 2754 // Compute a post-increment tripcounter.
2747 2755 Node *opaq = new (C) Opaque2Node( C, cle->incr() );
2748 2756 register_new_node( opaq, u_ctrl );
2749 2757 Node *neg_stride = _igvn.intcon(-cle->stride_con());
2750 2758 set_ctrl(neg_stride, C->root());
2751 2759 Node *post = new (C) AddINode( opaq, neg_stride);
2752 2760 register_new_node( post, u_ctrl );
2753 2761 _igvn.rehash_node_delayed(use);
2754 2762 for (uint j = 1; j < use->req(); j++) {
2755 2763 if (use->in(j) == phi)
2756 2764 use->set_req(j, post);
2757 2765 }
2758 2766 // Since DU info changed, rerun loop
2759 2767 progress = true;
2760 2768 break;
2761 2769 }
2762 2770 }
2763 2771
2764 2772 }
↓ open down ↓ |
214 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX