annotate src/share/vm/opto/loopnode.cpp @ 5546:3213ba4d3dff

8024069: replace_in_map() should operate on parent maps Summary: type information gets lost because replace_in_map() doesn't update parent maps Reviewed-by: kvn, twisti
author roland
date Sat, 19 Oct 2013 12:16:43 +0200
parents f2110083203d
children
rev   line source
duke@0 1 /*
sla@4802 2 * Copyright (c) 1998, 2013, Oracle and/or its affiliates. All rights reserved.
duke@0 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@0 4 *
duke@0 5 * This code is free software; you can redistribute it and/or modify it
duke@0 6 * under the terms of the GNU General Public License version 2 only, as
duke@0 7 * published by the Free Software Foundation.
duke@0 8 *
duke@0 9 * This code is distributed in the hope that it will be useful, but WITHOUT
duke@0 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@0 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@0 12 * version 2 for more details (a copy is included in the LICENSE file that
duke@0 13 * accompanied this code).
duke@0 14 *
duke@0 15 * You should have received a copy of the GNU General Public License version
duke@0 16 * 2 along with this work; if not, write to the Free Software Foundation,
duke@0 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@0 18 *
trims@1472 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1472 20 * or visit www.oracle.com if you need additional information or have any
trims@1472 21 * questions.
duke@0 22 *
duke@0 23 */
duke@0 24
stefank@1879 25 #include "precompiled.hpp"
stefank@1879 26 #include "ci/ciMethodData.hpp"
stefank@1879 27 #include "compiler/compileLog.hpp"
stefank@1879 28 #include "libadt/vectset.hpp"
stefank@1879 29 #include "memory/allocation.inline.hpp"
stefank@1879 30 #include "opto/addnode.hpp"
stefank@1879 31 #include "opto/callnode.hpp"
stefank@1879 32 #include "opto/connode.hpp"
stefank@1879 33 #include "opto/divnode.hpp"
stefank@1879 34 #include "opto/idealGraphPrinter.hpp"
stefank@1879 35 #include "opto/loopnode.hpp"
stefank@1879 36 #include "opto/mulnode.hpp"
stefank@1879 37 #include "opto/rootnode.hpp"
stefank@1879 38 #include "opto/superword.hpp"
duke@0 39
duke@0 40 //=============================================================================
duke@0 41 //------------------------------is_loop_iv-------------------------------------
duke@0 42 // Determine if a node is Counted loop induction variable.
duke@0 43 // The method is declared in node.hpp.
duke@0 44 const Node* Node::is_loop_iv() const {
duke@0 45 if (this->is_Phi() && !this->as_Phi()->is_copy() &&
duke@0 46 this->as_Phi()->region()->is_CountedLoop() &&
duke@0 47 this->as_Phi()->region()->as_CountedLoop()->phi() == this) {
duke@0 48 return this;
duke@0 49 } else {
duke@0 50 return NULL;
duke@0 51 }
duke@0 52 }
duke@0 53
duke@0 54 //=============================================================================
duke@0 55 //------------------------------dump_spec--------------------------------------
duke@0 56 // Dump special per-node info
duke@0 57 #ifndef PRODUCT
duke@0 58 void LoopNode::dump_spec(outputStream *st) const {
kvn@2230 59 if (is_inner_loop()) st->print( "inner " );
kvn@2230 60 if (is_partial_peel_loop()) st->print( "partial_peel " );
kvn@2230 61 if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
duke@0 62 }
duke@0 63 #endif
duke@0 64
kvn@2230 65 //------------------------------is_valid_counted_loop-------------------------
kvn@2230 66 bool LoopNode::is_valid_counted_loop() const {
kvn@2230 67 if (is_CountedLoop()) {
kvn@2230 68 CountedLoopNode* l = as_CountedLoop();
kvn@2230 69 CountedLoopEndNode* le = l->loopexit();
kvn@2230 70 if (le != NULL &&
kvn@2230 71 le->proj_out(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
kvn@2230 72 Node* phi = l->phi();
kvn@2230 73 Node* exit = le->proj_out(0 /* false */);
kvn@2230 74 if (exit != NULL && exit->Opcode() == Op_IfFalse &&
kvn@2230 75 phi != NULL && phi->is_Phi() &&
kvn@2230 76 phi->in(LoopNode::LoopBackControl) == l->incr() &&
kvn@2230 77 le->loopnode() == l && le->stride_is_con()) {
kvn@2230 78 return true;
kvn@2230 79 }
kvn@2230 80 }
kvn@2230 81 }
kvn@2230 82 return false;
kvn@2230 83 }
kvn@2230 84
duke@0 85 //------------------------------get_early_ctrl---------------------------------
duke@0 86 // Compute earliest legal control
duke@0 87 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
duke@0 88 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
duke@0 89 uint i;
duke@0 90 Node *early;
roland@4154 91 if (n->in(0) && !n->is_expensive()) {
duke@0 92 early = n->in(0);
roland@4154 93 if (!early->is_CFG()) // Might be a non-CFG multi-def
duke@0 94 early = get_ctrl(early); // So treat input as a straight data input
duke@0 95 i = 1;
duke@0 96 } else {
duke@0 97 early = get_ctrl(n->in(1));
duke@0 98 i = 2;
duke@0 99 }
duke@0 100 uint e_d = dom_depth(early);
duke@0 101 assert( early, "" );
roland@4154 102 for (; i < n->req(); i++) {
duke@0 103 Node *cin = get_ctrl(n->in(i));
duke@0 104 assert( cin, "" );
duke@0 105 // Keep deepest dominator depth
duke@0 106 uint c_d = dom_depth(cin);
roland@4154 107 if (c_d > e_d) { // Deeper guy?
duke@0 108 early = cin; // Keep deepest found so far
duke@0 109 e_d = c_d;
roland@4154 110 } else if (c_d == e_d && // Same depth?
roland@4154 111 early != cin) { // If not equal, must use slower algorithm
duke@0 112 // If same depth but not equal, one _must_ dominate the other
duke@0 113 // and we want the deeper (i.e., dominated) guy.
duke@0 114 Node *n1 = early;
duke@0 115 Node *n2 = cin;
roland@4154 116 while (1) {
duke@0 117 n1 = idom(n1); // Walk up until break cycle
duke@0 118 n2 = idom(n2);
roland@4154 119 if (n1 == cin || // Walked early up to cin
roland@4154 120 dom_depth(n2) < c_d)
duke@0 121 break; // early is deeper; keep him
roland@4154 122 if (n2 == early || // Walked cin up to early
roland@4154 123 dom_depth(n1) < c_d) {
duke@0 124 early = cin; // cin is deeper; keep him
duke@0 125 break;
duke@0 126 }
duke@0 127 }
duke@0 128 e_d = dom_depth(early); // Reset depth register cache
duke@0 129 }
duke@0 130 }
duke@0 131
duke@0 132 // Return earliest legal location
duke@0 133 assert(early == find_non_split_ctrl(early), "unexpected early control");
duke@0 134
roland@4154 135 if (n->is_expensive()) {
roland@4154 136 assert(n->in(0), "should have control input");
roland@4154 137 early = get_early_ctrl_for_expensive(n, early);
roland@4154 138 }
roland@4154 139
duke@0 140 return early;
duke@0 141 }
duke@0 142
roland@4154 143 //------------------------------get_early_ctrl_for_expensive---------------------------------
roland@4154 144 // Move node up the dominator tree as high as legal while still beneficial
roland@4154 145 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
roland@4154 146 assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
roland@4154 147 assert(OptimizeExpensiveOps, "optimization off?");
roland@4154 148
roland@4154 149 Node* ctl = n->in(0);
roland@4154 150 assert(ctl->is_CFG(), "expensive input 0 must be cfg");
roland@4154 151 uint min_dom_depth = dom_depth(earliest);
roland@4154 152 #ifdef ASSERT
roland@4154 153 if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {
roland@4154 154 dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);
roland@4154 155 assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
roland@4154 156 }
roland@4154 157 #endif
roland@4154 158 if (dom_depth(ctl) < min_dom_depth) {
roland@4154 159 return earliest;
roland@4154 160 }
roland@4154 161
roland@4154 162 while (1) {
roland@4154 163 Node *next = ctl;
roland@4154 164 // Moving the node out of a loop on the projection of a If
roland@4154 165 // confuses loop predication. So once we hit a Loop in a If branch
roland@4154 166 // that doesn't branch to an UNC, we stop. The code that process
roland@4154 167 // expensive nodes will notice the loop and skip over it to try to
roland@4154 168 // move the node further up.
roland@4154 169 if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) {
roland@5546 170 if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
roland@4154 171 break;
roland@4154 172 }
roland@4154 173 next = idom(ctl->in(1)->in(0));
roland@4154 174 } else if (ctl->is_Proj()) {
roland@4154 175 // We only move it up along a projection if the projection is
roland@4154 176 // the single control projection for its parent: same code path,
roland@4154 177 // if it's a If with UNC or fallthrough of a call.
roland@4154 178 Node* parent_ctl = ctl->in(0);
roland@4154 179 if (parent_ctl == NULL) {
roland@4154 180 break;
roland@4154 181 } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) {
roland@4154 182 next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
roland@4154 183 } else if (parent_ctl->is_If()) {
roland@5546 184 if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
roland@4154 185 break;
roland@4154 186 }
roland@4154 187 assert(idom(ctl) == parent_ctl, "strange");
roland@4154 188 next = idom(parent_ctl);
roland@4154 189 } else if (ctl->is_CatchProj()) {
roland@4154 190 if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {
roland@4154 191 break;
roland@4154 192 }
roland@4154 193 assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
roland@4154 194 next = parent_ctl->in(0)->in(0)->in(0);
roland@4154 195 } else {
roland@4154 196 // Check if parent control has a single projection (this
roland@4154 197 // control is the only possible successor of the parent
roland@4154 198 // control). If so, we can try to move the node above the
roland@4154 199 // parent control.
roland@4154 200 int nb_ctl_proj = 0;
roland@4154 201 for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {
roland@4154 202 Node *p = parent_ctl->fast_out(i);
roland@4154 203 if (p->is_Proj() && p->is_CFG()) {
roland@4154 204 nb_ctl_proj++;
roland@4154 205 if (nb_ctl_proj > 1) {
roland@4154 206 break;
roland@4154 207 }
roland@4154 208 }
roland@4154 209 }
roland@4154 210
roland@4154 211 if (nb_ctl_proj > 1) {
roland@4154 212 break;
roland@4154 213 }
roland@4154 214 assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call(), "unexpected node");
roland@4154 215 assert(idom(ctl) == parent_ctl, "strange");
roland@4154 216 next = idom(parent_ctl);
roland@4154 217 }
roland@4154 218 } else {
roland@4154 219 next = idom(ctl);
roland@4154 220 }
roland@4154 221 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
roland@4154 222 break;
roland@4154 223 }
roland@4154 224 ctl = next;
roland@4154 225 }
roland@4154 226
roland@4154 227 if (ctl != n->in(0)) {
roland@4154 228 _igvn.hash_delete(n);
roland@4154 229 n->set_req(0, ctl);
roland@4154 230 _igvn.hash_insert(n);
roland@4154 231 }
roland@4154 232
roland@4154 233 return ctl;
roland@4154 234 }
roland@4154 235
roland@4154 236
duke@0 237 //------------------------------set_early_ctrl---------------------------------
duke@0 238 // Set earliest legal control
duke@0 239 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
duke@0 240 Node *early = get_early_ctrl(n);
duke@0 241
duke@0 242 // Record earliest legal location
duke@0 243 set_ctrl(n, early);
duke@0 244 }
duke@0 245
duke@0 246 //------------------------------set_subtree_ctrl-------------------------------
duke@0 247 // set missing _ctrl entries on new nodes
duke@0 248 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
duke@0 249 // Already set? Get out.
duke@0 250 if( _nodes[n->_idx] ) return;
duke@0 251 // Recursively set _nodes array to indicate where the Node goes
duke@0 252 uint i;
duke@0 253 for( i = 0; i < n->req(); ++i ) {
duke@0 254 Node *m = n->in(i);
duke@0 255 if( m && m != C->root() )
duke@0 256 set_subtree_ctrl( m );
duke@0 257 }
duke@0 258
duke@0 259 // Fixup self
duke@0 260 set_early_ctrl( n );
duke@0 261 }
duke@0 262
duke@0 263 //------------------------------is_counted_loop--------------------------------
kvn@2230 264 bool PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) {
duke@0 265 PhaseGVN *gvn = &_igvn;
duke@0 266
duke@0 267 // Counted loop head must be a good RegionNode with only 3 not NULL
duke@0 268 // control input edges: Self, Entry, LoopBack.
kvn@2230 269 if (x->in(LoopNode::Self) == NULL || x->req() != 3)
kvn@2230 270 return false;
duke@0 271
duke@0 272 Node *init_control = x->in(LoopNode::EntryControl);
duke@0 273 Node *back_control = x->in(LoopNode::LoopBackControl);
kvn@2230 274 if (init_control == NULL || back_control == NULL) // Partially dead
kvn@2230 275 return false;
duke@0 276 // Must also check for TOP when looking for a dead loop
kvn@2230 277 if (init_control->is_top() || back_control->is_top())
kvn@2230 278 return false;
duke@0 279
duke@0 280 // Allow funny placement of Safepoint
kvn@2230 281 if (back_control->Opcode() == Op_SafePoint)
duke@0 282 back_control = back_control->in(TypeFunc::Control);
duke@0 283
duke@0 284 // Controlling test for loop
duke@0 285 Node *iftrue = back_control;
duke@0 286 uint iftrue_op = iftrue->Opcode();
kvn@2230 287 if (iftrue_op != Op_IfTrue &&
kvn@2230 288 iftrue_op != Op_IfFalse)
duke@0 289 // I have a weird back-control. Probably the loop-exit test is in
duke@0 290 // the middle of the loop and I am looking at some trailing control-flow
duke@0 291 // merge point. To fix this I would have to partially peel the loop.
kvn@2230 292 return false; // Obscure back-control
duke@0 293
duke@0 294 // Get boolean guarding loop-back test
duke@0 295 Node *iff = iftrue->in(0);
kvn@2230 296 if (get_loop(iff) != loop || !iff->in(1)->is_Bool())
kvn@2230 297 return false;
duke@0 298 BoolNode *test = iff->in(1)->as_Bool();
duke@0 299 BoolTest::mask bt = test->_test._test;
duke@0 300 float cl_prob = iff->as_If()->_prob;
kvn@2230 301 if (iftrue_op == Op_IfFalse) {
duke@0 302 bt = BoolTest(bt).negate();
duke@0 303 cl_prob = 1.0 - cl_prob;
duke@0 304 }
duke@0 305 // Get backedge compare
duke@0 306 Node *cmp = test->in(1);
duke@0 307 int cmp_op = cmp->Opcode();
kvn@2442 308 if (cmp_op != Op_CmpI)
kvn@2230 309 return false; // Avoid pointer & float compares
duke@0 310
duke@0 311 // Find the trip-counter increment & limit. Limit must be loop invariant.
duke@0 312 Node *incr = cmp->in(1);
duke@0 313 Node *limit = cmp->in(2);
duke@0 314
duke@0 315 // ---------
duke@0 316 // need 'loop()' test to tell if limit is loop invariant
duke@0 317 // ---------
duke@0 318
kvn@2230 319 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
kvn@2230 320 Node *tmp = incr; // Then reverse order into the CmpI
duke@0 321 incr = limit;
duke@0 322 limit = tmp;
duke@0 323 bt = BoolTest(bt).commute(); // And commute the exit test
duke@0 324 }
kvn@2230 325 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant
kvn@2230 326 return false;
kvn@2230 327 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
kvn@2230 328 return false;
duke@0 329
kvn@2230 330 Node* phi_incr = NULL;
duke@0 331 // Trip-counter increment must be commutative & associative.
kvn@2230 332 if (incr->is_Phi()) {
kvn@2230 333 if (incr->as_Phi()->region() != x || incr->req() != 3)
kvn@2230 334 return false; // Not simple trip counter expression
kvn@2230 335 phi_incr = incr;
kvn@2230 336 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
kvn@2230 337 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
kvn@2230 338 return false;
duke@0 339 }
kvn@2230 340
duke@0 341 Node* trunc1 = NULL;
duke@0 342 Node* trunc2 = NULL;
duke@0 343 const TypeInt* iv_trunc_t = NULL;
duke@0 344 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
kvn@2230 345 return false; // Funny increment opcode
duke@0 346 }
kvn@2230 347 assert(incr->Opcode() == Op_AddI, "wrong increment code");
duke@0 348
duke@0 349 // Get merge point
duke@0 350 Node *xphi = incr->in(1);
duke@0 351 Node *stride = incr->in(2);
kvn@2230 352 if (!stride->is_Con()) { // Oops, swap these
kvn@2230 353 if (!xphi->is_Con()) // Is the other guy a constant?
kvn@2230 354 return false; // Nope, unknown stride, bail out
duke@0 355 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
duke@0 356 xphi = stride;
duke@0 357 stride = tmp;
duke@0 358 }
kvn@2230 359 // Stride must be constant
kvn@2230 360 int stride_con = stride->get_int();
kvn@2442 361 if (stride_con == 0)
kvn@2442 362 return false; // missed some peephole opt
kvn@2230 363
kvn@2230 364 if (!xphi->is_Phi())
kvn@2230 365 return false; // Too much math on the trip counter
kvn@2230 366 if (phi_incr != NULL && phi_incr != xphi)
kvn@2230 367 return false;
duke@0 368 PhiNode *phi = xphi->as_Phi();
duke@0 369
duke@0 370 // Phi must be of loop header; backedge must wrap to increment
kvn@2230 371 if (phi->region() != x)
kvn@2230 372 return false;
kvn@2230 373 if (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr ||
kvn@2230 374 trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1) {
kvn@2230 375 return false;
duke@0 376 }
duke@0 377 Node *init_trip = phi->in(LoopNode::EntryControl);
duke@0 378
duke@0 379 // If iv trunc type is smaller than int, check for possible wrap.
duke@0 380 if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
duke@0 381 assert(trunc1 != NULL, "must have found some truncation");
duke@0 382
duke@0 383 // Get a better type for the phi (filtered thru if's)
duke@0 384 const TypeInt* phi_ft = filtered_type(phi);
duke@0 385
duke@0 386 // Can iv take on a value that will wrap?
duke@0 387 //
duke@0 388 // Ensure iv's limit is not within "stride" of the wrap value.
duke@0 389 //
duke@0 390 // Example for "short" type
duke@0 391 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
duke@0 392 // If the stride is +10, then the last value of the induction
duke@0 393 // variable before the increment (phi_ft->_hi) must be
duke@0 394 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
duke@0 395 // ensure no truncation occurs after the increment.
duke@0 396
duke@0 397 if (stride_con > 0) {
duke@0 398 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
duke@0 399 iv_trunc_t->_lo > phi_ft->_lo) {
kvn@2230 400 return false; // truncation may occur
duke@0 401 }
duke@0 402 } else if (stride_con < 0) {
duke@0 403 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
duke@0 404 iv_trunc_t->_hi < phi_ft->_hi) {
kvn@2230 405 return false; // truncation may occur
duke@0 406 }
duke@0 407 }
duke@0 408 // No possibility of wrap so truncation can be discarded
duke@0 409 // Promote iv type to Int
duke@0 410 } else {
duke@0 411 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
duke@0 412 }
duke@0 413
kvn@2230 414 // If the condition is inverted and we will be rolling
kvn@2230 415 // through MININT to MAXINT, then bail out.
kvn@2230 416 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
kvn@2230 417 // Odd stride
kvn@2230 418 bt == BoolTest::ne && stride_con != 1 && stride_con != -1 ||
kvn@2230 419 // Count down loop rolls through MAXINT
kvn@2230 420 (bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0 ||
kvn@2230 421 // Count up loop rolls through MININT
kvn@2442 422 (bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0) {
kvn@2230 423 return false; // Bail out
kvn@2230 424 }
kvn@2230 425
kvn@2230 426 const TypeInt* init_t = gvn->type(init_trip)->is_int();
kvn@2230 427 const TypeInt* limit_t = gvn->type(limit)->is_int();
kvn@2230 428
kvn@2230 429 if (stride_con > 0) {
vlivanov@3722 430 jlong init_p = (jlong)init_t->_lo + stride_con;
vlivanov@3722 431 if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi)
kvn@2230 432 return false; // cyclic loop or this loop trips only once
kvn@2230 433 } else {
vlivanov@3722 434 jlong init_p = (jlong)init_t->_hi + stride_con;
vlivanov@3722 435 if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo)
kvn@2230 436 return false; // cyclic loop or this loop trips only once
kvn@2230 437 }
kvn@2230 438
duke@0 439 // =================================================
duke@0 440 // ---- SUCCESS! Found A Trip-Counted Loop! -----
duke@0 441 //
kvn@2230 442 assert(x->Opcode() == Op_Loop, "regular loops only");
sla@4802 443 C->print_method(PHASE_BEFORE_CLOOPS, 3);
kvn@2442 444
kvn@3680 445 Node *hook = new (C) Node(6);
kvn@2442 446
kvn@2442 447 if (LoopLimitCheck) {
kvn@2442 448
kvn@2442 449 // ===================================================
kvn@2442 450 // Generate loop limit check to avoid integer overflow
kvn@2442 451 // in cases like next (cyclic loops):
kvn@2442 452 //
kvn@2442 453 // for (i=0; i <= max_jint; i++) {}
kvn@2442 454 // for (i=0; i < max_jint; i+=2) {}
kvn@2442 455 //
kvn@2442 456 //
kvn@2442 457 // Limit check predicate depends on the loop test:
kvn@2442 458 //
kvn@2442 459 // for(;i != limit; i++) --> limit <= (max_jint)
kvn@2442 460 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1)
kvn@2442 461 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride )
kvn@2442 462 //
kvn@2442 463
kvn@2442 464 // Check if limit is excluded to do more precise int overflow check.
kvn@2442 465 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
kvn@2442 466 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
kvn@2442 467
kvn@2442 468 // If compare points directly to the phi we need to adjust
kvn@2442 469 // the compare so that it points to the incr. Limit have
kvn@2442 470 // to be adjusted to keep trip count the same and the
kvn@2442 471 // adjusted limit should be checked for int overflow.
kvn@2442 472 if (phi_incr != NULL) {
kvn@2442 473 stride_m += stride_con;
kvn@2442 474 }
kvn@2442 475
kvn@2442 476 if (limit->is_Con()) {
kvn@2442 477 int limit_con = limit->get_int();
kvn@2442 478 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
kvn@2442 479 (stride_con < 0 && limit_con < (min_jint - stride_m))) {
kvn@2442 480 // Bailout: it could be integer overflow.
kvn@2442 481 return false;
kvn@2442 482 }
kvn@2442 483 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
kvn@2442 484 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
kvn@2442 485 // Limit's type may satisfy the condition, for example,
kvn@2442 486 // when it is an array length.
kvn@2442 487 } else {
kvn@2442 488 // Generate loop's limit check.
kvn@2442 489 // Loop limit check predicate should be near the loop.
kvn@2442 490 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
kvn@2442 491 if (!limit_check_proj) {
kvn@2442 492 // The limit check predicate is not generated if this method trapped here before.
kvn@2442 493 #ifdef ASSERT
kvn@2442 494 if (TraceLoopLimitCheck) {
kvn@2442 495 tty->print("missing loop limit check:");
kvn@2442 496 loop->dump_head();
kvn@2442 497 x->dump(1);
kvn@2442 498 }
kvn@2442 499 #endif
kvn@2442 500 return false;
kvn@2442 501 }
kvn@2442 502
kvn@2442 503 IfNode* check_iff = limit_check_proj->in(0)->as_If();
kvn@2442 504 Node* cmp_limit;
kvn@2442 505 Node* bol;
kvn@2442 506
kvn@2442 507 if (stride_con > 0) {
kvn@3680 508 cmp_limit = new (C) CmpINode(limit, _igvn.intcon(max_jint - stride_m));
kvn@3680 509 bol = new (C) BoolNode(cmp_limit, BoolTest::le);
kvn@2442 510 } else {
kvn@3680 511 cmp_limit = new (C) CmpINode(limit, _igvn.intcon(min_jint - stride_m));
kvn@3680 512 bol = new (C) BoolNode(cmp_limit, BoolTest::ge);
kvn@2442 513 }
kvn@2442 514 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
kvn@2442 515 bol = _igvn.register_new_node_with_optimizer(bol);
kvn@2442 516 set_subtree_ctrl(bol);
kvn@2442 517
kvn@2442 518 // Replace condition in original predicate but preserve Opaque node
kvn@2442 519 // so that previous predicates could be found.
kvn@2442 520 assert(check_iff->in(1)->Opcode() == Op_Conv2B &&
kvn@2442 521 check_iff->in(1)->in(1)->Opcode() == Op_Opaque1, "");
kvn@2442 522 Node* opq = check_iff->in(1)->in(1);
kvn@2442 523 _igvn.hash_delete(opq);
kvn@2442 524 opq->set_req(1, bol);
kvn@2442 525 // Update ctrl.
kvn@2442 526 set_ctrl(opq, check_iff->in(0));
kvn@2442 527 set_ctrl(check_iff->in(1), check_iff->in(0));
kvn@2442 528
kvn@2292 529 #ifndef PRODUCT
kvn@2442 530 // report that the loop predication has been actually performed
kvn@2442 531 // for this loop
kvn@2442 532 if (TraceLoopLimitCheck) {
kvn@2442 533 tty->print_cr("Counted Loop Limit Check generated:");
kvn@2442 534 debug_only( bol->dump(2); )
kvn@2442 535 }
kvn@2442 536 #endif
kvn@2292 537 }
kvn@2442 538
kvn@2442 539 if (phi_incr != NULL) {
kvn@2442 540 // If compare points directly to the phi we need to adjust
kvn@2442 541 // the compare so that it points to the incr. Limit have
kvn@2442 542 // to be adjusted to keep trip count the same and we
kvn@2442 543 // should avoid int overflow.
kvn@2442 544 //
kvn@2442 545 // i = init; do {} while(i++ < limit);
kvn@2442 546 // is converted to
kvn@2442 547 // i = init; do {} while(++i < limit+1);
kvn@2442 548 //
kvn@3680 549 limit = gvn->transform(new (C) AddINode(limit, stride));
kvn@2442 550 }
kvn@2442 551
kvn@2442 552 // Now we need to canonicalize loop condition.
kvn@2442 553 if (bt == BoolTest::ne) {
kvn@2442 554 assert(stride_con == 1 || stride_con == -1, "simple increment only");
kvn@2544 555 // 'ne' can be replaced with 'lt' only when init < limit.
kvn@2544 556 if (stride_con > 0 && init_t->_hi < limit_t->_lo)
kvn@2544 557 bt = BoolTest::lt;
kvn@2544 558 // 'ne' can be replaced with 'gt' only when init > limit.
kvn@2544 559 if (stride_con < 0 && init_t->_lo > limit_t->_hi)
kvn@2544 560 bt = BoolTest::gt;
kvn@2442 561 }
kvn@2442 562
kvn@2442 563 if (incl_limit) {
kvn@2442 564 // The limit check guaranties that 'limit <= (max_jint - stride)' so
kvn@2442 565 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
kvn@2442 566 //
kvn@2442 567 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
kvn@3680 568 limit = gvn->transform(new (C) AddINode(limit, one));
kvn@2442 569 if (bt == BoolTest::le)
kvn@2442 570 bt = BoolTest::lt;
kvn@2442 571 else if (bt == BoolTest::ge)
kvn@2442 572 bt = BoolTest::gt;
kvn@2442 573 else
kvn@2442 574 ShouldNotReachHere();
kvn@2442 575 }
kvn@2442 576 set_subtree_ctrl( limit );
kvn@2442 577
kvn@2442 578 } else { // LoopLimitCheck
kvn@2442 579
duke@0 580 // If compare points to incr, we are ok. Otherwise the compare
duke@0 581 // can directly point to the phi; in this case adjust the compare so that
twisti@605 582 // it points to the incr by adjusting the limit.
kvn@2230 583 if (cmp->in(1) == phi || cmp->in(2) == phi)
kvn@3680 584 limit = gvn->transform(new (C) AddINode(limit,stride));
duke@0 585
duke@0 586 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride.
duke@0 587 // Final value for iterator should be: trip_count * stride + init_trip.
duke@0 588 Node *one_p = gvn->intcon( 1);
duke@0 589 Node *one_m = gvn->intcon(-1);
duke@0 590
duke@0 591 Node *trip_count = NULL;
duke@0 592 switch( bt ) {
duke@0 593 case BoolTest::eq:
kvn@2230 594 ShouldNotReachHere();
duke@0 595 case BoolTest::ne: // Ahh, the case we desire
kvn@2230 596 if (stride_con == 1)
kvn@3680 597 trip_count = gvn->transform(new (C) SubINode(limit,init_trip));
kvn@2230 598 else if (stride_con == -1)
kvn@3680 599 trip_count = gvn->transform(new (C) SubINode(init_trip,limit));
duke@0 600 else
kvn@2230 601 ShouldNotReachHere();
kvn@2230 602 set_subtree_ctrl(trip_count);
duke@0 603 //_loop.map(trip_count->_idx,loop(limit));
duke@0 604 break;
duke@0 605 case BoolTest::le: // Maybe convert to '<' case
kvn@3680 606 limit = gvn->transform(new (C) AddINode(limit,one_p));
duke@0 607 set_subtree_ctrl( limit );
duke@0 608 hook->init_req(4, limit);
duke@0 609
duke@0 610 bt = BoolTest::lt;
duke@0 611 // Make the new limit be in the same loop nest as the old limit
duke@0 612 //_loop.map(limit->_idx,limit_loop);
duke@0 613 // Fall into next case
duke@0 614 case BoolTest::lt: { // Maybe convert to '!=' case
kvn@2230 615 if (stride_con < 0) // Count down loop rolls through MAXINT
kvn@2230 616 ShouldNotReachHere();
kvn@3680 617 Node *range = gvn->transform(new (C) SubINode(limit,init_trip));
duke@0 618 set_subtree_ctrl( range );
duke@0 619 hook->init_req(0, range);
duke@0 620
kvn@3680 621 Node *bias = gvn->transform(new (C) AddINode(range,stride));
duke@0 622 set_subtree_ctrl( bias );
duke@0 623 hook->init_req(1, bias);
duke@0 624
kvn@3680 625 Node *bias1 = gvn->transform(new (C) AddINode(bias,one_m));
duke@0 626 set_subtree_ctrl( bias1 );
duke@0 627 hook->init_req(2, bias1);
duke@0 628
kvn@3680 629 trip_count = gvn->transform(new (C) DivINode(0,bias1,stride));
duke@0 630 set_subtree_ctrl( trip_count );
duke@0 631 hook->init_req(3, trip_count);
duke@0 632 break;
duke@0 633 }
duke@0 634
duke@0 635 case BoolTest::ge: // Maybe convert to '>' case
kvn@3680 636 limit = gvn->transform(new (C) AddINode(limit,one_m));
duke@0 637 set_subtree_ctrl( limit );
duke@0 638 hook->init_req(4 ,limit);
duke@0 639
duke@0 640 bt = BoolTest::gt;
duke@0 641 // Make the new limit be in the same loop nest as the old limit
duke@0 642 //_loop.map(limit->_idx,limit_loop);
duke@0 643 // Fall into next case
duke@0 644 case BoolTest::gt: { // Maybe convert to '!=' case
kvn@2230 645 if (stride_con > 0) // count up loop rolls through MININT
kvn@2230 646 ShouldNotReachHere();
kvn@3680 647 Node *range = gvn->transform(new (C) SubINode(limit,init_trip));
duke@0 648 set_subtree_ctrl( range );
duke@0 649 hook->init_req(0, range);
duke@0 650
kvn@3680 651 Node *bias = gvn->transform(new (C) AddINode(range,stride));
duke@0 652 set_subtree_ctrl( bias );
duke@0 653 hook->init_req(1, bias);
duke@0 654
kvn@3680 655 Node *bias1 = gvn->transform(new (C) AddINode(bias,one_p));
duke@0 656 set_subtree_ctrl( bias1 );
duke@0 657 hook->init_req(2, bias1);
duke@0 658
kvn@3680 659 trip_count = gvn->transform(new (C) DivINode(0,bias1,stride));
duke@0 660 set_subtree_ctrl( trip_count );
duke@0 661 hook->init_req(3, trip_count);
duke@0 662 break;
duke@0 663 }
kvn@2230 664 } // switch( bt )
duke@0 665
kvn@3680 666 Node *span = gvn->transform(new (C) MulINode(trip_count,stride));
duke@0 667 set_subtree_ctrl( span );
duke@0 668 hook->init_req(5, span);
duke@0 669
kvn@3680 670 limit = gvn->transform(new (C) AddINode(span,init_trip));
duke@0 671 set_subtree_ctrl( limit );
duke@0 672
kvn@2442 673 } // LoopLimitCheck
kvn@2442 674
kvn@2230 675 // Check for SafePoint on backedge and remove
kvn@2230 676 Node *sfpt = x->in(LoopNode::LoopBackControl);
kvn@2230 677 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
kvn@2230 678 lazy_replace( sfpt, iftrue );
kvn@3588 679 if (loop->_safepts != NULL) {
kvn@3588 680 loop->_safepts->yank(sfpt);
kvn@3588 681 }
kvn@2230 682 loop->_tail = iftrue;
kvn@2230 683 }
kvn@2230 684
duke@0 685 // Build a canonical trip test.
duke@0 686 // Clone code, as old values may be in use.
duke@0 687 incr = incr->clone();
kvn@2700 688 incr->set_req(1,phi);
duke@0 689 incr->set_req(2,stride);
duke@0 690 incr = _igvn.register_new_node_with_optimizer(incr);
duke@0 691 set_early_ctrl( incr );
kvn@2700 692 _igvn.hash_delete(phi);
kvn@2700 693 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
kvn@2700 694
kvn@2700 695 // If phi type is more restrictive than Int, raise to
kvn@2700 696 // Int to prevent (almost) infinite recursion in igvn
kvn@2700 697 // which can only handle integer types for constants or minint..maxint.
kvn@2700 698 if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
kvn@2700 699 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
kvn@2700 700 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
kvn@2700 701 nphi = _igvn.register_new_node_with_optimizer(nphi);
kvn@2700 702 set_ctrl(nphi, get_ctrl(phi));
kvn@2700 703 _igvn.replace_node(phi, nphi);
kvn@2700 704 phi = nphi->as_Phi();
kvn@2700 705 }
duke@0 706 cmp = cmp->clone();
duke@0 707 cmp->set_req(1,incr);
duke@0 708 cmp->set_req(2,limit);
duke@0 709 cmp = _igvn.register_new_node_with_optimizer(cmp);
duke@0 710 set_ctrl(cmp, iff->in(0));
duke@0 711
kvn@2230 712 test = test->clone()->as_Bool();
kvn@2230 713 (*(BoolTest*)&test->_test)._test = bt;
duke@0 714 test->set_req(1,cmp);
duke@0 715 _igvn.register_new_node_with_optimizer(test);
duke@0 716 set_ctrl(test, iff->in(0));
duke@0 717
duke@0 718 // Replace the old IfNode with a new LoopEndNode
kvn@3680 719 Node *lex = _igvn.register_new_node_with_optimizer(new (C) CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
duke@0 720 IfNode *le = lex->as_If();
duke@0 721 uint dd = dom_depth(iff);
duke@0 722 set_idom(le, le->in(0), dd); // Update dominance for loop exit
duke@0 723 set_loop(le, loop);
duke@0 724
duke@0 725 // Get the loop-exit control
kvn@2230 726 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
duke@0 727
duke@0 728 // Need to swap loop-exit and loop-back control?
kvn@2230 729 if (iftrue_op == Op_IfFalse) {
kvn@3680 730 Node *ift2=_igvn.register_new_node_with_optimizer(new (C) IfTrueNode (le));
kvn@3680 731 Node *iff2=_igvn.register_new_node_with_optimizer(new (C) IfFalseNode(le));
duke@0 732
duke@0 733 loop->_tail = back_control = ift2;
duke@0 734 set_loop(ift2, loop);
kvn@2230 735 set_loop(iff2, get_loop(iffalse));
duke@0 736
duke@0 737 // Lazy update of 'get_ctrl' mechanism.
kvn@2230 738 lazy_replace_proj( iffalse, iff2 );
kvn@2230 739 lazy_replace_proj( iftrue, ift2 );
duke@0 740
duke@0 741 // Swap names
kvn@2230 742 iffalse = iff2;
kvn@2230 743 iftrue = ift2;
duke@0 744 } else {
kvn@2230 745 _igvn.hash_delete(iffalse);
duke@0 746 _igvn.hash_delete(iftrue);
kvn@2230 747 iffalse->set_req_X( 0, le, &_igvn );
kvn@2230 748 iftrue ->set_req_X( 0, le, &_igvn );
duke@0 749 }
duke@0 750
kvn@2230 751 set_idom(iftrue, le, dd+1);
kvn@2230 752 set_idom(iffalse, le, dd+1);
kvn@2230 753 assert(iff->outcnt() == 0, "should be dead now");
kvn@2230 754 lazy_replace( iff, le ); // fix 'get_ctrl'
duke@0 755
duke@0 756 // Now setup a new CountedLoopNode to replace the existing LoopNode
kvn@3680 757 CountedLoopNode *l = new (C) CountedLoopNode(init_control, back_control);
kvn@2230 758 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
duke@0 759 // The following assert is approximately true, and defines the intention
duke@0 760 // of can_be_counted_loop. It fails, however, because phase->type
duke@0 761 // is not yet initialized for this loop and its parts.
duke@0 762 //assert(l->can_be_counted_loop(this), "sanity");
duke@0 763 _igvn.register_new_node_with_optimizer(l);
duke@0 764 set_loop(l, loop);
duke@0 765 loop->_head = l;
duke@0 766 // Fix all data nodes placed at the old loop head.
duke@0 767 // Uses the lazy-update mechanism of 'get_ctrl'.
duke@0 768 lazy_replace( x, l );
duke@0 769 set_idom(l, init_control, dom_depth(x));
duke@0 770
twisti@605 771 // Check for immediately preceding SafePoint and remove
duke@0 772 Node *sfpt2 = le->in(0);
kvn@3588 773 if (sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2)) {
duke@0 774 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
kvn@3588 775 if (loop->_safepts != NULL) {
kvn@3588 776 loop->_safepts->yank(sfpt2);
kvn@3588 777 }
kvn@3588 778 }
duke@0 779
duke@0 780 // Free up intermediate goo
duke@0 781 _igvn.remove_dead_node(hook);
duke@0 782
kvn@2230 783 #ifdef ASSERT
kvn@2230 784 assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
kvn@2230 785 assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" );
kvn@2230 786 #endif
kvn@2442 787 #ifndef PRODUCT
kvn@2442 788 if (TraceLoopOpts) {
kvn@2442 789 tty->print("Counted ");
kvn@2442 790 loop->dump_head();
kvn@2442 791 }
kvn@2442 792 #endif
kvn@2230 793
sla@4802 794 C->print_method(PHASE_AFTER_CLOOPS, 3);
duke@0 795
kvn@2230 796 return true;
duke@0 797 }
duke@0 798
kvn@2442 799 //----------------------exact_limit-------------------------------------------
kvn@2442 800 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
kvn@2442 801 assert(loop->_head->is_CountedLoop(), "");
kvn@2442 802 CountedLoopNode *cl = loop->_head->as_CountedLoop();
kvn@2613 803 assert(cl->is_valid_counted_loop(), "");
kvn@2442 804
kvn@2442 805 if (!LoopLimitCheck || ABS(cl->stride_con()) == 1 ||
kvn@2442 806 cl->limit()->Opcode() == Op_LoopLimit) {
kvn@2442 807 // Old code has exact limit (it could be incorrect in case of int overflow).
kvn@2442 808 // Loop limit is exact with stride == 1. And loop may already have exact limit.
kvn@2442 809 return cl->limit();
kvn@2442 810 }
kvn@2442 811 Node *limit = NULL;
kvn@2442 812 #ifdef ASSERT
kvn@2442 813 BoolTest::mask bt = cl->loopexit()->test_trip();
kvn@2442 814 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
kvn@2442 815 #endif
kvn@2442 816 if (cl->has_exact_trip_count()) {
kvn@2442 817 // Simple case: loop has constant boundaries.
vlivanov@3722 818 // Use jlongs to avoid integer overflow.
kvn@2442 819 int stride_con = cl->stride_con();
vlivanov@3722 820 jlong init_con = cl->init_trip()->get_int();
vlivanov@3722 821 jlong limit_con = cl->limit()->get_int();
kvn@2442 822 julong trip_cnt = cl->trip_count();
vlivanov@3722 823 jlong final_con = init_con + trip_cnt*stride_con;
kvn@2442 824 int final_int = (int)final_con;
kvn@2442 825 // The final value should be in integer range since the loop
kvn@2442 826 // is counted and the limit was checked for overflow.
vlivanov@3722 827 assert(final_con == (jlong)final_int, "final value should be integer");
kvn@2442 828 limit = _igvn.intcon(final_int);
kvn@2442 829 } else {
kvn@2442 830 // Create new LoopLimit node to get exact limit (final iv value).
kvn@3680 831 limit = new (C) LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
kvn@2442 832 register_new_node(limit, cl->in(LoopNode::EntryControl));
kvn@2442 833 }
kvn@2442 834 assert(limit != NULL, "sanity");
kvn@2442 835 return limit;
kvn@2442 836 }
duke@0 837
duke@0 838 //------------------------------Ideal------------------------------------------
duke@0 839 // Return a node which is more "ideal" than the current node.
duke@0 840 // Attempt to convert into a counted-loop.
duke@0 841 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 842 if (!can_be_counted_loop(phase)) {
duke@0 843 phase->C->set_major_progress();
duke@0 844 }
duke@0 845 return RegionNode::Ideal(phase, can_reshape);
duke@0 846 }
duke@0 847
duke@0 848
duke@0 849 //=============================================================================
duke@0 850 //------------------------------Ideal------------------------------------------
duke@0 851 // Return a node which is more "ideal" than the current node.
duke@0 852 // Attempt to convert into a counted-loop.
duke@0 853 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 854 return RegionNode::Ideal(phase, can_reshape);
duke@0 855 }
duke@0 856
duke@0 857 //------------------------------dump_spec--------------------------------------
duke@0 858 // Dump special per-node info
duke@0 859 #ifndef PRODUCT
duke@0 860 void CountedLoopNode::dump_spec(outputStream *st) const {
duke@0 861 LoopNode::dump_spec(st);
kvn@2442 862 if (stride_is_con()) {
duke@0 863 st->print("stride: %d ",stride_con());
duke@0 864 }
kvn@2442 865 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
kvn@2442 866 if (is_main_loop()) st->print("main of N%d", _idx);
kvn@2442 867 if (is_post_loop()) st->print("post of N%d", _main_idx);
duke@0 868 }
duke@0 869 #endif
duke@0 870
duke@0 871 //=============================================================================
duke@0 872 int CountedLoopEndNode::stride_con() const {
duke@0 873 return stride()->bottom_type()->is_int()->get_con();
duke@0 874 }
duke@0 875
kvn@2442 876 //=============================================================================
kvn@2442 877 //------------------------------Value-----------------------------------------
kvn@2442 878 const Type *LoopLimitNode::Value( PhaseTransform *phase ) const {
kvn@2442 879 const Type* init_t = phase->type(in(Init));
kvn@2442 880 const Type* limit_t = phase->type(in(Limit));
kvn@2442 881 const Type* stride_t = phase->type(in(Stride));
kvn@2442 882 // Either input is TOP ==> the result is TOP
kvn@2442 883 if (init_t == Type::TOP) return Type::TOP;
kvn@2442 884 if (limit_t == Type::TOP) return Type::TOP;
kvn@2442 885 if (stride_t == Type::TOP) return Type::TOP;
kvn@2442 886
kvn@2442 887 int stride_con = stride_t->is_int()->get_con();
kvn@2442 888 if (stride_con == 1)
kvn@2442 889 return NULL; // Identity
kvn@2442 890
kvn@2442 891 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
vlivanov@3722 892 // Use jlongs to avoid integer overflow.
vlivanov@3722 893 jlong init_con = init_t->is_int()->get_con();
vlivanov@3722 894 jlong limit_con = limit_t->is_int()->get_con();
kvn@2442 895 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
vlivanov@3722 896 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
vlivanov@3722 897 jlong final_con = init_con + stride_con*trip_count;
kvn@2442 898 int final_int = (int)final_con;
kvn@2442 899 // The final value should be in integer range since the loop
kvn@2442 900 // is counted and the limit was checked for overflow.
vlivanov@3722 901 assert(final_con == (jlong)final_int, "final value should be integer");
kvn@2442 902 return TypeInt::make(final_int);
kvn@2442 903 }
kvn@2442 904
kvn@2442 905 return bottom_type(); // TypeInt::INT
kvn@2442 906 }
kvn@2442 907
kvn@2442 908 //------------------------------Ideal------------------------------------------
kvn@2442 909 // Return a node which is more "ideal" than the current node.
kvn@2442 910 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
kvn@2442 911 if (phase->type(in(Init)) == Type::TOP ||
kvn@2442 912 phase->type(in(Limit)) == Type::TOP ||
kvn@2442 913 phase->type(in(Stride)) == Type::TOP)
kvn@2442 914 return NULL; // Dead
kvn@2442 915
kvn@2442 916 int stride_con = phase->type(in(Stride))->is_int()->get_con();
kvn@2442 917 if (stride_con == 1)
kvn@2442 918 return NULL; // Identity
kvn@2442 919
kvn@2442 920 if (in(Init)->is_Con() && in(Limit)->is_Con())
kvn@2442 921 return NULL; // Value
kvn@2442 922
kvn@2442 923 // Delay following optimizations until all loop optimizations
kvn@2442 924 // done to keep Ideal graph simple.
kvn@2442 925 if (!can_reshape || phase->C->major_progress())
kvn@2442 926 return NULL;
kvn@2442 927
kvn@2442 928 const TypeInt* init_t = phase->type(in(Init) )->is_int();
kvn@2442 929 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
kvn@2442 930 int stride_p;
vlivanov@3722 931 jlong lim, ini;
kvn@2442 932 julong max;
kvn@2442 933 if (stride_con > 0) {
kvn@2442 934 stride_p = stride_con;
kvn@2442 935 lim = limit_t->_hi;
kvn@2442 936 ini = init_t->_lo;
kvn@2442 937 max = (julong)max_jint;
kvn@2442 938 } else {
kvn@2442 939 stride_p = -stride_con;
kvn@2442 940 lim = init_t->_hi;
kvn@2442 941 ini = limit_t->_lo;
kvn@2442 942 max = (julong)min_jint;
kvn@2442 943 }
kvn@2442 944 julong range = lim - ini + stride_p;
kvn@2442 945 if (range <= max) {
kvn@2442 946 // Convert to integer expression if it is not overflow.
kvn@2442 947 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
kvn@3680 948 Node *range = phase->transform(new (phase->C) SubINode(in(Limit), in(Init)));
kvn@3680 949 Node *bias = phase->transform(new (phase->C) AddINode(range, stride_m));
kvn@3680 950 Node *trip = phase->transform(new (phase->C) DivINode(0, bias, in(Stride)));
kvn@3680 951 Node *span = phase->transform(new (phase->C) MulINode(trip, in(Stride)));
kvn@3680 952 return new (phase->C) AddINode(span, in(Init)); // exact limit
kvn@2442 953 }
kvn@2442 954
kvn@2442 955 if (is_power_of_2(stride_p) || // divisor is 2^n
kvn@2442 956 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
kvn@2442 957 // Convert to long expression to avoid integer overflow
kvn@2442 958 // and let igvn optimizer convert this division.
kvn@2442 959 //
kvn@3680 960 Node* init = phase->transform( new (phase->C) ConvI2LNode(in(Init)));
kvn@3680 961 Node* limit = phase->transform( new (phase->C) ConvI2LNode(in(Limit)));
kvn@2442 962 Node* stride = phase->longcon(stride_con);
kvn@2442 963 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
kvn@2442 964
kvn@3680 965 Node *range = phase->transform(new (phase->C) SubLNode(limit, init));
kvn@3680 966 Node *bias = phase->transform(new (phase->C) AddLNode(range, stride_m));
kvn@2442 967 Node *span;
kvn@2442 968 if (stride_con > 0 && is_power_of_2(stride_p)) {
kvn@2442 969 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
kvn@2442 970 // and avoid generating rounding for division. Zero trip guard should
kvn@2442 971 // guarantee that init < limit but sometimes the guard is missing and
kvn@2442 972 // we can get situation when init > limit. Note, for the empty loop
kvn@2442 973 // optimization zero trip guard is generated explicitly which leaves
kvn@2442 974 // only RCE predicate where exact limit is used and the predicate
kvn@2442 975 // will simply fail forcing recompilation.
kvn@2442 976 Node* neg_stride = phase->longcon(-stride_con);
kvn@3680 977 span = phase->transform(new (phase->C) AndLNode(bias, neg_stride));
kvn@2442 978 } else {
kvn@3680 979 Node *trip = phase->transform(new (phase->C) DivLNode(0, bias, stride));
kvn@3680 980 span = phase->transform(new (phase->C) MulLNode(trip, stride));
kvn@2442 981 }
kvn@2442 982 // Convert back to int
kvn@3680 983 Node *span_int = phase->transform(new (phase->C) ConvL2INode(span));
kvn@3680 984 return new (phase->C) AddINode(span_int, in(Init)); // exact limit
kvn@2442 985 }
kvn@2442 986
kvn@2442 987 return NULL; // No progress
kvn@2442 988 }
kvn@2442 989
kvn@2442 990 //------------------------------Identity---------------------------------------
kvn@2442 991 // If stride == 1 return limit node.
kvn@2442 992 Node *LoopLimitNode::Identity( PhaseTransform *phase ) {
kvn@2442 993 int stride_con = phase->type(in(Stride))->is_int()->get_con();
kvn@2442 994 if (stride_con == 1 || stride_con == -1)
kvn@2442 995 return in(Limit);
kvn@2442 996 return this;
kvn@2442 997 }
kvn@2442 998
kvn@2442 999 //=============================================================================
duke@0 1000 //----------------------match_incr_with_optional_truncation--------------------
duke@0 1001 // Match increment with optional truncation:
duke@0 1002 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
duke@0 1003 // Return NULL for failure. Success returns the increment node.
duke@0 1004 Node* CountedLoopNode::match_incr_with_optional_truncation(
duke@0 1005 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
duke@0 1006 // Quick cutouts:
brutisso@3054 1007 if (expr == NULL || expr->req() != 3) return NULL;
duke@0 1008
duke@0 1009 Node *t1 = NULL;
duke@0 1010 Node *t2 = NULL;
duke@0 1011 const TypeInt* trunc_t = TypeInt::INT;
duke@0 1012 Node* n1 = expr;
duke@0 1013 int n1op = n1->Opcode();
duke@0 1014
duke@0 1015 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
duke@0 1016 if (n1op == Op_AndI &&
duke@0 1017 n1->in(2)->is_Con() &&
duke@0 1018 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
duke@0 1019 // %%% This check should match any mask of 2**K-1.
duke@0 1020 t1 = n1;
duke@0 1021 n1 = t1->in(1);
duke@0 1022 n1op = n1->Opcode();
duke@0 1023 trunc_t = TypeInt::CHAR;
duke@0 1024 } else if (n1op == Op_RShiftI &&
duke@0 1025 n1->in(1) != NULL &&
duke@0 1026 n1->in(1)->Opcode() == Op_LShiftI &&
duke@0 1027 n1->in(2) == n1->in(1)->in(2) &&
duke@0 1028 n1->in(2)->is_Con()) {
duke@0 1029 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
duke@0 1030 // %%% This check should match any shift in [1..31].
duke@0 1031 if (shift == 16 || shift == 8) {
duke@0 1032 t1 = n1;
duke@0 1033 t2 = t1->in(1);
duke@0 1034 n1 = t2->in(1);
duke@0 1035 n1op = n1->Opcode();
duke@0 1036 if (shift == 16) {
duke@0 1037 trunc_t = TypeInt::SHORT;
duke@0 1038 } else if (shift == 8) {
duke@0 1039 trunc_t = TypeInt::BYTE;
duke@0 1040 }
duke@0 1041 }
duke@0 1042 }
duke@0 1043
duke@0 1044 // If (maybe after stripping) it is an AddI, we won:
duke@0 1045 if (n1op == Op_AddI) {
duke@0 1046 *trunc1 = t1;
duke@0 1047 *trunc2 = t2;
duke@0 1048 *trunc_type = trunc_t;
duke@0 1049 return n1;
duke@0 1050 }
duke@0 1051
duke@0 1052 // failed
duke@0 1053 return NULL;
duke@0 1054 }
duke@0 1055
duke@0 1056
duke@0 1057 //------------------------------filtered_type--------------------------------
duke@0 1058 // Return a type based on condition control flow
duke@0 1059 // A successful return will be a type that is restricted due
duke@0 1060 // to a series of dominating if-tests, such as:
duke@0 1061 // if (i < 10) {
duke@0 1062 // if (i > 0) {
duke@0 1063 // here: "i" type is [1..10)
duke@0 1064 // }
duke@0 1065 // }
duke@0 1066 // or a control flow merge
duke@0 1067 // if (i < 10) {
duke@0 1068 // do {
duke@0 1069 // phi( , ) -- at top of loop type is [min_int..10)
duke@0 1070 // i = ?
duke@0 1071 // } while ( i < 10)
duke@0 1072 //
duke@0 1073 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
duke@0 1074 assert(n && n->bottom_type()->is_int(), "must be int");
duke@0 1075 const TypeInt* filtered_t = NULL;
duke@0 1076 if (!n->is_Phi()) {
duke@0 1077 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
duke@0 1078 filtered_t = filtered_type_from_dominators(n, n_ctrl);
duke@0 1079
duke@0 1080 } else {
duke@0 1081 Node* phi = n->as_Phi();
duke@0 1082 Node* region = phi->in(0);
duke@0 1083 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
duke@0 1084 if (region && region != C->top()) {
duke@0 1085 for (uint i = 1; i < phi->req(); i++) {
duke@0 1086 Node* val = phi->in(i);
duke@0 1087 Node* use_c = region->in(i);
duke@0 1088 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
duke@0 1089 if (val_t != NULL) {
duke@0 1090 if (filtered_t == NULL) {
duke@0 1091 filtered_t = val_t;
duke@0 1092 } else {
duke@0 1093 filtered_t = filtered_t->meet(val_t)->is_int();
duke@0 1094 }
duke@0 1095 }
duke@0 1096 }
duke@0 1097 }
duke@0 1098 }
duke@0 1099 const TypeInt* n_t = _igvn.type(n)->is_int();
duke@0 1100 if (filtered_t != NULL) {
duke@0 1101 n_t = n_t->join(filtered_t)->is_int();
duke@0 1102 }
duke@0 1103 return n_t;
duke@0 1104 }
duke@0 1105
duke@0 1106
duke@0 1107 //------------------------------filtered_type_from_dominators--------------------------------
duke@0 1108 // Return a possibly more restrictive type for val based on condition control flow of dominators
duke@0 1109 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
duke@0 1110 if (val->is_Con()) {
duke@0 1111 return val->bottom_type()->is_int();
duke@0 1112 }
duke@0 1113 uint if_limit = 10; // Max number of dominating if's visited
duke@0 1114 const TypeInt* rtn_t = NULL;
duke@0 1115
duke@0 1116 if (use_ctrl && use_ctrl != C->top()) {
duke@0 1117 Node* val_ctrl = get_ctrl(val);
duke@0 1118 uint val_dom_depth = dom_depth(val_ctrl);
duke@0 1119 Node* pred = use_ctrl;
duke@0 1120 uint if_cnt = 0;
duke@0 1121 while (if_cnt < if_limit) {
duke@0 1122 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
duke@0 1123 if_cnt++;
never@17 1124 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
duke@0 1125 if (if_t != NULL) {
duke@0 1126 if (rtn_t == NULL) {
duke@0 1127 rtn_t = if_t;
duke@0 1128 } else {
duke@0 1129 rtn_t = rtn_t->join(if_t)->is_int();
duke@0 1130 }
duke@0 1131 }
duke@0 1132 }
duke@0 1133 pred = idom(pred);
duke@0 1134 if (pred == NULL || pred == C->top()) {
duke@0 1135 break;
duke@0 1136 }
duke@0 1137 // Stop if going beyond definition block of val
duke@0 1138 if (dom_depth(pred) < val_dom_depth) {
duke@0 1139 break;
duke@0 1140 }
duke@0 1141 }
duke@0 1142 }
duke@0 1143 return rtn_t;
duke@0 1144 }
duke@0 1145
duke@0 1146
duke@0 1147 //------------------------------dump_spec--------------------------------------
duke@0 1148 // Dump special per-node info
duke@0 1149 #ifndef PRODUCT
duke@0 1150 void CountedLoopEndNode::dump_spec(outputStream *st) const {
duke@0 1151 if( in(TestValue)->is_Bool() ) {
duke@0 1152 BoolTest bt( test_trip()); // Added this for g++.
duke@0 1153
duke@0 1154 st->print("[");
duke@0 1155 bt.dump_on(st);
duke@0 1156 st->print("]");
duke@0 1157 }
duke@0 1158 st->print(" ");
duke@0 1159 IfNode::dump_spec(st);
duke@0 1160 }
duke@0 1161 #endif
duke@0 1162
duke@0 1163 //=============================================================================
duke@0 1164 //------------------------------is_member--------------------------------------
duke@0 1165 // Is 'l' a member of 'this'?
duke@0 1166 int IdealLoopTree::is_member( const IdealLoopTree *l ) const {
duke@0 1167 while( l->_nest > _nest ) l = l->_parent;
duke@0 1168 return l == this;
duke@0 1169 }
duke@0 1170
duke@0 1171 //------------------------------set_nest---------------------------------------
duke@0 1172 // Set loop tree nesting depth. Accumulate _has_call bits.
duke@0 1173 int IdealLoopTree::set_nest( uint depth ) {
duke@0 1174 _nest = depth;
duke@0 1175 int bits = _has_call;
duke@0 1176 if( _child ) bits |= _child->set_nest(depth+1);
duke@0 1177 if( bits ) _has_call = 1;
duke@0 1178 if( _next ) bits |= _next ->set_nest(depth );
duke@0 1179 return bits;
duke@0 1180 }
duke@0 1181
duke@0 1182 //------------------------------split_fall_in----------------------------------
duke@0 1183 // Split out multiple fall-in edges from the loop header. Move them to a
duke@0 1184 // private RegionNode before the loop. This becomes the loop landing pad.
duke@0 1185 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
duke@0 1186 PhaseIterGVN &igvn = phase->_igvn;
duke@0 1187 uint i;
duke@0 1188
duke@0 1189 // Make a new RegionNode to be the landing pad.
kvn@3680 1190 Node *landing_pad = new (phase->C) RegionNode( fall_in_cnt+1 );
duke@0 1191 phase->set_loop(landing_pad,_parent);
duke@0 1192 // Gather all the fall-in control paths into the landing pad
duke@0 1193 uint icnt = fall_in_cnt;
duke@0 1194 uint oreq = _head->req();
duke@0 1195 for( i = oreq-1; i>0; i-- )
duke@0 1196 if( !phase->is_member( this, _head->in(i) ) )
duke@0 1197 landing_pad->set_req(icnt--,_head->in(i));
duke@0 1198
duke@0 1199 // Peel off PhiNode edges as well
duke@0 1200 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
duke@0 1201 Node *oj = _head->fast_out(j);
duke@0 1202 if( oj->is_Phi() ) {
duke@0 1203 PhiNode* old_phi = oj->as_Phi();
duke@0 1204 assert( old_phi->region() == _head, "" );
duke@0 1205 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
duke@0 1206 Node *p = PhiNode::make_blank(landing_pad, old_phi);
duke@0 1207 uint icnt = fall_in_cnt;
duke@0 1208 for( i = oreq-1; i>0; i-- ) {
duke@0 1209 if( !phase->is_member( this, _head->in(i) ) ) {
duke@0 1210 p->init_req(icnt--, old_phi->in(i));
duke@0 1211 // Go ahead and clean out old edges from old phi
duke@0 1212 old_phi->del_req(i);
duke@0 1213 }
duke@0 1214 }
duke@0 1215 // Search for CSE's here, because ZKM.jar does a lot of
duke@0 1216 // loop hackery and we need to be a little incremental
duke@0 1217 // with the CSE to avoid O(N^2) node blow-up.
duke@0 1218 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
duke@0 1219 if( p2 ) { // Found CSE
duke@0 1220 p->destruct(); // Recover useless new node
duke@0 1221 p = p2; // Use old node
duke@0 1222 } else {
duke@0 1223 igvn.register_new_node_with_optimizer(p, old_phi);
duke@0 1224 }
duke@0 1225 // Make old Phi refer to new Phi.
duke@0 1226 old_phi->add_req(p);
duke@0 1227 // Check for the special case of making the old phi useless and
duke@0 1228 // disappear it. In JavaGrande I have a case where this useless
duke@0 1229 // Phi is the loop limit and prevents recognizing a CountedLoop
duke@0 1230 // which in turn prevents removing an empty loop.
duke@0 1231 Node *id_old_phi = old_phi->Identity( &igvn );
duke@0 1232 if( id_old_phi != old_phi ) { // Found a simple identity?
kvn@1541 1233 // Note that I cannot call 'replace_node' here, because
duke@0 1234 // that will yank the edge from old_phi to the Region and
duke@0 1235 // I'm mid-iteration over the Region's uses.
duke@0 1236 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
duke@0 1237 Node* use = old_phi->last_out(i);
kvn@3412 1238 igvn.rehash_node_delayed(use);
duke@0 1239 uint uses_found = 0;
duke@0 1240 for (uint j = 0; j < use->len(); j++) {
duke@0 1241 if (use->in(j) == old_phi) {
duke@0 1242 if (j < use->req()) use->set_req (j, id_old_phi);
duke@0 1243 else use->set_prec(j, id_old_phi);
duke@0 1244 uses_found++;
duke@0 1245 }
duke@0 1246 }
duke@0 1247 i -= uses_found; // we deleted 1 or more copies of this edge
duke@0 1248 }
duke@0 1249 }
duke@0 1250 igvn._worklist.push(old_phi);
duke@0 1251 }
duke@0 1252 }
duke@0 1253 // Finally clean out the fall-in edges from the RegionNode
duke@0 1254 for( i = oreq-1; i>0; i-- ) {
duke@0 1255 if( !phase->is_member( this, _head->in(i) ) ) {
duke@0 1256 _head->del_req(i);
duke@0 1257 }
duke@0 1258 }
duke@0 1259 // Transform landing pad
duke@0 1260 igvn.register_new_node_with_optimizer(landing_pad, _head);
duke@0 1261 // Insert landing pad into the header
duke@0 1262 _head->add_req(landing_pad);
duke@0 1263 }
duke@0 1264
duke@0 1265 //------------------------------split_outer_loop-------------------------------
duke@0 1266 // Split out the outermost loop from this shared header.
duke@0 1267 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
duke@0 1268 PhaseIterGVN &igvn = phase->_igvn;
duke@0 1269
duke@0 1270 // Find index of outermost loop; it should also be my tail.
duke@0 1271 uint outer_idx = 1;
duke@0 1272 while( _head->in(outer_idx) != _tail ) outer_idx++;
duke@0 1273
duke@0 1274 // Make a LoopNode for the outermost loop.
duke@0 1275 Node *ctl = _head->in(LoopNode::EntryControl);
kvn@3680 1276 Node *outer = new (phase->C) LoopNode( ctl, _head->in(outer_idx) );
duke@0 1277 outer = igvn.register_new_node_with_optimizer(outer, _head);
duke@0 1278 phase->set_created_loop_node();
kvn@2292 1279
duke@0 1280 // Outermost loop falls into '_head' loop
kvn@2608 1281 _head->set_req(LoopNode::EntryControl, outer);
duke@0 1282 _head->del_req(outer_idx);
duke@0 1283 // Split all the Phis up between '_head' loop and 'outer' loop.
duke@0 1284 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
duke@0 1285 Node *out = _head->fast_out(j);
duke@0 1286 if( out->is_Phi() ) {
duke@0 1287 PhiNode *old_phi = out->as_Phi();
duke@0 1288 assert( old_phi->region() == _head, "" );
duke@0 1289 Node *phi = PhiNode::make_blank(outer, old_phi);
duke@0 1290 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
duke@0 1291 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
duke@0 1292 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
duke@0 1293 // Make old Phi point to new Phi on the fall-in path
kvn@3412 1294 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
duke@0 1295 old_phi->del_req(outer_idx);
duke@0 1296 }
duke@0 1297 }
duke@0 1298
duke@0 1299 // Use the new loop head instead of the old shared one
duke@0 1300 _head = outer;
duke@0 1301 phase->set_loop(_head, this);
duke@0 1302 }
duke@0 1303
duke@0 1304 //------------------------------fix_parent-------------------------------------
duke@0 1305 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
duke@0 1306 loop->_parent = parent;
duke@0 1307 if( loop->_child ) fix_parent( loop->_child, loop );
duke@0 1308 if( loop->_next ) fix_parent( loop->_next , parent );
duke@0 1309 }
duke@0 1310
duke@0 1311 //------------------------------estimate_path_freq-----------------------------
duke@0 1312 static float estimate_path_freq( Node *n ) {
duke@0 1313 // Try to extract some path frequency info
duke@0 1314 IfNode *iff;
duke@0 1315 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
duke@0 1316 uint nop = n->Opcode();
duke@0 1317 if( nop == Op_SafePoint ) { // Skip any safepoint
duke@0 1318 n = n->in(0);
duke@0 1319 continue;
duke@0 1320 }
duke@0 1321 if( nop == Op_CatchProj ) { // Get count from a prior call
duke@0 1322 // Assume call does not always throw exceptions: means the call-site
duke@0 1323 // count is also the frequency of the fall-through path.
duke@0 1324 assert( n->is_CatchProj(), "" );
duke@0 1325 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
duke@0 1326 return 0.0f; // Assume call exception path is rare
duke@0 1327 Node *call = n->in(0)->in(0)->in(0);
duke@0 1328 assert( call->is_Call(), "expect a call here" );
duke@0 1329 const JVMState *jvms = ((CallNode*)call)->jvms();
duke@0 1330 ciMethodData* methodData = jvms->method()->method_data();
duke@0 1331 if (!methodData->is_mature()) return 0.0f; // No call-site data
duke@0 1332 ciProfileData* data = methodData->bci_to_data(jvms->bci());
duke@0 1333 if ((data == NULL) || !data->is_CounterData()) {
duke@0 1334 // no call profile available, try call's control input
duke@0 1335 n = n->in(0);
duke@0 1336 continue;
duke@0 1337 }
duke@0 1338 return data->as_CounterData()->count()/FreqCountInvocations;
duke@0 1339 }
duke@0 1340 // See if there's a gating IF test
duke@0 1341 Node *n_c = n->in(0);
duke@0 1342 if( !n_c->is_If() ) break; // No estimate available
duke@0 1343 iff = n_c->as_If();
duke@0 1344 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
duke@0 1345 // Compute how much count comes on this path
duke@0 1346 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
duke@0 1347 // Have no count info. Skip dull uncommon-trap like branches.
duke@0 1348 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
duke@0 1349 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
duke@0 1350 break;
duke@0 1351 // Skip through never-taken branch; look for a real loop exit.
duke@0 1352 n = iff->in(0);
duke@0 1353 }
duke@0 1354 return 0.0f; // No estimate available
duke@0 1355 }
duke@0 1356
duke@0 1357 //------------------------------merge_many_backedges---------------------------
duke@0 1358 // Merge all the backedges from the shared header into a private Region.
duke@0 1359 // Feed that region as the one backedge to this loop.
duke@0 1360 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
duke@0 1361 uint i;
duke@0 1362
duke@0 1363 // Scan for the top 2 hottest backedges
duke@0 1364 float hotcnt = 0.0f;
duke@0 1365 float warmcnt = 0.0f;
duke@0 1366 uint hot_idx = 0;
duke@0 1367 // Loop starts at 2 because slot 1 is the fall-in path
duke@0 1368 for( i = 2; i < _head->req(); i++ ) {
duke@0 1369 float cnt = estimate_path_freq(_head->in(i));
duke@0 1370 if( cnt > hotcnt ) { // Grab hottest path
duke@0 1371 warmcnt = hotcnt;
duke@0 1372 hotcnt = cnt;
duke@0 1373 hot_idx = i;
duke@0 1374 } else if( cnt > warmcnt ) { // And 2nd hottest path
duke@0 1375 warmcnt = cnt;
duke@0 1376 }
duke@0 1377 }
duke@0 1378
duke@0 1379 // See if the hottest backedge is worthy of being an inner loop
duke@0 1380 // by being much hotter than the next hottest backedge.
duke@0 1381 if( hotcnt <= 0.0001 ||
duke@0 1382 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
duke@0 1383
duke@0 1384 // Peel out the backedges into a private merge point; peel
duke@0 1385 // them all except optionally hot_idx.
duke@0 1386 PhaseIterGVN &igvn = phase->_igvn;
duke@0 1387
duke@0 1388 Node *hot_tail = NULL;
duke@0 1389 // Make a Region for the merge point
kvn@3680 1390 Node *r = new (phase->C) RegionNode(1);
duke@0 1391 for( i = 2; i < _head->req(); i++ ) {
duke@0 1392 if( i != hot_idx )
duke@0 1393 r->add_req( _head->in(i) );
duke@0 1394 else hot_tail = _head->in(i);
duke@0 1395 }
duke@0 1396 igvn.register_new_node_with_optimizer(r, _head);
duke@0 1397 // Plug region into end of loop _head, followed by hot_tail
duke@0 1398 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
duke@0 1399 _head->set_req(2, r);
duke@0 1400 if( hot_idx ) _head->add_req(hot_tail);
duke@0 1401
duke@0 1402 // Split all the Phis up between '_head' loop and the Region 'r'
duke@0 1403 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
duke@0 1404 Node *out = _head->fast_out(j);
duke@0 1405 if( out->is_Phi() ) {
duke@0 1406 PhiNode* n = out->as_Phi();
duke@0 1407 igvn.hash_delete(n); // Delete from hash before hacking edges
duke@0 1408 Node *hot_phi = NULL;
kvn@3680 1409 Node *phi = new (phase->C) PhiNode(r, n->type(), n->adr_type());
duke@0 1410 // Check all inputs for the ones to peel out
duke@0 1411 uint j = 1;
duke@0 1412 for( uint i = 2; i < n->req(); i++ ) {
duke@0 1413 if( i != hot_idx )
duke@0 1414 phi->set_req( j++, n->in(i) );
duke@0 1415 else hot_phi = n->in(i);
duke@0 1416 }
duke@0 1417 // Register the phi but do not transform until whole place transforms
duke@0 1418 igvn.register_new_node_with_optimizer(phi, n);
duke@0 1419 // Add the merge phi to the old Phi
duke@0 1420 while( n->req() > 3 ) n->del_req( n->req()-1 );
duke@0 1421 n->set_req(2, phi);
duke@0 1422 if( hot_idx ) n->add_req(hot_phi);
duke@0 1423 }
duke@0 1424 }
duke@0 1425
duke@0 1426
duke@0 1427 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
duke@0 1428 // of self loop tree. Turn self into a loop headed by _head and with
duke@0 1429 // tail being the new merge point.
duke@0 1430 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
duke@0 1431 phase->set_loop(_tail,ilt); // Adjust tail
duke@0 1432 _tail = r; // Self's tail is new merge point
duke@0 1433 phase->set_loop(r,this);
duke@0 1434 ilt->_child = _child; // New guy has my children
duke@0 1435 _child = ilt; // Self has new guy as only child
duke@0 1436 ilt->_parent = this; // new guy has self for parent
duke@0 1437 ilt->_nest = _nest; // Same nesting depth (for now)
duke@0 1438
duke@0 1439 // Starting with 'ilt', look for child loop trees using the same shared
duke@0 1440 // header. Flatten these out; they will no longer be loops in the end.
duke@0 1441 IdealLoopTree **pilt = &_child;
duke@0 1442 while( ilt ) {
duke@0 1443 if( ilt->_head == _head ) {
duke@0 1444 uint i;
duke@0 1445 for( i = 2; i < _head->req(); i++ )
duke@0 1446 if( _head->in(i) == ilt->_tail )
duke@0 1447 break; // Still a loop
duke@0 1448 if( i == _head->req() ) { // No longer a loop
duke@0 1449 // Flatten ilt. Hang ilt's "_next" list from the end of
duke@0 1450 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
duke@0 1451 IdealLoopTree **cp = &ilt->_child;
duke@0 1452 while( *cp ) cp = &(*cp)->_next; // Find end of child list
duke@0 1453 *cp = ilt->_next; // Hang next list at end of child list
duke@0 1454 *pilt = ilt->_child; // Move child up to replace ilt
duke@0 1455 ilt->_head = NULL; // Flag as a loop UNIONED into parent
duke@0 1456 ilt = ilt->_child; // Repeat using new ilt
duke@0 1457 continue; // do not advance over ilt->_child
duke@0 1458 }
duke@0 1459 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
duke@0 1460 phase->set_loop(_head,ilt);
duke@0 1461 }
duke@0 1462 pilt = &ilt->_child; // Advance to next
duke@0 1463 ilt = *pilt;
duke@0 1464 }
duke@0 1465
duke@0 1466 if( _child ) fix_parent( _child, this );
duke@0 1467 }
duke@0 1468
duke@0 1469 //------------------------------beautify_loops---------------------------------
duke@0 1470 // Split shared headers and insert loop landing pads.
duke@0 1471 // Insert a LoopNode to replace the RegionNode.
duke@0 1472 // Return TRUE if loop tree is structurally changed.
duke@0 1473 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
duke@0 1474 bool result = false;
duke@0 1475 // Cache parts in locals for easy
duke@0 1476 PhaseIterGVN &igvn = phase->_igvn;
duke@0 1477
duke@0 1478 igvn.hash_delete(_head); // Yank from hash before hacking edges
duke@0 1479
duke@0 1480 // Check for multiple fall-in paths. Peel off a landing pad if need be.
duke@0 1481 int fall_in_cnt = 0;
duke@0 1482 for( uint i = 1; i < _head->req(); i++ )
duke@0 1483 if( !phase->is_member( this, _head->in(i) ) )
duke@0 1484 fall_in_cnt++;
duke@0 1485 assert( fall_in_cnt, "at least 1 fall-in path" );
duke@0 1486 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
duke@0 1487 split_fall_in( phase, fall_in_cnt );
duke@0 1488
duke@0 1489 // Swap inputs to the _head and all Phis to move the fall-in edge to
duke@0 1490 // the left.
duke@0 1491 fall_in_cnt = 1;
duke@0 1492 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
duke@0 1493 fall_in_cnt++;
duke@0 1494 if( fall_in_cnt > 1 ) {
duke@0 1495 // Since I am just swapping inputs I do not need to update def-use info
duke@0 1496 Node *tmp = _head->in(1);
duke@0 1497 _head->set_req( 1, _head->in(fall_in_cnt) );
duke@0 1498 _head->set_req( fall_in_cnt, tmp );
duke@0 1499 // Swap also all Phis
duke@0 1500 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
duke@0 1501 Node* phi = _head->fast_out(i);
duke@0 1502 if( phi->is_Phi() ) {
duke@0 1503 igvn.hash_delete(phi); // Yank from hash before hacking edges
duke@0 1504 tmp = phi->in(1);
duke@0 1505 phi->set_req( 1, phi->in(fall_in_cnt) );
duke@0 1506 phi->set_req( fall_in_cnt, tmp );
duke@0 1507 }
duke@0 1508 }
duke@0 1509 }
duke@0 1510 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
duke@0 1511 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
duke@0 1512
duke@0 1513 // If I am a shared header (multiple backedges), peel off the many
duke@0 1514 // backedges into a private merge point and use the merge point as
duke@0 1515 // the one true backedge.
duke@0 1516 if( _head->req() > 3 ) {
kvn@2292 1517 // Merge the many backedges into a single backedge but leave
kvn@2292 1518 // the hottest backedge as separate edge for the following peel.
duke@0 1519 merge_many_backedges( phase );
duke@0 1520 result = true;
duke@0 1521 }
duke@0 1522
kvn@2292 1523 // If I have one hot backedge, peel off myself loop.
duke@0 1524 // I better be the outermost loop.
duke@0 1525 if( _head->req() > 3 ) {
duke@0 1526 split_outer_loop( phase );
duke@0 1527 result = true;
duke@0 1528
duke@0 1529 } else if( !_head->is_Loop() && !_irreducible ) {
duke@0 1530 // Make a new LoopNode to replace the old loop head
kvn@3680 1531 Node *l = new (phase->C) LoopNode( _head->in(1), _head->in(2) );
duke@0 1532 l = igvn.register_new_node_with_optimizer(l, _head);
duke@0 1533 phase->set_created_loop_node();
duke@0 1534 // Go ahead and replace _head
kvn@1541 1535 phase->_igvn.replace_node( _head, l );
duke@0 1536 _head = l;
duke@0 1537 phase->set_loop(_head, this);
duke@0 1538 }
duke@0 1539
duke@0 1540 // Now recursively beautify nested loops
duke@0 1541 if( _child ) result |= _child->beautify_loops( phase );
duke@0 1542 if( _next ) result |= _next ->beautify_loops( phase );
duke@0 1543 return result;
duke@0 1544 }
duke@0 1545
duke@0 1546 //------------------------------allpaths_check_safepts----------------------------
duke@0 1547 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
duke@0 1548 // encountered. Helper for check_safepts.
duke@0 1549 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
duke@0 1550 assert(stack.size() == 0, "empty stack");
duke@0 1551 stack.push(_tail);
duke@0 1552 visited.Clear();
duke@0 1553 visited.set(_tail->_idx);
duke@0 1554 while (stack.size() > 0) {
duke@0 1555 Node* n = stack.pop();
duke@0 1556 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
duke@0 1557 // Terminate this path
duke@0 1558 } else if (n->Opcode() == Op_SafePoint) {
duke@0 1559 if (_phase->get_loop(n) != this) {
duke@0 1560 if (_required_safept == NULL) _required_safept = new Node_List();
duke@0 1561 _required_safept->push(n); // save the one closest to the tail
duke@0 1562 }
duke@0 1563 // Terminate this path
duke@0 1564 } else {
duke@0 1565 uint start = n->is_Region() ? 1 : 0;
duke@0 1566 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
duke@0 1567 for (uint i = start; i < end; i++) {
duke@0 1568 Node* in = n->in(i);
duke@0 1569 assert(in->is_CFG(), "must be");
duke@0 1570 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
duke@0 1571 stack.push(in);
duke@0 1572 }
duke@0 1573 }
duke@0 1574 }
duke@0 1575 }
duke@0 1576 }
duke@0 1577
duke@0 1578 //------------------------------check_safepts----------------------------
duke@0 1579 // Given dominators, try to find loops with calls that must always be
duke@0 1580 // executed (call dominates loop tail). These loops do not need non-call
duke@0 1581 // safepoints (ncsfpt).
duke@0 1582 //
duke@0 1583 // A complication is that a safepoint in a inner loop may be needed
duke@0 1584 // by an outer loop. In the following, the inner loop sees it has a
duke@0 1585 // call (block 3) on every path from the head (block 2) to the
duke@0 1586 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
duke@0 1587 // in block 2, _but_ this leaves the outer loop without a safepoint.
duke@0 1588 //
duke@0 1589 // entry 0
duke@0 1590 // |
duke@0 1591 // v
duke@0 1592 // outer 1,2 +->1
duke@0 1593 // | |
duke@0 1594 // | v
duke@0 1595 // | 2<---+ ncsfpt in 2
duke@0 1596 // |_/|\ |
duke@0 1597 // | v |
duke@0 1598 // inner 2,3 / 3 | call in 3
duke@0 1599 // / | |
duke@0 1600 // v +--+
duke@0 1601 // exit 4
duke@0 1602 //
duke@0 1603 //
duke@0 1604 // This method creates a list (_required_safept) of ncsfpt nodes that must
duke@0 1605 // be protected is created for each loop. When a ncsfpt maybe deleted, it
duke@0 1606 // is first looked for in the lists for the outer loops of the current loop.
duke@0 1607 //
duke@0 1608 // The insights into the problem:
duke@0 1609 // A) counted loops are okay
duke@0 1610 // B) innermost loops are okay (only an inner loop can delete
duke@0 1611 // a ncsfpt needed by an outer loop)
duke@0 1612 // C) a loop is immune from an inner loop deleting a safepoint
duke@0 1613 // if the loop has a call on the idom-path
duke@0 1614 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
duke@0 1615 // idom-path that is not in a nested loop
duke@0 1616 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
duke@0 1617 // loop needs to be prevented from deletion by an inner loop
duke@0 1618 //
duke@0 1619 // There are two analyses:
duke@0 1620 // 1) The first, and cheaper one, scans the loop body from
duke@0 1621 // tail to head following the idom (immediate dominator)
duke@0 1622 // chain, looking for the cases (C,D,E) above.
duke@0 1623 // Since inner loops are scanned before outer loops, there is summary
duke@0 1624 // information about inner loops. Inner loops can be skipped over
duke@0 1625 // when the tail of an inner loop is encountered.
duke@0 1626 //
duke@0 1627 // 2) The second, invoked if the first fails to find a call or ncsfpt on
duke@0 1628 // the idom path (which is rare), scans all predecessor control paths
duke@0 1629 // from the tail to the head, terminating a path when a call or sfpt
duke@0 1630 // is encountered, to find the ncsfpt's that are closest to the tail.
duke@0 1631 //
duke@0 1632 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
duke@0 1633 // Bottom up traversal
duke@0 1634 IdealLoopTree* ch = _child;
kvn@3588 1635 if (_child) _child->check_safepts(visited, stack);
kvn@3588 1636 if (_next) _next ->check_safepts(visited, stack);
duke@0 1637
duke@0 1638 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
duke@0 1639 bool has_call = false; // call on dom-path
duke@0 1640 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
duke@0 1641 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
duke@0 1642 // Scan the dom-path nodes from tail to head
duke@0 1643 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
duke@0 1644 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
duke@0 1645 has_call = true;
duke@0 1646 _has_sfpt = 1; // Then no need for a safept!
duke@0 1647 break;
duke@0 1648 } else if (n->Opcode() == Op_SafePoint) {
duke@0 1649 if (_phase->get_loop(n) == this) {
duke@0 1650 has_local_ncsfpt = true;
duke@0 1651 break;
duke@0 1652 }
duke@0 1653 if (nonlocal_ncsfpt == NULL) {
duke@0 1654 nonlocal_ncsfpt = n; // save the one closest to the tail
duke@0 1655 }
duke@0 1656 } else {
duke@0 1657 IdealLoopTree* nlpt = _phase->get_loop(n);
duke@0 1658 if (this != nlpt) {
duke@0 1659 // If at an inner loop tail, see if the inner loop has already
duke@0 1660 // recorded seeing a call on the dom-path (and stop.) If not,
duke@0 1661 // jump to the head of the inner loop.
duke@0 1662 assert(is_member(nlpt), "nested loop");
duke@0 1663 Node* tail = nlpt->_tail;
duke@0 1664 if (tail->in(0)->is_If()) tail = tail->in(0);
duke@0 1665 if (n == tail) {
duke@0 1666 // If inner loop has call on dom-path, so does outer loop
duke@0 1667 if (nlpt->_has_sfpt) {
duke@0 1668 has_call = true;
duke@0 1669 _has_sfpt = 1;
duke@0 1670 break;
duke@0 1671 }
duke@0 1672 // Skip to head of inner loop
duke@0 1673 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
duke@0 1674 n = nlpt->_head;
duke@0 1675 }
duke@0 1676 }
duke@0 1677 }
duke@0 1678 }
duke@0 1679 // Record safept's that this loop needs preserved when an
duke@0 1680 // inner loop attempts to delete it's safepoints.
duke@0 1681 if (_child != NULL && !has_call && !has_local_ncsfpt) {
duke@0 1682 if (nonlocal_ncsfpt != NULL) {
duke@0 1683 if (_required_safept == NULL) _required_safept = new Node_List();
duke@0 1684 _required_safept->push(nonlocal_ncsfpt);
duke@0 1685 } else {
duke@0 1686 // Failed to find a suitable safept on the dom-path. Now use
duke@0 1687 // an all paths walk from tail to head, looking for safepoints to preserve.
duke@0 1688 allpaths_check_safepts(visited, stack);
duke@0 1689 }
duke@0 1690 }
duke@0 1691 }
duke@0 1692 }
duke@0 1693
duke@0 1694 //---------------------------is_deleteable_safept----------------------------
duke@0 1695 // Is safept not required by an outer loop?
duke@0 1696 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
duke@0 1697 assert(sfpt->Opcode() == Op_SafePoint, "");
duke@0 1698 IdealLoopTree* lp = get_loop(sfpt)->_parent;
duke@0 1699 while (lp != NULL) {
duke@0 1700 Node_List* sfpts = lp->_required_safept;
duke@0 1701 if (sfpts != NULL) {
duke@0 1702 for (uint i = 0; i < sfpts->size(); i++) {
duke@0 1703 if (sfpt == sfpts->at(i))
duke@0 1704 return false;
duke@0 1705 }
duke@0 1706 }
duke@0 1707 lp = lp->_parent;
duke@0 1708 }
duke@0 1709 return true;
duke@0 1710 }
duke@0 1711
kvn@2230 1712 //---------------------------replace_parallel_iv-------------------------------
kvn@2230 1713 // Replace parallel induction variable (parallel to trip counter)
kvn@2230 1714 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
kvn@2230 1715 assert(loop->_head->is_CountedLoop(), "");
kvn@2230 1716 CountedLoopNode *cl = loop->_head->as_CountedLoop();
kvn@2613 1717 if (!cl->is_valid_counted_loop())
kvn@2613 1718 return; // skip malformed counted loop
kvn@2230 1719 Node *incr = cl->incr();
kvn@2230 1720 if (incr == NULL)
kvn@2230 1721 return; // Dead loop?
kvn@2230 1722 Node *init = cl->init_trip();
kvn@2230 1723 Node *phi = cl->phi();
kvn@2230 1724 int stride_con = cl->stride_con();
kvn@2230 1725
kvn@2230 1726 // Visit all children, looking for Phis
kvn@2230 1727 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
kvn@2230 1728 Node *out = cl->out(i);
kvn@2230 1729 // Look for other phis (secondary IVs). Skip dead ones
kvn@2230 1730 if (!out->is_Phi() || out == phi || !has_node(out))
kvn@2230 1731 continue;
kvn@2230 1732 PhiNode* phi2 = out->as_Phi();
kvn@2230 1733 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
kvn@2230 1734 // Look for induction variables of the form: X += constant
kvn@2230 1735 if (phi2->region() != loop->_head ||
kvn@2230 1736 incr2->req() != 3 ||
kvn@2230 1737 incr2->in(1) != phi2 ||
kvn@2230 1738 incr2 == incr ||
kvn@2230 1739 incr2->Opcode() != Op_AddI ||
kvn@2230 1740 !incr2->in(2)->is_Con())
kvn@2230 1741 continue;
kvn@2230 1742
kvn@2230 1743 // Check for parallel induction variable (parallel to trip counter)
kvn@2230 1744 // via an affine function. In particular, count-down loops with
kvn@2230 1745 // count-up array indices are common. We only RCE references off
kvn@2230 1746 // the trip-counter, so we need to convert all these to trip-counter
kvn@2230 1747 // expressions.
kvn@2230 1748 Node *init2 = phi2->in( LoopNode::EntryControl );
kvn@2230 1749 int stride_con2 = incr2->in(2)->get_int();
kvn@2230 1750
kvn@2230 1751 // The general case here gets a little tricky. We want to find the
kvn@2230 1752 // GCD of all possible parallel IV's and make a new IV using this
kvn@2230 1753 // GCD for the loop. Then all possible IVs are simple multiples of
kvn@2230 1754 // the GCD. In practice, this will cover very few extra loops.
kvn@2230 1755 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
kvn@2230 1756 // where +/-1 is the common case, but other integer multiples are
kvn@2230 1757 // also easy to handle.
kvn@2230 1758 int ratio_con = stride_con2/stride_con;
kvn@2230 1759
kvn@2230 1760 if ((ratio_con * stride_con) == stride_con2) { // Check for exact
kvn@2700 1761 #ifndef PRODUCT
kvn@2700 1762 if (TraceLoopOpts) {
kvn@2700 1763 tty->print("Parallel IV: %d ", phi2->_idx);
kvn@2700 1764 loop->dump_head();
kvn@2700 1765 }
kvn@2700 1766 #endif
kvn@2230 1767 // Convert to using the trip counter. The parallel induction
kvn@2230 1768 // variable differs from the trip counter by a loop-invariant
kvn@2230 1769 // amount, the difference between their respective initial values.
kvn@2230 1770 // It is scaled by the 'ratio_con'.
kvn@2230 1771 Node* ratio = _igvn.intcon(ratio_con);
kvn@2230 1772 set_ctrl(ratio, C->root());
kvn@3680 1773 Node* ratio_init = new (C) MulINode(init, ratio);
kvn@2700 1774 _igvn.register_new_node_with_optimizer(ratio_init, init);
kvn@2700 1775 set_early_ctrl(ratio_init);
kvn@3680 1776 Node* diff = new (C) SubINode(init2, ratio_init);
kvn@2700 1777 _igvn.register_new_node_with_optimizer(diff, init2);
kvn@2700 1778 set_early_ctrl(diff);
kvn@3680 1779 Node* ratio_idx = new (C) MulINode(phi, ratio);
kvn@2700 1780 _igvn.register_new_node_with_optimizer(ratio_idx, phi);
kvn@2700 1781 set_ctrl(ratio_idx, cl);
kvn@3680 1782 Node* add = new (C) AddINode(ratio_idx, diff);
kvn@2700 1783 _igvn.register_new_node_with_optimizer(add);
kvn@2700 1784 set_ctrl(add, cl);
kvn@2230 1785 _igvn.replace_node( phi2, add );
kvn@2230 1786 // Sometimes an induction variable is unused
kvn@2230 1787 if (add->outcnt() == 0) {
kvn@2230 1788 _igvn.remove_dead_node(add);
kvn@2230 1789 }
kvn@2230 1790 --i; // deleted this phi; rescan starting with next position
kvn@2230 1791 continue;
kvn@2230 1792 }
kvn@2230 1793 }
kvn@2230 1794 }
kvn@2230 1795
duke@0 1796 //------------------------------counted_loop-----------------------------------
duke@0 1797 // Convert to counted loops where possible
duke@0 1798 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
duke@0 1799
duke@0 1800 // For grins, set the inner-loop flag here
kvn@2230 1801 if (!_child) {
kvn@2230 1802 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
duke@0 1803 }
duke@0 1804
kvn@2230 1805 if (_head->is_CountedLoop() ||
kvn@2230 1806 phase->is_counted_loop(_head, this)) {
duke@0 1807 _has_sfpt = 1; // Indicate we do not need a safepoint here
duke@0 1808
kvn@3588 1809 // Look for safepoints to remove.
kvn@3588 1810 Node_List* sfpts = _safepts;
kvn@3588 1811 if (sfpts != NULL) {
kvn@3588 1812 for (uint i = 0; i < sfpts->size(); i++) {
kvn@3588 1813 Node* n = sfpts->at(i);
kvn@3588 1814 assert(phase->get_loop(n) == this, "");
kvn@3588 1815 if (phase->is_deleteable_safept(n)) {
kvn@3588 1816 phase->lazy_replace(n, n->in(TypeFunc::Control));
kvn@3588 1817 }
kvn@3588 1818 }
kvn@3588 1819 }
duke@0 1820
kvn@2230 1821 // Look for induction variables
kvn@2230 1822 phase->replace_parallel_iv(this);
duke@0 1823
duke@0 1824 } else if (_parent != NULL && !_irreducible) {
duke@0 1825 // Not a counted loop.
kvn@3588 1826 // Look for a safepoint on the idom-path.
kvn@3588 1827 Node* sfpt = tail();
kvn@3588 1828 for (; sfpt != _head; sfpt = phase->idom(sfpt)) {
kvn@3588 1829 if (sfpt->Opcode() == Op_SafePoint && phase->get_loop(sfpt) == this)
kvn@3588 1830 break; // Found one
duke@0 1831 }
kvn@3588 1832 // Delete other safepoints in this loop.
kvn@3588 1833 Node_List* sfpts = _safepts;
kvn@3588 1834 if (sfpts != NULL && sfpt != _head && sfpt->Opcode() == Op_SafePoint) {
kvn@3588 1835 for (uint i = 0; i < sfpts->size(); i++) {
kvn@3588 1836 Node* n = sfpts->at(i);
kvn@3588 1837 assert(phase->get_loop(n) == this, "");
kvn@3588 1838 if (n != sfpt && phase->is_deleteable_safept(n)) {
kvn@3588 1839 phase->lazy_replace(n, n->in(TypeFunc::Control));
kvn@3588 1840 }
kvn@3588 1841 }
duke@0 1842 }
duke@0 1843 }
duke@0 1844
duke@0 1845 // Recursively
kvn@2230 1846 if (_child) _child->counted_loop( phase );
kvn@2230 1847 if (_next) _next ->counted_loop( phase );
duke@0 1848 }
duke@0 1849
duke@0 1850 #ifndef PRODUCT
duke@0 1851 //------------------------------dump_head--------------------------------------
duke@0 1852 // Dump 1 liner for loop header info
duke@0 1853 void IdealLoopTree::dump_head( ) const {
kvn@2230 1854 for (uint i=0; i<_nest; i++)
duke@0 1855 tty->print(" ");
duke@0 1856 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
kvn@2230 1857 if (_irreducible) tty->print(" IRREDUCIBLE");
kvn@2442 1858 Node* entry = _head->in(LoopNode::EntryControl);
kvn@2442 1859 if (LoopLimitCheck) {
kvn@2442 1860 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
kvn@2442 1861 if (predicate != NULL ) {
kvn@2442 1862 tty->print(" limit_check");
kvn@2442 1863 entry = entry->in(0)->in(0);
kvn@2442 1864 }
kvn@2442 1865 }
kvn@2230 1866 if (UseLoopPredicate) {
kvn@2442 1867 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
kvn@2292 1868 if (entry != NULL) {
kvn@2230 1869 tty->print(" predicated");
kvn@2230 1870 }
kvn@2230 1871 }
kvn@2230 1872 if (_head->is_CountedLoop()) {
duke@0 1873 CountedLoopNode *cl = _head->as_CountedLoop();
duke@0 1874 tty->print(" counted");
kvn@2312 1875
kvn@2312 1876 Node* init_n = cl->init_trip();
kvn@2312 1877 if (init_n != NULL && init_n->is_Con())
kvn@2312 1878 tty->print(" [%d,", cl->init_trip()->get_int());
kvn@2312 1879 else
kvn@2312 1880 tty->print(" [int,");
kvn@2312 1881 Node* limit_n = cl->limit();
kvn@2312 1882 if (limit_n != NULL && limit_n->is_Con())
kvn@2312 1883 tty->print("%d),", cl->limit()->get_int());
kvn@2312 1884 else
kvn@2312 1885 tty->print("int),");
kvn@2312 1886 int stride_con = cl->stride_con();
kvn@2312 1887 if (stride_con > 0) tty->print("+");
kvn@2312 1888 tty->print("%d", stride_con);
kvn@2312 1889
kvn@3566 1890 tty->print(" (%d iters) ", (int)cl->profile_trip_cnt());
kvn@3566 1891
kvn@2230 1892 if (cl->is_pre_loop ()) tty->print(" pre" );
kvn@2230 1893 if (cl->is_main_loop()) tty->print(" main");
kvn@2230 1894 if (cl->is_post_loop()) tty->print(" post");
duke@0 1895 }
duke@0 1896 tty->cr();
duke@0 1897 }
duke@0 1898
duke@0 1899 //------------------------------dump-------------------------------------------
duke@0 1900 // Dump loops by loop tree
duke@0 1901 void IdealLoopTree::dump( ) const {
duke@0 1902 dump_head();
kvn@2230 1903 if (_child) _child->dump();
kvn@2230 1904 if (_next) _next ->dump();
duke@0 1905 }
duke@0 1906
duke@0 1907 #endif
duke@0 1908
never@367 1909 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
never@367 1910 if (loop == root) {
never@367 1911 if (loop->_child != NULL) {
never@367 1912 log->begin_head("loop_tree");
never@367 1913 log->end_head();
never@367 1914 if( loop->_child ) log_loop_tree(root, loop->_child, log);
never@367 1915 log->tail("loop_tree");
never@367 1916 assert(loop->_next == NULL, "what?");
never@367 1917 }
never@367 1918 } else {
never@367 1919 Node* head = loop->_head;
never@367 1920 log->begin_head("loop");
never@367 1921 log->print(" idx='%d' ", head->_idx);
never@367 1922 if (loop->_irreducible) log->print("irreducible='1' ");
never@367 1923 if (head->is_Loop()) {
never@367 1924 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
never@367 1925 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
never@367 1926 }
never@367 1927 if (head->is_CountedLoop()) {
never@367 1928 CountedLoopNode* cl = head->as_CountedLoop();
never@367 1929 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
never@367 1930 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
never@367 1931 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
never@367 1932 }
never@367 1933 log->end_head();
never@367 1934 if( loop->_child ) log_loop_tree(root, loop->_child, log);
never@367 1935 log->tail("loop");
never@367 1936 if( loop->_next ) log_loop_tree(root, loop->_next, log);
never@367 1937 }
never@367 1938 }
never@367 1939
cfang@1172 1940 //---------------------collect_potentially_useful_predicates-----------------------
cfang@1172 1941 // Helper function to collect potentially useful predicates to prevent them from
cfang@1172 1942 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
cfang@1172 1943 void PhaseIdealLoop::collect_potentially_useful_predicates(
cfang@1172 1944 IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
cfang@1172 1945 if (loop->_child) { // child
cfang@1172 1946 collect_potentially_useful_predicates(loop->_child, useful_predicates);
cfang@1172 1947 }
cfang@1172 1948
cfang@1172 1949 // self (only loops that we can apply loop predication may use their predicates)
kvn@2230 1950 if (loop->_head->is_Loop() &&
kvn@2230 1951 !loop->_irreducible &&
cfang@1172 1952 !loop->tail()->is_top()) {
kvn@2230 1953 LoopNode* lpn = loop->_head->as_Loop();
cfang@1172 1954 Node* entry = lpn->in(LoopNode::EntryControl);
kvn@2442 1955 Node* predicate_proj = find_predicate(entry); // loop_limit_check first
cfang@1172 1956 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
kvn@2230 1957 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
cfang@1172 1958 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
kvn@2442 1959 entry = entry->in(0)->in(0);
kvn@2442 1960 }
kvn@2442 1961 predicate_proj = find_predicate(entry); // Predicate
kvn@2442 1962 if (predicate_proj != NULL ) {
kvn@2442 1963 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
cfang@1172 1964 }
cfang@1172 1965 }
cfang@1172 1966
kvn@2230 1967 if (loop->_next) { // sibling
cfang@1172 1968 collect_potentially_useful_predicates(loop->_next, useful_predicates);
cfang@1172 1969 }
cfang@1172 1970 }
cfang@1172 1971
cfang@1172 1972 //------------------------eliminate_useless_predicates-----------------------------
cfang@1172 1973 // Eliminate all inserted predicates if they could not be used by loop predication.
kvn@2442 1974 // Note: it will also eliminates loop limits check predicate since it also uses
kvn@2442 1975 // Opaque1 node (see Parse::add_predicate()).
cfang@1172 1976 void PhaseIdealLoop::eliminate_useless_predicates() {
kvn@2230 1977 if (C->predicate_count() == 0)
kvn@2230 1978 return; // no predicate left
cfang@1172 1979
cfang@1172 1980 Unique_Node_List useful_predicates; // to store useful predicates
cfang@1172 1981 if (C->has_loops()) {
cfang@1172 1982 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
cfang@1172 1983 }
cfang@1172 1984
cfang@1172 1985 for (int i = C->predicate_count(); i > 0; i--) {
cfang@1172 1986 Node * n = C->predicate_opaque1_node(i-1);
cfang@1172 1987 assert(n->Opcode() == Op_Opaque1, "must be");
cfang@1172 1988 if (!useful_predicates.member(n)) { // not in the useful list
cfang@1172 1989 _igvn.replace_node(n, n->in(1));
cfang@1172 1990 }
cfang@1172 1991 }
cfang@1172 1992 }
cfang@1172 1993
roland@4154 1994 //------------------------process_expensive_nodes-----------------------------
roland@4154 1995 // Expensive nodes have their control input set to prevent the GVN
roland@4154 1996 // from commoning them and as a result forcing the resulting node to
roland@4154 1997 // be in a more frequent path. Use CFG information here, to change the
roland@4154 1998 // control inputs so that some expensive nodes can be commoned while
roland@4154 1999 // not executed more frequently.
roland@4154 2000 bool PhaseIdealLoop::process_expensive_nodes() {
roland@4154 2001 assert(OptimizeExpensiveOps, "optimization off?");
roland@4154 2002
roland@4154 2003 // Sort nodes to bring similar nodes together
roland@4154 2004 C->sort_expensive_nodes();
roland@4154 2005
roland@4154 2006 bool progress = false;
roland@4154 2007
roland@4154 2008 for (int i = 0; i < C->expensive_count(); ) {
roland@4154 2009 Node* n = C->expensive_node(i);
roland@4154 2010 int start = i;
roland@4154 2011 // Find nodes similar to n
roland@4154 2012 i++;
roland@4154 2013 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
roland@4154 2014 int end = i;
roland@4154 2015 // And compare them two by two
roland@4154 2016 for (int j = start; j < end; j++) {
roland@4154 2017 Node* n1 = C->expensive_node(j);
roland@4154 2018 if (is_node_unreachable(n1)) {
roland@4154 2019 continue;
roland@4154 2020 }
roland@4154 2021 for (int k = j+1; k < end; k++) {
roland@4154 2022 Node* n2 = C->expensive_node(k);
roland@4154 2023 if (is_node_unreachable(n2)) {
roland@4154 2024 continue;
roland@4154 2025 }
roland@4154 2026
roland@4154 2027 assert(n1 != n2, "should be pair of nodes");
roland@4154 2028
roland@4154 2029 Node* c1 = n1->in(0);
roland@4154 2030 Node* c2 = n2->in(0);
roland@4154 2031
roland@4154 2032 Node* parent_c1 = c1;
roland@4154 2033 Node* parent_c2 = c2;
roland@4154 2034
roland@4154 2035 // The call to get_early_ctrl_for_expensive() moves the
roland@4154 2036 // expensive nodes up but stops at loops that are in a if
roland@4154 2037 // branch. See whether we can exit the loop and move above the
roland@4154 2038 // If.
roland@4154 2039 if (c1->is_Loop()) {
roland@4154 2040 parent_c1 = c1->in(1);
roland@4154 2041 }
roland@4154 2042 if (c2->is_Loop()) {
roland@4154 2043 parent_c2 = c2->in(1);
roland@4154 2044 }
roland@4154 2045
roland@4154 2046 if (parent_c1 == parent_c2) {
roland@4154 2047 _igvn._worklist.push(n1);
roland@4154 2048 _igvn._worklist.push(n2);
roland@4154 2049 continue;
roland@4154 2050 }
roland@4154 2051
roland@4154 2052 // Look for identical expensive node up the dominator chain.
roland@4154 2053 if (is_dominator(c1, c2)) {
roland@4154 2054 c2 = c1;
roland@4154 2055 } else if (is_dominator(c2, c1)) {
roland@4154 2056 c1 = c2;
roland@4154 2057 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
roland@4154 2058 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
roland@4154 2059 // Both branches have the same expensive node so move it up
roland@4154 2060 // before the if.
roland@4154 2061 c1 = c2 = idom(parent_c1->in(0));
roland@4154 2062 }
roland@4154 2063 // Do the actual moves
roland@4154 2064 if (n1->in(0) != c1) {
roland@4154 2065 _igvn.hash_delete(n1);
roland@4154 2066 n1->set_req(0, c1);
roland@4154 2067 _igvn.hash_insert(n1);
roland@4154 2068 _igvn._worklist.push(n1);
roland@4154 2069 progress = true;
roland@4154 2070 }
roland@4154 2071 if (n2->in(0) != c2) {
roland@4154 2072 _igvn.hash_delete(n2);
roland@4154 2073 n2->set_req(0, c2);
roland@4154 2074 _igvn.hash_insert(n2);
roland@4154 2075 _igvn._worklist.push(n2);
roland@4154 2076 progress = true;
roland@4154 2077 }
roland@4154 2078 }
roland@4154 2079 }
roland@4154 2080 }
roland@4154 2081
roland@4154 2082 return progress;
roland@4154 2083 }
roland@4154 2084
roland@4154 2085
duke@0 2086 //=============================================================================
never@921 2087 //----------------------------build_and_optimize-------------------------------
duke@0 2088 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
duke@0 2089 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
kvn@2825 2090 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool skip_loop_opts) {
kvn@2120 2091 ResourceMark rm;
kvn@2120 2092
never@921 2093 int old_progress = C->major_progress();
never@2250 2094 uint orig_worklist_size = _igvn._worklist.size();
never@921 2095
duke@0 2096 // Reset major-progress flag for the driver's heuristics
duke@0 2097 C->clear_major_progress();
duke@0 2098
duke@0 2099 #ifndef PRODUCT
duke@0 2100 // Capture for later assert
duke@0 2101 uint unique = C->unique();
duke@0 2102 _loop_invokes++;
duke@0 2103 _loop_work += unique;
duke@0 2104 #endif
duke@0 2105
duke@0 2106 // True if the method has at least 1 irreducible loop
duke@0 2107 _has_irreducible_loops = false;
duke@0 2108
duke@0 2109 _created_loop_node = false;
duke@0 2110
duke@0 2111 Arena *a = Thread::current()->resource_area();
duke@0 2112 VectorSet visited(a);
duke@0 2113 // Pre-grow the mapping from Nodes to IdealLoopTrees.
duke@0 2114 _nodes.map(C->unique(), NULL);
duke@0 2115 memset(_nodes.adr(), 0, wordSize * C->unique());
duke@0 2116
duke@0 2117 // Pre-build the top-level outermost loop tree entry
duke@0 2118 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
duke@0 2119 // Do not need a safepoint at the top level
duke@0 2120 _ltree_root->_has_sfpt = 1;
duke@0 2121
kvn@2292 2122 // Initialize Dominators.
kvn@2292 2123 // Checked in clone_loop_predicate() during beautify_loops().
kvn@2292 2124 _idom_size = 0;
kvn@2292 2125 _idom = NULL;
kvn@2292 2126 _dom_depth = NULL;
kvn@2292 2127 _dom_stk = NULL;
kvn@2292 2128
duke@0 2129 // Empty pre-order array
duke@0 2130 allocate_preorders();
duke@0 2131
duke@0 2132 // Build a loop tree on the fly. Build a mapping from CFG nodes to
duke@0 2133 // IdealLoopTree entries. Data nodes are NOT walked.
duke@0 2134 build_loop_tree();
duke@0 2135 // Check for bailout, and return
duke@0 2136 if (C->failing()) {
duke@0 2137 return;
duke@0 2138 }
duke@0 2139
duke@0 2140 // No loops after all
never@921 2141 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
duke@0 2142
duke@0 2143 // There should always be an outer loop containing the Root and Return nodes.
duke@0 2144 // If not, we have a degenerate empty program. Bail out in this case.
duke@0 2145 if (!has_node(C->root())) {
never@921 2146 if (!_verify_only) {
never@921 2147 C->clear_major_progress();
never@921 2148 C->record_method_not_compilable("empty program detected during loop optimization");
never@921 2149 }
duke@0 2150 return;
duke@0 2151 }
duke@0 2152
duke@0 2153 // Nothing to do, so get out
roland@4154 2154 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only;
roland@4154 2155 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
roland@4154 2156 if (stop_early && !do_expensive_nodes) {
duke@0 2157 _igvn.optimize(); // Cleanup NeverBranches
duke@0 2158 return;
duke@0 2159 }
duke@0 2160
duke@0 2161 // Set loop nesting depth
duke@0 2162 _ltree_root->set_nest( 0 );
duke@0 2163
duke@0 2164 // Split shared headers and insert loop landing pads.
duke@0 2165 // Do not bother doing this on the Root loop of course.
never@921 2166 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
sla@4802 2167 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
duke@0 2168 if( _ltree_root->_child->beautify_loops( this ) ) {
duke@0 2169 // Re-build loop tree!
duke@0 2170 _ltree_root->_child = NULL;
duke@0 2171 _nodes.clear();
duke@0 2172 reallocate_preorders();
duke@0 2173 build_loop_tree();
duke@0 2174 // Check for bailout, and return
duke@0 2175 if (C->failing()) {
duke@0 2176 return;
duke@0 2177 }
duke@0 2178 // Reset loop nesting depth
duke@0 2179 _ltree_root->set_nest( 0 );
never@222 2180
sla@4802 2181 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
duke@0 2182 }
duke@0 2183 }
duke@0 2184
duke@0 2185 // Build Dominators for elision of NULL checks & loop finding.
duke@0 2186 // Since nodes do not have a slot for immediate dominator, make
twisti@605 2187 // a persistent side array for that info indexed on node->_idx.
duke@0 2188 _idom_size = C->unique();
duke@0 2189 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
duke@0 2190 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
duke@0 2191 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
duke@0 2192 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
duke@0 2193
duke@0 2194 Dominators();
duke@0 2195
never@921 2196 if (!_verify_only) {
never@921 2197 // As a side effect, Dominators removed any unreachable CFG paths
never@921 2198 // into RegionNodes. It doesn't do this test against Root, so
never@921 2199 // we do it here.
never@921 2200 for( uint i = 1; i < C->root()->req(); i++ ) {
never@921 2201 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
kvn@3412 2202 _igvn.delete_input_of(C->root(), i);
never@921 2203 i--; // Rerun same iteration on compressed edges
never@921 2204 }
duke@0 2205 }
never@921 2206
never@921 2207 // Given dominators, try to find inner loops with calls that must
never@921 2208 // always be executed (call dominates loop tail). These loops do
never@921 2209 // not need a separate safepoint.
never@921 2210 Node_List cisstack(a);
never@921 2211 _ltree_root->check_safepts(visited, cisstack);
duke@0 2212 }
duke@0 2213
duke@0 2214 // Walk the DATA nodes and place into loops. Find earliest control
duke@0 2215 // node. For CFG nodes, the _nodes array starts out and remains
duke@0 2216 // holding the associated IdealLoopTree pointer. For DATA nodes, the
duke@0 2217 // _nodes array holds the earliest legal controlling CFG node.
duke@0 2218
duke@0 2219 // Allocate stack with enough space to avoid frequent realloc
duke@0 2220 int stack_size = (C->unique() >> 1) + 16; // (unique>>1)+16 from Java2D stats
duke@0 2221 Node_Stack nstack( a, stack_size );
duke@0 2222
duke@0 2223 visited.Clear();
duke@0 2224 Node_List worklist(a);
duke@0 2225 // Don't need C->root() on worklist since
duke@0 2226 // it will be processed among C->top() inputs
duke@0 2227 worklist.push( C->top() );
duke@0 2228 visited.set( C->top()->_idx ); // Set C->top() as visited now
never@921 2229 build_loop_early( visited, worklist, nstack );
duke@0 2230
duke@0 2231 // Given early legal placement, try finding counted loops. This placement
duke@0 2232 // is good enough to discover most loop invariants.
never@921 2233 if( !_verify_me && !_verify_only )
duke@0 2234 _ltree_root->counted_loop( this );
duke@0 2235
duke@0 2236 // Find latest loop placement. Find ideal loop placement.
duke@0 2237 visited.Clear();
duke@0 2238 init_dom_lca_tags();
duke@0 2239 // Need C->root() on worklist when processing outs
duke@0 2240 worklist.push( C->root() );
duke@0 2241 NOT_PRODUCT( C->verify_graph_edges(); )
duke@0 2242 worklist.push( C->top() );
never@921 2243 build_loop_late( visited, worklist, nstack );
never@921 2244
never@921 2245 if (_verify_only) {
never@921 2246 // restore major progress flag
never@921 2247 for (int i = 0; i < old_progress; i++)
never@921 2248 C->set_major_progress();
never@921 2249 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
never@2250 2250 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
never@921 2251 return;
never@921 2252 }
duke@0 2253
kvn@4433 2254 // clear out the dead code after build_loop_late
kvn@4433 2255 while (_deadlist.size()) {
kvn@4433 2256 _igvn.remove_globally_dead_node(_deadlist.pop());
kvn@4433 2257 }
kvn@4433 2258
roland@4154 2259 if (stop_early) {
roland@4154 2260 assert(do_expensive_nodes, "why are we here?");
roland@4154 2261 if (process_expensive_nodes()) {
roland@4154 2262 // If we made some progress when processing expensive nodes then
roland@4154 2263 // the IGVN may modify the graph in a way that will allow us to
roland@4154 2264 // make some more progress: we need to try processing expensive
roland@4154 2265 // nodes again.
roland@4154 2266 C->set_major_progress();
roland@4154 2267 }
roland@4154 2268 _igvn.optimize();
roland@4154 2269 return;
roland@4154 2270 }
roland@4154 2271
kvn@2292 2272 // Some parser-inserted loop predicates could never be used by loop
kvn@2292 2273 // predication or they were moved away from loop during some optimizations.
kvn@2292 2274 // For example, peeling. Eliminate them before next loop optimizations.
kvn@2442 2275 if (UseLoopPredicate || LoopLimitCheck) {
cfang@1172 2276 eliminate_useless_predicates();
cfang@1172 2277 }
cfang@1172 2278
duke@0 2279 #ifndef PRODUCT
duke@0 2280 C->verify_graph_edges();
kvn@2230 2281 if (_verify_me) { // Nested verify pass?
duke@0 2282 // Check to see if the verify mode is broken
duke@0 2283 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
duke@0 2284 return;
duke@0 2285 }
kvn@2230 2286 if(VerifyLoopOptimizations) verify();
kvn@2230 2287 if(TraceLoopOpts && C->has_loops()) {
kvn@2230 2288 _ltree_root->dump();
kvn@2230 2289 }
duke@0 2290 #endif
duke@0 2291
kvn@2825 2292 if (skip_loop_opts) {
kvn@2825 2293 // Cleanup any modified bits
kvn@2825 2294 _igvn.optimize();
kvn@2825 2295
kvn@2825 2296 if (C->log() != NULL) {
kvn@2825 2297 log_loop_tree(_ltree_root, _ltree_root, C->log());
kvn@2825 2298 }
kvn@2825 2299 return;
kvn@2825 2300 }
kvn@2825 2301
duke@0 2302 if (ReassociateInvariants) {
duke@0 2303 // Reassociate invariants and prep for split_thru_phi
duke@0 2304 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
duke@0 2305 IdealLoopTree* lpt = iter.current();
duke@0 2306 if (!lpt->is_counted() || !lpt->is_inner()) continue;
duke@0 2307
duke@0 2308 lpt->reassociate_invariants(this);
duke@0 2309
duke@0 2310 // Because RCE opportunities can be masked by split_thru_phi,
duke@0 2311 // look for RCE candidates and inhibit split_thru_phi
duke@0 2312 // on just their loop-phi's for this pass of loop opts
cfang@1172 2313 if (SplitIfBlocks && do_split_ifs) {
duke@0 2314 if (lpt->policy_range_check(this)) {
kvn@39 2315 lpt->_rce_candidate = 1; // = true
duke@0 2316 }
duke@0 2317 }
duke@0 2318 }
duke@0 2319 }
duke@0 2320
duke@0 2321 // Check for aggressive application of split-if and other transforms
duke@0 2322 // that require basic-block info (like cloning through Phi's)
duke@0 2323 if( SplitIfBlocks && do_split_ifs ) {
duke@0 2324 visited.Clear();
duke@0 2325 split_if_with_blocks( visited, nstack );
duke@0 2326 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
duke@0 2327 }
duke@0 2328
roland@4154 2329 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
roland@4154 2330 C->set_major_progress();
roland@4154 2331 }
roland@4154 2332
cfang@1172 2333 // Perform loop predication before iteration splitting
kvn@2292 2334 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
cfang@1172 2335 _ltree_root->_child->loop_predication(this);
cfang@1172 2336 }
cfang@1172 2337
never@1683 2338 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
never@1683 2339 if (do_intrinsify_fill()) {
never@1683 2340 C->set_major_progress();
never@1683 2341 }
never@1683 2342 }
never@1683 2343
duke@0 2344 // Perform iteration-splitting on inner loops. Split iterations to avoid
duke@0 2345 // range checks or one-shot null checks.
duke@0 2346
duke@0 2347 // If split-if's didn't hack the graph too bad (no CFG changes)
duke@0 2348 // then do loop opts.
cfang@1172 2349 if (C->has_loops() && !C->major_progress()) {
duke@0 2350 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
duke@0 2351 _ltree_root->_child->iteration_split( this, worklist );
duke@0 2352 // No verify after peeling! GCM has hoisted code out of the loop.
duke@0 2353 // After peeling, the hoisted code could sink inside the peeled area.
duke@0 2354 // The peeling code does not try to recompute the best location for
duke@0 2355 // all the code before the peeled area, so the verify pass will always
duke@0 2356 // complain about it.
duke@0 2357 }
duke@0 2358 // Do verify graph edges in any case
duke@0 2359 NOT_PRODUCT( C->verify_graph_edges(); );
duke@0 2360
cfang@1172 2361 if (!do_split_ifs) {
duke@0 2362 // We saw major progress in Split-If to get here. We forced a
duke@0 2363 // pass with unrolling and not split-if, however more split-if's
duke@0 2364 // might make progress. If the unrolling didn't make progress
duke@0 2365 // then the major-progress flag got cleared and we won't try
duke@0 2366 // another round of Split-If. In particular the ever-common
duke@0 2367 // instance-of/check-cast pattern requires at least 2 rounds of
duke@0 2368 // Split-If to clear out.
duke@0 2369 C->set_major_progress();
duke@0 2370 }
duke@0 2371
duke@0 2372 // Repeat loop optimizations if new loops were seen
duke@0 2373 if (created_loop_node()) {
duke@0 2374 C->set_major_progress();
duke@0 2375 }
duke@0 2376
kvn@2292 2377 // Keep loop predicates and perform optimizations with them
kvn@2292 2378 // until no more loop optimizations could be done.
kvn@2292 2379 // After that switch predicates off and do more loop optimizations.
kvn@2292 2380 if (!C->major_progress() && (C->predicate_count() > 0)) {
kvn@2292 2381 C->cleanup_loop_predicates(_igvn);
kvn@2292 2382 #ifndef PRODUCT
kvn@2292 2383 if (TraceLoopOpts) {
kvn@2292 2384 tty->print_cr("PredicatesOff");
kvn@2292 2385 }
kvn@2292 2386 #endif
kvn@2292 2387 C->set_major_progress();
kvn@2292 2388 }
duke@0 2389
kvn@2292 2390 // Convert scalar to superword operations at the end of all loop opts.
duke@0 2391 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
duke@0 2392 // SuperWord transform
duke@0 2393 SuperWord sw(this);
duke@0 2394 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
duke@0 2395 IdealLoopTree* lpt = iter.current();
duke@0 2396 if (lpt->is_counted()) {
duke@0 2397 sw.transform_loop(lpt);
duke@0 2398 }
duke@0 2399 }
duke@0 2400 }
duke@0 2401
duke@0 2402 // Cleanup any modified bits
duke@0 2403 _igvn.optimize();
duke@0 2404
never@367 2405 // disable assert until issue with split_flow_path is resolved (6742111)
never@367 2406 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
never@367 2407 // "shouldn't introduce irreducible loops");
never@367 2408
never@367 2409 if (C->log() != NULL) {
never@367 2410 log_loop_tree(_ltree_root, _ltree_root, C->log());
never@367 2411 }
duke@0 2412 }
duke@0 2413
duke@0 2414 #ifndef PRODUCT
duke@0 2415 //------------------------------print_statistics-------------------------------
duke@0 2416 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
duke@0 2417 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
duke@0 2418 void PhaseIdealLoop::print_statistics() {
duke@0 2419 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
duke@0 2420 }
duke@0 2421
duke@0 2422 //------------------------------verify-----------------------------------------
duke@0 2423 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
duke@0 2424 static int fail; // debug only, so its multi-thread dont care
duke@0 2425 void PhaseIdealLoop::verify() const {
duke@0 2426 int old_progress = C->major_progress();
duke@0 2427 ResourceMark rm;
never@921 2428 PhaseIdealLoop loop_verify( _igvn, this );
duke@0 2429 VectorSet visited(Thread::current()->resource_area());
duke@0 2430
duke@0 2431 fail = 0;
duke@0 2432 verify_compare( C->root(), &loop_verify, visited );
duke@0 2433 assert( fail == 0, "verify loops failed" );
duke@0 2434 // Verify loop structure is the same
duke@0 2435 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
duke@0 2436 // Reset major-progress. It was cleared by creating a verify version of
duke@0 2437 // PhaseIdealLoop.
duke@0 2438 for( int i=0; i<old_progress; i++ )
duke@0 2439 C->set_major_progress();
duke@0 2440 }
duke@0 2441
duke@0 2442 //------------------------------verify_compare---------------------------------
duke@0 2443 // Make sure me and the given PhaseIdealLoop agree on key data structures
duke@0 2444 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
duke@0 2445 if( !n ) return;
duke@0 2446 if( visited.test_set( n->_idx ) ) return;
duke@0 2447 if( !_nodes[n->_idx] ) { // Unreachable
duke@0 2448 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
duke@0 2449 return;
duke@0 2450 }
duke@0 2451
duke@0 2452 uint i;
duke@0 2453 for( i = 0; i < n->req(); i++ )
duke@0 2454 verify_compare( n->in(i), loop_verify, visited );
duke@0 2455
duke@0 2456 // Check the '_nodes' block/loop structure
duke@0 2457 i = n->_idx;
duke@0 2458 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
duke@0 2459 if( _nodes[i] != loop_verify->_nodes[i] &&
duke@0 2460 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
duke@0 2461 tty->print("Mismatched control setting for: ");
duke@0 2462 n->dump();
duke@0 2463 if( fail++ > 10 ) return;
duke@0 2464 Node *c = get_ctrl_no_update(n);
duke@0 2465 tty->print("We have it as: ");
duke@0 2466 if( c->in(0) ) c->dump();
duke@0 2467 else tty->print_cr("N%d",c->_idx);
duke@0 2468 tty->print("Verify thinks: ");
duke@0 2469 if( loop_verify->has_ctrl(n) )
duke@0 2470 loop_verify->get_ctrl_no_update(n)->dump();
duke@0 2471 else
duke@0 2472 loop_verify->get_loop_idx(n)->dump();
duke@0 2473 tty->cr();
duke@0 2474 }
duke@0 2475 } else { // We have a loop
duke@0 2476 IdealLoopTree *us = get_loop_idx(n);
duke@0 2477 if( loop_verify->has_ctrl(n) ) {
duke@0 2478 tty->print("Mismatched loop setting for: ");
duke@0 2479 n->dump();
duke@0 2480 if( fail++ > 10 ) return;
duke@0 2481 tty->print("We have it as: ");
duke@0 2482 us->dump();
duke@0 2483 tty->print("Verify thinks: ");
duke@0 2484 loop_verify->get_ctrl_no_update(n)->dump();
duke@0 2485 tty->cr();
duke@0 2486 } else if (!C->major_progress()) {
duke@0 2487 // Loop selection can be messed up if we did a major progress
duke@0 2488 // operation, like split-if. Do not verify in that case.
duke@0 2489 IdealLoopTree *them = loop_verify->get_loop_idx(n);
duke@0 2490 if( us->_head != them->_head || us->_tail != them->_tail ) {
duke@0 2491 tty->print("Unequals loops for: ");
duke@0 2492 n->dump();
duke@0 2493 if( fail++ > 10 ) return;
duke@0 2494 tty->print("We have it as: ");
duke@0 2495 us->dump();
duke@0 2496 tty->print("Verify thinks: ");
duke@0 2497 them->dump();
duke@0 2498 tty->cr();
duke@0 2499 }
duke@0 2500 }
duke@0 2501 }
duke@0 2502
duke@0 2503 // Check for immediate dominators being equal
duke@0 2504 if( i >= _idom_size ) {
duke@0 2505 if( !n->is_CFG() ) return;
duke@0 2506 tty->print("CFG Node with no idom: ");
duke@0 2507 n->dump();
duke@0 2508 return;
duke@0 2509 }
duke@0 2510 if( !n->is_CFG() ) return;
duke@0 2511 if( n == C->root() ) return; // No IDOM here
duke@0 2512
duke@0 2513 assert(n->_idx == i, "sanity");
duke@0 2514 Node *id = idom_no_update(n);
duke@0 2515 if( id != loop_verify->idom_no_update(n) ) {
duke@0 2516 tty->print("Unequals idoms for: ");
duke@0 2517 n->dump();
duke@0 2518 if( fail++ > 10 ) return;
duke@0 2519 tty->print("We have it as: ");
duke@0 2520 id->dump();
duke@0 2521 tty->print("Verify thinks: ");
duke@0 2522 loop_verify->idom_no_update(n)->dump();
duke@0 2523 tty->cr();
duke@0 2524 }
duke@0 2525
duke@0 2526 }
duke@0 2527
duke@0 2528 //------------------------------verify_tree------------------------------------
duke@0 2529 // Verify that tree structures match. Because the CFG can change, siblings
duke@0 2530 // within the loop tree can be reordered. We attempt to deal with that by
duke@0 2531 // reordering the verify's loop tree if possible.
duke@0 2532 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
duke@0 2533 assert( _parent == parent, "Badly formed loop tree" );
duke@0 2534
duke@0 2535 // Siblings not in same order? Attempt to re-order.
duke@0 2536 if( _head != loop->_head ) {
duke@0 2537 // Find _next pointer to update
duke@0 2538 IdealLoopTree **pp = &loop->_parent->_child;
duke@0 2539 while( *pp != loop )
duke@0 2540 pp = &((*pp)->_next);
duke@0 2541 // Find proper sibling to be next
duke@0 2542 IdealLoopTree **nn = &loop->_next;
duke@0 2543 while( (*nn) && (*nn)->_head != _head )
duke@0 2544 nn = &((*nn)->_next);
duke@0 2545
duke@0 2546 // Check for no match.
duke@0 2547 if( !(*nn) ) {
duke@0 2548 // Annoyingly, irreducible loops can pick different headers
duke@0 2549 // after a major_progress operation, so the rest of the loop
duke@0 2550 // tree cannot be matched.
duke@0 2551 if (_irreducible && Compile::current()->major_progress()) return;
duke@0 2552 assert( 0, "failed to match loop tree" );
duke@0 2553 }
duke@0 2554
duke@0 2555 // Move (*nn) to (*pp)
duke@0 2556 IdealLoopTree *hit = *nn;
duke@0 2557 *nn = hit->_next;
duke@0 2558 hit->_next = loop;
duke@0 2559 *pp = loop;
duke@0 2560 loop = hit;
duke@0 2561 // Now try again to verify
duke@0 2562 }
duke@0 2563
duke@0 2564 assert( _head == loop->_head , "mismatched loop head" );
duke@0 2565 Node *tail = _tail; // Inline a non-updating version of
duke@0 2566 while( !tail->in(0) ) // the 'tail()' call.
duke@0 2567 tail = tail->in(1);
duke@0 2568 assert( tail == loop->_tail, "mismatched loop tail" );
duke@0 2569
duke@0 2570 // Counted loops that are guarded should be able to find their guards
duke@0 2571 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
duke@0 2572 CountedLoopNode *cl = _head->as_CountedLoop();
duke@0 2573 Node *init = cl->init_trip();
duke@0 2574 Node *ctrl = cl->in(LoopNode::EntryControl);
duke@0 2575 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
duke@0 2576 Node *iff = ctrl->in(0);
duke@0 2577 assert( iff->Opcode() == Op_If, "" );
duke@0 2578 Node *bol = iff->in(1);
duke@0 2579 assert( bol->Opcode() == Op_Bool, "" );
duke@0 2580 Node *cmp = bol->in(1);
duke@0 2581 assert( cmp->Opcode() == Op_CmpI, "" );
duke@0 2582 Node *add = cmp->in(1);
duke@0 2583 Node *opaq;
duke@0 2584 if( add->Opcode() == Op_Opaque1 ) {
duke@0 2585 opaq = add;
duke@0 2586 } else {
duke@0 2587 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
duke@0 2588 assert( add == init, "" );
duke@0 2589 opaq = cmp->in(2);
duke@0 2590 }
duke@0 2591 assert( opaq->Opcode() == Op_Opaque1, "" );
duke@0 2592
duke@0 2593 }
duke@0 2594
duke@0 2595 if (_child != NULL) _child->verify_tree(loop->_child, this);
duke@0 2596 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
duke@0 2597 // Innermost loops need to verify loop bodies,
duke@0 2598 // but only if no 'major_progress'
duke@0 2599 int fail = 0;
duke@0 2600 if (!Compile::current()->major_progress() && _child == NULL) {
duke@0 2601 for( uint i = 0; i < _body.size(); i++ ) {
duke@0 2602 Node *n = _body.at(i);
duke@0 2603 if (n->outcnt() == 0) continue; // Ignore dead
duke@0 2604 uint j;
duke@0 2605 for( j = 0; j < loop->_body.size(); j++ )
duke@0 2606 if( loop->_body.at(j) == n )
duke@0 2607 break;
duke@0 2608 if( j == loop->_body.size() ) { // Not found in loop body
duke@0 2609 // Last ditch effort to avoid assertion: Its possible that we
duke@0 2610 // have some users (so outcnt not zero) but are still dead.
duke@0 2611 // Try to find from root.
duke@0 2612 if (Compile::current()->root()->find(n->_idx)) {
duke@0 2613 fail++;
duke@0 2614 tty->print("We have that verify does not: ");
duke@0 2615 n->dump();
duke@0 2616 }
duke@0 2617 }
duke@0 2618 }
duke@0 2619 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
duke@0 2620 Node *n = loop->_body.at(i2);
duke@0 2621 if (n->outcnt() == 0) continue; // Ignore dead
duke@0 2622 uint j;
duke@0 2623 for( j = 0; j < _body.size(); j++ )
duke@0 2624 if( _body.at(j) == n )
duke@0 2625 break;
duke@0 2626 if( j == _body.size() ) { // Not found in loop body
duke@0 2627 // Last ditch effort to avoid assertion: Its possible that we
duke@0 2628 // have some users (so outcnt not zero) but are still dead.
duke@0 2629 // Try to find from root.
duke@0 2630 if (Compile::current()->root()->find(n->_idx)) {
duke@0 2631 fail++;
duke@0 2632 tty->print("Verify has that we do not: ");
duke@0 2633 n->dump();
duke@0 2634 }
duke@0 2635 }
duke@0 2636 }
duke@0 2637 assert( !fail, "loop body mismatch" );
duke@0 2638 }
duke@0 2639 }
duke@0 2640
duke@0 2641 #endif
duke@0 2642
duke@0 2643 //------------------------------set_idom---------------------------------------
duke@0 2644 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
duke@0 2645 uint idx = d->_idx;
duke@0 2646 if (idx >= _idom_size) {
duke@0 2647 uint newsize = _idom_size<<1;
duke@0 2648 while( idx >= newsize ) {
duke@0 2649 newsize <<= 1;
duke@0 2650 }
duke@0 2651 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
duke@0 2652 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
duke@0 2653 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
duke@0 2654 _idom_size = newsize;
duke@0 2655 }
duke@0 2656 _idom[idx] = n;
duke@0 2657 _dom_depth[idx] = dom_depth;
duke@0 2658 }
duke@0 2659
duke@0 2660 //------------------------------recompute_dom_depth---------------------------------------
duke@0 2661 // The dominator tree is constructed with only parent pointers.
duke@0 2662 // This recomputes the depth in the tree by first tagging all
duke@0 2663 // nodes as "no depth yet" marker. The next pass then runs up
duke@0 2664 // the dom tree from each node marked "no depth yet", and computes
duke@0 2665 // the depth on the way back down.
duke@0 2666 void PhaseIdealLoop::recompute_dom_depth() {
duke@0 2667 uint no_depth_marker = C->unique();
duke@0 2668 uint i;
duke@0 2669 // Initialize depth to "no depth yet"
duke@0 2670 for (i = 0; i < _idom_size; i++) {
duke@0 2671 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
duke@0 2672 _dom_depth[i] = no_depth_marker;
duke@0 2673 }
duke@0 2674 }
duke@0 2675 if (_dom_stk == NULL) {
duke@0 2676 uint init_size = C->unique() / 100; // Guess that 1/100 is a reasonable initial size.
duke@0 2677 if (init_size < 10) init_size = 10;
kvn@2120 2678 _dom_stk = new GrowableArray<uint>(init_size);
duke@0 2679 }
duke@0 2680 // Compute new depth for each node.
duke@0 2681 for (i = 0; i < _idom_size; i++) {
duke@0 2682 uint j = i;
duke@0 2683 // Run up the dom tree to find a node with a depth
duke@0 2684 while (_dom_depth[j] == no_depth_marker) {
duke@0 2685 _dom_stk->push(j);
duke@0 2686 j = _idom[j]->_idx;
duke@0 2687 }
duke@0 2688 // Compute the depth on the way back down this tree branch
duke@0 2689 uint dd = _dom_depth[j] + 1;
duke@0 2690 while (_dom_stk->length() > 0) {
duke@0 2691 uint j = _dom_stk->pop();
duke@0 2692 _dom_depth[j] = dd;
duke@0 2693 dd++;
duke@0 2694 }
duke@0 2695 }
duke@0 2696 }
duke@0 2697
duke@0 2698 //------------------------------sort-------------------------------------------
duke@0 2699 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
duke@0 2700 // loop tree, not the root.
duke@0 2701 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
duke@0 2702 if( !innermost ) return loop; // New innermost loop
duke@0 2703
duke@0 2704 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
duke@0 2705 assert( loop_preorder, "not yet post-walked loop" );
duke@0 2706 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
duke@0 2707 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
duke@0 2708
duke@0 2709 // Insert at start of list
duke@0 2710 while( l ) { // Insertion sort based on pre-order
duke@0 2711 if( l == loop ) return innermost; // Already on list!
duke@0 2712 int l_preorder = get_preorder(l->_head); // Cache pre-order number
duke@0 2713 assert( l_preorder, "not yet post-walked l" );
duke@0 2714 // Check header pre-order number to figure proper nesting
duke@0 2715 if( loop_preorder > l_preorder )
duke@0 2716 break; // End of insertion
duke@0 2717 // If headers tie (e.g., shared headers) check tail pre-order numbers.
duke@0 2718 // Since I split shared headers, you'd think this could not happen.
duke@0 2719 // BUT: I must first do the preorder numbering before I can discover I
duke@0 2720 // have shared headers, so the split headers all get the same preorder
duke@0 2721 // number as the RegionNode they split from.
duke@0 2722 if( loop_preorder == l_preorder &&
duke@0 2723 get_preorder(loop->_tail) < get_preorder(l->_tail) )
duke@0 2724 break; // Also check for shared headers (same pre#)
duke@0 2725 pp = &l->_parent; // Chain up list
duke@0 2726 l = *pp;
duke@0 2727 }
duke@0 2728 // Link into list
duke@0 2729 // Point predecessor to me
duke@0 2730 *pp = loop;
duke@0 2731 // Point me to successor
duke@0 2732 IdealLoopTree *p = loop->_parent;
duke@0 2733 loop->_parent = l; // Point me to successor
duke@0 2734 if( p ) sort( p, innermost ); // Insert my parents into list as well
duke@0 2735 return innermost;
duke@0 2736 }
duke@0 2737
duke@0 2738 //------------------------------build_loop_tree--------------------------------
duke@0 2739 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
duke@0 2740 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
duke@0 2741 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
duke@0 2742 // tightest enclosing IdealLoopTree for post-walked.
duke@0 2743 //
duke@0 2744 // During my forward walk I do a short 1-layer lookahead to see if I can find
duke@0 2745 // a loop backedge with that doesn't have any work on the backedge. This
duke@0 2746 // helps me construct nested loops with shared headers better.
duke@0 2747 //
duke@0 2748 // Once I've done the forward recursion, I do the post-work. For each child
duke@0 2749 // I check to see if there is a backedge. Backedges define a loop! I
duke@0 2750 // insert an IdealLoopTree at the target of the backedge.
duke@0 2751 //
duke@0 2752 // During the post-work I also check to see if I have several children
duke@0 2753 // belonging to different loops. If so, then this Node is a decision point
duke@0 2754 // where control flow can choose to change loop nests. It is at this
duke@0 2755 // decision point where I can figure out how loops are nested. At this
duke@0 2756 // time I can properly order the different loop nests from my children.
duke@0 2757 // Note that there may not be any backedges at the decision point!
duke@0 2758 //
duke@0 2759 // Since the decision point can be far removed from the backedges, I can't
duke@0 2760 // order my loops at the time I discover them. Thus at the decision point
duke@0 2761 // I need to inspect loop header pre-order numbers to properly nest my
duke@0 2762 // loops. This means I need to sort my childrens' loops by pre-order.
duke@0 2763 // The sort is of size number-of-control-children, which generally limits
duke@0 2764 // it to size 2 (i.e., I just choose between my 2 target loops).
duke@0 2765 void PhaseIdealLoop::build_loop_tree() {
duke@0 2766 // Allocate stack of size C->unique()/2 to avoid frequent realloc
duke@0 2767 GrowableArray <Node *> bltstack(C->unique() >> 1);
duke@0 2768 Node *n = C->root();
duke@0 2769 bltstack.push(n);
duke@0 2770 int pre_order = 1;
duke@0 2771 int stack_size;
duke@0 2772
duke@0 2773 while ( ( stack_size = bltstack.length() ) != 0 ) {
duke@0 2774 n = bltstack.top(); // Leave node on stack
duke@0 2775 if ( !is_visited(n) ) {
duke@0 2776 // ---- Pre-pass Work ----
duke@0 2777 // Pre-walked but not post-walked nodes need a pre_order number.
duke@0 2778
duke@0 2779 set_preorder_visited( n, pre_order ); // set as visited
duke@0 2780
duke@0 2781 // ---- Scan over children ----
duke@0 2782 // Scan first over control projections that lead to loop headers.
duke@0 2783 // This helps us find inner-to-outer loops with shared headers better.
duke@0 2784
duke@0 2785 // Scan children's children for loop headers.
duke@0 2786 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
duke@0 2787 Node* m = n->raw_out(i); // Child
duke@0 2788 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
duke@0 2789 // Scan over children's children to find loop
duke@0 2790 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
duke@0 2791 Node* l = m->fast_out(j);
duke@0 2792 if( is_visited(l) && // Been visited?
duke@0 2793 !is_postvisited(l) && // But not post-visited
duke@0 2794 get_preorder(l) < pre_order ) { // And smaller pre-order
duke@0 2795 // Found! Scan the DFS down this path before doing other paths
duke@0 2796 bltstack.push(m);
duke@0 2797 break;
duke@0 2798 }
duke@0 2799 }
duke@0 2800 }
duke@0 2801 }
duke@0 2802 pre_order++;
duke@0 2803 }
duke@0 2804 else if ( !is_postvisited(n) ) {
duke@0 2805 // Note: build_loop_tree_impl() adds out edges on rare occasions,
duke@0 2806 // such as com.sun.rsasign.am::a.
duke@0 2807 // For non-recursive version, first, process current children.
duke@0 2808 // On next iteration, check if additional children were added.
duke@0 2809 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
duke@0 2810 Node* u = n->raw_out(k);
duke@0 2811 if ( u->is_CFG() && !is_visited(u) ) {
duke@0 2812 bltstack.push(u);
duke@0 2813 }
duke@0 2814 }
duke@0 2815 if ( bltstack.length() == stack_size ) {
duke@0 2816 // There were no additional children, post visit node now
duke@0 2817 (void)bltstack.pop(); // Remove node from stack
duke@0 2818 pre_order = build_loop_tree_impl( n, pre_order );
duke@0 2819 // Check for bailout
duke@0 2820 if (C->failing()) {
duke@0 2821 return;
duke@0 2822 }
duke@0 2823 // Check to grow _preorders[] array for the case when
duke@0 2824 // build_loop_tree_impl() adds new nodes.
duke@0 2825 check_grow_preorders();
duke@0 2826 }
duke@0 2827 }
duke@0 2828 else {
duke@0 2829 (void)bltstack.pop(); // Remove post-visited node from stack
duke@0 2830 }
duke@0 2831 }
duke@0 2832 }
duke@0 2833
duke@0 2834 //------------------------------build_loop_tree_impl---------------------------
duke@0 2835 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
duke@0 2836 // ---- Post-pass Work ----
duke@0 2837 // Pre-walked but not post-walked nodes need a pre_order number.
duke@0 2838
duke@0 2839 // Tightest enclosing loop for this Node
duke@0 2840 IdealLoopTree *innermost = NULL;
duke@0 2841
duke@0 2842 // For all children, see if any edge is a backedge. If so, make a loop
duke@0 2843 // for it. Then find the tightest enclosing loop for the self Node.
duke@0 2844 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
duke@0 2845 Node* m = n->fast_out(i); // Child
duke@0 2846 if( n == m ) continue; // Ignore control self-cycles
duke@0 2847 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
duke@0 2848
duke@0 2849 IdealLoopTree *l; // Child's loop
duke@0 2850 if( !is_postvisited(m) ) { // Child visited but not post-visited?
duke@0 2851 // Found a backedge
duke@0 2852 assert( get_preorder(m) < pre_order, "should be backedge" );
duke@0 2853 // Check for the RootNode, which is already a LoopNode and is allowed
duke@0 2854 // to have multiple "backedges".
duke@0 2855 if( m == C->root()) { // Found the root?
duke@0 2856 l = _ltree_root; // Root is the outermost LoopNode
duke@0 2857 } else { // Else found a nested loop
duke@0 2858 // Insert a LoopNode to mark this loop.
duke@0 2859 l = new IdealLoopTree(this, m, n);
duke@0 2860 } // End of Else found a nested loop
duke@0 2861 if( !has_loop(m) ) // If 'm' does not already have a loop set
duke@0 2862 set_loop(m, l); // Set loop header to loop now
duke@0 2863
duke@0 2864 } else { // Else not a nested loop
duke@0 2865 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
duke@0 2866 l = get_loop(m); // Get previously determined loop
duke@0 2867 // If successor is header of a loop (nest), move up-loop till it
duke@0 2868 // is a member of some outer enclosing loop. Since there are no
duke@0 2869 // shared headers (I've split them already) I only need to go up
duke@0 2870 // at most 1 level.
duke@0 2871 while( l && l->_head == m ) // Successor heads loop?
duke@0 2872 l = l->_parent; // Move up 1 for me
duke@0 2873 // If this loop is not properly parented, then this loop
duke@0 2874 // has no exit path out, i.e. its an infinite loop.
duke@0 2875 if( !l ) {
duke@0 2876 // Make loop "reachable" from root so the CFG is reachable. Basically
duke@0 2877 // insert a bogus loop exit that is never taken. 'm', the loop head,
duke@0 2878 // points to 'n', one (of possibly many) fall-in paths. There may be
duke@0 2879 // many backedges as well.
duke@0 2880
duke@0 2881 // Here I set the loop to be the root loop. I could have, after
duke@0 2882 // inserting a bogus loop exit, restarted the recursion and found my
duke@0 2883 // new loop exit. This would make the infinite loop a first-class
duke@0 2884 // loop and it would then get properly optimized. What's the use of
duke@0 2885 // optimizing an infinite loop?
duke@0 2886 l = _ltree_root; // Oops, found infinite loop
duke@0 2887
never@921 2888 if (!_verify_only) {
never@921 2889 // Insert the NeverBranch between 'm' and it's control user.
kvn@3680 2890 NeverBranchNode *iff = new (C) NeverBranchNode( m );
never@921 2891 _igvn.register_new_node_with_optimizer(iff);
never@921 2892 set_loop(iff, l);
kvn@3680 2893 Node *if_t = new (C) CProjNode( iff, 0 );
never@921 2894 _igvn.register_new_node_with_optimizer(if_t);
never@921 2895 set_loop(if_t, l);
duke@0 2896
never@921 2897 Node* cfg = NULL; // Find the One True Control User of m
never@921 2898 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
never@921 2899 Node* x = m->fast_out(j);
never@921 2900 if (x->is_CFG() && x != m && x != iff)
never@921 2901 { cfg = x; break; }
never@921 2902 }
never@921 2903 assert(cfg != NULL, "must find the control user of m");
never@921 2904 uint k = 0; // Probably cfg->in(0)
never@921 2905 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
never@921 2906 cfg->set_req( k, if_t ); // Now point to NeverBranch
never@921 2907
never@921 2908 // Now create the never-taken loop exit
kvn@3680 2909 Node *if_f = new (C) CProjNode( iff, 1 );
never@921 2910 _igvn.register_new_node_with_optimizer(if_f);
never@921 2911 set_loop(if_f, l);
never@921 2912 // Find frame ptr for Halt. Relies on the optimizer
never@921 2913 // V-N'ing. Easier and quicker than searching through
never@921 2914 // the program structure.
kvn@3680 2915 Node *frame = new (C) ParmNode( C->start(), TypeFunc::FramePtr );
never@921 2916 _igvn.register_new_node_with_optimizer(frame);
never@921 2917 // Halt & Catch Fire
kvn@3680 2918 Node *halt = new (C) HaltNode( if_f, frame );
never@921 2919 _igvn.register_new_node_with_optimizer(halt);
never@921 2920 set_loop(halt, l);
never@921 2921 C->root()->add_req(halt);
duke@0 2922 }
duke@0 2923 set_loop(C->root(), _ltree_root);
duke@0 2924 }
duke@0 2925 }
duke@0 2926 // Weeny check for irreducible. This child was already visited (this
duke@0 2927 // IS the post-work phase). Is this child's loop header post-visited
duke@0 2928 // as well? If so, then I found another entry into the loop.
never@921 2929 if (!_verify_only) {
never@921 2930 while( is_postvisited(l->_head) ) {
never@921 2931 // found irreducible
never@921 2932 l->_irreducible = 1; // = true
never@921 2933 l = l->_parent;
never@921 2934 _has_irreducible_loops = true;
never@921 2935 // Check for bad CFG here to prevent crash, and bailout of compile
never@921 2936 if (l == NULL) {
never@921 2937 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
never@921 2938 return pre_order;
never@921 2939 }
duke@0 2940 }
duke@0 2941 }
duke@0 2942
duke@0 2943 // This Node might be a decision point for loops. It is only if
duke@0 2944 // it's children belong to several different loops. The sort call
duke@0 2945 // does a trivial amount of work if there is only 1 child or all
duke@0 2946 // children belong to the same loop. If however, the children
duke@0 2947 // belong to different loops, the sort call will properly set the
duke@0 2948 // _parent pointers to show how the loops nest.
duke@0 2949 //
duke@0 2950 // In any case, it returns the tightest enclosing loop.
duke@0 2951 innermost = sort( l, innermost );
duke@0 2952 }
duke@0 2953
duke@0 2954 // Def-use info will have some dead stuff; dead stuff will have no
duke@0 2955 // loop decided on.
duke@0 2956
duke@0 2957 // Am I a loop header? If so fix up my parent's child and next ptrs.
duke@0 2958 if( innermost && innermost->_head == n ) {
duke@0 2959 assert( get_loop(n) == innermost, "" );
duke@0 2960 IdealLoopTree *p = innermost->_parent;
duke@0 2961 IdealLoopTree *l = innermost;
duke@0 2962 while( p && l->_head == n ) {
duke@0 2963 l->_next = p->_child; // Put self on parents 'next child'
duke@0 2964 p->_child = l; // Make self as first child of parent
duke@0 2965 l = p; // Now walk up the parent chain
duke@0 2966 p = l->_parent;
duke@0 2967 }
duke@0 2968 } else {
duke@0 2969 // Note that it is possible for a LoopNode to reach here, if the
duke@0 2970 // backedge has been made unreachable (hence the LoopNode no longer
duke@0 2971 // denotes a Loop, and will eventually be removed).
duke@0 2972
duke@0 2973 // Record tightest enclosing loop for self. Mark as post-visited.
duke@0 2974 set_loop(n, innermost);
duke@0 2975 // Also record has_call flag early on
duke@0 2976 if( innermost ) {
duke@0 2977 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
duke@0 2978 // Do not count uncommon calls
duke@0 2979 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
duke@0 2980 Node *iff = n->in(0)->in(0);
kvn@3447 2981 // No any calls for vectorized loops.
kvn@3447 2982 if( UseSuperWord || !iff->is_If() ||
duke@0 2983 (n->in(0)->Opcode() == Op_IfFalse &&
duke@0 2984 (1.0 - iff->as_If()->_prob) >= 0.01) ||
duke@0 2985 (iff->as_If()->_prob >= 0.01) )
duke@0 2986 innermost->_has_call = 1;
duke@0 2987 }
kvn@39 2988 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
kvn@39 2989 // Disable loop optimizations if the loop has a scalar replaceable
kvn@39 2990 // allocation. This disabling may cause a potential performance lost
kvn@39 2991 // if the allocation is not eliminated for some reason.
kvn@39 2992 innermost->_allow_optimizations = false;
kvn@39 2993 innermost->_has_call = 1; // = true
kvn@3588 2994 } else if (n->Opcode() == Op_SafePoint) {
kvn@3588 2995 // Record all safepoints in this loop.
kvn@3588 2996 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List();
kvn@3588 2997 innermost->_safepts->push(n);
duke@0 2998 }
duke@0 2999 }
duke@0 3000 }
duke@0 3001
duke@0 3002 // Flag as post-visited now
duke@0 3003 set_postvisited(n);
duke@0 3004 return pre_order;
duke@0 3005 }
duke@0 3006
duke@0 3007
duke@0 3008 //------------------------------build_loop_early-------------------------------
duke@0 3009 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
duke@0 3010 // First pass computes the earliest controlling node possible. This is the
duke@0 3011 // controlling input with the deepest dominating depth.
never@921 3012 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
duke@0 3013 while (worklist.size() != 0) {
duke@0 3014 // Use local variables nstack_top_n & nstack_top_i to cache values
duke@0 3015 // on nstack's top.
duke@0 3016 Node *nstack_top_n = worklist.pop();
duke@0 3017 uint nstack_top_i = 0;
duke@0 3018 //while_nstack_nonempty:
duke@0 3019 while (true) {
duke@0 3020 // Get parent node and next input's index from stack's top.
duke@0 3021 Node *n = nstack_top_n;
duke@0 3022 uint i = nstack_top_i;
duke@0 3023 uint cnt = n->req(); // Count of inputs
duke@0 3024 if (i == 0) { // Pre-process the node.
duke@0 3025 if( has_node(n) && // Have either loop or control already?
duke@0 3026 !has_ctrl(n) ) { // Have loop picked out already?
duke@0 3027 // During "merge_many_backedges" we fold up several nested loops
duke@0 3028 // into a single loop. This makes the members of the original
duke@0 3029 // loop bodies pointing to dead loops; they need to move up
duke@0 3030 // to the new UNION'd larger loop. I set the _head field of these
duke@0 3031 // dead loops to NULL and the _parent field points to the owning
duke@0 3032 // loop. Shades of UNION-FIND algorithm.
duke@0 3033 IdealLoopTree *ilt;
duke@0 3034 while( !(ilt = get_loop(n))->_head ) {
duke@0 3035 // Normally I would use a set_loop here. But in this one special
duke@0 3036 // case, it is legal (and expected) to change what loop a Node
duke@0 3037 // belongs to.
duke@0 3038 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
duke@0 3039 }
duke@0 3040 // Remove safepoints ONLY if I've already seen I don't need one.
duke@0 3041 // (the old code here would yank a 2nd safepoint after seeing a
duke@0 3042 // first one, even though the 1st did not dominate in the loop body
duke@0 3043 // and thus could be avoided indefinitely)
never@921 3044 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
duke@0 3045 is_deleteable_safept(n)) {
duke@0 3046 Node *in = n->in(TypeFunc::Control);
duke@0 3047 lazy_replace(n,in); // Pull safepoint now
kvn@3588 3048 if (ilt->_safepts != NULL) {
kvn@3588 3049 ilt->_safepts->yank(n);
kvn@3588 3050 }
duke@0 3051 // Carry on with the recursion "as if" we are walking
duke@0 3052 // only the control input
duke@0 3053 if( !visited.test_set( in->_idx ) ) {
duke@0 3054 worklist.push(in); // Visit this guy later, using worklist
duke@0 3055 }
duke@0 3056 // Get next node from nstack:
duke@0 3057 // - skip n's inputs processing by setting i > cnt;
duke@0 3058 // - we also will not call set_early_ctrl(n) since
duke@0 3059 // has_node(n) == true (see the condition above).
duke@0 3060 i = cnt + 1;
duke@0 3061 }
duke@0 3062 }
duke@0 3063 } // if (i == 0)
duke@0 3064
duke@0 3065 // Visit all inputs
duke@0 3066 bool done = true; // Assume all n's inputs will be processed
duke@0 3067 while (i < cnt) {
duke@0 3068 Node *in = n->in(i);
duke@0 3069 ++i;
duke@0 3070 if (in == NULL) continue;
duke@0 3071 if (in->pinned() && !in->is_CFG())
duke@0 3072 set_ctrl(in, in->in(0));
duke@0 3073 int is_visited = visited.test_set( in->_idx );
duke@0 3074 if (!has_node(in)) { // No controlling input yet?
duke@0 3075 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
duke@0 3076 assert( !is_visited, "visit only once" );
duke@0 3077 nstack.push(n, i); // Save parent node and next input's index.
duke@0 3078 nstack_top_n = in; // Process current input now.
duke@0 3079 nstack_top_i = 0;
duke@0 3080 done = false; // Not all n's inputs processed.
duke@0 3081 break; // continue while_nstack_nonempty;
duke@0 3082 } else if (!is_visited) {
duke@0 3083 // This guy has a location picked out for him, but has not yet
duke@0 3084 // been visited. Happens to all CFG nodes, for instance.
duke@0 3085 // Visit him using the worklist instead of recursion, to break
duke@0 3086 // cycles. Since he has a location already we do not need to
duke@0 3087 // find his location before proceeding with the current Node.
duke@0 3088 worklist.push(in); // Visit this guy later, using worklist
duke@0 3089 }
duke@0 3090 }
duke@0 3091 if (done) {
duke@0 3092 // All of n's inputs have been processed, complete post-processing.
duke@0 3093
twisti@605 3094 // Compute earliest point this Node can go.
duke@0 3095 // CFG, Phi, pinned nodes already know their controlling input.
duke@0 3096 if (!has_node(n)) {
duke@0 3097 // Record earliest legal location
duke@0 3098 set_early_ctrl( n );
duke@0 3099 }
duke@0 3100 if (nstack.is_empty()) {
duke@0 3101 // Finished all nodes on stack.
duke@0 3102 // Process next node on the worklist.
duke@0 3103 break;
duke@0 3104 }
duke@0 3105 // Get saved parent node and next input's index.
duke@0 3106 nstack_top_n = nstack.node();
duke@0 3107 nstack_top_i = nstack.index();
duke@0 3108 nstack.pop();
duke@0 3109 }
duke@0 3110 } // while (true)
duke@0 3111 }
duke@0 3112 }
duke@0 3113
duke@0 3114 //------------------------------dom_lca_internal--------------------------------
duke@0 3115 // Pair-wise LCA
duke@0 3116 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
duke@0 3117 if( !n1 ) return n2; // Handle NULL original LCA
duke@0 3118 assert( n1->is_CFG(), "" );
duke@0 3119 assert( n2->is_CFG(), "" );
duke@0 3120 // find LCA of all uses
duke@0 3121 uint d1 = dom_depth(n1);
duke@0 3122 uint d2 = dom_depth(n2);
duke@0 3123 while (n1 != n2) {
duke@0 3124 if (d1 > d2) {
duke@0 3125 n1 = idom(n1);
duke@0 3126 d1 = dom_depth(n1);
duke@0 3127 } else if (d1 < d2) {
duke@0 3128 n2 = idom(n2);
duke@0 3129 d2 = dom_depth(n2);
duke@0 3130 } else {
duke@0 3131 // Here d1 == d2. Due to edits of the dominator-tree, sections
duke@0 3132 // of the tree might have the same depth. These sections have
duke@0 3133 // to be searched more carefully.
duke@0 3134
duke@0 3135 // Scan up all the n1's with equal depth, looking for n2.
duke@0 3136 Node *t1 = idom(n1);
duke@0 3137 while (dom_depth(t1) == d1) {
duke@0 3138 if (t1 == n2) return n2;
duke@0 3139 t1 = idom(t1);
duke@0 3140 }
duke@0 3141 // Scan up all the n2's with equal depth, looking for n1.
duke@0 3142 Node *t2 = idom(n2);
duke@0 3143 while (dom_depth(t2) == d2) {
duke@0 3144 if (t2 == n1) return n1;
duke@0 3145 t2 = idom(t2);
duke@0 3146 }
duke@0 3147 // Move up to a new dominator-depth value as well as up the dom-tree.
duke@0 3148 n1 = t1;
duke@0 3149 n2 = t2;
duke@0 3150 d1 = dom_depth(n1);
duke@0 3151 d2 = dom_depth(n2);
duke@0 3152 }
duke@0 3153 }
duke@0 3154 return n1;
duke@0 3155 }
duke@0 3156
duke@0 3157 //------------------------------compute_idom-----------------------------------
duke@0 3158 // Locally compute IDOM using dom_lca call. Correct only if the incoming
duke@0 3159 // IDOMs are correct.
duke@0 3160 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
duke@0 3161 assert( region->is_Region(), "" );
duke@0 3162 Node *LCA = NULL;
duke@0 3163 for( uint i = 1; i < region->req(); i++ ) {
duke@0 3164 if( region->in(i) != C->top() )
duke@0 3165 LCA = dom_lca( LCA, region->in(i) );
duke@0 3166 }
duke@0 3167 return LCA;
duke@0 3168 }
duke@0 3169
never@921 3170 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
never@921 3171 bool had_error = false;
never@921 3172 #ifdef ASSERT
never@921 3173 if (early != C->root()) {
never@921 3174 // Make sure that there's a dominance path from use to LCA
never@921 3175 Node* d = use;
never@921 3176 while (d != LCA) {
never@921 3177 d = idom(d);
never@921 3178 if (d == C->root()) {
never@921 3179 tty->print_cr("*** Use %d isn't dominated by def %s", use->_idx, n->_idx);
never@921 3180 n->dump();
never@921 3181 use->dump();
never@921 3182 had_error = true;
never@921 3183 break;
never@921 3184 }
never@921 3185 }
never@921 3186 }
never@921 3187 #endif
never@921 3188 return had_error;
never@921 3189 }
duke@0 3190
never@921 3191
never@921 3192 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
duke@0 3193 // Compute LCA over list of uses
never@921 3194 bool had_error = false;
duke@0 3195 Node *LCA = NULL;
duke@0 3196 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
duke@0 3197 Node* c = n->fast_out(i);
duke@0 3198 if (_nodes[c->_idx] == NULL)
duke@0 3199 continue; // Skip the occasional dead node
duke@0 3200 if( c->is_Phi() ) { // For Phis, we must land above on the path
duke@0 3201 for( uint j=1; j<c->req(); j++ ) {// For all inputs
duke@0 3202 if( c->in(j) == n ) { // Found matching input?
duke@0 3203 Node *use = c->in(0)->in(j);
never@921 3204 if (_verify_only && use->is_top()) continue;
duke@0 3205 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
never@921 3206 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
duke@0 3207 }
duke@0 3208 }
duke@0 3209 } else {
duke@0 3210 // For CFG data-users, use is in the block just prior
duke@0 3211 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
duke@0 3212 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
never@921 3213 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
duke@0 3214 }
duke@0 3215 }
never@921 3216 assert(!had_error, "bad dominance");
never@921 3217 return LCA;
never@921 3218 }
never@921 3219
never@921 3220 //------------------------------get_late_ctrl----------------------------------
never@921 3221 // Compute latest legal control.
never@921 3222 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
never@921 3223 assert(early != NULL, "early control should not be NULL");
never@921 3224
never@921 3225 Node* LCA = compute_lca_of_uses(n, early);
never@921 3226 #ifdef ASSERT
never@921 3227 if (LCA == C->root() && LCA != early) {
never@921 3228 // def doesn't dominate uses so print some useful debugging output
never@921 3229 compute_lca_of_uses(n, early, true);
never@921 3230 }
never@921 3231 #endif
duke@0 3232
duke@0 3233 // if this is a load, check for anti-dependent stores
duke@0 3234 // We use a conservative algorithm to identify potential interfering
duke@0 3235 // instructions and for rescheduling the load. The users of the memory
duke@0 3236 // input of this load are examined. Any use which is not a load and is
duke@0 3237 // dominated by early is considered a potentially interfering store.
duke@0 3238 // This can produce false positives.
duke@0 3239 if (n->is_Load() && LCA != early) {
duke@0 3240 Node_List worklist;
duke@0 3241
duke@0 3242 Node *mem = n->in(MemNode::Memory);
duke@0 3243 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
duke@0 3244 Node* s = mem->fast_out(i);
duke@0 3245 worklist.push(s);
duke@0 3246 }
duke@0 3247 while(worklist.size() != 0 && LCA != early) {
duke@0 3248 Node* s = worklist.pop();
duke@0 3249 if (s->is_Load()) {
duke@0 3250 continue;
duke@0 3251 } else if (s->is_MergeMem()) {
duke@0 3252 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
duke@0 3253 Node* s1 = s->fast_out(i);
duke@0 3254 worklist.push(s1);
duke@0 3255 }
duke@0 3256 } else {
duke@0 3257 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
duke@0 3258 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
duke@0 3259 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
duke@0 3260 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
duke@0 3261 }
duke@0 3262 }
duke@0 3263 }
duke@0 3264 }
duke@0 3265
duke@0 3266 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
duke@0 3267 return LCA;
duke@0 3268 }
duke@0 3269
duke@0 3270 // true if CFG node d dominates CFG node n
duke@0 3271 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
duke@0 3272 if (d == n)
duke@0 3273 return true;
duke@0 3274 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
duke@0 3275 uint dd = dom_depth(d);
duke@0 3276 while (dom_depth(n) >= dd) {
duke@0 3277 if (n == d)
duke@0 3278 return true;
duke@0 3279 n = idom(n);
duke@0 3280 }
duke@0 3281 return false;
duke@0 3282 }
duke@0 3283
duke@0 3284 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
duke@0 3285 // Pair-wise LCA with tags.
duke@0 3286 // Tag each index with the node 'tag' currently being processed
duke@0 3287 // before advancing up the dominator chain using idom().
duke@0 3288 // Later calls that find a match to 'tag' know that this path has already
duke@0 3289 // been considered in the current LCA (which is input 'n1' by convention).
duke@0 3290 // Since get_late_ctrl() is only called once for each node, the tag array
duke@0 3291 // does not need to be cleared between calls to get_late_ctrl().
duke@0 3292 // Algorithm trades a larger constant factor for better asymptotic behavior
duke@0 3293 //
duke@0 3294 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
duke@0 3295 uint d1 = dom_depth(n1);
duke@0 3296 uint d2 = dom_depth(n2);
duke@0 3297
duke@0 3298 do {
duke@0 3299 if (d1 > d2) {
duke@0 3300 // current lca is deeper than n2
duke@0 3301 _dom_lca_tags.map(n1->_idx, tag);
duke@0 3302 n1 = idom(n1);
duke@0 3303 d1 = dom_depth(n1);
duke@0 3304 } else if (d1 < d2) {
duke@0 3305 // n2 is deeper than current lca
duke@0 3306 Node *memo = _dom_lca_tags[n2->_idx];
duke@0 3307 if( memo == tag ) {
duke@0 3308 return n1; // Return the current LCA
duke@0 3309 }
duke@0 3310 _dom_lca_tags.map(n2->_idx, tag);
duke@0 3311 n2 = idom(n2);
duke@0 3312 d2 = dom_depth(n2);
duke@0 3313 } else {
duke@0 3314 // Here d1 == d2. Due to edits of the dominator-tree, sections
duke@0 3315 // of the tree might have the same depth. These sections have
duke@0 3316 // to be searched more carefully.
duke@0 3317
duke@0 3318 // Scan up all the n1's with equal depth, looking for n2.
duke@0 3319 _dom_lca_tags.map(n1->_idx, tag);
duke@0 3320 Node *t1 = idom(n1);
duke@0 3321 while (dom_depth(t1) == d1) {
duke@0 3322 if (t1 == n2) return n2;
duke@0 3323 _dom_lca_tags.map(t1->_idx, tag);
duke@0 3324 t1 = idom(t1);
duke@0 3325 }
duke@0 3326 // Scan up all the n2's with equal depth, looking for n1.
duke@0 3327 _dom_lca_tags.map(n2->_idx, tag);
duke@0 3328 Node *t2 = idom(n2);
duke@0 3329 while (dom_depth(t2) == d2) {
duke@0 3330 if (t2 == n1) return n1;
duke@0 3331 _dom_lca_tags.map(t2->_idx, tag);
duke@0 3332 t2 = idom(t2);
duke@0 3333 }
duke@0 3334 // Move up to a new dominator-depth value as well as up the dom-tree.
duke@0 3335 n1 = t1;
duke@0 3336 n2 = t2;
duke@0 3337 d1 = dom_depth(n1);
duke@0 3338 d2 = dom_depth(n2);
duke@0 3339 }
duke@0 3340 } while (n1 != n2);
duke@0 3341 return n1;
duke@0 3342 }
duke@0 3343
duke@0 3344 //------------------------------init_dom_lca_tags------------------------------
duke@0 3345 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
duke@0 3346 // Intended use does not involve any growth for the array, so it could
duke@0 3347 // be of fixed size.
duke@0 3348 void PhaseIdealLoop::init_dom_lca_tags() {
duke@0 3349 uint limit = C->unique() + 1;
duke@0 3350 _dom_lca_tags.map( limit, NULL );
duke@0 3351 #ifdef ASSERT
duke@0 3352 for( uint i = 0; i < limit; ++i ) {
duke@0 3353 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
duke@0 3354 }
duke@0 3355 #endif // ASSERT
duke@0 3356 }
duke@0 3357
duke@0 3358 //------------------------------clear_dom_lca_tags------------------------------
duke@0 3359 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
duke@0 3360 // Intended use does not involve any growth for the array, so it could
duke@0 3361 // be of fixed size.
duke@0 3362 void PhaseIdealLoop::clear_dom_lca_tags() {
duke@0 3363 uint limit = C->unique() + 1;
duke@0 3364 _dom_lca_tags.map( limit, NULL );
duke@0 3365 _dom_lca_tags.clear();
duke@0 3366 #ifdef ASSERT
duke@0 3367 for( uint i = 0; i < limit; ++i ) {
duke@0 3368 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
duke@0 3369 }
duke@0 3370 #endif // ASSERT
duke@0 3371 }
duke@0 3372
duke@0 3373 //------------------------------build_loop_late--------------------------------
duke@0 3374 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
duke@0 3375 // Second pass finds latest legal placement, and ideal loop placement.
never@921 3376 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
duke@0 3377 while (worklist.size() != 0) {
duke@0 3378 Node *n = worklist.pop();
duke@0 3379 // Only visit once
duke@0 3380 if (visited.test_set(n->_idx)) continue;
duke@0 3381 uint cnt = n->outcnt();
duke@0 3382 uint i = 0;
duke@0 3383 while (true) {
duke@0 3384 assert( _nodes[n->_idx], "no dead nodes" );
duke@0 3385 // Visit all children
duke@0 3386 if (i < cnt) {
duke@0 3387 Node* use = n->raw_out(i);
duke@0 3388 ++i;
duke@0 3389 // Check for dead uses. Aggressively prune such junk. It might be
duke@0 3390 // dead in the global sense, but still have local uses so I cannot
duke@0 3391 // easily call 'remove_dead_node'.
duke@0 3392 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
duke@0 3393 // Due to cycles, we might not hit the same fixed point in the verify
duke@0 3394 // pass as we do in the regular pass. Instead, visit such phis as
duke@0 3395 // simple uses of the loop head.
duke@0 3396 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
duke@0 3397 if( !visited.test(use->_idx) )
duke@0 3398 worklist.push(use);
duke@0 3399 } else if( !visited.test_set(use->_idx) ) {
duke@0 3400 nstack.push(n, i); // Save parent and next use's index.
duke@0 3401 n = use; // Process all children of current use.
duke@0 3402 cnt = use->outcnt();
duke@0 3403 i = 0;
duke@0 3404 }
duke@0 3405 } else {
duke@0 3406 // Do not visit around the backedge of loops via data edges.
duke@0 3407 // push dead code onto a worklist
duke@0 3408 _deadlist.push(use);
duke@0 3409 }
duke@0 3410 } else {
duke@0 3411 // All of n's children have been processed, complete post-processing.
never@921 3412 build_loop_late_post(n);
duke@0 3413 if (nstack.is_empty()) {
duke@0 3414 // Finished all nodes on stack.
duke@0 3415 // Process next node on the worklist.
duke@0 3416 break;
duke@0 3417 }
duke@0 3418 // Get saved parent node and next use's index. Visit the rest of uses.
duke@0 3419 n = nstack.node();
duke@0 3420 cnt = n->outcnt();
duke@0 3421 i = nstack.index();
duke@0 3422 nstack.pop();
duke@0 3423 }
duke@0 3424 }
duke@0 3425 }
duke@0 3426 }
duke@0 3427
duke@0 3428 //------------------------------build_loop_late_post---------------------------
duke@0 3429 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
duke@0 3430 // Second pass finds latest legal placement, and ideal loop placement.
never@921 3431 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
duke@0 3432
never@921 3433 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress() && !_verify_only) {
duke@0 3434 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
duke@0 3435 }
duke@0 3436
duke@0 3437 // CFG and pinned nodes already handled
duke@0 3438 if( n->in(0) ) {
duke@0 3439 if( n->in(0)->is_top() ) return; // Dead?
duke@0 3440
duke@0 3441 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
duke@0 3442 // _must_ be pinned (they have to observe their control edge of course).
duke@0 3443 // Unlike Stores (which modify an unallocable resource, the memory
duke@0 3444 // state), Mods/Loads can float around. So free them up.
duke@0 3445 bool pinned = true;
duke@0 3446 switch( n->Opcode() ) {
duke@0 3447 case Op_DivI:
duke@0 3448 case Op_DivF:
duke@0 3449 case Op_DivD:
duke@0 3450 case Op_ModI:
duke@0 3451 case Op_ModF:
duke@0 3452 case Op_ModD:
duke@0 3453 case Op_LoadB: // Same with Loads; they can sink
kvn@3447 3454 case Op_LoadUB: // during loop optimizations.
kvn@3447 3455 case Op_LoadUS:
duke@0 3456 case Op_LoadD:
duke@0 3457 case Op_LoadF:
duke@0 3458 case Op_LoadI:
duke@0 3459 case Op_LoadKlass:
kvn@293 3460 case Op_LoadNKlass:
duke@0 3461 case Op_LoadL:
duke@0 3462 case Op_LoadS:
duke@0 3463 case Op_LoadP:
kvn@293 3464 case Op_LoadN:
duke@0 3465 case Op_LoadRange:
duke@0 3466 case Op_LoadD_unaligned:
duke@0 3467 case Op_LoadL_unaligned:
duke@0 3468 case Op_StrComp: // Does a bunch of load-like effects
cfang@681 3469 case Op_StrEquals:
cfang@681 3470 case Op_StrIndexOf:
rasbold@169 3471 case Op_AryEq:
duke@0 3472 pinned = false;
duke@0 3473 }
duke@0 3474 if( pinned ) {
twisti@605 3475 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
twisti@605 3476 if( !chosen_loop->_child ) // Inner loop?
twisti@605 3477 chosen_loop->_body.push(n); // Collect inner loops
duke@0 3478 return;
duke@0 3479 }
duke@0 3480 } else { // No slot zero
duke@0 3481 if( n->is_CFG() ) { // CFG with no slot 0 is dead
duke@0 3482 _nodes.map(n->_idx,0); // No block setting, it's globally dead
duke@0 3483 return;
duke@0 3484 }
duke@0 3485 assert(!n->is_CFG() || n->outcnt() == 0, "");
duke@0 3486 }
duke@0 3487
duke@0 3488 // Do I have a "safe range" I can select over?
duke@0 3489 Node *early = get_ctrl(n);// Early location already computed
duke@0 3490
duke@0 3491 // Compute latest point this Node can go
duke@0 3492 Node *LCA = get_late_ctrl( n, early );
duke@0 3493 // LCA is NULL due to uses being dead
duke@0 3494 if( LCA == NULL ) {
duke@0 3495 #ifdef ASSERT
duke@0 3496 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
duke@0 3497 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
duke@0 3498 }
duke@0 3499 #endif
duke@0 3500 _nodes.map(n->_idx, 0); // This node is useless
duke@0 3501 _deadlist.push(n);
duke@0 3502 return;
duke@0 3503 }
duke@0 3504 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
duke@0 3505
duke@0 3506 Node *legal = LCA; // Walk 'legal' up the IDOM chain
duke@0 3507 Node *least = legal; // Best legal position so far
duke@0 3508 while( early != legal ) { // While not at earliest legal
cfang@1172 3509 #ifdef ASSERT
cfang@1172 3510 if (legal->is_Start() && !early->is_Root()) {
cfang@1172 3511 // Bad graph. Print idom path and fail.
roland@4154 3512 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
cfang@1172 3513 assert(false, "Bad graph detected in build_loop_late");
cfang@1172 3514 }
cfang@1172 3515 #endif
duke@0 3516 // Find least loop nesting depth
duke@0 3517 legal = idom(legal); // Bump up the IDOM tree
duke@0 3518 // Check for lower nesting depth
duke@0 3519 if( get_loop(legal)->_nest < get_loop(least)->_nest )
duke@0 3520 least = legal;
duke@0 3521 }
never@921 3522 assert(early == legal || legal != C->root(), "bad dominance of inputs");
duke@0 3523
duke@0 3524 // Try not to place code on a loop entry projection
duke@0 3525 // which can inhibit range check elimination.
duke@0 3526 if (least != early) {
duke@0 3527 Node* ctrl_out = least->unique_ctrl_out();
duke@0 3528 if (ctrl_out && ctrl_out->is_CountedLoop() &&
duke@0 3529 least == ctrl_out->in(LoopNode::EntryControl)) {
duke@0 3530 Node* least_dom = idom(least);
duke@0 3531 if (get_loop(least_dom)->is_member(get_loop(least))) {
duke@0 3532 least = least_dom;
duke@0 3533 }
duke@0 3534 }
duke@0 3535 }
duke@0 3536
duke@0 3537 #ifdef ASSERT
duke@0 3538 // If verifying, verify that 'verify_me' has a legal location
duke@0 3539 // and choose it as our location.
never@921 3540 if( _verify_me ) {
never@921 3541 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
duke@0 3542 Node *legal = LCA;
duke@0 3543 while( early != legal ) { // While not at earliest legal
duke@0 3544 if( legal == v_ctrl ) break; // Check for prior good location
duke@0 3545 legal = idom(legal) ;// Bump up the IDOM tree
duke@0 3546 }
duke@0 3547 // Check for prior good location
duke@0 3548 if( legal == v_ctrl ) least = legal; // Keep prior if found
duke@0 3549 }
duke@0 3550 #endif
duke@0 3551
duke@0 3552 // Assign discovered "here or above" point
duke@0 3553 least = find_non_split_ctrl(least);
duke@0 3554 set_ctrl(n, least);
duke@0 3555
duke@0 3556 // Collect inner loop bodies
twisti@605 3557 IdealLoopTree *chosen_loop = get_loop(least);
twisti@605 3558 if( !chosen_loop->_child ) // Inner loop?
twisti@605 3559 chosen_loop->_body.push(n);// Collect inner loops
duke@0 3560 }
duke@0 3561
kvn@2973 3562 #ifdef ASSERT
roland@4154 3563 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
roland@4154 3564 tty->print_cr(msg);
kvn@2973 3565 tty->print("n: "); n->dump();
kvn@2973 3566 tty->print("early(n): "); early->dump();
kvn@2973 3567 if (n->in(0) != NULL && !n->in(0)->is_top() &&