Mercurial > hg > truffle
annotate src/share/vm/opto/loopnode.cpp @ 367:194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
Reviewed-by: rasbold, jrose
| author | never |
|---|---|
| date | Wed, 17 Sep 2008 12:59:52 -0700 |
| parents | c3e045194476 |
| children | 3b5ac9e7e6ea |
| rev | line source |
|---|---|
| 0 | 1 /* |
| 196 | 2 * Copyright 1998-2008 Sun Microsystems, Inc. All Rights Reserved. |
| 0 | 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
| 4 * | |
| 5 * This code is free software; you can redistribute it and/or modify it | |
| 6 * under the terms of the GNU General Public License version 2 only, as | |
| 7 * published by the Free Software Foundation. | |
| 8 * | |
| 9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
| 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
| 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
| 12 * version 2 for more details (a copy is included in the LICENSE file that | |
| 13 * accompanied this code). | |
| 14 * | |
| 15 * You should have received a copy of the GNU General Public License version | |
| 16 * 2 along with this work; if not, write to the Free Software Foundation, | |
| 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
| 18 * | |
| 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
| 20 * CA 95054 USA or visit www.sun.com if you need additional information or | |
| 21 * have any questions. | |
| 22 * | |
| 23 */ | |
| 24 | |
| 25 #include "incls/_precompiled.incl" | |
| 26 #include "incls/_loopnode.cpp.incl" | |
| 27 | |
| 28 //============================================================================= | |
| 29 //------------------------------is_loop_iv------------------------------------- | |
| 30 // Determine if a node is Counted loop induction variable. | |
| 31 // The method is declared in node.hpp. | |
| 32 const Node* Node::is_loop_iv() const { | |
| 33 if (this->is_Phi() && !this->as_Phi()->is_copy() && | |
| 34 this->as_Phi()->region()->is_CountedLoop() && | |
| 35 this->as_Phi()->region()->as_CountedLoop()->phi() == this) { | |
| 36 return this; | |
| 37 } else { | |
| 38 return NULL; | |
| 39 } | |
| 40 } | |
| 41 | |
| 42 //============================================================================= | |
| 43 //------------------------------dump_spec-------------------------------------- | |
| 44 // Dump special per-node info | |
| 45 #ifndef PRODUCT | |
| 46 void LoopNode::dump_spec(outputStream *st) const { | |
| 47 if( is_inner_loop () ) st->print( "inner " ); | |
| 48 if( is_partial_peel_loop () ) st->print( "partial_peel " ); | |
| 49 if( partial_peel_has_failed () ) st->print( "partial_peel_failed " ); | |
| 50 } | |
| 51 #endif | |
| 52 | |
| 53 //------------------------------get_early_ctrl--------------------------------- | |
| 54 // Compute earliest legal control | |
| 55 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) { | |
| 56 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" ); | |
| 57 uint i; | |
| 58 Node *early; | |
| 59 if( n->in(0) ) { | |
| 60 early = n->in(0); | |
| 61 if( !early->is_CFG() ) // Might be a non-CFG multi-def | |
| 62 early = get_ctrl(early); // So treat input as a straight data input | |
| 63 i = 1; | |
| 64 } else { | |
| 65 early = get_ctrl(n->in(1)); | |
| 66 i = 2; | |
| 67 } | |
| 68 uint e_d = dom_depth(early); | |
| 69 assert( early, "" ); | |
| 70 for( ; i < n->req(); i++ ) { | |
| 71 Node *cin = get_ctrl(n->in(i)); | |
| 72 assert( cin, "" ); | |
| 73 // Keep deepest dominator depth | |
| 74 uint c_d = dom_depth(cin); | |
| 75 if( c_d > e_d ) { // Deeper guy? | |
| 76 early = cin; // Keep deepest found so far | |
| 77 e_d = c_d; | |
| 78 } else if( c_d == e_d && // Same depth? | |
| 79 early != cin ) { // If not equal, must use slower algorithm | |
| 80 // If same depth but not equal, one _must_ dominate the other | |
| 81 // and we want the deeper (i.e., dominated) guy. | |
| 82 Node *n1 = early; | |
| 83 Node *n2 = cin; | |
| 84 while( 1 ) { | |
| 85 n1 = idom(n1); // Walk up until break cycle | |
| 86 n2 = idom(n2); | |
| 87 if( n1 == cin || // Walked early up to cin | |
| 88 dom_depth(n2) < c_d ) | |
| 89 break; // early is deeper; keep him | |
| 90 if( n2 == early || // Walked cin up to early | |
| 91 dom_depth(n1) < c_d ) { | |
| 92 early = cin; // cin is deeper; keep him | |
| 93 break; | |
| 94 } | |
| 95 } | |
| 96 e_d = dom_depth(early); // Reset depth register cache | |
| 97 } | |
| 98 } | |
| 99 | |
| 100 // Return earliest legal location | |
| 101 assert(early == find_non_split_ctrl(early), "unexpected early control"); | |
| 102 | |
| 103 return early; | |
| 104 } | |
| 105 | |
| 106 //------------------------------set_early_ctrl--------------------------------- | |
| 107 // Set earliest legal control | |
| 108 void PhaseIdealLoop::set_early_ctrl( Node *n ) { | |
| 109 Node *early = get_early_ctrl(n); | |
| 110 | |
| 111 // Record earliest legal location | |
| 112 set_ctrl(n, early); | |
| 113 } | |
| 114 | |
| 115 //------------------------------set_subtree_ctrl------------------------------- | |
| 116 // set missing _ctrl entries on new nodes | |
| 117 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) { | |
| 118 // Already set? Get out. | |
| 119 if( _nodes[n->_idx] ) return; | |
| 120 // Recursively set _nodes array to indicate where the Node goes | |
| 121 uint i; | |
| 122 for( i = 0; i < n->req(); ++i ) { | |
| 123 Node *m = n->in(i); | |
| 124 if( m && m != C->root() ) | |
| 125 set_subtree_ctrl( m ); | |
| 126 } | |
| 127 | |
| 128 // Fixup self | |
| 129 set_early_ctrl( n ); | |
| 130 } | |
| 131 | |
| 132 //------------------------------is_counted_loop-------------------------------- | |
| 133 Node *PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) { | |
| 134 PhaseGVN *gvn = &_igvn; | |
| 135 | |
| 136 // Counted loop head must be a good RegionNode with only 3 not NULL | |
| 137 // control input edges: Self, Entry, LoopBack. | |
| 138 if ( x->in(LoopNode::Self) == NULL || x->req() != 3 ) | |
| 139 return NULL; | |
| 140 | |
| 141 Node *init_control = x->in(LoopNode::EntryControl); | |
| 142 Node *back_control = x->in(LoopNode::LoopBackControl); | |
| 143 if( init_control == NULL || back_control == NULL ) // Partially dead | |
| 144 return NULL; | |
| 145 // Must also check for TOP when looking for a dead loop | |
| 146 if( init_control->is_top() || back_control->is_top() ) | |
| 147 return NULL; | |
| 148 | |
| 149 // Allow funny placement of Safepoint | |
| 150 if( back_control->Opcode() == Op_SafePoint ) | |
| 151 back_control = back_control->in(TypeFunc::Control); | |
| 152 | |
| 153 // Controlling test for loop | |
| 154 Node *iftrue = back_control; | |
| 155 uint iftrue_op = iftrue->Opcode(); | |
| 156 if( iftrue_op != Op_IfTrue && | |
| 157 iftrue_op != Op_IfFalse ) | |
| 158 // I have a weird back-control. Probably the loop-exit test is in | |
| 159 // the middle of the loop and I am looking at some trailing control-flow | |
| 160 // merge point. To fix this I would have to partially peel the loop. | |
| 161 return NULL; // Obscure back-control | |
| 162 | |
| 163 // Get boolean guarding loop-back test | |
| 164 Node *iff = iftrue->in(0); | |
| 165 if( get_loop(iff) != loop || !iff->in(1)->is_Bool() ) return NULL; | |
| 166 BoolNode *test = iff->in(1)->as_Bool(); | |
| 167 BoolTest::mask bt = test->_test._test; | |
| 168 float cl_prob = iff->as_If()->_prob; | |
| 169 if( iftrue_op == Op_IfFalse ) { | |
| 170 bt = BoolTest(bt).negate(); | |
| 171 cl_prob = 1.0 - cl_prob; | |
| 172 } | |
| 173 // Get backedge compare | |
| 174 Node *cmp = test->in(1); | |
| 175 int cmp_op = cmp->Opcode(); | |
| 176 if( cmp_op != Op_CmpI ) | |
| 177 return NULL; // Avoid pointer & float compares | |
| 178 | |
| 179 // Find the trip-counter increment & limit. Limit must be loop invariant. | |
| 180 Node *incr = cmp->in(1); | |
| 181 Node *limit = cmp->in(2); | |
| 182 | |
| 183 // --------- | |
| 184 // need 'loop()' test to tell if limit is loop invariant | |
| 185 // --------- | |
| 186 | |
| 187 if( !is_member( loop, get_ctrl(incr) ) ) { // Swapped trip counter and limit? | |
| 188 Node *tmp = incr; // Then reverse order into the CmpI | |
| 189 incr = limit; | |
| 190 limit = tmp; | |
| 191 bt = BoolTest(bt).commute(); // And commute the exit test | |
| 192 } | |
| 193 if( is_member( loop, get_ctrl(limit) ) ) // Limit must loop-invariant | |
| 194 return NULL; | |
| 195 | |
| 196 // Trip-counter increment must be commutative & associative. | |
| 197 uint incr_op = incr->Opcode(); | |
| 198 if( incr_op == Op_Phi && incr->req() == 3 ) { | |
| 199 incr = incr->in(2); // Assume incr is on backedge of Phi | |
| 200 incr_op = incr->Opcode(); | |
| 201 } | |
| 202 Node* trunc1 = NULL; | |
| 203 Node* trunc2 = NULL; | |
| 204 const TypeInt* iv_trunc_t = NULL; | |
| 205 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) { | |
| 206 return NULL; // Funny increment opcode | |
| 207 } | |
| 208 | |
| 209 // Get merge point | |
| 210 Node *xphi = incr->in(1); | |
| 211 Node *stride = incr->in(2); | |
| 212 if( !stride->is_Con() ) { // Oops, swap these | |
| 213 if( !xphi->is_Con() ) // Is the other guy a constant? | |
| 214 return NULL; // Nope, unknown stride, bail out | |
| 215 Node *tmp = xphi; // 'incr' is commutative, so ok to swap | |
| 216 xphi = stride; | |
| 217 stride = tmp; | |
| 218 } | |
| 219 //if( loop(xphi) != l) return NULL;// Merge point is in inner loop?? | |
| 220 if( !xphi->is_Phi() ) return NULL; // Too much math on the trip counter | |
| 221 PhiNode *phi = xphi->as_Phi(); | |
| 222 | |
| 223 // Stride must be constant | |
| 224 const Type *stride_t = stride->bottom_type(); | |
| 225 int stride_con = stride_t->is_int()->get_con(); | |
| 226 assert( stride_con, "missed some peephole opt" ); | |
| 227 | |
| 228 // Phi must be of loop header; backedge must wrap to increment | |
| 229 if( phi->region() != x ) return NULL; | |
| 230 if( trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr || | |
| 231 trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1 ) { | |
| 232 return NULL; | |
| 233 } | |
| 234 Node *init_trip = phi->in(LoopNode::EntryControl); | |
| 235 //if (!init_trip->is_Con()) return NULL; // avoid rolling over MAXINT/MININT | |
| 236 | |
| 237 // If iv trunc type is smaller than int, check for possible wrap. | |
| 238 if (!TypeInt::INT->higher_equal(iv_trunc_t)) { | |
| 239 assert(trunc1 != NULL, "must have found some truncation"); | |
| 240 | |
| 241 // Get a better type for the phi (filtered thru if's) | |
| 242 const TypeInt* phi_ft = filtered_type(phi); | |
| 243 | |
| 244 // Can iv take on a value that will wrap? | |
| 245 // | |
| 246 // Ensure iv's limit is not within "stride" of the wrap value. | |
| 247 // | |
| 248 // Example for "short" type | |
| 249 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t) | |
| 250 // If the stride is +10, then the last value of the induction | |
| 251 // variable before the increment (phi_ft->_hi) must be | |
| 252 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to | |
| 253 // ensure no truncation occurs after the increment. | |
| 254 | |
| 255 if (stride_con > 0) { | |
| 256 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con || | |
| 257 iv_trunc_t->_lo > phi_ft->_lo) { | |
| 258 return NULL; // truncation may occur | |
| 259 } | |
| 260 } else if (stride_con < 0) { | |
| 261 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con || | |
| 262 iv_trunc_t->_hi < phi_ft->_hi) { | |
| 263 return NULL; // truncation may occur | |
| 264 } | |
| 265 } | |
| 266 // No possibility of wrap so truncation can be discarded | |
| 267 // Promote iv type to Int | |
| 268 } else { | |
| 269 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int"); | |
| 270 } | |
| 271 | |
| 272 // ================================================= | |
| 273 // ---- SUCCESS! Found A Trip-Counted Loop! ----- | |
| 274 // | |
| 275 // Canonicalize the condition on the test. If we can exactly determine | |
| 276 // the trip-counter exit value, then set limit to that value and use | |
| 277 // a '!=' test. Otherwise use conditon '<' for count-up loops and | |
| 278 // '>' for count-down loops. If the condition is inverted and we will | |
| 279 // be rolling through MININT to MAXINT, then bail out. | |
| 280 | |
| 281 C->print_method("Before CountedLoop", 3); | |
| 282 | |
| 283 // Check for SafePoint on backedge and remove | |
| 284 Node *sfpt = x->in(LoopNode::LoopBackControl); | |
| 285 if( sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) { | |
| 286 lazy_replace( sfpt, iftrue ); | |
| 287 loop->_tail = iftrue; | |
| 288 } | |
| 289 | |
| 290 | |
| 291 // If compare points to incr, we are ok. Otherwise the compare | |
| 292 // can directly point to the phi; in this case adjust the compare so that | |
| 293 // it points to the incr by adusting the limit. | |
| 294 if( cmp->in(1) == phi || cmp->in(2) == phi ) | |
| 295 limit = gvn->transform(new (C, 3) AddINode(limit,stride)); | |
| 296 | |
| 297 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride. | |
| 298 // Final value for iterator should be: trip_count * stride + init_trip. | |
| 299 const Type *limit_t = limit->bottom_type(); | |
| 300 const Type *init_t = init_trip->bottom_type(); | |
| 301 Node *one_p = gvn->intcon( 1); | |
| 302 Node *one_m = gvn->intcon(-1); | |
| 303 | |
| 304 Node *trip_count = NULL; | |
| 305 Node *hook = new (C, 6) Node(6); | |
| 306 switch( bt ) { | |
| 307 case BoolTest::eq: | |
| 308 return NULL; // Bail out, but this loop trips at most twice! | |
| 309 case BoolTest::ne: // Ahh, the case we desire | |
| 310 if( stride_con == 1 ) | |
| 311 trip_count = gvn->transform(new (C, 3) SubINode(limit,init_trip)); | |
| 312 else if( stride_con == -1 ) | |
| 313 trip_count = gvn->transform(new (C, 3) SubINode(init_trip,limit)); | |
| 314 else | |
| 315 return NULL; // Odd stride; must prove we hit limit exactly | |
| 316 set_subtree_ctrl( trip_count ); | |
| 317 //_loop.map(trip_count->_idx,loop(limit)); | |
| 318 break; | |
| 319 case BoolTest::le: // Maybe convert to '<' case | |
| 320 limit = gvn->transform(new (C, 3) AddINode(limit,one_p)); | |
| 321 set_subtree_ctrl( limit ); | |
| 322 hook->init_req(4, limit); | |
| 323 | |
| 324 bt = BoolTest::lt; | |
| 325 // Make the new limit be in the same loop nest as the old limit | |
| 326 //_loop.map(limit->_idx,limit_loop); | |
| 327 // Fall into next case | |
| 328 case BoolTest::lt: { // Maybe convert to '!=' case | |
| 329 if( stride_con < 0 ) return NULL; // Count down loop rolls through MAXINT | |
| 330 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip)); | |
| 331 set_subtree_ctrl( range ); | |
| 332 hook->init_req(0, range); | |
| 333 | |
| 334 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride)); | |
| 335 set_subtree_ctrl( bias ); | |
| 336 hook->init_req(1, bias); | |
| 337 | |
| 338 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_m)); | |
| 339 set_subtree_ctrl( bias1 ); | |
| 340 hook->init_req(2, bias1); | |
| 341 | |
| 342 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride)); | |
| 343 set_subtree_ctrl( trip_count ); | |
| 344 hook->init_req(3, trip_count); | |
| 345 break; | |
| 346 } | |
| 347 | |
| 348 case BoolTest::ge: // Maybe convert to '>' case | |
| 349 limit = gvn->transform(new (C, 3) AddINode(limit,one_m)); | |
| 350 set_subtree_ctrl( limit ); | |
| 351 hook->init_req(4 ,limit); | |
| 352 | |
| 353 bt = BoolTest::gt; | |
| 354 // Make the new limit be in the same loop nest as the old limit | |
| 355 //_loop.map(limit->_idx,limit_loop); | |
| 356 // Fall into next case | |
| 357 case BoolTest::gt: { // Maybe convert to '!=' case | |
| 358 if( stride_con > 0 ) return NULL; // count up loop rolls through MININT | |
| 359 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip)); | |
| 360 set_subtree_ctrl( range ); | |
| 361 hook->init_req(0, range); | |
| 362 | |
| 363 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride)); | |
| 364 set_subtree_ctrl( bias ); | |
| 365 hook->init_req(1, bias); | |
| 366 | |
| 367 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_p)); | |
| 368 set_subtree_ctrl( bias1 ); | |
| 369 hook->init_req(2, bias1); | |
| 370 | |
| 371 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride)); | |
| 372 set_subtree_ctrl( trip_count ); | |
| 373 hook->init_req(3, trip_count); | |
| 374 break; | |
| 375 } | |
| 376 } | |
| 377 | |
| 378 Node *span = gvn->transform(new (C, 3) MulINode(trip_count,stride)); | |
| 379 set_subtree_ctrl( span ); | |
| 380 hook->init_req(5, span); | |
| 381 | |
| 382 limit = gvn->transform(new (C, 3) AddINode(span,init_trip)); | |
| 383 set_subtree_ctrl( limit ); | |
| 384 | |
| 385 // Build a canonical trip test. | |
| 386 // Clone code, as old values may be in use. | |
| 387 incr = incr->clone(); | |
| 388 incr->set_req(1,phi); | |
| 389 incr->set_req(2,stride); | |
| 390 incr = _igvn.register_new_node_with_optimizer(incr); | |
| 391 set_early_ctrl( incr ); | |
| 392 _igvn.hash_delete(phi); | |
| 393 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn ); | |
| 394 | |
| 395 // If phi type is more restrictive than Int, raise to | |
| 396 // Int to prevent (almost) infinite recursion in igvn | |
| 397 // which can only handle integer types for constants or minint..maxint. | |
| 398 if (!TypeInt::INT->higher_equal(phi->bottom_type())) { | |
| 399 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT); | |
| 400 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl)); | |
| 401 nphi = _igvn.register_new_node_with_optimizer(nphi); | |
| 402 set_ctrl(nphi, get_ctrl(phi)); | |
| 403 _igvn.subsume_node(phi, nphi); | |
| 404 phi = nphi->as_Phi(); | |
| 405 } | |
| 406 cmp = cmp->clone(); | |
| 407 cmp->set_req(1,incr); | |
| 408 cmp->set_req(2,limit); | |
| 409 cmp = _igvn.register_new_node_with_optimizer(cmp); | |
| 410 set_ctrl(cmp, iff->in(0)); | |
| 411 | |
| 412 Node *tmp = test->clone(); | |
| 413 assert( tmp->is_Bool(), "" ); | |
| 414 test = (BoolNode*)tmp; | |
| 415 (*(BoolTest*)&test->_test)._test = bt; //BoolTest::ne; | |
| 416 test->set_req(1,cmp); | |
| 417 _igvn.register_new_node_with_optimizer(test); | |
| 418 set_ctrl(test, iff->in(0)); | |
| 419 // If the exit test is dead, STOP! | |
| 420 if( test == NULL ) return NULL; | |
| 421 _igvn.hash_delete(iff); | |
| 422 iff->set_req_X( 1, test, &_igvn ); | |
| 423 | |
| 424 // Replace the old IfNode with a new LoopEndNode | |
| 425 Node *lex = _igvn.register_new_node_with_optimizer(new (C, 2) CountedLoopEndNode( iff->in(0), iff->in(1), cl_prob, iff->as_If()->_fcnt )); | |
| 426 IfNode *le = lex->as_If(); | |
| 427 uint dd = dom_depth(iff); | |
| 428 set_idom(le, le->in(0), dd); // Update dominance for loop exit | |
| 429 set_loop(le, loop); | |
| 430 | |
| 431 // Get the loop-exit control | |
| 432 Node *if_f = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue)); | |
| 433 | |
| 434 // Need to swap loop-exit and loop-back control? | |
| 435 if( iftrue_op == Op_IfFalse ) { | |
| 436 Node *ift2=_igvn.register_new_node_with_optimizer(new (C, 1) IfTrueNode (le)); | |
| 437 Node *iff2=_igvn.register_new_node_with_optimizer(new (C, 1) IfFalseNode(le)); | |
| 438 | |
| 439 loop->_tail = back_control = ift2; | |
| 440 set_loop(ift2, loop); | |
| 441 set_loop(iff2, get_loop(if_f)); | |
| 442 | |
| 443 // Lazy update of 'get_ctrl' mechanism. | |
| 444 lazy_replace_proj( if_f , iff2 ); | |
| 445 lazy_replace_proj( iftrue, ift2 ); | |
| 446 | |
| 447 // Swap names | |
| 448 if_f = iff2; | |
| 449 iftrue = ift2; | |
| 450 } else { | |
| 451 _igvn.hash_delete(if_f ); | |
| 452 _igvn.hash_delete(iftrue); | |
| 453 if_f ->set_req_X( 0, le, &_igvn ); | |
| 454 iftrue->set_req_X( 0, le, &_igvn ); | |
| 455 } | |
| 456 | |
| 457 set_idom(iftrue, le, dd+1); | |
| 458 set_idom(if_f, le, dd+1); | |
| 459 | |
| 460 // Now setup a new CountedLoopNode to replace the existing LoopNode | |
| 461 CountedLoopNode *l = new (C, 3) CountedLoopNode(init_control, back_control); | |
| 462 // The following assert is approximately true, and defines the intention | |
| 463 // of can_be_counted_loop. It fails, however, because phase->type | |
| 464 // is not yet initialized for this loop and its parts. | |
| 465 //assert(l->can_be_counted_loop(this), "sanity"); | |
| 466 _igvn.register_new_node_with_optimizer(l); | |
| 467 set_loop(l, loop); | |
| 468 loop->_head = l; | |
| 469 // Fix all data nodes placed at the old loop head. | |
| 470 // Uses the lazy-update mechanism of 'get_ctrl'. | |
| 471 lazy_replace( x, l ); | |
| 472 set_idom(l, init_control, dom_depth(x)); | |
| 473 | |
| 474 // Check for immediately preceeding SafePoint and remove | |
| 475 Node *sfpt2 = le->in(0); | |
| 476 if( sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2)) | |
| 477 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control)); | |
| 478 | |
| 479 // Free up intermediate goo | |
| 480 _igvn.remove_dead_node(hook); | |
| 481 | |
| 482 C->print_method("After CountedLoop", 3); | |
| 483 | |
| 484 // Return trip counter | |
| 485 return trip_count; | |
| 486 } | |
| 487 | |
| 488 | |
| 489 //------------------------------Ideal------------------------------------------ | |
| 490 // Return a node which is more "ideal" than the current node. | |
| 491 // Attempt to convert into a counted-loop. | |
| 492 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
| 493 if (!can_be_counted_loop(phase)) { | |
| 494 phase->C->set_major_progress(); | |
| 495 } | |
| 496 return RegionNode::Ideal(phase, can_reshape); | |
| 497 } | |
| 498 | |
| 499 | |
| 500 //============================================================================= | |
| 501 //------------------------------Ideal------------------------------------------ | |
| 502 // Return a node which is more "ideal" than the current node. | |
| 503 // Attempt to convert into a counted-loop. | |
| 504 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
| 505 return RegionNode::Ideal(phase, can_reshape); | |
| 506 } | |
| 507 | |
| 508 //------------------------------dump_spec-------------------------------------- | |
| 509 // Dump special per-node info | |
| 510 #ifndef PRODUCT | |
| 511 void CountedLoopNode::dump_spec(outputStream *st) const { | |
| 512 LoopNode::dump_spec(st); | |
| 513 if( stride_is_con() ) { | |
| 514 st->print("stride: %d ",stride_con()); | |
| 515 } else { | |
| 516 st->print("stride: not constant "); | |
| 517 } | |
| 518 if( is_pre_loop () ) st->print("pre of N%d" , _main_idx ); | |
| 519 if( is_main_loop() ) st->print("main of N%d", _idx ); | |
| 520 if( is_post_loop() ) st->print("post of N%d", _main_idx ); | |
| 521 } | |
| 522 #endif | |
| 523 | |
| 524 //============================================================================= | |
| 525 int CountedLoopEndNode::stride_con() const { | |
| 526 return stride()->bottom_type()->is_int()->get_con(); | |
| 527 } | |
| 528 | |
| 529 | |
| 530 //----------------------match_incr_with_optional_truncation-------------------- | |
| 531 // Match increment with optional truncation: | |
| 532 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16 | |
| 533 // Return NULL for failure. Success returns the increment node. | |
| 534 Node* CountedLoopNode::match_incr_with_optional_truncation( | |
| 535 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) { | |
| 536 // Quick cutouts: | |
| 537 if (expr == NULL || expr->req() != 3) return false; | |
| 538 | |
| 539 Node *t1 = NULL; | |
| 540 Node *t2 = NULL; | |
| 541 const TypeInt* trunc_t = TypeInt::INT; | |
| 542 Node* n1 = expr; | |
| 543 int n1op = n1->Opcode(); | |
| 544 | |
| 545 // Try to strip (n1 & M) or (n1 << N >> N) from n1. | |
| 546 if (n1op == Op_AndI && | |
| 547 n1->in(2)->is_Con() && | |
| 548 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) { | |
| 549 // %%% This check should match any mask of 2**K-1. | |
| 550 t1 = n1; | |
| 551 n1 = t1->in(1); | |
| 552 n1op = n1->Opcode(); | |
| 553 trunc_t = TypeInt::CHAR; | |
| 554 } else if (n1op == Op_RShiftI && | |
| 555 n1->in(1) != NULL && | |
| 556 n1->in(1)->Opcode() == Op_LShiftI && | |
| 557 n1->in(2) == n1->in(1)->in(2) && | |
| 558 n1->in(2)->is_Con()) { | |
| 559 jint shift = n1->in(2)->bottom_type()->is_int()->get_con(); | |
| 560 // %%% This check should match any shift in [1..31]. | |
| 561 if (shift == 16 || shift == 8) { | |
| 562 t1 = n1; | |
| 563 t2 = t1->in(1); | |
| 564 n1 = t2->in(1); | |
| 565 n1op = n1->Opcode(); | |
| 566 if (shift == 16) { | |
| 567 trunc_t = TypeInt::SHORT; | |
| 568 } else if (shift == 8) { | |
| 569 trunc_t = TypeInt::BYTE; | |
| 570 } | |
| 571 } | |
| 572 } | |
| 573 | |
| 574 // If (maybe after stripping) it is an AddI, we won: | |
| 575 if (n1op == Op_AddI) { | |
| 576 *trunc1 = t1; | |
| 577 *trunc2 = t2; | |
| 578 *trunc_type = trunc_t; | |
| 579 return n1; | |
| 580 } | |
| 581 | |
| 582 // failed | |
| 583 return NULL; | |
| 584 } | |
| 585 | |
| 586 | |
| 587 //------------------------------filtered_type-------------------------------- | |
| 588 // Return a type based on condition control flow | |
| 589 // A successful return will be a type that is restricted due | |
| 590 // to a series of dominating if-tests, such as: | |
| 591 // if (i < 10) { | |
| 592 // if (i > 0) { | |
| 593 // here: "i" type is [1..10) | |
| 594 // } | |
| 595 // } | |
| 596 // or a control flow merge | |
| 597 // if (i < 10) { | |
| 598 // do { | |
| 599 // phi( , ) -- at top of loop type is [min_int..10) | |
| 600 // i = ? | |
| 601 // } while ( i < 10) | |
| 602 // | |
| 603 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) { | |
| 604 assert(n && n->bottom_type()->is_int(), "must be int"); | |
| 605 const TypeInt* filtered_t = NULL; | |
| 606 if (!n->is_Phi()) { | |
| 607 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control"); | |
| 608 filtered_t = filtered_type_from_dominators(n, n_ctrl); | |
| 609 | |
| 610 } else { | |
| 611 Node* phi = n->as_Phi(); | |
| 612 Node* region = phi->in(0); | |
| 613 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region"); | |
| 614 if (region && region != C->top()) { | |
| 615 for (uint i = 1; i < phi->req(); i++) { | |
| 616 Node* val = phi->in(i); | |
| 617 Node* use_c = region->in(i); | |
| 618 const TypeInt* val_t = filtered_type_from_dominators(val, use_c); | |
| 619 if (val_t != NULL) { | |
| 620 if (filtered_t == NULL) { | |
| 621 filtered_t = val_t; | |
| 622 } else { | |
| 623 filtered_t = filtered_t->meet(val_t)->is_int(); | |
| 624 } | |
| 625 } | |
| 626 } | |
| 627 } | |
| 628 } | |
| 629 const TypeInt* n_t = _igvn.type(n)->is_int(); | |
| 630 if (filtered_t != NULL) { | |
| 631 n_t = n_t->join(filtered_t)->is_int(); | |
| 632 } | |
| 633 return n_t; | |
| 634 } | |
| 635 | |
| 636 | |
| 637 //------------------------------filtered_type_from_dominators-------------------------------- | |
| 638 // Return a possibly more restrictive type for val based on condition control flow of dominators | |
| 639 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) { | |
| 640 if (val->is_Con()) { | |
| 641 return val->bottom_type()->is_int(); | |
| 642 } | |
| 643 uint if_limit = 10; // Max number of dominating if's visited | |
| 644 const TypeInt* rtn_t = NULL; | |
| 645 | |
| 646 if (use_ctrl && use_ctrl != C->top()) { | |
| 647 Node* val_ctrl = get_ctrl(val); | |
| 648 uint val_dom_depth = dom_depth(val_ctrl); | |
| 649 Node* pred = use_ctrl; | |
| 650 uint if_cnt = 0; | |
| 651 while (if_cnt < if_limit) { | |
| 652 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) { | |
| 653 if_cnt++; | |
|
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
654 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred); |
| 0 | 655 if (if_t != NULL) { |
| 656 if (rtn_t == NULL) { | |
| 657 rtn_t = if_t; | |
| 658 } else { | |
| 659 rtn_t = rtn_t->join(if_t)->is_int(); | |
| 660 } | |
| 661 } | |
| 662 } | |
| 663 pred = idom(pred); | |
| 664 if (pred == NULL || pred == C->top()) { | |
| 665 break; | |
| 666 } | |
| 667 // Stop if going beyond definition block of val | |
| 668 if (dom_depth(pred) < val_dom_depth) { | |
| 669 break; | |
| 670 } | |
| 671 } | |
| 672 } | |
| 673 return rtn_t; | |
| 674 } | |
| 675 | |
| 676 | |
| 677 //------------------------------dump_spec-------------------------------------- | |
| 678 // Dump special per-node info | |
| 679 #ifndef PRODUCT | |
| 680 void CountedLoopEndNode::dump_spec(outputStream *st) const { | |
| 681 if( in(TestValue)->is_Bool() ) { | |
| 682 BoolTest bt( test_trip()); // Added this for g++. | |
| 683 | |
| 684 st->print("["); | |
| 685 bt.dump_on(st); | |
| 686 st->print("]"); | |
| 687 } | |
| 688 st->print(" "); | |
| 689 IfNode::dump_spec(st); | |
| 690 } | |
| 691 #endif | |
| 692 | |
| 693 //============================================================================= | |
| 694 //------------------------------is_member-------------------------------------- | |
| 695 // Is 'l' a member of 'this'? | |
| 696 int IdealLoopTree::is_member( const IdealLoopTree *l ) const { | |
| 697 while( l->_nest > _nest ) l = l->_parent; | |
| 698 return l == this; | |
| 699 } | |
| 700 | |
| 701 //------------------------------set_nest--------------------------------------- | |
| 702 // Set loop tree nesting depth. Accumulate _has_call bits. | |
| 703 int IdealLoopTree::set_nest( uint depth ) { | |
| 704 _nest = depth; | |
| 705 int bits = _has_call; | |
| 706 if( _child ) bits |= _child->set_nest(depth+1); | |
| 707 if( bits ) _has_call = 1; | |
| 708 if( _next ) bits |= _next ->set_nest(depth ); | |
| 709 return bits; | |
| 710 } | |
| 711 | |
| 712 //------------------------------split_fall_in---------------------------------- | |
| 713 // Split out multiple fall-in edges from the loop header. Move them to a | |
| 714 // private RegionNode before the loop. This becomes the loop landing pad. | |
| 715 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) { | |
| 716 PhaseIterGVN &igvn = phase->_igvn; | |
| 717 uint i; | |
| 718 | |
| 719 // Make a new RegionNode to be the landing pad. | |
| 720 Node *landing_pad = new (phase->C, fall_in_cnt+1) RegionNode( fall_in_cnt+1 ); | |
| 721 phase->set_loop(landing_pad,_parent); | |
| 722 // Gather all the fall-in control paths into the landing pad | |
| 723 uint icnt = fall_in_cnt; | |
| 724 uint oreq = _head->req(); | |
| 725 for( i = oreq-1; i>0; i-- ) | |
| 726 if( !phase->is_member( this, _head->in(i) ) ) | |
| 727 landing_pad->set_req(icnt--,_head->in(i)); | |
| 728 | |
| 729 // Peel off PhiNode edges as well | |
| 730 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { | |
| 731 Node *oj = _head->fast_out(j); | |
| 732 if( oj->is_Phi() ) { | |
| 733 PhiNode* old_phi = oj->as_Phi(); | |
| 734 assert( old_phi->region() == _head, "" ); | |
| 735 igvn.hash_delete(old_phi); // Yank from hash before hacking edges | |
| 736 Node *p = PhiNode::make_blank(landing_pad, old_phi); | |
| 737 uint icnt = fall_in_cnt; | |
| 738 for( i = oreq-1; i>0; i-- ) { | |
| 739 if( !phase->is_member( this, _head->in(i) ) ) { | |
| 740 p->init_req(icnt--, old_phi->in(i)); | |
| 741 // Go ahead and clean out old edges from old phi | |
| 742 old_phi->del_req(i); | |
| 743 } | |
| 744 } | |
| 745 // Search for CSE's here, because ZKM.jar does a lot of | |
| 746 // loop hackery and we need to be a little incremental | |
| 747 // with the CSE to avoid O(N^2) node blow-up. | |
| 748 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE | |
| 749 if( p2 ) { // Found CSE | |
| 750 p->destruct(); // Recover useless new node | |
| 751 p = p2; // Use old node | |
| 752 } else { | |
| 753 igvn.register_new_node_with_optimizer(p, old_phi); | |
| 754 } | |
| 755 // Make old Phi refer to new Phi. | |
| 756 old_phi->add_req(p); | |
| 757 // Check for the special case of making the old phi useless and | |
| 758 // disappear it. In JavaGrande I have a case where this useless | |
| 759 // Phi is the loop limit and prevents recognizing a CountedLoop | |
| 760 // which in turn prevents removing an empty loop. | |
| 761 Node *id_old_phi = old_phi->Identity( &igvn ); | |
| 762 if( id_old_phi != old_phi ) { // Found a simple identity? | |
| 763 // Note that I cannot call 'subsume_node' here, because | |
| 764 // that will yank the edge from old_phi to the Region and | |
| 765 // I'm mid-iteration over the Region's uses. | |
| 766 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) { | |
| 767 Node* use = old_phi->last_out(i); | |
| 768 igvn.hash_delete(use); | |
| 769 igvn._worklist.push(use); | |
| 770 uint uses_found = 0; | |
| 771 for (uint j = 0; j < use->len(); j++) { | |
| 772 if (use->in(j) == old_phi) { | |
| 773 if (j < use->req()) use->set_req (j, id_old_phi); | |
| 774 else use->set_prec(j, id_old_phi); | |
| 775 uses_found++; | |
| 776 } | |
| 777 } | |
| 778 i -= uses_found; // we deleted 1 or more copies of this edge | |
| 779 } | |
| 780 } | |
| 781 igvn._worklist.push(old_phi); | |
| 782 } | |
| 783 } | |
| 784 // Finally clean out the fall-in edges from the RegionNode | |
| 785 for( i = oreq-1; i>0; i-- ) { | |
| 786 if( !phase->is_member( this, _head->in(i) ) ) { | |
| 787 _head->del_req(i); | |
| 788 } | |
| 789 } | |
| 790 // Transform landing pad | |
| 791 igvn.register_new_node_with_optimizer(landing_pad, _head); | |
| 792 // Insert landing pad into the header | |
| 793 _head->add_req(landing_pad); | |
| 794 } | |
| 795 | |
| 796 //------------------------------split_outer_loop------------------------------- | |
| 797 // Split out the outermost loop from this shared header. | |
| 798 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) { | |
| 799 PhaseIterGVN &igvn = phase->_igvn; | |
| 800 | |
| 801 // Find index of outermost loop; it should also be my tail. | |
| 802 uint outer_idx = 1; | |
| 803 while( _head->in(outer_idx) != _tail ) outer_idx++; | |
| 804 | |
| 805 // Make a LoopNode for the outermost loop. | |
| 806 Node *ctl = _head->in(LoopNode::EntryControl); | |
| 807 Node *outer = new (phase->C, 3) LoopNode( ctl, _head->in(outer_idx) ); | |
| 808 outer = igvn.register_new_node_with_optimizer(outer, _head); | |
| 809 phase->set_created_loop_node(); | |
| 810 // Outermost loop falls into '_head' loop | |
| 811 _head->set_req(LoopNode::EntryControl, outer); | |
| 812 _head->del_req(outer_idx); | |
| 813 // Split all the Phis up between '_head' loop and 'outer' loop. | |
| 814 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { | |
| 815 Node *out = _head->fast_out(j); | |
| 816 if( out->is_Phi() ) { | |
| 817 PhiNode *old_phi = out->as_Phi(); | |
| 818 assert( old_phi->region() == _head, "" ); | |
| 819 Node *phi = PhiNode::make_blank(outer, old_phi); | |
| 820 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl)); | |
| 821 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx)); | |
| 822 phi = igvn.register_new_node_with_optimizer(phi, old_phi); | |
| 823 // Make old Phi point to new Phi on the fall-in path | |
| 824 igvn.hash_delete(old_phi); | |
| 825 old_phi->set_req(LoopNode::EntryControl, phi); | |
| 826 old_phi->del_req(outer_idx); | |
| 827 igvn._worklist.push(old_phi); | |
| 828 } | |
| 829 } | |
| 830 | |
| 831 // Use the new loop head instead of the old shared one | |
| 832 _head = outer; | |
| 833 phase->set_loop(_head, this); | |
| 834 } | |
| 835 | |
| 836 //------------------------------fix_parent------------------------------------- | |
| 837 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) { | |
| 838 loop->_parent = parent; | |
| 839 if( loop->_child ) fix_parent( loop->_child, loop ); | |
| 840 if( loop->_next ) fix_parent( loop->_next , parent ); | |
| 841 } | |
| 842 | |
| 843 //------------------------------estimate_path_freq----------------------------- | |
| 844 static float estimate_path_freq( Node *n ) { | |
| 845 // Try to extract some path frequency info | |
| 846 IfNode *iff; | |
| 847 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests | |
| 848 uint nop = n->Opcode(); | |
| 849 if( nop == Op_SafePoint ) { // Skip any safepoint | |
| 850 n = n->in(0); | |
| 851 continue; | |
| 852 } | |
| 853 if( nop == Op_CatchProj ) { // Get count from a prior call | |
| 854 // Assume call does not always throw exceptions: means the call-site | |
| 855 // count is also the frequency of the fall-through path. | |
| 856 assert( n->is_CatchProj(), "" ); | |
| 857 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index ) | |
| 858 return 0.0f; // Assume call exception path is rare | |
| 859 Node *call = n->in(0)->in(0)->in(0); | |
| 860 assert( call->is_Call(), "expect a call here" ); | |
| 861 const JVMState *jvms = ((CallNode*)call)->jvms(); | |
| 862 ciMethodData* methodData = jvms->method()->method_data(); | |
| 863 if (!methodData->is_mature()) return 0.0f; // No call-site data | |
| 864 ciProfileData* data = methodData->bci_to_data(jvms->bci()); | |
| 865 if ((data == NULL) || !data->is_CounterData()) { | |
| 866 // no call profile available, try call's control input | |
| 867 n = n->in(0); | |
| 868 continue; | |
| 869 } | |
| 870 return data->as_CounterData()->count()/FreqCountInvocations; | |
| 871 } | |
| 872 // See if there's a gating IF test | |
| 873 Node *n_c = n->in(0); | |
| 874 if( !n_c->is_If() ) break; // No estimate available | |
| 875 iff = n_c->as_If(); | |
| 876 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count? | |
| 877 // Compute how much count comes on this path | |
| 878 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt; | |
| 879 // Have no count info. Skip dull uncommon-trap like branches. | |
| 880 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) || | |
| 881 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) ) | |
| 882 break; | |
| 883 // Skip through never-taken branch; look for a real loop exit. | |
| 884 n = iff->in(0); | |
| 885 } | |
| 886 return 0.0f; // No estimate available | |
| 887 } | |
| 888 | |
| 889 //------------------------------merge_many_backedges--------------------------- | |
| 890 // Merge all the backedges from the shared header into a private Region. | |
| 891 // Feed that region as the one backedge to this loop. | |
| 892 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) { | |
| 893 uint i; | |
| 894 | |
| 895 // Scan for the top 2 hottest backedges | |
| 896 float hotcnt = 0.0f; | |
| 897 float warmcnt = 0.0f; | |
| 898 uint hot_idx = 0; | |
| 899 // Loop starts at 2 because slot 1 is the fall-in path | |
| 900 for( i = 2; i < _head->req(); i++ ) { | |
| 901 float cnt = estimate_path_freq(_head->in(i)); | |
| 902 if( cnt > hotcnt ) { // Grab hottest path | |
| 903 warmcnt = hotcnt; | |
| 904 hotcnt = cnt; | |
| 905 hot_idx = i; | |
| 906 } else if( cnt > warmcnt ) { // And 2nd hottest path | |
| 907 warmcnt = cnt; | |
| 908 } | |
| 909 } | |
| 910 | |
| 911 // See if the hottest backedge is worthy of being an inner loop | |
| 912 // by being much hotter than the next hottest backedge. | |
| 913 if( hotcnt <= 0.0001 || | |
| 914 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge | |
| 915 | |
| 916 // Peel out the backedges into a private merge point; peel | |
| 917 // them all except optionally hot_idx. | |
| 918 PhaseIterGVN &igvn = phase->_igvn; | |
| 919 | |
| 920 Node *hot_tail = NULL; | |
| 921 // Make a Region for the merge point | |
| 922 Node *r = new (phase->C, 1) RegionNode(1); | |
| 923 for( i = 2; i < _head->req(); i++ ) { | |
| 924 if( i != hot_idx ) | |
| 925 r->add_req( _head->in(i) ); | |
| 926 else hot_tail = _head->in(i); | |
| 927 } | |
| 928 igvn.register_new_node_with_optimizer(r, _head); | |
| 929 // Plug region into end of loop _head, followed by hot_tail | |
| 930 while( _head->req() > 3 ) _head->del_req( _head->req()-1 ); | |
| 931 _head->set_req(2, r); | |
| 932 if( hot_idx ) _head->add_req(hot_tail); | |
| 933 | |
| 934 // Split all the Phis up between '_head' loop and the Region 'r' | |
| 935 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { | |
| 936 Node *out = _head->fast_out(j); | |
| 937 if( out->is_Phi() ) { | |
| 938 PhiNode* n = out->as_Phi(); | |
| 939 igvn.hash_delete(n); // Delete from hash before hacking edges | |
| 940 Node *hot_phi = NULL; | |
| 941 Node *phi = new (phase->C, r->req()) PhiNode(r, n->type(), n->adr_type()); | |
| 942 // Check all inputs for the ones to peel out | |
| 943 uint j = 1; | |
| 944 for( uint i = 2; i < n->req(); i++ ) { | |
| 945 if( i != hot_idx ) | |
| 946 phi->set_req( j++, n->in(i) ); | |
| 947 else hot_phi = n->in(i); | |
| 948 } | |
| 949 // Register the phi but do not transform until whole place transforms | |
| 950 igvn.register_new_node_with_optimizer(phi, n); | |
| 951 // Add the merge phi to the old Phi | |
| 952 while( n->req() > 3 ) n->del_req( n->req()-1 ); | |
| 953 n->set_req(2, phi); | |
| 954 if( hot_idx ) n->add_req(hot_phi); | |
| 955 } | |
| 956 } | |
| 957 | |
| 958 | |
| 959 // Insert a new IdealLoopTree inserted below me. Turn it into a clone | |
| 960 // of self loop tree. Turn self into a loop headed by _head and with | |
| 961 // tail being the new merge point. | |
| 962 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail ); | |
| 963 phase->set_loop(_tail,ilt); // Adjust tail | |
| 964 _tail = r; // Self's tail is new merge point | |
| 965 phase->set_loop(r,this); | |
| 966 ilt->_child = _child; // New guy has my children | |
| 967 _child = ilt; // Self has new guy as only child | |
| 968 ilt->_parent = this; // new guy has self for parent | |
| 969 ilt->_nest = _nest; // Same nesting depth (for now) | |
| 970 | |
| 971 // Starting with 'ilt', look for child loop trees using the same shared | |
| 972 // header. Flatten these out; they will no longer be loops in the end. | |
| 973 IdealLoopTree **pilt = &_child; | |
| 974 while( ilt ) { | |
| 975 if( ilt->_head == _head ) { | |
| 976 uint i; | |
| 977 for( i = 2; i < _head->req(); i++ ) | |
| 978 if( _head->in(i) == ilt->_tail ) | |
| 979 break; // Still a loop | |
| 980 if( i == _head->req() ) { // No longer a loop | |
| 981 // Flatten ilt. Hang ilt's "_next" list from the end of | |
| 982 // ilt's '_child' list. Move the ilt's _child up to replace ilt. | |
| 983 IdealLoopTree **cp = &ilt->_child; | |
| 984 while( *cp ) cp = &(*cp)->_next; // Find end of child list | |
| 985 *cp = ilt->_next; // Hang next list at end of child list | |
| 986 *pilt = ilt->_child; // Move child up to replace ilt | |
| 987 ilt->_head = NULL; // Flag as a loop UNIONED into parent | |
| 988 ilt = ilt->_child; // Repeat using new ilt | |
| 989 continue; // do not advance over ilt->_child | |
| 990 } | |
| 991 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" ); | |
| 992 phase->set_loop(_head,ilt); | |
| 993 } | |
| 994 pilt = &ilt->_child; // Advance to next | |
| 995 ilt = *pilt; | |
| 996 } | |
| 997 | |
| 998 if( _child ) fix_parent( _child, this ); | |
| 999 } | |
| 1000 | |
| 1001 //------------------------------beautify_loops--------------------------------- | |
| 1002 // Split shared headers and insert loop landing pads. | |
| 1003 // Insert a LoopNode to replace the RegionNode. | |
| 1004 // Return TRUE if loop tree is structurally changed. | |
| 1005 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) { | |
| 1006 bool result = false; | |
| 1007 // Cache parts in locals for easy | |
| 1008 PhaseIterGVN &igvn = phase->_igvn; | |
| 1009 | |
| 1010 phase->C->print_method("Before beautify loops", 3); | |
| 1011 | |
| 1012 igvn.hash_delete(_head); // Yank from hash before hacking edges | |
| 1013 | |
| 1014 // Check for multiple fall-in paths. Peel off a landing pad if need be. | |
| 1015 int fall_in_cnt = 0; | |
| 1016 for( uint i = 1; i < _head->req(); i++ ) | |
| 1017 if( !phase->is_member( this, _head->in(i) ) ) | |
| 1018 fall_in_cnt++; | |
| 1019 assert( fall_in_cnt, "at least 1 fall-in path" ); | |
| 1020 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins | |
| 1021 split_fall_in( phase, fall_in_cnt ); | |
| 1022 | |
| 1023 // Swap inputs to the _head and all Phis to move the fall-in edge to | |
| 1024 // the left. | |
| 1025 fall_in_cnt = 1; | |
| 1026 while( phase->is_member( this, _head->in(fall_in_cnt) ) ) | |
| 1027 fall_in_cnt++; | |
| 1028 if( fall_in_cnt > 1 ) { | |
| 1029 // Since I am just swapping inputs I do not need to update def-use info | |
| 1030 Node *tmp = _head->in(1); | |
| 1031 _head->set_req( 1, _head->in(fall_in_cnt) ); | |
| 1032 _head->set_req( fall_in_cnt, tmp ); | |
| 1033 // Swap also all Phis | |
| 1034 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) { | |
| 1035 Node* phi = _head->fast_out(i); | |
| 1036 if( phi->is_Phi() ) { | |
| 1037 igvn.hash_delete(phi); // Yank from hash before hacking edges | |
| 1038 tmp = phi->in(1); | |
| 1039 phi->set_req( 1, phi->in(fall_in_cnt) ); | |
| 1040 phi->set_req( fall_in_cnt, tmp ); | |
| 1041 } | |
| 1042 } | |
| 1043 } | |
| 1044 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" ); | |
| 1045 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" ); | |
| 1046 | |
| 1047 // If I am a shared header (multiple backedges), peel off the many | |
| 1048 // backedges into a private merge point and use the merge point as | |
| 1049 // the one true backedge. | |
| 1050 if( _head->req() > 3 ) { | |
| 1051 // Merge the many backedges into a single backedge. | |
| 1052 merge_many_backedges( phase ); | |
| 1053 result = true; | |
| 1054 } | |
| 1055 | |
| 1056 // If I am a shared header (multiple backedges), peel off myself loop. | |
| 1057 // I better be the outermost loop. | |
| 1058 if( _head->req() > 3 ) { | |
| 1059 split_outer_loop( phase ); | |
| 1060 result = true; | |
| 1061 | |
| 1062 } else if( !_head->is_Loop() && !_irreducible ) { | |
| 1063 // Make a new LoopNode to replace the old loop head | |
| 1064 Node *l = new (phase->C, 3) LoopNode( _head->in(1), _head->in(2) ); | |
| 1065 l = igvn.register_new_node_with_optimizer(l, _head); | |
| 1066 phase->set_created_loop_node(); | |
| 1067 // Go ahead and replace _head | |
| 1068 phase->_igvn.subsume_node( _head, l ); | |
| 1069 _head = l; | |
| 1070 phase->set_loop(_head, this); | |
| 1071 for (DUIterator_Fast imax, i = l->fast_outs(imax); i < imax; i++) | |
| 1072 phase->_igvn.add_users_to_worklist(l->fast_out(i)); | |
| 1073 } | |
| 1074 | |
| 1075 // Now recursively beautify nested loops | |
| 1076 if( _child ) result |= _child->beautify_loops( phase ); | |
| 1077 if( _next ) result |= _next ->beautify_loops( phase ); | |
| 1078 return result; | |
| 1079 } | |
| 1080 | |
| 1081 //------------------------------allpaths_check_safepts---------------------------- | |
| 1082 // Allpaths backwards scan from loop tail, terminating each path at first safepoint | |
| 1083 // encountered. Helper for check_safepts. | |
| 1084 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) { | |
| 1085 assert(stack.size() == 0, "empty stack"); | |
| 1086 stack.push(_tail); | |
| 1087 visited.Clear(); | |
| 1088 visited.set(_tail->_idx); | |
| 1089 while (stack.size() > 0) { | |
| 1090 Node* n = stack.pop(); | |
| 1091 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { | |
| 1092 // Terminate this path | |
| 1093 } else if (n->Opcode() == Op_SafePoint) { | |
| 1094 if (_phase->get_loop(n) != this) { | |
| 1095 if (_required_safept == NULL) _required_safept = new Node_List(); | |
| 1096 _required_safept->push(n); // save the one closest to the tail | |
| 1097 } | |
| 1098 // Terminate this path | |
| 1099 } else { | |
| 1100 uint start = n->is_Region() ? 1 : 0; | |
| 1101 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1; | |
| 1102 for (uint i = start; i < end; i++) { | |
| 1103 Node* in = n->in(i); | |
| 1104 assert(in->is_CFG(), "must be"); | |
| 1105 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) { | |
| 1106 stack.push(in); | |
| 1107 } | |
| 1108 } | |
| 1109 } | |
| 1110 } | |
| 1111 } | |
| 1112 | |
| 1113 //------------------------------check_safepts---------------------------- | |
| 1114 // Given dominators, try to find loops with calls that must always be | |
| 1115 // executed (call dominates loop tail). These loops do not need non-call | |
| 1116 // safepoints (ncsfpt). | |
| 1117 // | |
| 1118 // A complication is that a safepoint in a inner loop may be needed | |
| 1119 // by an outer loop. In the following, the inner loop sees it has a | |
| 1120 // call (block 3) on every path from the head (block 2) to the | |
| 1121 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint) | |
| 1122 // in block 2, _but_ this leaves the outer loop without a safepoint. | |
| 1123 // | |
| 1124 // entry 0 | |
| 1125 // | | |
| 1126 // v | |
| 1127 // outer 1,2 +->1 | |
| 1128 // | | | |
| 1129 // | v | |
| 1130 // | 2<---+ ncsfpt in 2 | |
| 1131 // |_/|\ | | |
| 1132 // | v | | |
| 1133 // inner 2,3 / 3 | call in 3 | |
| 1134 // / | | | |
| 1135 // v +--+ | |
| 1136 // exit 4 | |
| 1137 // | |
| 1138 // | |
| 1139 // This method creates a list (_required_safept) of ncsfpt nodes that must | |
| 1140 // be protected is created for each loop. When a ncsfpt maybe deleted, it | |
| 1141 // is first looked for in the lists for the outer loops of the current loop. | |
| 1142 // | |
| 1143 // The insights into the problem: | |
| 1144 // A) counted loops are okay | |
| 1145 // B) innermost loops are okay (only an inner loop can delete | |
| 1146 // a ncsfpt needed by an outer loop) | |
| 1147 // C) a loop is immune from an inner loop deleting a safepoint | |
| 1148 // if the loop has a call on the idom-path | |
| 1149 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the | |
| 1150 // idom-path that is not in a nested loop | |
| 1151 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner | |
| 1152 // loop needs to be prevented from deletion by an inner loop | |
| 1153 // | |
| 1154 // There are two analyses: | |
| 1155 // 1) The first, and cheaper one, scans the loop body from | |
| 1156 // tail to head following the idom (immediate dominator) | |
| 1157 // chain, looking for the cases (C,D,E) above. | |
| 1158 // Since inner loops are scanned before outer loops, there is summary | |
| 1159 // information about inner loops. Inner loops can be skipped over | |
| 1160 // when the tail of an inner loop is encountered. | |
| 1161 // | |
| 1162 // 2) The second, invoked if the first fails to find a call or ncsfpt on | |
| 1163 // the idom path (which is rare), scans all predecessor control paths | |
| 1164 // from the tail to the head, terminating a path when a call or sfpt | |
| 1165 // is encountered, to find the ncsfpt's that are closest to the tail. | |
| 1166 // | |
| 1167 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) { | |
| 1168 // Bottom up traversal | |
| 1169 IdealLoopTree* ch = _child; | |
| 1170 while (ch != NULL) { | |
| 1171 ch->check_safepts(visited, stack); | |
| 1172 ch = ch->_next; | |
| 1173 } | |
| 1174 | |
| 1175 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) { | |
| 1176 bool has_call = false; // call on dom-path | |
| 1177 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth | |
| 1178 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth | |
| 1179 // Scan the dom-path nodes from tail to head | |
| 1180 for (Node* n = tail(); n != _head; n = _phase->idom(n)) { | |
| 1181 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { | |
| 1182 has_call = true; | |
| 1183 _has_sfpt = 1; // Then no need for a safept! | |
| 1184 break; | |
| 1185 } else if (n->Opcode() == Op_SafePoint) { | |
| 1186 if (_phase->get_loop(n) == this) { | |
| 1187 has_local_ncsfpt = true; | |
| 1188 break; | |
| 1189 } | |
| 1190 if (nonlocal_ncsfpt == NULL) { | |
| 1191 nonlocal_ncsfpt = n; // save the one closest to the tail | |
| 1192 } | |
| 1193 } else { | |
| 1194 IdealLoopTree* nlpt = _phase->get_loop(n); | |
| 1195 if (this != nlpt) { | |
| 1196 // If at an inner loop tail, see if the inner loop has already | |
| 1197 // recorded seeing a call on the dom-path (and stop.) If not, | |
| 1198 // jump to the head of the inner loop. | |
| 1199 assert(is_member(nlpt), "nested loop"); | |
| 1200 Node* tail = nlpt->_tail; | |
| 1201 if (tail->in(0)->is_If()) tail = tail->in(0); | |
| 1202 if (n == tail) { | |
| 1203 // If inner loop has call on dom-path, so does outer loop | |
| 1204 if (nlpt->_has_sfpt) { | |
| 1205 has_call = true; | |
| 1206 _has_sfpt = 1; | |
| 1207 break; | |
| 1208 } | |
| 1209 // Skip to head of inner loop | |
| 1210 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head"); | |
| 1211 n = nlpt->_head; | |
| 1212 } | |
| 1213 } | |
| 1214 } | |
| 1215 } | |
| 1216 // Record safept's that this loop needs preserved when an | |
| 1217 // inner loop attempts to delete it's safepoints. | |
| 1218 if (_child != NULL && !has_call && !has_local_ncsfpt) { | |
| 1219 if (nonlocal_ncsfpt != NULL) { | |
| 1220 if (_required_safept == NULL) _required_safept = new Node_List(); | |
| 1221 _required_safept->push(nonlocal_ncsfpt); | |
| 1222 } else { | |
| 1223 // Failed to find a suitable safept on the dom-path. Now use | |
| 1224 // an all paths walk from tail to head, looking for safepoints to preserve. | |
| 1225 allpaths_check_safepts(visited, stack); | |
| 1226 } | |
| 1227 } | |
| 1228 } | |
| 1229 } | |
| 1230 | |
| 1231 //---------------------------is_deleteable_safept---------------------------- | |
| 1232 // Is safept not required by an outer loop? | |
| 1233 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) { | |
| 1234 assert(sfpt->Opcode() == Op_SafePoint, ""); | |
| 1235 IdealLoopTree* lp = get_loop(sfpt)->_parent; | |
| 1236 while (lp != NULL) { | |
| 1237 Node_List* sfpts = lp->_required_safept; | |
| 1238 if (sfpts != NULL) { | |
| 1239 for (uint i = 0; i < sfpts->size(); i++) { | |
| 1240 if (sfpt == sfpts->at(i)) | |
| 1241 return false; | |
| 1242 } | |
| 1243 } | |
| 1244 lp = lp->_parent; | |
| 1245 } | |
| 1246 return true; | |
| 1247 } | |
| 1248 | |
| 1249 //------------------------------counted_loop----------------------------------- | |
| 1250 // Convert to counted loops where possible | |
| 1251 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) { | |
| 1252 | |
| 1253 // For grins, set the inner-loop flag here | |
| 1254 if( !_child ) { | |
| 1255 if( _head->is_Loop() ) _head->as_Loop()->set_inner_loop(); | |
| 1256 } | |
| 1257 | |
| 1258 if( _head->is_CountedLoop() || | |
| 1259 phase->is_counted_loop( _head, this ) ) { | |
| 1260 _has_sfpt = 1; // Indicate we do not need a safepoint here | |
| 1261 | |
| 1262 // Look for a safepoint to remove | |
| 1263 for (Node* n = tail(); n != _head; n = phase->idom(n)) | |
| 1264 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this && | |
| 1265 phase->is_deleteable_safept(n)) | |
| 1266 phase->lazy_replace(n,n->in(TypeFunc::Control)); | |
| 1267 | |
| 1268 CountedLoopNode *cl = _head->as_CountedLoop(); | |
| 1269 Node *incr = cl->incr(); | |
| 1270 if( !incr ) return; // Dead loop? | |
| 1271 Node *init = cl->init_trip(); | |
| 1272 Node *phi = cl->phi(); | |
| 1273 // protect against stride not being a constant | |
| 1274 if( !cl->stride_is_con() ) return; | |
| 1275 int stride_con = cl->stride_con(); | |
| 1276 | |
| 1277 // Look for induction variables | |
| 1278 | |
| 1279 // Visit all children, looking for Phis | |
| 1280 for (DUIterator i = cl->outs(); cl->has_out(i); i++) { | |
| 1281 Node *out = cl->out(i); | |
|
367
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1282 if (!out->is_Phi() || out == phi) continue; // Looking for other phis |
| 0 | 1283 PhiNode* phi2 = out->as_Phi(); |
| 1284 Node *incr2 = phi2->in( LoopNode::LoopBackControl ); | |
| 1285 // Look for induction variables of the form: X += constant | |
| 1286 if( phi2->region() != _head || | |
| 1287 incr2->req() != 3 || | |
| 1288 incr2->in(1) != phi2 || | |
| 1289 incr2 == incr || | |
| 1290 incr2->Opcode() != Op_AddI || | |
| 1291 !incr2->in(2)->is_Con() ) | |
| 1292 continue; | |
| 1293 | |
| 1294 // Check for parallel induction variable (parallel to trip counter) | |
| 1295 // via an affine function. In particular, count-down loops with | |
| 1296 // count-up array indices are common. We only RCE references off | |
| 1297 // the trip-counter, so we need to convert all these to trip-counter | |
| 1298 // expressions. | |
| 1299 Node *init2 = phi2->in( LoopNode::EntryControl ); | |
| 1300 int stride_con2 = incr2->in(2)->get_int(); | |
| 1301 | |
| 1302 // The general case here gets a little tricky. We want to find the | |
| 1303 // GCD of all possible parallel IV's and make a new IV using this | |
| 1304 // GCD for the loop. Then all possible IVs are simple multiples of | |
| 1305 // the GCD. In practice, this will cover very few extra loops. | |
| 1306 // Instead we require 'stride_con2' to be a multiple of 'stride_con', | |
| 1307 // where +/-1 is the common case, but other integer multiples are | |
| 1308 // also easy to handle. | |
| 1309 int ratio_con = stride_con2/stride_con; | |
| 1310 | |
| 1311 if( ratio_con * stride_con == stride_con2 ) { // Check for exact | |
| 1312 // Convert to using the trip counter. The parallel induction | |
| 1313 // variable differs from the trip counter by a loop-invariant | |
| 1314 // amount, the difference between their respective initial values. | |
| 1315 // It is scaled by the 'ratio_con'. | |
| 1316 Compile* C = phase->C; | |
| 1317 Node* ratio = phase->_igvn.intcon(ratio_con); | |
| 1318 phase->set_ctrl(ratio, C->root()); | |
| 1319 Node* ratio_init = new (C, 3) MulINode(init, ratio); | |
| 1320 phase->_igvn.register_new_node_with_optimizer(ratio_init, init); | |
| 1321 phase->set_early_ctrl(ratio_init); | |
| 1322 Node* diff = new (C, 3) SubINode(init2, ratio_init); | |
| 1323 phase->_igvn.register_new_node_with_optimizer(diff, init2); | |
| 1324 phase->set_early_ctrl(diff); | |
| 1325 Node* ratio_idx = new (C, 3) MulINode(phi, ratio); | |
| 1326 phase->_igvn.register_new_node_with_optimizer(ratio_idx, phi); | |
| 1327 phase->set_ctrl(ratio_idx, cl); | |
| 1328 Node* add = new (C, 3) AddINode(ratio_idx, diff); | |
| 1329 phase->_igvn.register_new_node_with_optimizer(add); | |
| 1330 phase->set_ctrl(add, cl); | |
| 1331 phase->_igvn.hash_delete( phi2 ); | |
| 1332 phase->_igvn.subsume_node( phi2, add ); | |
| 1333 // Sometimes an induction variable is unused | |
| 1334 if (add->outcnt() == 0) { | |
| 1335 phase->_igvn.remove_dead_node(add); | |
| 1336 } | |
| 1337 --i; // deleted this phi; rescan starting with next position | |
| 1338 continue; | |
| 1339 } | |
| 1340 } | |
| 1341 } else if (_parent != NULL && !_irreducible) { | |
| 1342 // Not a counted loop. | |
| 1343 // Look for a safepoint on the idom-path to remove, preserving the first one | |
| 1344 bool found = false; | |
| 1345 Node* n = tail(); | |
| 1346 for (; n != _head && !found; n = phase->idom(n)) { | |
| 1347 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this) | |
| 1348 found = true; // Found one | |
| 1349 } | |
| 1350 // Skip past it and delete the others | |
| 1351 for (; n != _head; n = phase->idom(n)) { | |
| 1352 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this && | |
| 1353 phase->is_deleteable_safept(n)) | |
| 1354 phase->lazy_replace(n,n->in(TypeFunc::Control)); | |
| 1355 } | |
| 1356 } | |
| 1357 | |
| 1358 // Recursively | |
| 1359 if( _child ) _child->counted_loop( phase ); | |
| 1360 if( _next ) _next ->counted_loop( phase ); | |
| 1361 } | |
| 1362 | |
| 1363 #ifndef PRODUCT | |
| 1364 //------------------------------dump_head-------------------------------------- | |
| 1365 // Dump 1 liner for loop header info | |
| 1366 void IdealLoopTree::dump_head( ) const { | |
| 1367 for( uint i=0; i<_nest; i++ ) | |
| 1368 tty->print(" "); | |
| 1369 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx); | |
| 1370 if( _irreducible ) tty->print(" IRREDUCIBLE"); | |
| 1371 if( _head->is_CountedLoop() ) { | |
| 1372 CountedLoopNode *cl = _head->as_CountedLoop(); | |
| 1373 tty->print(" counted"); | |
| 1374 if( cl->is_pre_loop () ) tty->print(" pre" ); | |
| 1375 if( cl->is_main_loop() ) tty->print(" main"); | |
| 1376 if( cl->is_post_loop() ) tty->print(" post"); | |
| 1377 } | |
| 1378 tty->cr(); | |
| 1379 } | |
| 1380 | |
| 1381 //------------------------------dump------------------------------------------- | |
| 1382 // Dump loops by loop tree | |
| 1383 void IdealLoopTree::dump( ) const { | |
| 1384 dump_head(); | |
| 1385 if( _child ) _child->dump(); | |
| 1386 if( _next ) _next ->dump(); | |
| 1387 } | |
| 1388 | |
| 1389 #endif | |
| 1390 | |
|
367
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1391 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) { |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1392 if (loop == root) { |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1393 if (loop->_child != NULL) { |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1394 log->begin_head("loop_tree"); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1395 log->end_head(); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1396 if( loop->_child ) log_loop_tree(root, loop->_child, log); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1397 log->tail("loop_tree"); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1398 assert(loop->_next == NULL, "what?"); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1399 } |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1400 } else { |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1401 Node* head = loop->_head; |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1402 log->begin_head("loop"); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1403 log->print(" idx='%d' ", head->_idx); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1404 if (loop->_irreducible) log->print("irreducible='1' "); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1405 if (head->is_Loop()) { |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1406 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' "); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1407 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' "); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1408 } |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1409 if (head->is_CountedLoop()) { |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1410 CountedLoopNode* cl = head->as_CountedLoop(); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1411 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx()); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1412 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1413 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx()); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1414 } |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1415 log->end_head(); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1416 if( loop->_child ) log_loop_tree(root, loop->_child, log); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1417 log->tail("loop"); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1418 if( loop->_next ) log_loop_tree(root, loop->_next, log); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1419 } |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1420 } |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1421 |
| 0 | 1422 //============================================================================= |
| 1423 //------------------------------PhaseIdealLoop--------------------------------- | |
| 1424 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to | |
| 1425 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups. | |
| 1426 PhaseIdealLoop::PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me, bool do_split_ifs ) | |
| 1427 : PhaseTransform(Ideal_Loop), | |
| 1428 _igvn(igvn), | |
| 1429 _dom_lca_tags(C->comp_arena()) { | |
| 1430 // Reset major-progress flag for the driver's heuristics | |
| 1431 C->clear_major_progress(); | |
| 1432 | |
| 1433 #ifndef PRODUCT | |
| 1434 // Capture for later assert | |
| 1435 uint unique = C->unique(); | |
| 1436 _loop_invokes++; | |
| 1437 _loop_work += unique; | |
| 1438 #endif | |
| 1439 | |
| 1440 // True if the method has at least 1 irreducible loop | |
| 1441 _has_irreducible_loops = false; | |
| 1442 | |
| 1443 _created_loop_node = false; | |
| 1444 | |
| 1445 Arena *a = Thread::current()->resource_area(); | |
| 1446 VectorSet visited(a); | |
| 1447 // Pre-grow the mapping from Nodes to IdealLoopTrees. | |
| 1448 _nodes.map(C->unique(), NULL); | |
| 1449 memset(_nodes.adr(), 0, wordSize * C->unique()); | |
| 1450 | |
| 1451 // Pre-build the top-level outermost loop tree entry | |
| 1452 _ltree_root = new IdealLoopTree( this, C->root(), C->root() ); | |
| 1453 // Do not need a safepoint at the top level | |
| 1454 _ltree_root->_has_sfpt = 1; | |
| 1455 | |
| 1456 // Empty pre-order array | |
| 1457 allocate_preorders(); | |
| 1458 | |
| 1459 // Build a loop tree on the fly. Build a mapping from CFG nodes to | |
| 1460 // IdealLoopTree entries. Data nodes are NOT walked. | |
| 1461 build_loop_tree(); | |
| 1462 // Check for bailout, and return | |
| 1463 if (C->failing()) { | |
| 1464 return; | |
| 1465 } | |
| 1466 | |
| 1467 // No loops after all | |
| 1468 if( !_ltree_root->_child ) C->set_has_loops(false); | |
| 1469 | |
| 1470 // There should always be an outer loop containing the Root and Return nodes. | |
| 1471 // If not, we have a degenerate empty program. Bail out in this case. | |
| 1472 if (!has_node(C->root())) { | |
| 1473 C->clear_major_progress(); | |
| 1474 C->record_method_not_compilable("empty program detected during loop optimization"); | |
| 1475 return; | |
| 1476 } | |
| 1477 | |
| 1478 // Nothing to do, so get out | |
| 1479 if( !C->has_loops() && !do_split_ifs && !verify_me) { | |
| 1480 _igvn.optimize(); // Cleanup NeverBranches | |
| 1481 return; | |
| 1482 } | |
| 1483 | |
| 1484 // Set loop nesting depth | |
| 1485 _ltree_root->set_nest( 0 ); | |
| 1486 | |
| 1487 // Split shared headers and insert loop landing pads. | |
| 1488 // Do not bother doing this on the Root loop of course. | |
| 1489 if( !verify_me && _ltree_root->_child ) { | |
| 1490 if( _ltree_root->_child->beautify_loops( this ) ) { | |
| 1491 // Re-build loop tree! | |
| 1492 _ltree_root->_child = NULL; | |
| 1493 _nodes.clear(); | |
| 1494 reallocate_preorders(); | |
| 1495 build_loop_tree(); | |
| 1496 // Check for bailout, and return | |
| 1497 if (C->failing()) { | |
| 1498 return; | |
| 1499 } | |
| 1500 // Reset loop nesting depth | |
| 1501 _ltree_root->set_nest( 0 ); | |
| 222 | 1502 |
| 1503 C->print_method("After beautify loops", 3); | |
| 0 | 1504 } |
| 1505 } | |
| 1506 | |
| 1507 // Build Dominators for elision of NULL checks & loop finding. | |
| 1508 // Since nodes do not have a slot for immediate dominator, make | |
| 1509 // a persistant side array for that info indexed on node->_idx. | |
| 1510 _idom_size = C->unique(); | |
| 1511 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size ); | |
| 1512 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size ); | |
| 1513 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth | |
| 1514 memset( _dom_depth, 0, _idom_size * sizeof(uint) ); | |
| 1515 | |
| 1516 Dominators(); | |
| 1517 | |
| 1518 // As a side effect, Dominators removed any unreachable CFG paths | |
| 1519 // into RegionNodes. It doesn't do this test against Root, so | |
| 1520 // we do it here. | |
| 1521 for( uint i = 1; i < C->root()->req(); i++ ) { | |
| 1522 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root? | |
| 1523 _igvn.hash_delete(C->root()); | |
| 1524 C->root()->del_req(i); | |
| 1525 _igvn._worklist.push(C->root()); | |
| 1526 i--; // Rerun same iteration on compressed edges | |
| 1527 } | |
| 1528 } | |
| 1529 | |
| 1530 // Given dominators, try to find inner loops with calls that must | |
| 1531 // always be executed (call dominates loop tail). These loops do | |
| 1532 // not need a seperate safepoint. | |
| 1533 Node_List cisstack(a); | |
| 1534 _ltree_root->check_safepts(visited, cisstack); | |
| 1535 | |
| 1536 // Walk the DATA nodes and place into loops. Find earliest control | |
| 1537 // node. For CFG nodes, the _nodes array starts out and remains | |
| 1538 // holding the associated IdealLoopTree pointer. For DATA nodes, the | |
| 1539 // _nodes array holds the earliest legal controlling CFG node. | |
| 1540 | |
| 1541 // Allocate stack with enough space to avoid frequent realloc | |
| 1542 int stack_size = (C->unique() >> 1) + 16; // (unique>>1)+16 from Java2D stats | |
| 1543 Node_Stack nstack( a, stack_size ); | |
| 1544 | |
| 1545 visited.Clear(); | |
| 1546 Node_List worklist(a); | |
| 1547 // Don't need C->root() on worklist since | |
| 1548 // it will be processed among C->top() inputs | |
| 1549 worklist.push( C->top() ); | |
| 1550 visited.set( C->top()->_idx ); // Set C->top() as visited now | |
| 1551 build_loop_early( visited, worklist, nstack, verify_me ); | |
| 1552 | |
| 1553 // Given early legal placement, try finding counted loops. This placement | |
| 1554 // is good enough to discover most loop invariants. | |
| 1555 if( !verify_me ) | |
| 1556 _ltree_root->counted_loop( this ); | |
| 1557 | |
| 1558 // Find latest loop placement. Find ideal loop placement. | |
| 1559 visited.Clear(); | |
| 1560 init_dom_lca_tags(); | |
| 1561 // Need C->root() on worklist when processing outs | |
| 1562 worklist.push( C->root() ); | |
| 1563 NOT_PRODUCT( C->verify_graph_edges(); ) | |
| 1564 worklist.push( C->top() ); | |
| 1565 build_loop_late( visited, worklist, nstack, verify_me ); | |
| 1566 | |
| 1567 // clear out the dead code | |
| 1568 while(_deadlist.size()) { | |
| 1569 igvn.remove_globally_dead_node(_deadlist.pop()); | |
| 1570 } | |
| 1571 | |
| 1572 #ifndef PRODUCT | |
| 1573 C->verify_graph_edges(); | |
| 1574 if( verify_me ) { // Nested verify pass? | |
| 1575 // Check to see if the verify mode is broken | |
| 1576 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?"); | |
| 1577 return; | |
| 1578 } | |
| 1579 if( VerifyLoopOptimizations ) verify(); | |
| 1580 #endif | |
| 1581 | |
| 1582 if (ReassociateInvariants) { | |
| 1583 // Reassociate invariants and prep for split_thru_phi | |
| 1584 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { | |
| 1585 IdealLoopTree* lpt = iter.current(); | |
| 1586 if (!lpt->is_counted() || !lpt->is_inner()) continue; | |
| 1587 | |
| 1588 lpt->reassociate_invariants(this); | |
| 1589 | |
| 1590 // Because RCE opportunities can be masked by split_thru_phi, | |
| 1591 // look for RCE candidates and inhibit split_thru_phi | |
| 1592 // on just their loop-phi's for this pass of loop opts | |
| 1593 if( SplitIfBlocks && do_split_ifs ) { | |
| 1594 if (lpt->policy_range_check(this)) { | |
|
39
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
17
diff
changeset
|
1595 lpt->_rce_candidate = 1; // = true |
| 0 | 1596 } |
| 1597 } | |
| 1598 } | |
| 1599 } | |
| 1600 | |
| 1601 // Check for aggressive application of split-if and other transforms | |
| 1602 // that require basic-block info (like cloning through Phi's) | |
| 1603 if( SplitIfBlocks && do_split_ifs ) { | |
| 1604 visited.Clear(); | |
| 1605 split_if_with_blocks( visited, nstack ); | |
| 1606 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); ); | |
| 1607 } | |
| 1608 | |
| 1609 // Perform iteration-splitting on inner loops. Split iterations to avoid | |
| 1610 // range checks or one-shot null checks. | |
| 1611 | |
| 1612 // If split-if's didn't hack the graph too bad (no CFG changes) | |
| 1613 // then do loop opts. | |
| 1614 if( C->has_loops() && !C->major_progress() ) { | |
| 1615 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) ); | |
| 1616 _ltree_root->_child->iteration_split( this, worklist ); | |
| 1617 // No verify after peeling! GCM has hoisted code out of the loop. | |
| 1618 // After peeling, the hoisted code could sink inside the peeled area. | |
| 1619 // The peeling code does not try to recompute the best location for | |
| 1620 // all the code before the peeled area, so the verify pass will always | |
| 1621 // complain about it. | |
| 1622 } | |
| 1623 // Do verify graph edges in any case | |
| 1624 NOT_PRODUCT( C->verify_graph_edges(); ); | |
| 1625 | |
| 1626 if( !do_split_ifs ) { | |
| 1627 // We saw major progress in Split-If to get here. We forced a | |
| 1628 // pass with unrolling and not split-if, however more split-if's | |
| 1629 // might make progress. If the unrolling didn't make progress | |
| 1630 // then the major-progress flag got cleared and we won't try | |
| 1631 // another round of Split-If. In particular the ever-common | |
| 1632 // instance-of/check-cast pattern requires at least 2 rounds of | |
| 1633 // Split-If to clear out. | |
| 1634 C->set_major_progress(); | |
| 1635 } | |
| 1636 | |
| 1637 // Repeat loop optimizations if new loops were seen | |
| 1638 if (created_loop_node()) { | |
| 1639 C->set_major_progress(); | |
| 1640 } | |
| 1641 | |
| 1642 // Convert scalar to superword operations | |
| 1643 | |
| 1644 if (UseSuperWord && C->has_loops() && !C->major_progress()) { | |
| 1645 // SuperWord transform | |
| 1646 SuperWord sw(this); | |
| 1647 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { | |
| 1648 IdealLoopTree* lpt = iter.current(); | |
| 1649 if (lpt->is_counted()) { | |
| 1650 sw.transform_loop(lpt); | |
| 1651 } | |
| 1652 } | |
| 1653 } | |
| 1654 | |
| 1655 // Cleanup any modified bits | |
| 1656 _igvn.optimize(); | |
| 1657 | |
|
367
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1658 // disable assert until issue with split_flow_path is resolved (6742111) |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1659 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(), |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1660 // "shouldn't introduce irreducible loops"); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1661 |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1662 if (C->log() != NULL) { |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1663 log_loop_tree(_ltree_root, _ltree_root, C->log()); |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
1664 } |
| 0 | 1665 } |
| 1666 | |
| 1667 #ifndef PRODUCT | |
| 1668 //------------------------------print_statistics------------------------------- | |
| 1669 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes | |
| 1670 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique | |
| 1671 void PhaseIdealLoop::print_statistics() { | |
| 1672 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work); | |
| 1673 } | |
| 1674 | |
| 1675 //------------------------------verify----------------------------------------- | |
| 1676 // Build a verify-only PhaseIdealLoop, and see that it agrees with me. | |
| 1677 static int fail; // debug only, so its multi-thread dont care | |
| 1678 void PhaseIdealLoop::verify() const { | |
| 1679 int old_progress = C->major_progress(); | |
| 1680 ResourceMark rm; | |
| 1681 PhaseIdealLoop loop_verify( _igvn, this, false ); | |
| 1682 VectorSet visited(Thread::current()->resource_area()); | |
| 1683 | |
| 1684 fail = 0; | |
| 1685 verify_compare( C->root(), &loop_verify, visited ); | |
| 1686 assert( fail == 0, "verify loops failed" ); | |
| 1687 // Verify loop structure is the same | |
| 1688 _ltree_root->verify_tree(loop_verify._ltree_root, NULL); | |
| 1689 // Reset major-progress. It was cleared by creating a verify version of | |
| 1690 // PhaseIdealLoop. | |
| 1691 for( int i=0; i<old_progress; i++ ) | |
| 1692 C->set_major_progress(); | |
| 1693 } | |
| 1694 | |
| 1695 //------------------------------verify_compare--------------------------------- | |
| 1696 // Make sure me and the given PhaseIdealLoop agree on key data structures | |
| 1697 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const { | |
| 1698 if( !n ) return; | |
| 1699 if( visited.test_set( n->_idx ) ) return; | |
| 1700 if( !_nodes[n->_idx] ) { // Unreachable | |
| 1701 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" ); | |
| 1702 return; | |
| 1703 } | |
| 1704 | |
| 1705 uint i; | |
| 1706 for( i = 0; i < n->req(); i++ ) | |
| 1707 verify_compare( n->in(i), loop_verify, visited ); | |
| 1708 | |
| 1709 // Check the '_nodes' block/loop structure | |
| 1710 i = n->_idx; | |
| 1711 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl | |
| 1712 if( _nodes[i] != loop_verify->_nodes[i] && | |
| 1713 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) { | |
| 1714 tty->print("Mismatched control setting for: "); | |
| 1715 n->dump(); | |
| 1716 if( fail++ > 10 ) return; | |
| 1717 Node *c = get_ctrl_no_update(n); | |
| 1718 tty->print("We have it as: "); | |
| 1719 if( c->in(0) ) c->dump(); | |
| 1720 else tty->print_cr("N%d",c->_idx); | |
| 1721 tty->print("Verify thinks: "); | |
| 1722 if( loop_verify->has_ctrl(n) ) | |
| 1723 loop_verify->get_ctrl_no_update(n)->dump(); | |
| 1724 else | |
| 1725 loop_verify->get_loop_idx(n)->dump(); | |
| 1726 tty->cr(); | |
| 1727 } | |
| 1728 } else { // We have a loop | |
| 1729 IdealLoopTree *us = get_loop_idx(n); | |
| 1730 if( loop_verify->has_ctrl(n) ) { | |
| 1731 tty->print("Mismatched loop setting for: "); | |
| 1732 n->dump(); | |
| 1733 if( fail++ > 10 ) return; | |
| 1734 tty->print("We have it as: "); | |
| 1735 us->dump(); | |
| 1736 tty->print("Verify thinks: "); | |
| 1737 loop_verify->get_ctrl_no_update(n)->dump(); | |
| 1738 tty->cr(); | |
| 1739 } else if (!C->major_progress()) { | |
| 1740 // Loop selection can be messed up if we did a major progress | |
| 1741 // operation, like split-if. Do not verify in that case. | |
| 1742 IdealLoopTree *them = loop_verify->get_loop_idx(n); | |
| 1743 if( us->_head != them->_head || us->_tail != them->_tail ) { | |
| 1744 tty->print("Unequals loops for: "); | |
| 1745 n->dump(); | |
| 1746 if( fail++ > 10 ) return; | |
| 1747 tty->print("We have it as: "); | |
| 1748 us->dump(); | |
| 1749 tty->print("Verify thinks: "); | |
| 1750 them->dump(); | |
| 1751 tty->cr(); | |
| 1752 } | |
| 1753 } | |
| 1754 } | |
| 1755 | |
| 1756 // Check for immediate dominators being equal | |
| 1757 if( i >= _idom_size ) { | |
| 1758 if( !n->is_CFG() ) return; | |
| 1759 tty->print("CFG Node with no idom: "); | |
| 1760 n->dump(); | |
| 1761 return; | |
| 1762 } | |
| 1763 if( !n->is_CFG() ) return; | |
| 1764 if( n == C->root() ) return; // No IDOM here | |
| 1765 | |
| 1766 assert(n->_idx == i, "sanity"); | |
| 1767 Node *id = idom_no_update(n); | |
| 1768 if( id != loop_verify->idom_no_update(n) ) { | |
| 1769 tty->print("Unequals idoms for: "); | |
| 1770 n->dump(); | |
| 1771 if( fail++ > 10 ) return; | |
| 1772 tty->print("We have it as: "); | |
| 1773 id->dump(); | |
| 1774 tty->print("Verify thinks: "); | |
| 1775 loop_verify->idom_no_update(n)->dump(); | |
| 1776 tty->cr(); | |
| 1777 } | |
| 1778 | |
| 1779 } | |
| 1780 | |
| 1781 //------------------------------verify_tree------------------------------------ | |
| 1782 // Verify that tree structures match. Because the CFG can change, siblings | |
| 1783 // within the loop tree can be reordered. We attempt to deal with that by | |
| 1784 // reordering the verify's loop tree if possible. | |
| 1785 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const { | |
| 1786 assert( _parent == parent, "Badly formed loop tree" ); | |
| 1787 | |
| 1788 // Siblings not in same order? Attempt to re-order. | |
| 1789 if( _head != loop->_head ) { | |
| 1790 // Find _next pointer to update | |
| 1791 IdealLoopTree **pp = &loop->_parent->_child; | |
| 1792 while( *pp != loop ) | |
| 1793 pp = &((*pp)->_next); | |
| 1794 // Find proper sibling to be next | |
| 1795 IdealLoopTree **nn = &loop->_next; | |
| 1796 while( (*nn) && (*nn)->_head != _head ) | |
| 1797 nn = &((*nn)->_next); | |
| 1798 | |
| 1799 // Check for no match. | |
| 1800 if( !(*nn) ) { | |
| 1801 // Annoyingly, irreducible loops can pick different headers | |
| 1802 // after a major_progress operation, so the rest of the loop | |
| 1803 // tree cannot be matched. | |
| 1804 if (_irreducible && Compile::current()->major_progress()) return; | |
| 1805 assert( 0, "failed to match loop tree" ); | |
| 1806 } | |
| 1807 | |
| 1808 // Move (*nn) to (*pp) | |
| 1809 IdealLoopTree *hit = *nn; | |
| 1810 *nn = hit->_next; | |
| 1811 hit->_next = loop; | |
| 1812 *pp = loop; | |
| 1813 loop = hit; | |
| 1814 // Now try again to verify | |
| 1815 } | |
| 1816 | |
| 1817 assert( _head == loop->_head , "mismatched loop head" ); | |
| 1818 Node *tail = _tail; // Inline a non-updating version of | |
| 1819 while( !tail->in(0) ) // the 'tail()' call. | |
| 1820 tail = tail->in(1); | |
| 1821 assert( tail == loop->_tail, "mismatched loop tail" ); | |
| 1822 | |
| 1823 // Counted loops that are guarded should be able to find their guards | |
| 1824 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) { | |
| 1825 CountedLoopNode *cl = _head->as_CountedLoop(); | |
| 1826 Node *init = cl->init_trip(); | |
| 1827 Node *ctrl = cl->in(LoopNode::EntryControl); | |
| 1828 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); | |
| 1829 Node *iff = ctrl->in(0); | |
| 1830 assert( iff->Opcode() == Op_If, "" ); | |
| 1831 Node *bol = iff->in(1); | |
| 1832 assert( bol->Opcode() == Op_Bool, "" ); | |
| 1833 Node *cmp = bol->in(1); | |
| 1834 assert( cmp->Opcode() == Op_CmpI, "" ); | |
| 1835 Node *add = cmp->in(1); | |
| 1836 Node *opaq; | |
| 1837 if( add->Opcode() == Op_Opaque1 ) { | |
| 1838 opaq = add; | |
| 1839 } else { | |
| 1840 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" ); | |
| 1841 assert( add == init, "" ); | |
| 1842 opaq = cmp->in(2); | |
| 1843 } | |
| 1844 assert( opaq->Opcode() == Op_Opaque1, "" ); | |
| 1845 | |
| 1846 } | |
| 1847 | |
| 1848 if (_child != NULL) _child->verify_tree(loop->_child, this); | |
| 1849 if (_next != NULL) _next ->verify_tree(loop->_next, parent); | |
| 1850 // Innermost loops need to verify loop bodies, | |
| 1851 // but only if no 'major_progress' | |
| 1852 int fail = 0; | |
| 1853 if (!Compile::current()->major_progress() && _child == NULL) { | |
| 1854 for( uint i = 0; i < _body.size(); i++ ) { | |
| 1855 Node *n = _body.at(i); | |
| 1856 if (n->outcnt() == 0) continue; // Ignore dead | |
| 1857 uint j; | |
| 1858 for( j = 0; j < loop->_body.size(); j++ ) | |
| 1859 if( loop->_body.at(j) == n ) | |
| 1860 break; | |
| 1861 if( j == loop->_body.size() ) { // Not found in loop body | |
| 1862 // Last ditch effort to avoid assertion: Its possible that we | |
| 1863 // have some users (so outcnt not zero) but are still dead. | |
| 1864 // Try to find from root. | |
| 1865 if (Compile::current()->root()->find(n->_idx)) { | |
| 1866 fail++; | |
| 1867 tty->print("We have that verify does not: "); | |
| 1868 n->dump(); | |
| 1869 } | |
| 1870 } | |
| 1871 } | |
| 1872 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) { | |
| 1873 Node *n = loop->_body.at(i2); | |
| 1874 if (n->outcnt() == 0) continue; // Ignore dead | |
| 1875 uint j; | |
| 1876 for( j = 0; j < _body.size(); j++ ) | |
| 1877 if( _body.at(j) == n ) | |
| 1878 break; | |
| 1879 if( j == _body.size() ) { // Not found in loop body | |
| 1880 // Last ditch effort to avoid assertion: Its possible that we | |
| 1881 // have some users (so outcnt not zero) but are still dead. | |
| 1882 // Try to find from root. | |
| 1883 if (Compile::current()->root()->find(n->_idx)) { | |
| 1884 fail++; | |
| 1885 tty->print("Verify has that we do not: "); | |
| 1886 n->dump(); | |
| 1887 } | |
| 1888 } | |
| 1889 } | |
| 1890 assert( !fail, "loop body mismatch" ); | |
| 1891 } | |
| 1892 } | |
| 1893 | |
| 1894 #endif | |
| 1895 | |
| 1896 //------------------------------set_idom--------------------------------------- | |
| 1897 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) { | |
| 1898 uint idx = d->_idx; | |
| 1899 if (idx >= _idom_size) { | |
| 1900 uint newsize = _idom_size<<1; | |
| 1901 while( idx >= newsize ) { | |
| 1902 newsize <<= 1; | |
| 1903 } | |
| 1904 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize); | |
| 1905 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize); | |
| 1906 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) ); | |
| 1907 _idom_size = newsize; | |
| 1908 } | |
| 1909 _idom[idx] = n; | |
| 1910 _dom_depth[idx] = dom_depth; | |
| 1911 } | |
| 1912 | |
| 1913 //------------------------------recompute_dom_depth--------------------------------------- | |
| 1914 // The dominator tree is constructed with only parent pointers. | |
| 1915 // This recomputes the depth in the tree by first tagging all | |
| 1916 // nodes as "no depth yet" marker. The next pass then runs up | |
| 1917 // the dom tree from each node marked "no depth yet", and computes | |
| 1918 // the depth on the way back down. | |
| 1919 void PhaseIdealLoop::recompute_dom_depth() { | |
| 1920 uint no_depth_marker = C->unique(); | |
| 1921 uint i; | |
| 1922 // Initialize depth to "no depth yet" | |
| 1923 for (i = 0; i < _idom_size; i++) { | |
| 1924 if (_dom_depth[i] > 0 && _idom[i] != NULL) { | |
| 1925 _dom_depth[i] = no_depth_marker; | |
| 1926 } | |
| 1927 } | |
| 1928 if (_dom_stk == NULL) { | |
| 1929 uint init_size = C->unique() / 100; // Guess that 1/100 is a reasonable initial size. | |
| 1930 if (init_size < 10) init_size = 10; | |
| 1931 _dom_stk = new (C->node_arena()) GrowableArray<uint>(C->node_arena(), init_size, 0, 0); | |
| 1932 } | |
| 1933 // Compute new depth for each node. | |
| 1934 for (i = 0; i < _idom_size; i++) { | |
| 1935 uint j = i; | |
| 1936 // Run up the dom tree to find a node with a depth | |
| 1937 while (_dom_depth[j] == no_depth_marker) { | |
| 1938 _dom_stk->push(j); | |
| 1939 j = _idom[j]->_idx; | |
| 1940 } | |
| 1941 // Compute the depth on the way back down this tree branch | |
| 1942 uint dd = _dom_depth[j] + 1; | |
| 1943 while (_dom_stk->length() > 0) { | |
| 1944 uint j = _dom_stk->pop(); | |
| 1945 _dom_depth[j] = dd; | |
| 1946 dd++; | |
| 1947 } | |
| 1948 } | |
| 1949 } | |
| 1950 | |
| 1951 //------------------------------sort------------------------------------------- | |
| 1952 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the | |
| 1953 // loop tree, not the root. | |
| 1954 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) { | |
| 1955 if( !innermost ) return loop; // New innermost loop | |
| 1956 | |
| 1957 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number | |
| 1958 assert( loop_preorder, "not yet post-walked loop" ); | |
| 1959 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer | |
| 1960 IdealLoopTree *l = *pp; // Do I go before or after 'l'? | |
| 1961 | |
| 1962 // Insert at start of list | |
| 1963 while( l ) { // Insertion sort based on pre-order | |
| 1964 if( l == loop ) return innermost; // Already on list! | |
| 1965 int l_preorder = get_preorder(l->_head); // Cache pre-order number | |
| 1966 assert( l_preorder, "not yet post-walked l" ); | |
| 1967 // Check header pre-order number to figure proper nesting | |
| 1968 if( loop_preorder > l_preorder ) | |
| 1969 break; // End of insertion | |
| 1970 // If headers tie (e.g., shared headers) check tail pre-order numbers. | |
| 1971 // Since I split shared headers, you'd think this could not happen. | |
| 1972 // BUT: I must first do the preorder numbering before I can discover I | |
| 1973 // have shared headers, so the split headers all get the same preorder | |
| 1974 // number as the RegionNode they split from. | |
| 1975 if( loop_preorder == l_preorder && | |
| 1976 get_preorder(loop->_tail) < get_preorder(l->_tail) ) | |
| 1977 break; // Also check for shared headers (same pre#) | |
| 1978 pp = &l->_parent; // Chain up list | |
| 1979 l = *pp; | |
| 1980 } | |
| 1981 // Link into list | |
| 1982 // Point predecessor to me | |
| 1983 *pp = loop; | |
| 1984 // Point me to successor | |
| 1985 IdealLoopTree *p = loop->_parent; | |
| 1986 loop->_parent = l; // Point me to successor | |
| 1987 if( p ) sort( p, innermost ); // Insert my parents into list as well | |
| 1988 return innermost; | |
| 1989 } | |
| 1990 | |
| 1991 //------------------------------build_loop_tree-------------------------------- | |
| 1992 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit | |
| 1993 // bits. The _nodes[] array is mapped by Node index and holds a NULL for | |
| 1994 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the | |
| 1995 // tightest enclosing IdealLoopTree for post-walked. | |
| 1996 // | |
| 1997 // During my forward walk I do a short 1-layer lookahead to see if I can find | |
| 1998 // a loop backedge with that doesn't have any work on the backedge. This | |
| 1999 // helps me construct nested loops with shared headers better. | |
| 2000 // | |
| 2001 // Once I've done the forward recursion, I do the post-work. For each child | |
| 2002 // I check to see if there is a backedge. Backedges define a loop! I | |
| 2003 // insert an IdealLoopTree at the target of the backedge. | |
| 2004 // | |
| 2005 // During the post-work I also check to see if I have several children | |
| 2006 // belonging to different loops. If so, then this Node is a decision point | |
| 2007 // where control flow can choose to change loop nests. It is at this | |
| 2008 // decision point where I can figure out how loops are nested. At this | |
| 2009 // time I can properly order the different loop nests from my children. | |
| 2010 // Note that there may not be any backedges at the decision point! | |
| 2011 // | |
| 2012 // Since the decision point can be far removed from the backedges, I can't | |
| 2013 // order my loops at the time I discover them. Thus at the decision point | |
| 2014 // I need to inspect loop header pre-order numbers to properly nest my | |
| 2015 // loops. This means I need to sort my childrens' loops by pre-order. | |
| 2016 // The sort is of size number-of-control-children, which generally limits | |
| 2017 // it to size 2 (i.e., I just choose between my 2 target loops). | |
| 2018 void PhaseIdealLoop::build_loop_tree() { | |
| 2019 // Allocate stack of size C->unique()/2 to avoid frequent realloc | |
| 2020 GrowableArray <Node *> bltstack(C->unique() >> 1); | |
| 2021 Node *n = C->root(); | |
| 2022 bltstack.push(n); | |
| 2023 int pre_order = 1; | |
| 2024 int stack_size; | |
| 2025 | |
| 2026 while ( ( stack_size = bltstack.length() ) != 0 ) { | |
| 2027 n = bltstack.top(); // Leave node on stack | |
| 2028 if ( !is_visited(n) ) { | |
| 2029 // ---- Pre-pass Work ---- | |
| 2030 // Pre-walked but not post-walked nodes need a pre_order number. | |
| 2031 | |
| 2032 set_preorder_visited( n, pre_order ); // set as visited | |
| 2033 | |
| 2034 // ---- Scan over children ---- | |
| 2035 // Scan first over control projections that lead to loop headers. | |
| 2036 // This helps us find inner-to-outer loops with shared headers better. | |
| 2037 | |
| 2038 // Scan children's children for loop headers. | |
| 2039 for ( int i = n->outcnt() - 1; i >= 0; --i ) { | |
| 2040 Node* m = n->raw_out(i); // Child | |
| 2041 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children | |
| 2042 // Scan over children's children to find loop | |
| 2043 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { | |
| 2044 Node* l = m->fast_out(j); | |
| 2045 if( is_visited(l) && // Been visited? | |
| 2046 !is_postvisited(l) && // But not post-visited | |
| 2047 get_preorder(l) < pre_order ) { // And smaller pre-order | |
| 2048 // Found! Scan the DFS down this path before doing other paths | |
| 2049 bltstack.push(m); | |
| 2050 break; | |
| 2051 } | |
| 2052 } | |
| 2053 } | |
| 2054 } | |
| 2055 pre_order++; | |
| 2056 } | |
| 2057 else if ( !is_postvisited(n) ) { | |
| 2058 // Note: build_loop_tree_impl() adds out edges on rare occasions, | |
| 2059 // such as com.sun.rsasign.am::a. | |
| 2060 // For non-recursive version, first, process current children. | |
| 2061 // On next iteration, check if additional children were added. | |
| 2062 for ( int k = n->outcnt() - 1; k >= 0; --k ) { | |
| 2063 Node* u = n->raw_out(k); | |
| 2064 if ( u->is_CFG() && !is_visited(u) ) { | |
| 2065 bltstack.push(u); | |
| 2066 } | |
| 2067 } | |
| 2068 if ( bltstack.length() == stack_size ) { | |
| 2069 // There were no additional children, post visit node now | |
| 2070 (void)bltstack.pop(); // Remove node from stack | |
| 2071 pre_order = build_loop_tree_impl( n, pre_order ); | |
| 2072 // Check for bailout | |
| 2073 if (C->failing()) { | |
| 2074 return; | |
| 2075 } | |
| 2076 // Check to grow _preorders[] array for the case when | |
| 2077 // build_loop_tree_impl() adds new nodes. | |
| 2078 check_grow_preorders(); | |
| 2079 } | |
| 2080 } | |
| 2081 else { | |
| 2082 (void)bltstack.pop(); // Remove post-visited node from stack | |
| 2083 } | |
| 2084 } | |
| 2085 } | |
| 2086 | |
| 2087 //------------------------------build_loop_tree_impl--------------------------- | |
| 2088 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) { | |
| 2089 // ---- Post-pass Work ---- | |
| 2090 // Pre-walked but not post-walked nodes need a pre_order number. | |
| 2091 | |
| 2092 // Tightest enclosing loop for this Node | |
| 2093 IdealLoopTree *innermost = NULL; | |
| 2094 | |
| 2095 // For all children, see if any edge is a backedge. If so, make a loop | |
| 2096 // for it. Then find the tightest enclosing loop for the self Node. | |
| 2097 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { | |
| 2098 Node* m = n->fast_out(i); // Child | |
| 2099 if( n == m ) continue; // Ignore control self-cycles | |
| 2100 if( !m->is_CFG() ) continue;// Ignore non-CFG edges | |
| 2101 | |
| 2102 IdealLoopTree *l; // Child's loop | |
| 2103 if( !is_postvisited(m) ) { // Child visited but not post-visited? | |
| 2104 // Found a backedge | |
| 2105 assert( get_preorder(m) < pre_order, "should be backedge" ); | |
| 2106 // Check for the RootNode, which is already a LoopNode and is allowed | |
| 2107 // to have multiple "backedges". | |
| 2108 if( m == C->root()) { // Found the root? | |
| 2109 l = _ltree_root; // Root is the outermost LoopNode | |
| 2110 } else { // Else found a nested loop | |
| 2111 // Insert a LoopNode to mark this loop. | |
| 2112 l = new IdealLoopTree(this, m, n); | |
| 2113 } // End of Else found a nested loop | |
| 2114 if( !has_loop(m) ) // If 'm' does not already have a loop set | |
| 2115 set_loop(m, l); // Set loop header to loop now | |
| 2116 | |
| 2117 } else { // Else not a nested loop | |
| 2118 if( !_nodes[m->_idx] ) continue; // Dead code has no loop | |
| 2119 l = get_loop(m); // Get previously determined loop | |
| 2120 // If successor is header of a loop (nest), move up-loop till it | |
| 2121 // is a member of some outer enclosing loop. Since there are no | |
| 2122 // shared headers (I've split them already) I only need to go up | |
| 2123 // at most 1 level. | |
| 2124 while( l && l->_head == m ) // Successor heads loop? | |
| 2125 l = l->_parent; // Move up 1 for me | |
| 2126 // If this loop is not properly parented, then this loop | |
| 2127 // has no exit path out, i.e. its an infinite loop. | |
| 2128 if( !l ) { | |
| 2129 // Make loop "reachable" from root so the CFG is reachable. Basically | |
| 2130 // insert a bogus loop exit that is never taken. 'm', the loop head, | |
| 2131 // points to 'n', one (of possibly many) fall-in paths. There may be | |
| 2132 // many backedges as well. | |
| 2133 | |
| 2134 // Here I set the loop to be the root loop. I could have, after | |
| 2135 // inserting a bogus loop exit, restarted the recursion and found my | |
| 2136 // new loop exit. This would make the infinite loop a first-class | |
| 2137 // loop and it would then get properly optimized. What's the use of | |
| 2138 // optimizing an infinite loop? | |
| 2139 l = _ltree_root; // Oops, found infinite loop | |
| 2140 | |
| 2141 // Insert the NeverBranch between 'm' and it's control user. | |
| 2142 NeverBranchNode *iff = new (C, 1) NeverBranchNode( m ); | |
| 2143 _igvn.register_new_node_with_optimizer(iff); | |
| 2144 set_loop(iff, l); | |
| 2145 Node *if_t = new (C, 1) CProjNode( iff, 0 ); | |
| 2146 _igvn.register_new_node_with_optimizer(if_t); | |
| 2147 set_loop(if_t, l); | |
| 2148 | |
| 2149 Node* cfg = NULL; // Find the One True Control User of m | |
| 2150 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { | |
| 2151 Node* x = m->fast_out(j); | |
| 2152 if (x->is_CFG() && x != m && x != iff) | |
| 2153 { cfg = x; break; } | |
| 2154 } | |
| 2155 assert(cfg != NULL, "must find the control user of m"); | |
| 2156 uint k = 0; // Probably cfg->in(0) | |
| 2157 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region | |
| 2158 cfg->set_req( k, if_t ); // Now point to NeverBranch | |
| 2159 | |
| 2160 // Now create the never-taken loop exit | |
| 2161 Node *if_f = new (C, 1) CProjNode( iff, 1 ); | |
| 2162 _igvn.register_new_node_with_optimizer(if_f); | |
| 2163 set_loop(if_f, l); | |
| 2164 // Find frame ptr for Halt. Relies on the optimizer | |
| 2165 // V-N'ing. Easier and quicker than searching through | |
| 2166 // the program structure. | |
| 2167 Node *frame = new (C, 1) ParmNode( C->start(), TypeFunc::FramePtr ); | |
| 2168 _igvn.register_new_node_with_optimizer(frame); | |
| 2169 // Halt & Catch Fire | |
| 2170 Node *halt = new (C, TypeFunc::Parms) HaltNode( if_f, frame ); | |
| 2171 _igvn.register_new_node_with_optimizer(halt); | |
| 2172 set_loop(halt, l); | |
| 2173 C->root()->add_req(halt); | |
| 2174 set_loop(C->root(), _ltree_root); | |
| 2175 } | |
| 2176 } | |
| 2177 // Weeny check for irreducible. This child was already visited (this | |
| 2178 // IS the post-work phase). Is this child's loop header post-visited | |
| 2179 // as well? If so, then I found another entry into the loop. | |
| 2180 while( is_postvisited(l->_head) ) { | |
| 2181 // found irreducible | |
|
39
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
17
diff
changeset
|
2182 l->_irreducible = 1; // = true |
| 0 | 2183 l = l->_parent; |
| 2184 _has_irreducible_loops = true; | |
| 2185 // Check for bad CFG here to prevent crash, and bailout of compile | |
| 2186 if (l == NULL) { | |
| 2187 C->record_method_not_compilable("unhandled CFG detected during loop optimization"); | |
| 2188 return pre_order; | |
| 2189 } | |
| 2190 } | |
| 2191 | |
| 2192 // This Node might be a decision point for loops. It is only if | |
| 2193 // it's children belong to several different loops. The sort call | |
| 2194 // does a trivial amount of work if there is only 1 child or all | |
| 2195 // children belong to the same loop. If however, the children | |
| 2196 // belong to different loops, the sort call will properly set the | |
| 2197 // _parent pointers to show how the loops nest. | |
| 2198 // | |
| 2199 // In any case, it returns the tightest enclosing loop. | |
| 2200 innermost = sort( l, innermost ); | |
| 2201 } | |
| 2202 | |
| 2203 // Def-use info will have some dead stuff; dead stuff will have no | |
| 2204 // loop decided on. | |
| 2205 | |
| 2206 // Am I a loop header? If so fix up my parent's child and next ptrs. | |
| 2207 if( innermost && innermost->_head == n ) { | |
| 2208 assert( get_loop(n) == innermost, "" ); | |
| 2209 IdealLoopTree *p = innermost->_parent; | |
| 2210 IdealLoopTree *l = innermost; | |
| 2211 while( p && l->_head == n ) { | |
| 2212 l->_next = p->_child; // Put self on parents 'next child' | |
| 2213 p->_child = l; // Make self as first child of parent | |
| 2214 l = p; // Now walk up the parent chain | |
| 2215 p = l->_parent; | |
| 2216 } | |
| 2217 } else { | |
| 2218 // Note that it is possible for a LoopNode to reach here, if the | |
| 2219 // backedge has been made unreachable (hence the LoopNode no longer | |
| 2220 // denotes a Loop, and will eventually be removed). | |
| 2221 | |
| 2222 // Record tightest enclosing loop for self. Mark as post-visited. | |
| 2223 set_loop(n, innermost); | |
| 2224 // Also record has_call flag early on | |
| 2225 if( innermost ) { | |
| 2226 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) { | |
| 2227 // Do not count uncommon calls | |
| 2228 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) { | |
| 2229 Node *iff = n->in(0)->in(0); | |
| 2230 if( !iff->is_If() || | |
| 2231 (n->in(0)->Opcode() == Op_IfFalse && | |
| 2232 (1.0 - iff->as_If()->_prob) >= 0.01) || | |
| 2233 (iff->as_If()->_prob >= 0.01) ) | |
| 2234 innermost->_has_call = 1; | |
| 2235 } | |
|
39
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
17
diff
changeset
|
2236 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) { |
|
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
17
diff
changeset
|
2237 // Disable loop optimizations if the loop has a scalar replaceable |
|
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
17
diff
changeset
|
2238 // allocation. This disabling may cause a potential performance lost |
|
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
17
diff
changeset
|
2239 // if the allocation is not eliminated for some reason. |
|
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
17
diff
changeset
|
2240 innermost->_allow_optimizations = false; |
|
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
17
diff
changeset
|
2241 innermost->_has_call = 1; // = true |
| 0 | 2242 } |
| 2243 } | |
| 2244 } | |
| 2245 | |
| 2246 // Flag as post-visited now | |
| 2247 set_postvisited(n); | |
| 2248 return pre_order; | |
| 2249 } | |
| 2250 | |
| 2251 | |
| 2252 //------------------------------build_loop_early------------------------------- | |
| 2253 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. | |
| 2254 // First pass computes the earliest controlling node possible. This is the | |
| 2255 // controlling input with the deepest dominating depth. | |
| 2256 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack, const PhaseIdealLoop *verify_me ) { | |
| 2257 while (worklist.size() != 0) { | |
| 2258 // Use local variables nstack_top_n & nstack_top_i to cache values | |
| 2259 // on nstack's top. | |
| 2260 Node *nstack_top_n = worklist.pop(); | |
| 2261 uint nstack_top_i = 0; | |
| 2262 //while_nstack_nonempty: | |
| 2263 while (true) { | |
| 2264 // Get parent node and next input's index from stack's top. | |
| 2265 Node *n = nstack_top_n; | |
| 2266 uint i = nstack_top_i; | |
| 2267 uint cnt = n->req(); // Count of inputs | |
| 2268 if (i == 0) { // Pre-process the node. | |
| 2269 if( has_node(n) && // Have either loop or control already? | |
| 2270 !has_ctrl(n) ) { // Have loop picked out already? | |
| 2271 // During "merge_many_backedges" we fold up several nested loops | |
| 2272 // into a single loop. This makes the members of the original | |
| 2273 // loop bodies pointing to dead loops; they need to move up | |
| 2274 // to the new UNION'd larger loop. I set the _head field of these | |
| 2275 // dead loops to NULL and the _parent field points to the owning | |
| 2276 // loop. Shades of UNION-FIND algorithm. | |
| 2277 IdealLoopTree *ilt; | |
| 2278 while( !(ilt = get_loop(n))->_head ) { | |
| 2279 // Normally I would use a set_loop here. But in this one special | |
| 2280 // case, it is legal (and expected) to change what loop a Node | |
| 2281 // belongs to. | |
| 2282 _nodes.map(n->_idx, (Node*)(ilt->_parent) ); | |
| 2283 } | |
| 2284 // Remove safepoints ONLY if I've already seen I don't need one. | |
| 2285 // (the old code here would yank a 2nd safepoint after seeing a | |
| 2286 // first one, even though the 1st did not dominate in the loop body | |
| 2287 // and thus could be avoided indefinitely) | |
| 2288 if( !verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint && | |
| 2289 is_deleteable_safept(n)) { | |
| 2290 Node *in = n->in(TypeFunc::Control); | |
| 2291 lazy_replace(n,in); // Pull safepoint now | |
| 2292 // Carry on with the recursion "as if" we are walking | |
| 2293 // only the control input | |
| 2294 if( !visited.test_set( in->_idx ) ) { | |
| 2295 worklist.push(in); // Visit this guy later, using worklist | |
| 2296 } | |
| 2297 // Get next node from nstack: | |
| 2298 // - skip n's inputs processing by setting i > cnt; | |
| 2299 // - we also will not call set_early_ctrl(n) since | |
| 2300 // has_node(n) == true (see the condition above). | |
| 2301 i = cnt + 1; | |
| 2302 } | |
| 2303 } | |
| 2304 } // if (i == 0) | |
| 2305 | |
| 2306 // Visit all inputs | |
| 2307 bool done = true; // Assume all n's inputs will be processed | |
| 2308 while (i < cnt) { | |
| 2309 Node *in = n->in(i); | |
| 2310 ++i; | |
| 2311 if (in == NULL) continue; | |
| 2312 if (in->pinned() && !in->is_CFG()) | |
| 2313 set_ctrl(in, in->in(0)); | |
| 2314 int is_visited = visited.test_set( in->_idx ); | |
| 2315 if (!has_node(in)) { // No controlling input yet? | |
| 2316 assert( !in->is_CFG(), "CFG Node with no controlling input?" ); | |
| 2317 assert( !is_visited, "visit only once" ); | |
| 2318 nstack.push(n, i); // Save parent node and next input's index. | |
| 2319 nstack_top_n = in; // Process current input now. | |
| 2320 nstack_top_i = 0; | |
| 2321 done = false; // Not all n's inputs processed. | |
| 2322 break; // continue while_nstack_nonempty; | |
| 2323 } else if (!is_visited) { | |
| 2324 // This guy has a location picked out for him, but has not yet | |
| 2325 // been visited. Happens to all CFG nodes, for instance. | |
| 2326 // Visit him using the worklist instead of recursion, to break | |
| 2327 // cycles. Since he has a location already we do not need to | |
| 2328 // find his location before proceeding with the current Node. | |
| 2329 worklist.push(in); // Visit this guy later, using worklist | |
| 2330 } | |
| 2331 } | |
| 2332 if (done) { | |
| 2333 // All of n's inputs have been processed, complete post-processing. | |
| 2334 | |
| 2335 // Compute earilest point this Node can go. | |
| 2336 // CFG, Phi, pinned nodes already know their controlling input. | |
| 2337 if (!has_node(n)) { | |
| 2338 // Record earliest legal location | |
| 2339 set_early_ctrl( n ); | |
| 2340 } | |
| 2341 if (nstack.is_empty()) { | |
| 2342 // Finished all nodes on stack. | |
| 2343 // Process next node on the worklist. | |
| 2344 break; | |
| 2345 } | |
| 2346 // Get saved parent node and next input's index. | |
| 2347 nstack_top_n = nstack.node(); | |
| 2348 nstack_top_i = nstack.index(); | |
| 2349 nstack.pop(); | |
| 2350 } | |
| 2351 } // while (true) | |
| 2352 } | |
| 2353 } | |
| 2354 | |
| 2355 //------------------------------dom_lca_internal-------------------------------- | |
| 2356 // Pair-wise LCA | |
| 2357 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const { | |
| 2358 if( !n1 ) return n2; // Handle NULL original LCA | |
| 2359 assert( n1->is_CFG(), "" ); | |
| 2360 assert( n2->is_CFG(), "" ); | |
| 2361 // find LCA of all uses | |
| 2362 uint d1 = dom_depth(n1); | |
| 2363 uint d2 = dom_depth(n2); | |
| 2364 while (n1 != n2) { | |
| 2365 if (d1 > d2) { | |
| 2366 n1 = idom(n1); | |
| 2367 d1 = dom_depth(n1); | |
| 2368 } else if (d1 < d2) { | |
| 2369 n2 = idom(n2); | |
| 2370 d2 = dom_depth(n2); | |
| 2371 } else { | |
| 2372 // Here d1 == d2. Due to edits of the dominator-tree, sections | |
| 2373 // of the tree might have the same depth. These sections have | |
| 2374 // to be searched more carefully. | |
| 2375 | |
| 2376 // Scan up all the n1's with equal depth, looking for n2. | |
| 2377 Node *t1 = idom(n1); | |
| 2378 while (dom_depth(t1) == d1) { | |
| 2379 if (t1 == n2) return n2; | |
| 2380 t1 = idom(t1); | |
| 2381 } | |
| 2382 // Scan up all the n2's with equal depth, looking for n1. | |
| 2383 Node *t2 = idom(n2); | |
| 2384 while (dom_depth(t2) == d2) { | |
| 2385 if (t2 == n1) return n1; | |
| 2386 t2 = idom(t2); | |
| 2387 } | |
| 2388 // Move up to a new dominator-depth value as well as up the dom-tree. | |
| 2389 n1 = t1; | |
| 2390 n2 = t2; | |
| 2391 d1 = dom_depth(n1); | |
| 2392 d2 = dom_depth(n2); | |
| 2393 } | |
| 2394 } | |
| 2395 return n1; | |
| 2396 } | |
| 2397 | |
| 2398 //------------------------------compute_idom----------------------------------- | |
| 2399 // Locally compute IDOM using dom_lca call. Correct only if the incoming | |
| 2400 // IDOMs are correct. | |
| 2401 Node *PhaseIdealLoop::compute_idom( Node *region ) const { | |
| 2402 assert( region->is_Region(), "" ); | |
| 2403 Node *LCA = NULL; | |
| 2404 for( uint i = 1; i < region->req(); i++ ) { | |
| 2405 if( region->in(i) != C->top() ) | |
| 2406 LCA = dom_lca( LCA, region->in(i) ); | |
| 2407 } | |
| 2408 return LCA; | |
| 2409 } | |
| 2410 | |
| 2411 //------------------------------get_late_ctrl---------------------------------- | |
| 2412 // Compute latest legal control. | |
| 2413 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) { | |
| 2414 assert(early != NULL, "early control should not be NULL"); | |
| 2415 | |
| 2416 // Compute LCA over list of uses | |
| 2417 Node *LCA = NULL; | |
| 2418 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) { | |
| 2419 Node* c = n->fast_out(i); | |
| 2420 if (_nodes[c->_idx] == NULL) | |
| 2421 continue; // Skip the occasional dead node | |
| 2422 if( c->is_Phi() ) { // For Phis, we must land above on the path | |
| 2423 for( uint j=1; j<c->req(); j++ ) {// For all inputs | |
| 2424 if( c->in(j) == n ) { // Found matching input? | |
| 2425 Node *use = c->in(0)->in(j); | |
| 2426 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); | |
| 2427 } | |
| 2428 } | |
| 2429 } else { | |
| 2430 // For CFG data-users, use is in the block just prior | |
| 2431 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0); | |
| 2432 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); | |
| 2433 } | |
| 2434 } | |
| 2435 | |
| 2436 // if this is a load, check for anti-dependent stores | |
| 2437 // We use a conservative algorithm to identify potential interfering | |
| 2438 // instructions and for rescheduling the load. The users of the memory | |
| 2439 // input of this load are examined. Any use which is not a load and is | |
| 2440 // dominated by early is considered a potentially interfering store. | |
| 2441 // This can produce false positives. | |
| 2442 if (n->is_Load() && LCA != early) { | |
| 2443 Node_List worklist; | |
| 2444 | |
| 2445 Node *mem = n->in(MemNode::Memory); | |
| 2446 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { | |
| 2447 Node* s = mem->fast_out(i); | |
| 2448 worklist.push(s); | |
| 2449 } | |
| 2450 while(worklist.size() != 0 && LCA != early) { | |
| 2451 Node* s = worklist.pop(); | |
| 2452 if (s->is_Load()) { | |
| 2453 continue; | |
| 2454 } else if (s->is_MergeMem()) { | |
| 2455 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) { | |
| 2456 Node* s1 = s->fast_out(i); | |
| 2457 worklist.push(s1); | |
| 2458 } | |
| 2459 } else { | |
| 2460 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0); | |
| 2461 assert(sctrl != NULL || s->outcnt() == 0, "must have control"); | |
| 2462 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) { | |
| 2463 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n); | |
| 2464 } | |
| 2465 } | |
| 2466 } | |
| 2467 } | |
| 2468 | |
| 2469 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control"); | |
| 2470 return LCA; | |
| 2471 } | |
| 2472 | |
| 2473 // true if CFG node d dominates CFG node n | |
| 2474 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) { | |
| 2475 if (d == n) | |
| 2476 return true; | |
| 2477 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); | |
| 2478 uint dd = dom_depth(d); | |
| 2479 while (dom_depth(n) >= dd) { | |
| 2480 if (n == d) | |
| 2481 return true; | |
| 2482 n = idom(n); | |
| 2483 } | |
| 2484 return false; | |
| 2485 } | |
| 2486 | |
| 2487 //------------------------------dom_lca_for_get_late_ctrl_internal------------- | |
| 2488 // Pair-wise LCA with tags. | |
| 2489 // Tag each index with the node 'tag' currently being processed | |
| 2490 // before advancing up the dominator chain using idom(). | |
| 2491 // Later calls that find a match to 'tag' know that this path has already | |
| 2492 // been considered in the current LCA (which is input 'n1' by convention). | |
| 2493 // Since get_late_ctrl() is only called once for each node, the tag array | |
| 2494 // does not need to be cleared between calls to get_late_ctrl(). | |
| 2495 // Algorithm trades a larger constant factor for better asymptotic behavior | |
| 2496 // | |
| 2497 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) { | |
| 2498 uint d1 = dom_depth(n1); | |
| 2499 uint d2 = dom_depth(n2); | |
| 2500 | |
| 2501 do { | |
| 2502 if (d1 > d2) { | |
| 2503 // current lca is deeper than n2 | |
| 2504 _dom_lca_tags.map(n1->_idx, tag); | |
| 2505 n1 = idom(n1); | |
| 2506 d1 = dom_depth(n1); | |
| 2507 } else if (d1 < d2) { | |
| 2508 // n2 is deeper than current lca | |
| 2509 Node *memo = _dom_lca_tags[n2->_idx]; | |
| 2510 if( memo == tag ) { | |
| 2511 return n1; // Return the current LCA | |
| 2512 } | |
| 2513 _dom_lca_tags.map(n2->_idx, tag); | |
| 2514 n2 = idom(n2); | |
| 2515 d2 = dom_depth(n2); | |
| 2516 } else { | |
| 2517 // Here d1 == d2. Due to edits of the dominator-tree, sections | |
| 2518 // of the tree might have the same depth. These sections have | |
| 2519 // to be searched more carefully. | |
| 2520 | |
| 2521 // Scan up all the n1's with equal depth, looking for n2. | |
| 2522 _dom_lca_tags.map(n1->_idx, tag); | |
| 2523 Node *t1 = idom(n1); | |
| 2524 while (dom_depth(t1) == d1) { | |
| 2525 if (t1 == n2) return n2; | |
| 2526 _dom_lca_tags.map(t1->_idx, tag); | |
| 2527 t1 = idom(t1); | |
| 2528 } | |
| 2529 // Scan up all the n2's with equal depth, looking for n1. | |
| 2530 _dom_lca_tags.map(n2->_idx, tag); | |
| 2531 Node *t2 = idom(n2); | |
| 2532 while (dom_depth(t2) == d2) { | |
| 2533 if (t2 == n1) return n1; | |
| 2534 _dom_lca_tags.map(t2->_idx, tag); | |
| 2535 t2 = idom(t2); | |
| 2536 } | |
| 2537 // Move up to a new dominator-depth value as well as up the dom-tree. | |
| 2538 n1 = t1; | |
| 2539 n2 = t2; | |
| 2540 d1 = dom_depth(n1); | |
| 2541 d2 = dom_depth(n2); | |
| 2542 } | |
| 2543 } while (n1 != n2); | |
| 2544 return n1; | |
| 2545 } | |
| 2546 | |
| 2547 //------------------------------init_dom_lca_tags------------------------------ | |
| 2548 // Tag could be a node's integer index, 32bits instead of 64bits in some cases | |
| 2549 // Intended use does not involve any growth for the array, so it could | |
| 2550 // be of fixed size. | |
| 2551 void PhaseIdealLoop::init_dom_lca_tags() { | |
| 2552 uint limit = C->unique() + 1; | |
| 2553 _dom_lca_tags.map( limit, NULL ); | |
| 2554 #ifdef ASSERT | |
| 2555 for( uint i = 0; i < limit; ++i ) { | |
| 2556 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); | |
| 2557 } | |
| 2558 #endif // ASSERT | |
| 2559 } | |
| 2560 | |
| 2561 //------------------------------clear_dom_lca_tags------------------------------ | |
| 2562 // Tag could be a node's integer index, 32bits instead of 64bits in some cases | |
| 2563 // Intended use does not involve any growth for the array, so it could | |
| 2564 // be of fixed size. | |
| 2565 void PhaseIdealLoop::clear_dom_lca_tags() { | |
| 2566 uint limit = C->unique() + 1; | |
| 2567 _dom_lca_tags.map( limit, NULL ); | |
| 2568 _dom_lca_tags.clear(); | |
| 2569 #ifdef ASSERT | |
| 2570 for( uint i = 0; i < limit; ++i ) { | |
| 2571 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); | |
| 2572 } | |
| 2573 #endif // ASSERT | |
| 2574 } | |
| 2575 | |
| 2576 //------------------------------build_loop_late-------------------------------- | |
| 2577 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. | |
| 2578 // Second pass finds latest legal placement, and ideal loop placement. | |
| 2579 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack, const PhaseIdealLoop *verify_me ) { | |
| 2580 while (worklist.size() != 0) { | |
| 2581 Node *n = worklist.pop(); | |
| 2582 // Only visit once | |
| 2583 if (visited.test_set(n->_idx)) continue; | |
| 2584 uint cnt = n->outcnt(); | |
| 2585 uint i = 0; | |
| 2586 while (true) { | |
| 2587 assert( _nodes[n->_idx], "no dead nodes" ); | |
| 2588 // Visit all children | |
| 2589 if (i < cnt) { | |
| 2590 Node* use = n->raw_out(i); | |
| 2591 ++i; | |
| 2592 // Check for dead uses. Aggressively prune such junk. It might be | |
| 2593 // dead in the global sense, but still have local uses so I cannot | |
| 2594 // easily call 'remove_dead_node'. | |
| 2595 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead? | |
| 2596 // Due to cycles, we might not hit the same fixed point in the verify | |
| 2597 // pass as we do in the regular pass. Instead, visit such phis as | |
| 2598 // simple uses of the loop head. | |
| 2599 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) { | |
| 2600 if( !visited.test(use->_idx) ) | |
| 2601 worklist.push(use); | |
| 2602 } else if( !visited.test_set(use->_idx) ) { | |
| 2603 nstack.push(n, i); // Save parent and next use's index. | |
| 2604 n = use; // Process all children of current use. | |
| 2605 cnt = use->outcnt(); | |
| 2606 i = 0; | |
| 2607 } | |
| 2608 } else { | |
| 2609 // Do not visit around the backedge of loops via data edges. | |
| 2610 // push dead code onto a worklist | |
| 2611 _deadlist.push(use); | |
| 2612 } | |
| 2613 } else { | |
| 2614 // All of n's children have been processed, complete post-processing. | |
| 2615 build_loop_late_post(n, verify_me); | |
| 2616 if (nstack.is_empty()) { | |
| 2617 // Finished all nodes on stack. | |
| 2618 // Process next node on the worklist. | |
| 2619 break; | |
| 2620 } | |
| 2621 // Get saved parent node and next use's index. Visit the rest of uses. | |
| 2622 n = nstack.node(); | |
| 2623 cnt = n->outcnt(); | |
| 2624 i = nstack.index(); | |
| 2625 nstack.pop(); | |
| 2626 } | |
| 2627 } | |
| 2628 } | |
| 2629 } | |
| 2630 | |
| 2631 //------------------------------build_loop_late_post--------------------------- | |
| 2632 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. | |
| 2633 // Second pass finds latest legal placement, and ideal loop placement. | |
| 2634 void PhaseIdealLoop::build_loop_late_post( Node *n, const PhaseIdealLoop *verify_me ) { | |
| 2635 | |
| 2636 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress()) { | |
| 2637 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops. | |
| 2638 } | |
| 2639 | |
| 2640 // CFG and pinned nodes already handled | |
| 2641 if( n->in(0) ) { | |
| 2642 if( n->in(0)->is_top() ) return; // Dead? | |
| 2643 | |
| 2644 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads | |
| 2645 // _must_ be pinned (they have to observe their control edge of course). | |
| 2646 // Unlike Stores (which modify an unallocable resource, the memory | |
| 2647 // state), Mods/Loads can float around. So free them up. | |
| 2648 bool pinned = true; | |
| 2649 switch( n->Opcode() ) { | |
| 2650 case Op_DivI: | |
| 2651 case Op_DivF: | |
| 2652 case Op_DivD: | |
| 2653 case Op_ModI: | |
| 2654 case Op_ModF: | |
| 2655 case Op_ModD: | |
| 2656 case Op_LoadB: // Same with Loads; they can sink | |
| 2657 case Op_LoadC: // during loop optimizations. | |
| 2658 case Op_LoadD: | |
| 2659 case Op_LoadF: | |
| 2660 case Op_LoadI: | |
| 2661 case Op_LoadKlass: | |
|
293
c3e045194476
6731641: assert(m->adr_type() == mach->adr_type(),"matcher should not change adr type")
kvn
parents:
235
diff
changeset
|
2662 case Op_LoadNKlass: |
| 0 | 2663 case Op_LoadL: |
| 2664 case Op_LoadS: | |
| 2665 case Op_LoadP: | |
|
293
c3e045194476
6731641: assert(m->adr_type() == mach->adr_type(),"matcher should not change adr type")
kvn
parents:
235
diff
changeset
|
2666 case Op_LoadN: |
| 0 | 2667 case Op_LoadRange: |
| 2668 case Op_LoadD_unaligned: | |
| 2669 case Op_LoadL_unaligned: | |
| 2670 case Op_StrComp: // Does a bunch of load-like effects | |
|
169
9148c65abefc
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
39
diff
changeset
|
2671 case Op_AryEq: |
| 0 | 2672 pinned = false; |
| 2673 } | |
| 2674 if( pinned ) { | |
| 2675 IdealLoopTree *choosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n)); | |
| 2676 if( !choosen_loop->_child ) // Inner loop? | |
| 2677 choosen_loop->_body.push(n); // Collect inner loops | |
| 2678 return; | |
| 2679 } | |
| 2680 } else { // No slot zero | |
| 2681 if( n->is_CFG() ) { // CFG with no slot 0 is dead | |
| 2682 _nodes.map(n->_idx,0); // No block setting, it's globally dead | |
| 2683 return; | |
| 2684 } | |
| 2685 assert(!n->is_CFG() || n->outcnt() == 0, ""); | |
| 2686 } | |
| 2687 | |
| 2688 // Do I have a "safe range" I can select over? | |
| 2689 Node *early = get_ctrl(n);// Early location already computed | |
| 2690 | |
| 2691 // Compute latest point this Node can go | |
| 2692 Node *LCA = get_late_ctrl( n, early ); | |
| 2693 // LCA is NULL due to uses being dead | |
| 2694 if( LCA == NULL ) { | |
| 2695 #ifdef ASSERT | |
| 2696 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) { | |
| 2697 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead"); | |
| 2698 } | |
| 2699 #endif | |
| 2700 _nodes.map(n->_idx, 0); // This node is useless | |
| 2701 _deadlist.push(n); | |
| 2702 return; | |
| 2703 } | |
| 2704 assert(LCA != NULL && !LCA->is_top(), "no dead nodes"); | |
| 2705 | |
| 2706 Node *legal = LCA; // Walk 'legal' up the IDOM chain | |
| 2707 Node *least = legal; // Best legal position so far | |
| 2708 while( early != legal ) { // While not at earliest legal | |
| 2709 // Find least loop nesting depth | |
| 2710 legal = idom(legal); // Bump up the IDOM tree | |
| 2711 // Check for lower nesting depth | |
| 2712 if( get_loop(legal)->_nest < get_loop(least)->_nest ) | |
| 2713 least = legal; | |
| 2714 } | |
| 2715 | |
| 2716 // Try not to place code on a loop entry projection | |
| 2717 // which can inhibit range check elimination. | |
| 2718 if (least != early) { | |
| 2719 Node* ctrl_out = least->unique_ctrl_out(); | |
| 2720 if (ctrl_out && ctrl_out->is_CountedLoop() && | |
| 2721 least == ctrl_out->in(LoopNode::EntryControl)) { | |
| 2722 Node* least_dom = idom(least); | |
| 2723 if (get_loop(least_dom)->is_member(get_loop(least))) { | |
| 2724 least = least_dom; | |
| 2725 } | |
| 2726 } | |
| 2727 } | |
| 2728 | |
| 2729 #ifdef ASSERT | |
| 2730 // If verifying, verify that 'verify_me' has a legal location | |
| 2731 // and choose it as our location. | |
| 2732 if( verify_me ) { | |
| 2733 Node *v_ctrl = verify_me->get_ctrl_no_update(n); | |
| 2734 Node *legal = LCA; | |
| 2735 while( early != legal ) { // While not at earliest legal | |
| 2736 if( legal == v_ctrl ) break; // Check for prior good location | |
| 2737 legal = idom(legal) ;// Bump up the IDOM tree | |
| 2738 } | |
| 2739 // Check for prior good location | |
| 2740 if( legal == v_ctrl ) least = legal; // Keep prior if found | |
| 2741 } | |
| 2742 #endif | |
| 2743 | |
| 2744 // Assign discovered "here or above" point | |
| 2745 least = find_non_split_ctrl(least); | |
| 2746 set_ctrl(n, least); | |
| 2747 | |
| 2748 // Collect inner loop bodies | |
| 2749 IdealLoopTree *choosen_loop = get_loop(least); | |
| 2750 if( !choosen_loop->_child ) // Inner loop? | |
| 2751 choosen_loop->_body.push(n);// Collect inner loops | |
| 2752 } | |
| 2753 | |
| 2754 #ifndef PRODUCT | |
| 2755 //------------------------------dump------------------------------------------- | |
| 2756 void PhaseIdealLoop::dump( ) const { | |
| 2757 ResourceMark rm; | |
| 2758 Arena* arena = Thread::current()->resource_area(); | |
| 2759 Node_Stack stack(arena, C->unique() >> 2); | |
| 2760 Node_List rpo_list; | |
| 2761 VectorSet visited(arena); | |
| 2762 visited.set(C->top()->_idx); | |
| 2763 rpo( C->root(), stack, visited, rpo_list ); | |
| 2764 // Dump root loop indexed by last element in PO order | |
| 2765 dump( _ltree_root, rpo_list.size(), rpo_list ); | |
| 2766 } | |
| 2767 | |
| 2768 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const { | |
|
367
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
293
diff
changeset
|
2769 loop->dump_head(); |
| 0 | 2770 |
| 2771 // Now scan for CFG nodes in the same loop | |
| 2772 for( uint j=idx; j > 0; j-- ) { | |
| 2773 Node *n = rpo_list[j-1]; | |
| 2774 if( !_nodes[n->_idx] ) // Skip dead nodes | |
| 2775 continue; | |
| 2776 if( get_loop(n) != loop ) { // Wrong loop nest | |
| 2777 if( get_loop(n)->_head == n && // Found nested loop? | |
| 2778 get_loop(n)->_parent == loop ) | |
| 2779 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly | |
| 2780 continue; | |
| 2781 } | |
| 2782 | |
| 2783 // Dump controlling node | |
| 2784 for( uint x = 0; x < loop->_nest; x++ ) | |
| 2785 tty->print(" "); | |
| 2786 tty->print("C"); | |
| 2787 if( n == C->root() ) { | |
| 2788 n->dump(); | |
| 2789 } else { | |
| 2790 Node* cached_idom = idom_no_update(n); | |
| 2791 Node *computed_idom = n->in(0); | |
| 2792 if( n->is_Region() ) { | |
| 2793 computed_idom = compute_idom(n); | |
| 2794 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or | |
| 2795 // any MultiBranch ctrl node), so apply a similar transform to | |
| 2796 // the cached idom returned from idom_no_update. | |
| 2797 cached_idom = find_non_split_ctrl(cached_idom); | |
| 2798 } | |
| 2799 tty->print(" ID:%d",computed_idom->_idx); | |
| 2800 n->dump(); | |
| 2801 if( cached_idom != computed_idom ) { | |
| 2802 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d", | |
| 2803 computed_idom->_idx, cached_idom->_idx); | |
| 2804 } | |
| 2805 } | |
| 2806 // Dump nodes it controls | |
| 2807 for( uint k = 0; k < _nodes.Size(); k++ ) { | |
| 2808 // (k < C->unique() && get_ctrl(find(k)) == n) | |
| 2809 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) { | |
| 2810 Node *m = C->root()->find(k); | |
| 2811 if( m && m->outcnt() > 0 ) { | |
| 2812 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) { | |
| 2813 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p", | |
| 2814 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL); | |
| 2815 } | |
| 2816 for( uint j = 0; j < loop->_nest; j++ ) | |
| 2817 tty->print(" "); | |
| 2818 tty->print(" "); | |
| 2819 m->dump(); | |
| 2820 } | |
| 2821 } | |
| 2822 } | |
| 2823 } | |
| 2824 } | |
| 2825 | |
| 2826 // Collect a R-P-O for the whole CFG. | |
| 2827 // Result list is in post-order (scan backwards for RPO) | |
| 2828 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const { | |
| 2829 stk.push(start, 0); | |
| 2830 visited.set(start->_idx); | |
| 2831 | |
| 2832 while (stk.is_nonempty()) { | |
| 2833 Node* m = stk.node(); | |
| 2834 uint idx = stk.index(); | |
| 2835 if (idx < m->outcnt()) { | |
| 2836 stk.set_index(idx + 1); | |
| 2837 Node* n = m->raw_out(idx); | |
| 2838 if (n->is_CFG() && !visited.test_set(n->_idx)) { | |
| 2839 stk.push(n, 0); | |
| 2840 } | |
| 2841 } else { | |
| 2842 rpo_list.push(m); | |
| 2843 stk.pop(); | |
| 2844 } | |
| 2845 } | |
| 2846 } | |
| 2847 #endif | |
| 2848 | |
| 2849 | |
| 2850 //============================================================================= | |
| 2851 //------------------------------LoopTreeIterator----------------------------------- | |
| 2852 | |
| 2853 // Advance to next loop tree using a preorder, left-to-right traversal. | |
| 2854 void LoopTreeIterator::next() { | |
| 2855 assert(!done(), "must not be done."); | |
| 2856 if (_curnt->_child != NULL) { | |
| 2857 _curnt = _curnt->_child; | |
| 2858 } else if (_curnt->_next != NULL) { | |
| 2859 _curnt = _curnt->_next; | |
| 2860 } else { | |
| 2861 while (_curnt != _root && _curnt->_next == NULL) { | |
| 2862 _curnt = _curnt->_parent; | |
| 2863 } | |
| 2864 if (_curnt == _root) { | |
| 2865 _curnt = NULL; | |
| 2866 assert(done(), "must be done."); | |
| 2867 } else { | |
| 2868 assert(_curnt->_next != NULL, "must be more to do"); | |
| 2869 _curnt = _curnt->_next; | |
| 2870 } | |
| 2871 } | |
| 2872 } |