Thu, 20 Sep 2012 16:49:17 +0200
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
Summary: use shorter instruction sequences for atomic add and atomic exchange when possible.
Reviewed-by: kvn, jrose
duke@435 | 1 | /* |
brutisso@3489 | 2 | * Copyright (c) 1998, 2012, Oracle and/or its affiliates. All rights reserved. |
duke@435 | 3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
duke@435 | 4 | * |
duke@435 | 5 | * This code is free software; you can redistribute it and/or modify it |
duke@435 | 6 | * under the terms of the GNU General Public License version 2 only, as |
duke@435 | 7 | * published by the Free Software Foundation. |
duke@435 | 8 | * |
duke@435 | 9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
duke@435 | 10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
duke@435 | 11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
duke@435 | 12 | * version 2 for more details (a copy is included in the LICENSE file that |
duke@435 | 13 | * accompanied this code). |
duke@435 | 14 | * |
duke@435 | 15 | * You should have received a copy of the GNU General Public License version |
duke@435 | 16 | * 2 along with this work; if not, write to the Free Software Foundation, |
duke@435 | 17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
duke@435 | 18 | * |
trims@1907 | 19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
trims@1907 | 20 | * or visit www.oracle.com if you need additional information or have any |
trims@1907 | 21 | * questions. |
duke@435 | 22 | * |
duke@435 | 23 | */ |
duke@435 | 24 | |
stefank@2314 | 25 | #include "precompiled.hpp" |
stefank@2314 | 26 | #include "ci/ciMethodData.hpp" |
stefank@2314 | 27 | #include "compiler/compileLog.hpp" |
stefank@2314 | 28 | #include "libadt/vectset.hpp" |
stefank@2314 | 29 | #include "memory/allocation.inline.hpp" |
stefank@2314 | 30 | #include "opto/addnode.hpp" |
stefank@2314 | 31 | #include "opto/callnode.hpp" |
stefank@2314 | 32 | #include "opto/connode.hpp" |
stefank@2314 | 33 | #include "opto/divnode.hpp" |
stefank@2314 | 34 | #include "opto/idealGraphPrinter.hpp" |
stefank@2314 | 35 | #include "opto/loopnode.hpp" |
stefank@2314 | 36 | #include "opto/mulnode.hpp" |
stefank@2314 | 37 | #include "opto/rootnode.hpp" |
stefank@2314 | 38 | #include "opto/superword.hpp" |
duke@435 | 39 | |
duke@435 | 40 | //============================================================================= |
duke@435 | 41 | //------------------------------is_loop_iv------------------------------------- |
duke@435 | 42 | // Determine if a node is Counted loop induction variable. |
duke@435 | 43 | // The method is declared in node.hpp. |
duke@435 | 44 | const Node* Node::is_loop_iv() const { |
duke@435 | 45 | if (this->is_Phi() && !this->as_Phi()->is_copy() && |
duke@435 | 46 | this->as_Phi()->region()->is_CountedLoop() && |
duke@435 | 47 | this->as_Phi()->region()->as_CountedLoop()->phi() == this) { |
duke@435 | 48 | return this; |
duke@435 | 49 | } else { |
duke@435 | 50 | return NULL; |
duke@435 | 51 | } |
duke@435 | 52 | } |
duke@435 | 53 | |
duke@435 | 54 | //============================================================================= |
duke@435 | 55 | //------------------------------dump_spec-------------------------------------- |
duke@435 | 56 | // Dump special per-node info |
duke@435 | 57 | #ifndef PRODUCT |
duke@435 | 58 | void LoopNode::dump_spec(outputStream *st) const { |
kvn@2665 | 59 | if (is_inner_loop()) st->print( "inner " ); |
kvn@2665 | 60 | if (is_partial_peel_loop()) st->print( "partial_peel " ); |
kvn@2665 | 61 | if (partial_peel_has_failed()) st->print( "partial_peel_failed " ); |
duke@435 | 62 | } |
duke@435 | 63 | #endif |
duke@435 | 64 | |
kvn@2665 | 65 | //------------------------------is_valid_counted_loop------------------------- |
kvn@2665 | 66 | bool LoopNode::is_valid_counted_loop() const { |
kvn@2665 | 67 | if (is_CountedLoop()) { |
kvn@2665 | 68 | CountedLoopNode* l = as_CountedLoop(); |
kvn@2665 | 69 | CountedLoopEndNode* le = l->loopexit(); |
kvn@2665 | 70 | if (le != NULL && |
kvn@2665 | 71 | le->proj_out(1 /* true */) == l->in(LoopNode::LoopBackControl)) { |
kvn@2665 | 72 | Node* phi = l->phi(); |
kvn@2665 | 73 | Node* exit = le->proj_out(0 /* false */); |
kvn@2665 | 74 | if (exit != NULL && exit->Opcode() == Op_IfFalse && |
kvn@2665 | 75 | phi != NULL && phi->is_Phi() && |
kvn@2665 | 76 | phi->in(LoopNode::LoopBackControl) == l->incr() && |
kvn@2665 | 77 | le->loopnode() == l && le->stride_is_con()) { |
kvn@2665 | 78 | return true; |
kvn@2665 | 79 | } |
kvn@2665 | 80 | } |
kvn@2665 | 81 | } |
kvn@2665 | 82 | return false; |
kvn@2665 | 83 | } |
kvn@2665 | 84 | |
duke@435 | 85 | //------------------------------get_early_ctrl--------------------------------- |
duke@435 | 86 | // Compute earliest legal control |
duke@435 | 87 | Node *PhaseIdealLoop::get_early_ctrl( Node *n ) { |
duke@435 | 88 | assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" ); |
duke@435 | 89 | uint i; |
duke@435 | 90 | Node *early; |
duke@435 | 91 | if( n->in(0) ) { |
duke@435 | 92 | early = n->in(0); |
duke@435 | 93 | if( !early->is_CFG() ) // Might be a non-CFG multi-def |
duke@435 | 94 | early = get_ctrl(early); // So treat input as a straight data input |
duke@435 | 95 | i = 1; |
duke@435 | 96 | } else { |
duke@435 | 97 | early = get_ctrl(n->in(1)); |
duke@435 | 98 | i = 2; |
duke@435 | 99 | } |
duke@435 | 100 | uint e_d = dom_depth(early); |
duke@435 | 101 | assert( early, "" ); |
duke@435 | 102 | for( ; i < n->req(); i++ ) { |
duke@435 | 103 | Node *cin = get_ctrl(n->in(i)); |
duke@435 | 104 | assert( cin, "" ); |
duke@435 | 105 | // Keep deepest dominator depth |
duke@435 | 106 | uint c_d = dom_depth(cin); |
duke@435 | 107 | if( c_d > e_d ) { // Deeper guy? |
duke@435 | 108 | early = cin; // Keep deepest found so far |
duke@435 | 109 | e_d = c_d; |
duke@435 | 110 | } else if( c_d == e_d && // Same depth? |
duke@435 | 111 | early != cin ) { // If not equal, must use slower algorithm |
duke@435 | 112 | // If same depth but not equal, one _must_ dominate the other |
duke@435 | 113 | // and we want the deeper (i.e., dominated) guy. |
duke@435 | 114 | Node *n1 = early; |
duke@435 | 115 | Node *n2 = cin; |
duke@435 | 116 | while( 1 ) { |
duke@435 | 117 | n1 = idom(n1); // Walk up until break cycle |
duke@435 | 118 | n2 = idom(n2); |
duke@435 | 119 | if( n1 == cin || // Walked early up to cin |
duke@435 | 120 | dom_depth(n2) < c_d ) |
duke@435 | 121 | break; // early is deeper; keep him |
duke@435 | 122 | if( n2 == early || // Walked cin up to early |
duke@435 | 123 | dom_depth(n1) < c_d ) { |
duke@435 | 124 | early = cin; // cin is deeper; keep him |
duke@435 | 125 | break; |
duke@435 | 126 | } |
duke@435 | 127 | } |
duke@435 | 128 | e_d = dom_depth(early); // Reset depth register cache |
duke@435 | 129 | } |
duke@435 | 130 | } |
duke@435 | 131 | |
duke@435 | 132 | // Return earliest legal location |
duke@435 | 133 | assert(early == find_non_split_ctrl(early), "unexpected early control"); |
duke@435 | 134 | |
duke@435 | 135 | return early; |
duke@435 | 136 | } |
duke@435 | 137 | |
duke@435 | 138 | //------------------------------set_early_ctrl--------------------------------- |
duke@435 | 139 | // Set earliest legal control |
duke@435 | 140 | void PhaseIdealLoop::set_early_ctrl( Node *n ) { |
duke@435 | 141 | Node *early = get_early_ctrl(n); |
duke@435 | 142 | |
duke@435 | 143 | // Record earliest legal location |
duke@435 | 144 | set_ctrl(n, early); |
duke@435 | 145 | } |
duke@435 | 146 | |
duke@435 | 147 | //------------------------------set_subtree_ctrl------------------------------- |
duke@435 | 148 | // set missing _ctrl entries on new nodes |
duke@435 | 149 | void PhaseIdealLoop::set_subtree_ctrl( Node *n ) { |
duke@435 | 150 | // Already set? Get out. |
duke@435 | 151 | if( _nodes[n->_idx] ) return; |
duke@435 | 152 | // Recursively set _nodes array to indicate where the Node goes |
duke@435 | 153 | uint i; |
duke@435 | 154 | for( i = 0; i < n->req(); ++i ) { |
duke@435 | 155 | Node *m = n->in(i); |
duke@435 | 156 | if( m && m != C->root() ) |
duke@435 | 157 | set_subtree_ctrl( m ); |
duke@435 | 158 | } |
duke@435 | 159 | |
duke@435 | 160 | // Fixup self |
duke@435 | 161 | set_early_ctrl( n ); |
duke@435 | 162 | } |
duke@435 | 163 | |
duke@435 | 164 | //------------------------------is_counted_loop-------------------------------- |
kvn@2665 | 165 | bool PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) { |
duke@435 | 166 | PhaseGVN *gvn = &_igvn; |
duke@435 | 167 | |
duke@435 | 168 | // Counted loop head must be a good RegionNode with only 3 not NULL |
duke@435 | 169 | // control input edges: Self, Entry, LoopBack. |
kvn@2665 | 170 | if (x->in(LoopNode::Self) == NULL || x->req() != 3) |
kvn@2665 | 171 | return false; |
duke@435 | 172 | |
duke@435 | 173 | Node *init_control = x->in(LoopNode::EntryControl); |
duke@435 | 174 | Node *back_control = x->in(LoopNode::LoopBackControl); |
kvn@2665 | 175 | if (init_control == NULL || back_control == NULL) // Partially dead |
kvn@2665 | 176 | return false; |
duke@435 | 177 | // Must also check for TOP when looking for a dead loop |
kvn@2665 | 178 | if (init_control->is_top() || back_control->is_top()) |
kvn@2665 | 179 | return false; |
duke@435 | 180 | |
duke@435 | 181 | // Allow funny placement of Safepoint |
kvn@2665 | 182 | if (back_control->Opcode() == Op_SafePoint) |
duke@435 | 183 | back_control = back_control->in(TypeFunc::Control); |
duke@435 | 184 | |
duke@435 | 185 | // Controlling test for loop |
duke@435 | 186 | Node *iftrue = back_control; |
duke@435 | 187 | uint iftrue_op = iftrue->Opcode(); |
kvn@2665 | 188 | if (iftrue_op != Op_IfTrue && |
kvn@2665 | 189 | iftrue_op != Op_IfFalse) |
duke@435 | 190 | // I have a weird back-control. Probably the loop-exit test is in |
duke@435 | 191 | // the middle of the loop and I am looking at some trailing control-flow |
duke@435 | 192 | // merge point. To fix this I would have to partially peel the loop. |
kvn@2665 | 193 | return false; // Obscure back-control |
duke@435 | 194 | |
duke@435 | 195 | // Get boolean guarding loop-back test |
duke@435 | 196 | Node *iff = iftrue->in(0); |
kvn@2665 | 197 | if (get_loop(iff) != loop || !iff->in(1)->is_Bool()) |
kvn@2665 | 198 | return false; |
duke@435 | 199 | BoolNode *test = iff->in(1)->as_Bool(); |
duke@435 | 200 | BoolTest::mask bt = test->_test._test; |
duke@435 | 201 | float cl_prob = iff->as_If()->_prob; |
kvn@2665 | 202 | if (iftrue_op == Op_IfFalse) { |
duke@435 | 203 | bt = BoolTest(bt).negate(); |
duke@435 | 204 | cl_prob = 1.0 - cl_prob; |
duke@435 | 205 | } |
duke@435 | 206 | // Get backedge compare |
duke@435 | 207 | Node *cmp = test->in(1); |
duke@435 | 208 | int cmp_op = cmp->Opcode(); |
kvn@2877 | 209 | if (cmp_op != Op_CmpI) |
kvn@2665 | 210 | return false; // Avoid pointer & float compares |
duke@435 | 211 | |
duke@435 | 212 | // Find the trip-counter increment & limit. Limit must be loop invariant. |
duke@435 | 213 | Node *incr = cmp->in(1); |
duke@435 | 214 | Node *limit = cmp->in(2); |
duke@435 | 215 | |
duke@435 | 216 | // --------- |
duke@435 | 217 | // need 'loop()' test to tell if limit is loop invariant |
duke@435 | 218 | // --------- |
duke@435 | 219 | |
kvn@2665 | 220 | if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit? |
kvn@2665 | 221 | Node *tmp = incr; // Then reverse order into the CmpI |
duke@435 | 222 | incr = limit; |
duke@435 | 223 | limit = tmp; |
duke@435 | 224 | bt = BoolTest(bt).commute(); // And commute the exit test |
duke@435 | 225 | } |
kvn@2665 | 226 | if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant |
kvn@2665 | 227 | return false; |
kvn@2665 | 228 | if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant |
kvn@2665 | 229 | return false; |
duke@435 | 230 | |
kvn@2665 | 231 | Node* phi_incr = NULL; |
duke@435 | 232 | // Trip-counter increment must be commutative & associative. |
kvn@2665 | 233 | if (incr->is_Phi()) { |
kvn@2665 | 234 | if (incr->as_Phi()->region() != x || incr->req() != 3) |
kvn@2665 | 235 | return false; // Not simple trip counter expression |
kvn@2665 | 236 | phi_incr = incr; |
kvn@2665 | 237 | incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi |
kvn@2665 | 238 | if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant |
kvn@2665 | 239 | return false; |
duke@435 | 240 | } |
kvn@2665 | 241 | |
duke@435 | 242 | Node* trunc1 = NULL; |
duke@435 | 243 | Node* trunc2 = NULL; |
duke@435 | 244 | const TypeInt* iv_trunc_t = NULL; |
duke@435 | 245 | if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) { |
kvn@2665 | 246 | return false; // Funny increment opcode |
duke@435 | 247 | } |
kvn@2665 | 248 | assert(incr->Opcode() == Op_AddI, "wrong increment code"); |
duke@435 | 249 | |
duke@435 | 250 | // Get merge point |
duke@435 | 251 | Node *xphi = incr->in(1); |
duke@435 | 252 | Node *stride = incr->in(2); |
kvn@2665 | 253 | if (!stride->is_Con()) { // Oops, swap these |
kvn@2665 | 254 | if (!xphi->is_Con()) // Is the other guy a constant? |
kvn@2665 | 255 | return false; // Nope, unknown stride, bail out |
duke@435 | 256 | Node *tmp = xphi; // 'incr' is commutative, so ok to swap |
duke@435 | 257 | xphi = stride; |
duke@435 | 258 | stride = tmp; |
duke@435 | 259 | } |
kvn@2665 | 260 | // Stride must be constant |
kvn@2665 | 261 | int stride_con = stride->get_int(); |
kvn@2877 | 262 | if (stride_con == 0) |
kvn@2877 | 263 | return false; // missed some peephole opt |
kvn@2665 | 264 | |
kvn@2665 | 265 | if (!xphi->is_Phi()) |
kvn@2665 | 266 | return false; // Too much math on the trip counter |
kvn@2665 | 267 | if (phi_incr != NULL && phi_incr != xphi) |
kvn@2665 | 268 | return false; |
duke@435 | 269 | PhiNode *phi = xphi->as_Phi(); |
duke@435 | 270 | |
duke@435 | 271 | // Phi must be of loop header; backedge must wrap to increment |
kvn@2665 | 272 | if (phi->region() != x) |
kvn@2665 | 273 | return false; |
kvn@2665 | 274 | if (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr || |
kvn@2665 | 275 | trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1) { |
kvn@2665 | 276 | return false; |
duke@435 | 277 | } |
duke@435 | 278 | Node *init_trip = phi->in(LoopNode::EntryControl); |
duke@435 | 279 | |
duke@435 | 280 | // If iv trunc type is smaller than int, check for possible wrap. |
duke@435 | 281 | if (!TypeInt::INT->higher_equal(iv_trunc_t)) { |
duke@435 | 282 | assert(trunc1 != NULL, "must have found some truncation"); |
duke@435 | 283 | |
duke@435 | 284 | // Get a better type for the phi (filtered thru if's) |
duke@435 | 285 | const TypeInt* phi_ft = filtered_type(phi); |
duke@435 | 286 | |
duke@435 | 287 | // Can iv take on a value that will wrap? |
duke@435 | 288 | // |
duke@435 | 289 | // Ensure iv's limit is not within "stride" of the wrap value. |
duke@435 | 290 | // |
duke@435 | 291 | // Example for "short" type |
duke@435 | 292 | // Truncation ensures value is in the range -32768..32767 (iv_trunc_t) |
duke@435 | 293 | // If the stride is +10, then the last value of the induction |
duke@435 | 294 | // variable before the increment (phi_ft->_hi) must be |
duke@435 | 295 | // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to |
duke@435 | 296 | // ensure no truncation occurs after the increment. |
duke@435 | 297 | |
duke@435 | 298 | if (stride_con > 0) { |
duke@435 | 299 | if (iv_trunc_t->_hi - phi_ft->_hi < stride_con || |
duke@435 | 300 | iv_trunc_t->_lo > phi_ft->_lo) { |
kvn@2665 | 301 | return false; // truncation may occur |
duke@435 | 302 | } |
duke@435 | 303 | } else if (stride_con < 0) { |
duke@435 | 304 | if (iv_trunc_t->_lo - phi_ft->_lo > stride_con || |
duke@435 | 305 | iv_trunc_t->_hi < phi_ft->_hi) { |
kvn@2665 | 306 | return false; // truncation may occur |
duke@435 | 307 | } |
duke@435 | 308 | } |
duke@435 | 309 | // No possibility of wrap so truncation can be discarded |
duke@435 | 310 | // Promote iv type to Int |
duke@435 | 311 | } else { |
duke@435 | 312 | assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int"); |
duke@435 | 313 | } |
duke@435 | 314 | |
kvn@2665 | 315 | // If the condition is inverted and we will be rolling |
kvn@2665 | 316 | // through MININT to MAXINT, then bail out. |
kvn@2665 | 317 | if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice! |
kvn@2665 | 318 | // Odd stride |
kvn@2665 | 319 | bt == BoolTest::ne && stride_con != 1 && stride_con != -1 || |
kvn@2665 | 320 | // Count down loop rolls through MAXINT |
kvn@2665 | 321 | (bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0 || |
kvn@2665 | 322 | // Count up loop rolls through MININT |
kvn@2877 | 323 | (bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0) { |
kvn@2665 | 324 | return false; // Bail out |
kvn@2665 | 325 | } |
kvn@2665 | 326 | |
kvn@2665 | 327 | const TypeInt* init_t = gvn->type(init_trip)->is_int(); |
kvn@2665 | 328 | const TypeInt* limit_t = gvn->type(limit)->is_int(); |
kvn@2665 | 329 | |
kvn@2665 | 330 | if (stride_con > 0) { |
kvn@2665 | 331 | long init_p = (long)init_t->_lo + stride_con; |
kvn@2665 | 332 | if (init_p > (long)max_jint || init_p > (long)limit_t->_hi) |
kvn@2665 | 333 | return false; // cyclic loop or this loop trips only once |
kvn@2665 | 334 | } else { |
kvn@2665 | 335 | long init_p = (long)init_t->_hi + stride_con; |
kvn@2665 | 336 | if (init_p < (long)min_jint || init_p < (long)limit_t->_lo) |
kvn@2665 | 337 | return false; // cyclic loop or this loop trips only once |
kvn@2665 | 338 | } |
kvn@2665 | 339 | |
duke@435 | 340 | // ================================================= |
duke@435 | 341 | // ---- SUCCESS! Found A Trip-Counted Loop! ----- |
duke@435 | 342 | // |
kvn@2665 | 343 | assert(x->Opcode() == Op_Loop, "regular loops only"); |
duke@435 | 344 | C->print_method("Before CountedLoop", 3); |
kvn@2877 | 345 | |
kvn@2877 | 346 | Node *hook = new (C, 6) Node(6); |
kvn@2877 | 347 | |
kvn@2877 | 348 | if (LoopLimitCheck) { |
kvn@2877 | 349 | |
kvn@2877 | 350 | // =================================================== |
kvn@2877 | 351 | // Generate loop limit check to avoid integer overflow |
kvn@2877 | 352 | // in cases like next (cyclic loops): |
kvn@2877 | 353 | // |
kvn@2877 | 354 | // for (i=0; i <= max_jint; i++) {} |
kvn@2877 | 355 | // for (i=0; i < max_jint; i+=2) {} |
kvn@2877 | 356 | // |
kvn@2877 | 357 | // |
kvn@2877 | 358 | // Limit check predicate depends on the loop test: |
kvn@2877 | 359 | // |
kvn@2877 | 360 | // for(;i != limit; i++) --> limit <= (max_jint) |
kvn@2877 | 361 | // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1) |
kvn@2877 | 362 | // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride ) |
kvn@2877 | 363 | // |
kvn@2877 | 364 | |
kvn@2877 | 365 | // Check if limit is excluded to do more precise int overflow check. |
kvn@2877 | 366 | bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge); |
kvn@2877 | 367 | int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1)); |
kvn@2877 | 368 | |
kvn@2877 | 369 | // If compare points directly to the phi we need to adjust |
kvn@2877 | 370 | // the compare so that it points to the incr. Limit have |
kvn@2877 | 371 | // to be adjusted to keep trip count the same and the |
kvn@2877 | 372 | // adjusted limit should be checked for int overflow. |
kvn@2877 | 373 | if (phi_incr != NULL) { |
kvn@2877 | 374 | stride_m += stride_con; |
kvn@2877 | 375 | } |
kvn@2877 | 376 | |
kvn@2877 | 377 | if (limit->is_Con()) { |
kvn@2877 | 378 | int limit_con = limit->get_int(); |
kvn@2877 | 379 | if ((stride_con > 0 && limit_con > (max_jint - stride_m)) || |
kvn@2877 | 380 | (stride_con < 0 && limit_con < (min_jint - stride_m))) { |
kvn@2877 | 381 | // Bailout: it could be integer overflow. |
kvn@2877 | 382 | return false; |
kvn@2877 | 383 | } |
kvn@2877 | 384 | } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) || |
kvn@2877 | 385 | (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) { |
kvn@2877 | 386 | // Limit's type may satisfy the condition, for example, |
kvn@2877 | 387 | // when it is an array length. |
kvn@2877 | 388 | } else { |
kvn@2877 | 389 | // Generate loop's limit check. |
kvn@2877 | 390 | // Loop limit check predicate should be near the loop. |
kvn@2877 | 391 | ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check); |
kvn@2877 | 392 | if (!limit_check_proj) { |
kvn@2877 | 393 | // The limit check predicate is not generated if this method trapped here before. |
kvn@2877 | 394 | #ifdef ASSERT |
kvn@2877 | 395 | if (TraceLoopLimitCheck) { |
kvn@2877 | 396 | tty->print("missing loop limit check:"); |
kvn@2877 | 397 | loop->dump_head(); |
kvn@2877 | 398 | x->dump(1); |
kvn@2877 | 399 | } |
kvn@2877 | 400 | #endif |
kvn@2877 | 401 | return false; |
kvn@2877 | 402 | } |
kvn@2877 | 403 | |
kvn@2877 | 404 | IfNode* check_iff = limit_check_proj->in(0)->as_If(); |
kvn@2877 | 405 | Node* cmp_limit; |
kvn@2877 | 406 | Node* bol; |
kvn@2877 | 407 | |
kvn@2877 | 408 | if (stride_con > 0) { |
kvn@2877 | 409 | cmp_limit = new (C, 3) CmpINode(limit, _igvn.intcon(max_jint - stride_m)); |
kvn@2877 | 410 | bol = new (C, 2) BoolNode(cmp_limit, BoolTest::le); |
kvn@2877 | 411 | } else { |
kvn@2877 | 412 | cmp_limit = new (C, 3) CmpINode(limit, _igvn.intcon(min_jint - stride_m)); |
kvn@2877 | 413 | bol = new (C, 2) BoolNode(cmp_limit, BoolTest::ge); |
kvn@2877 | 414 | } |
kvn@2877 | 415 | cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit); |
kvn@2877 | 416 | bol = _igvn.register_new_node_with_optimizer(bol); |
kvn@2877 | 417 | set_subtree_ctrl(bol); |
kvn@2877 | 418 | |
kvn@2877 | 419 | // Replace condition in original predicate but preserve Opaque node |
kvn@2877 | 420 | // so that previous predicates could be found. |
kvn@2877 | 421 | assert(check_iff->in(1)->Opcode() == Op_Conv2B && |
kvn@2877 | 422 | check_iff->in(1)->in(1)->Opcode() == Op_Opaque1, ""); |
kvn@2877 | 423 | Node* opq = check_iff->in(1)->in(1); |
kvn@2877 | 424 | _igvn.hash_delete(opq); |
kvn@2877 | 425 | opq->set_req(1, bol); |
kvn@2877 | 426 | // Update ctrl. |
kvn@2877 | 427 | set_ctrl(opq, check_iff->in(0)); |
kvn@2877 | 428 | set_ctrl(check_iff->in(1), check_iff->in(0)); |
kvn@2877 | 429 | |
kvn@2727 | 430 | #ifndef PRODUCT |
kvn@2877 | 431 | // report that the loop predication has been actually performed |
kvn@2877 | 432 | // for this loop |
kvn@2877 | 433 | if (TraceLoopLimitCheck) { |
kvn@2877 | 434 | tty->print_cr("Counted Loop Limit Check generated:"); |
kvn@2877 | 435 | debug_only( bol->dump(2); ) |
kvn@2877 | 436 | } |
kvn@2877 | 437 | #endif |
kvn@2727 | 438 | } |
kvn@2877 | 439 | |
kvn@2877 | 440 | if (phi_incr != NULL) { |
kvn@2877 | 441 | // If compare points directly to the phi we need to adjust |
kvn@2877 | 442 | // the compare so that it points to the incr. Limit have |
kvn@2877 | 443 | // to be adjusted to keep trip count the same and we |
kvn@2877 | 444 | // should avoid int overflow. |
kvn@2877 | 445 | // |
kvn@2877 | 446 | // i = init; do {} while(i++ < limit); |
kvn@2877 | 447 | // is converted to |
kvn@2877 | 448 | // i = init; do {} while(++i < limit+1); |
kvn@2877 | 449 | // |
kvn@2877 | 450 | limit = gvn->transform(new (C, 3) AddINode(limit, stride)); |
kvn@2877 | 451 | } |
kvn@2877 | 452 | |
kvn@2877 | 453 | // Now we need to canonicalize loop condition. |
kvn@2877 | 454 | if (bt == BoolTest::ne) { |
kvn@2877 | 455 | assert(stride_con == 1 || stride_con == -1, "simple increment only"); |
kvn@2979 | 456 | // 'ne' can be replaced with 'lt' only when init < limit. |
kvn@2979 | 457 | if (stride_con > 0 && init_t->_hi < limit_t->_lo) |
kvn@2979 | 458 | bt = BoolTest::lt; |
kvn@2979 | 459 | // 'ne' can be replaced with 'gt' only when init > limit. |
kvn@2979 | 460 | if (stride_con < 0 && init_t->_lo > limit_t->_hi) |
kvn@2979 | 461 | bt = BoolTest::gt; |
kvn@2877 | 462 | } |
kvn@2877 | 463 | |
kvn@2877 | 464 | if (incl_limit) { |
kvn@2877 | 465 | // The limit check guaranties that 'limit <= (max_jint - stride)' so |
kvn@2877 | 466 | // we can convert 'i <= limit' to 'i < limit+1' since stride != 0. |
kvn@2877 | 467 | // |
kvn@2877 | 468 | Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1); |
kvn@2877 | 469 | limit = gvn->transform(new (C, 3) AddINode(limit, one)); |
kvn@2877 | 470 | if (bt == BoolTest::le) |
kvn@2877 | 471 | bt = BoolTest::lt; |
kvn@2877 | 472 | else if (bt == BoolTest::ge) |
kvn@2877 | 473 | bt = BoolTest::gt; |
kvn@2877 | 474 | else |
kvn@2877 | 475 | ShouldNotReachHere(); |
kvn@2877 | 476 | } |
kvn@2877 | 477 | set_subtree_ctrl( limit ); |
kvn@2877 | 478 | |
kvn@2877 | 479 | } else { // LoopLimitCheck |
kvn@2877 | 480 | |
duke@435 | 481 | // If compare points to incr, we are ok. Otherwise the compare |
duke@435 | 482 | // can directly point to the phi; in this case adjust the compare so that |
twisti@1040 | 483 | // it points to the incr by adjusting the limit. |
kvn@2665 | 484 | if (cmp->in(1) == phi || cmp->in(2) == phi) |
duke@435 | 485 | limit = gvn->transform(new (C, 3) AddINode(limit,stride)); |
duke@435 | 486 | |
duke@435 | 487 | // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride. |
duke@435 | 488 | // Final value for iterator should be: trip_count * stride + init_trip. |
duke@435 | 489 | Node *one_p = gvn->intcon( 1); |
duke@435 | 490 | Node *one_m = gvn->intcon(-1); |
duke@435 | 491 | |
duke@435 | 492 | Node *trip_count = NULL; |
duke@435 | 493 | switch( bt ) { |
duke@435 | 494 | case BoolTest::eq: |
kvn@2665 | 495 | ShouldNotReachHere(); |
duke@435 | 496 | case BoolTest::ne: // Ahh, the case we desire |
kvn@2665 | 497 | if (stride_con == 1) |
duke@435 | 498 | trip_count = gvn->transform(new (C, 3) SubINode(limit,init_trip)); |
kvn@2665 | 499 | else if (stride_con == -1) |
duke@435 | 500 | trip_count = gvn->transform(new (C, 3) SubINode(init_trip,limit)); |
duke@435 | 501 | else |
kvn@2665 | 502 | ShouldNotReachHere(); |
kvn@2665 | 503 | set_subtree_ctrl(trip_count); |
duke@435 | 504 | //_loop.map(trip_count->_idx,loop(limit)); |
duke@435 | 505 | break; |
duke@435 | 506 | case BoolTest::le: // Maybe convert to '<' case |
duke@435 | 507 | limit = gvn->transform(new (C, 3) AddINode(limit,one_p)); |
duke@435 | 508 | set_subtree_ctrl( limit ); |
duke@435 | 509 | hook->init_req(4, limit); |
duke@435 | 510 | |
duke@435 | 511 | bt = BoolTest::lt; |
duke@435 | 512 | // Make the new limit be in the same loop nest as the old limit |
duke@435 | 513 | //_loop.map(limit->_idx,limit_loop); |
duke@435 | 514 | // Fall into next case |
duke@435 | 515 | case BoolTest::lt: { // Maybe convert to '!=' case |
kvn@2665 | 516 | if (stride_con < 0) // Count down loop rolls through MAXINT |
kvn@2665 | 517 | ShouldNotReachHere(); |
duke@435 | 518 | Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip)); |
duke@435 | 519 | set_subtree_ctrl( range ); |
duke@435 | 520 | hook->init_req(0, range); |
duke@435 | 521 | |
duke@435 | 522 | Node *bias = gvn->transform(new (C, 3) AddINode(range,stride)); |
duke@435 | 523 | set_subtree_ctrl( bias ); |
duke@435 | 524 | hook->init_req(1, bias); |
duke@435 | 525 | |
duke@435 | 526 | Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_m)); |
duke@435 | 527 | set_subtree_ctrl( bias1 ); |
duke@435 | 528 | hook->init_req(2, bias1); |
duke@435 | 529 | |
duke@435 | 530 | trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride)); |
duke@435 | 531 | set_subtree_ctrl( trip_count ); |
duke@435 | 532 | hook->init_req(3, trip_count); |
duke@435 | 533 | break; |
duke@435 | 534 | } |
duke@435 | 535 | |
duke@435 | 536 | case BoolTest::ge: // Maybe convert to '>' case |
duke@435 | 537 | limit = gvn->transform(new (C, 3) AddINode(limit,one_m)); |
duke@435 | 538 | set_subtree_ctrl( limit ); |
duke@435 | 539 | hook->init_req(4 ,limit); |
duke@435 | 540 | |
duke@435 | 541 | bt = BoolTest::gt; |
duke@435 | 542 | // Make the new limit be in the same loop nest as the old limit |
duke@435 | 543 | //_loop.map(limit->_idx,limit_loop); |
duke@435 | 544 | // Fall into next case |
duke@435 | 545 | case BoolTest::gt: { // Maybe convert to '!=' case |
kvn@2665 | 546 | if (stride_con > 0) // count up loop rolls through MININT |
kvn@2665 | 547 | ShouldNotReachHere(); |
duke@435 | 548 | Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip)); |
duke@435 | 549 | set_subtree_ctrl( range ); |
duke@435 | 550 | hook->init_req(0, range); |
duke@435 | 551 | |
duke@435 | 552 | Node *bias = gvn->transform(new (C, 3) AddINode(range,stride)); |
duke@435 | 553 | set_subtree_ctrl( bias ); |
duke@435 | 554 | hook->init_req(1, bias); |
duke@435 | 555 | |
duke@435 | 556 | Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_p)); |
duke@435 | 557 | set_subtree_ctrl( bias1 ); |
duke@435 | 558 | hook->init_req(2, bias1); |
duke@435 | 559 | |
duke@435 | 560 | trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride)); |
duke@435 | 561 | set_subtree_ctrl( trip_count ); |
duke@435 | 562 | hook->init_req(3, trip_count); |
duke@435 | 563 | break; |
duke@435 | 564 | } |
kvn@2665 | 565 | } // switch( bt ) |
duke@435 | 566 | |
duke@435 | 567 | Node *span = gvn->transform(new (C, 3) MulINode(trip_count,stride)); |
duke@435 | 568 | set_subtree_ctrl( span ); |
duke@435 | 569 | hook->init_req(5, span); |
duke@435 | 570 | |
duke@435 | 571 | limit = gvn->transform(new (C, 3) AddINode(span,init_trip)); |
duke@435 | 572 | set_subtree_ctrl( limit ); |
duke@435 | 573 | |
kvn@2877 | 574 | } // LoopLimitCheck |
kvn@2877 | 575 | |
kvn@2665 | 576 | // Check for SafePoint on backedge and remove |
kvn@2665 | 577 | Node *sfpt = x->in(LoopNode::LoopBackControl); |
kvn@2665 | 578 | if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) { |
kvn@2665 | 579 | lazy_replace( sfpt, iftrue ); |
kvn@4023 | 580 | if (loop->_safepts != NULL) { |
kvn@4023 | 581 | loop->_safepts->yank(sfpt); |
kvn@4023 | 582 | } |
kvn@2665 | 583 | loop->_tail = iftrue; |
kvn@2665 | 584 | } |
kvn@2665 | 585 | |
duke@435 | 586 | // Build a canonical trip test. |
duke@435 | 587 | // Clone code, as old values may be in use. |
duke@435 | 588 | incr = incr->clone(); |
kvn@3135 | 589 | incr->set_req(1,phi); |
duke@435 | 590 | incr->set_req(2,stride); |
duke@435 | 591 | incr = _igvn.register_new_node_with_optimizer(incr); |
duke@435 | 592 | set_early_ctrl( incr ); |
kvn@3135 | 593 | _igvn.hash_delete(phi); |
kvn@3135 | 594 | phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn ); |
kvn@3135 | 595 | |
kvn@3135 | 596 | // If phi type is more restrictive than Int, raise to |
kvn@3135 | 597 | // Int to prevent (almost) infinite recursion in igvn |
kvn@3135 | 598 | // which can only handle integer types for constants or minint..maxint. |
kvn@3135 | 599 | if (!TypeInt::INT->higher_equal(phi->bottom_type())) { |
kvn@3135 | 600 | Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT); |
kvn@3135 | 601 | nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl)); |
kvn@3135 | 602 | nphi = _igvn.register_new_node_with_optimizer(nphi); |
kvn@3135 | 603 | set_ctrl(nphi, get_ctrl(phi)); |
kvn@3135 | 604 | _igvn.replace_node(phi, nphi); |
kvn@3135 | 605 | phi = nphi->as_Phi(); |
kvn@3135 | 606 | } |
duke@435 | 607 | cmp = cmp->clone(); |
duke@435 | 608 | cmp->set_req(1,incr); |
duke@435 | 609 | cmp->set_req(2,limit); |
duke@435 | 610 | cmp = _igvn.register_new_node_with_optimizer(cmp); |
duke@435 | 611 | set_ctrl(cmp, iff->in(0)); |
duke@435 | 612 | |
kvn@2665 | 613 | test = test->clone()->as_Bool(); |
kvn@2665 | 614 | (*(BoolTest*)&test->_test)._test = bt; |
duke@435 | 615 | test->set_req(1,cmp); |
duke@435 | 616 | _igvn.register_new_node_with_optimizer(test); |
duke@435 | 617 | set_ctrl(test, iff->in(0)); |
duke@435 | 618 | |
duke@435 | 619 | // Replace the old IfNode with a new LoopEndNode |
kvn@2665 | 620 | Node *lex = _igvn.register_new_node_with_optimizer(new (C, 2) CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt )); |
duke@435 | 621 | IfNode *le = lex->as_If(); |
duke@435 | 622 | uint dd = dom_depth(iff); |
duke@435 | 623 | set_idom(le, le->in(0), dd); // Update dominance for loop exit |
duke@435 | 624 | set_loop(le, loop); |
duke@435 | 625 | |
duke@435 | 626 | // Get the loop-exit control |
kvn@2665 | 627 | Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue)); |
duke@435 | 628 | |
duke@435 | 629 | // Need to swap loop-exit and loop-back control? |
kvn@2665 | 630 | if (iftrue_op == Op_IfFalse) { |
duke@435 | 631 | Node *ift2=_igvn.register_new_node_with_optimizer(new (C, 1) IfTrueNode (le)); |
duke@435 | 632 | Node *iff2=_igvn.register_new_node_with_optimizer(new (C, 1) IfFalseNode(le)); |
duke@435 | 633 | |
duke@435 | 634 | loop->_tail = back_control = ift2; |
duke@435 | 635 | set_loop(ift2, loop); |
kvn@2665 | 636 | set_loop(iff2, get_loop(iffalse)); |
duke@435 | 637 | |
duke@435 | 638 | // Lazy update of 'get_ctrl' mechanism. |
kvn@2665 | 639 | lazy_replace_proj( iffalse, iff2 ); |
kvn@2665 | 640 | lazy_replace_proj( iftrue, ift2 ); |
duke@435 | 641 | |
duke@435 | 642 | // Swap names |
kvn@2665 | 643 | iffalse = iff2; |
kvn@2665 | 644 | iftrue = ift2; |
duke@435 | 645 | } else { |
kvn@2665 | 646 | _igvn.hash_delete(iffalse); |
duke@435 | 647 | _igvn.hash_delete(iftrue); |
kvn@2665 | 648 | iffalse->set_req_X( 0, le, &_igvn ); |
kvn@2665 | 649 | iftrue ->set_req_X( 0, le, &_igvn ); |
duke@435 | 650 | } |
duke@435 | 651 | |
kvn@2665 | 652 | set_idom(iftrue, le, dd+1); |
kvn@2665 | 653 | set_idom(iffalse, le, dd+1); |
kvn@2665 | 654 | assert(iff->outcnt() == 0, "should be dead now"); |
kvn@2665 | 655 | lazy_replace( iff, le ); // fix 'get_ctrl' |
duke@435 | 656 | |
duke@435 | 657 | // Now setup a new CountedLoopNode to replace the existing LoopNode |
duke@435 | 658 | CountedLoopNode *l = new (C, 3) CountedLoopNode(init_control, back_control); |
kvn@2665 | 659 | l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve |
duke@435 | 660 | // The following assert is approximately true, and defines the intention |
duke@435 | 661 | // of can_be_counted_loop. It fails, however, because phase->type |
duke@435 | 662 | // is not yet initialized for this loop and its parts. |
duke@435 | 663 | //assert(l->can_be_counted_loop(this), "sanity"); |
duke@435 | 664 | _igvn.register_new_node_with_optimizer(l); |
duke@435 | 665 | set_loop(l, loop); |
duke@435 | 666 | loop->_head = l; |
duke@435 | 667 | // Fix all data nodes placed at the old loop head. |
duke@435 | 668 | // Uses the lazy-update mechanism of 'get_ctrl'. |
duke@435 | 669 | lazy_replace( x, l ); |
duke@435 | 670 | set_idom(l, init_control, dom_depth(x)); |
duke@435 | 671 | |
twisti@1040 | 672 | // Check for immediately preceding SafePoint and remove |
duke@435 | 673 | Node *sfpt2 = le->in(0); |
kvn@4023 | 674 | if (sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2)) { |
duke@435 | 675 | lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control)); |
kvn@4023 | 676 | if (loop->_safepts != NULL) { |
kvn@4023 | 677 | loop->_safepts->yank(sfpt2); |
kvn@4023 | 678 | } |
kvn@4023 | 679 | } |
duke@435 | 680 | |
duke@435 | 681 | // Free up intermediate goo |
duke@435 | 682 | _igvn.remove_dead_node(hook); |
duke@435 | 683 | |
kvn@2665 | 684 | #ifdef ASSERT |
kvn@2665 | 685 | assert(l->is_valid_counted_loop(), "counted loop shape is messed up"); |
kvn@2665 | 686 | assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" ); |
kvn@2665 | 687 | #endif |
kvn@2877 | 688 | #ifndef PRODUCT |
kvn@2877 | 689 | if (TraceLoopOpts) { |
kvn@2877 | 690 | tty->print("Counted "); |
kvn@2877 | 691 | loop->dump_head(); |
kvn@2877 | 692 | } |
kvn@2877 | 693 | #endif |
kvn@2665 | 694 | |
duke@435 | 695 | C->print_method("After CountedLoop", 3); |
duke@435 | 696 | |
kvn@2665 | 697 | return true; |
duke@435 | 698 | } |
duke@435 | 699 | |
kvn@2877 | 700 | //----------------------exact_limit------------------------------------------- |
kvn@2877 | 701 | Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) { |
kvn@2877 | 702 | assert(loop->_head->is_CountedLoop(), ""); |
kvn@2877 | 703 | CountedLoopNode *cl = loop->_head->as_CountedLoop(); |
kvn@3048 | 704 | assert(cl->is_valid_counted_loop(), ""); |
kvn@2877 | 705 | |
kvn@2877 | 706 | if (!LoopLimitCheck || ABS(cl->stride_con()) == 1 || |
kvn@2877 | 707 | cl->limit()->Opcode() == Op_LoopLimit) { |
kvn@2877 | 708 | // Old code has exact limit (it could be incorrect in case of int overflow). |
kvn@2877 | 709 | // Loop limit is exact with stride == 1. And loop may already have exact limit. |
kvn@2877 | 710 | return cl->limit(); |
kvn@2877 | 711 | } |
kvn@2877 | 712 | Node *limit = NULL; |
kvn@2877 | 713 | #ifdef ASSERT |
kvn@2877 | 714 | BoolTest::mask bt = cl->loopexit()->test_trip(); |
kvn@2877 | 715 | assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected"); |
kvn@2877 | 716 | #endif |
kvn@2877 | 717 | if (cl->has_exact_trip_count()) { |
kvn@2877 | 718 | // Simple case: loop has constant boundaries. |
kvn@2877 | 719 | // Use longs to avoid integer overflow. |
kvn@2877 | 720 | int stride_con = cl->stride_con(); |
kvn@2877 | 721 | long init_con = cl->init_trip()->get_int(); |
kvn@2877 | 722 | long limit_con = cl->limit()->get_int(); |
kvn@2877 | 723 | julong trip_cnt = cl->trip_count(); |
kvn@2877 | 724 | long final_con = init_con + trip_cnt*stride_con; |
kvn@2877 | 725 | int final_int = (int)final_con; |
kvn@2877 | 726 | // The final value should be in integer range since the loop |
kvn@2877 | 727 | // is counted and the limit was checked for overflow. |
kvn@2877 | 728 | assert(final_con == (long)final_int, "final value should be integer"); |
kvn@2877 | 729 | limit = _igvn.intcon(final_int); |
kvn@2877 | 730 | } else { |
kvn@2877 | 731 | // Create new LoopLimit node to get exact limit (final iv value). |
kvn@2877 | 732 | limit = new (C, 4) LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride()); |
kvn@2877 | 733 | register_new_node(limit, cl->in(LoopNode::EntryControl)); |
kvn@2877 | 734 | } |
kvn@2877 | 735 | assert(limit != NULL, "sanity"); |
kvn@2877 | 736 | return limit; |
kvn@2877 | 737 | } |
duke@435 | 738 | |
duke@435 | 739 | //------------------------------Ideal------------------------------------------ |
duke@435 | 740 | // Return a node which is more "ideal" than the current node. |
duke@435 | 741 | // Attempt to convert into a counted-loop. |
duke@435 | 742 | Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
duke@435 | 743 | if (!can_be_counted_loop(phase)) { |
duke@435 | 744 | phase->C->set_major_progress(); |
duke@435 | 745 | } |
duke@435 | 746 | return RegionNode::Ideal(phase, can_reshape); |
duke@435 | 747 | } |
duke@435 | 748 | |
duke@435 | 749 | |
duke@435 | 750 | //============================================================================= |
duke@435 | 751 | //------------------------------Ideal------------------------------------------ |
duke@435 | 752 | // Return a node which is more "ideal" than the current node. |
duke@435 | 753 | // Attempt to convert into a counted-loop. |
duke@435 | 754 | Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
duke@435 | 755 | return RegionNode::Ideal(phase, can_reshape); |
duke@435 | 756 | } |
duke@435 | 757 | |
duke@435 | 758 | //------------------------------dump_spec-------------------------------------- |
duke@435 | 759 | // Dump special per-node info |
duke@435 | 760 | #ifndef PRODUCT |
duke@435 | 761 | void CountedLoopNode::dump_spec(outputStream *st) const { |
duke@435 | 762 | LoopNode::dump_spec(st); |
kvn@2877 | 763 | if (stride_is_con()) { |
duke@435 | 764 | st->print("stride: %d ",stride_con()); |
duke@435 | 765 | } |
kvn@2877 | 766 | if (is_pre_loop ()) st->print("pre of N%d" , _main_idx); |
kvn@2877 | 767 | if (is_main_loop()) st->print("main of N%d", _idx); |
kvn@2877 | 768 | if (is_post_loop()) st->print("post of N%d", _main_idx); |
duke@435 | 769 | } |
duke@435 | 770 | #endif |
duke@435 | 771 | |
duke@435 | 772 | //============================================================================= |
duke@435 | 773 | int CountedLoopEndNode::stride_con() const { |
duke@435 | 774 | return stride()->bottom_type()->is_int()->get_con(); |
duke@435 | 775 | } |
duke@435 | 776 | |
kvn@2877 | 777 | //============================================================================= |
kvn@2877 | 778 | //------------------------------Value----------------------------------------- |
kvn@2877 | 779 | const Type *LoopLimitNode::Value( PhaseTransform *phase ) const { |
kvn@2877 | 780 | const Type* init_t = phase->type(in(Init)); |
kvn@2877 | 781 | const Type* limit_t = phase->type(in(Limit)); |
kvn@2877 | 782 | const Type* stride_t = phase->type(in(Stride)); |
kvn@2877 | 783 | // Either input is TOP ==> the result is TOP |
kvn@2877 | 784 | if (init_t == Type::TOP) return Type::TOP; |
kvn@2877 | 785 | if (limit_t == Type::TOP) return Type::TOP; |
kvn@2877 | 786 | if (stride_t == Type::TOP) return Type::TOP; |
kvn@2877 | 787 | |
kvn@2877 | 788 | int stride_con = stride_t->is_int()->get_con(); |
kvn@2877 | 789 | if (stride_con == 1) |
kvn@2877 | 790 | return NULL; // Identity |
kvn@2877 | 791 | |
kvn@2877 | 792 | if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) { |
kvn@2877 | 793 | // Use longs to avoid integer overflow. |
kvn@2877 | 794 | long init_con = init_t->is_int()->get_con(); |
kvn@2877 | 795 | long limit_con = limit_t->is_int()->get_con(); |
kvn@2877 | 796 | int stride_m = stride_con - (stride_con > 0 ? 1 : -1); |
kvn@2877 | 797 | long trip_count = (limit_con - init_con + stride_m)/stride_con; |
kvn@2877 | 798 | long final_con = init_con + stride_con*trip_count; |
kvn@2877 | 799 | int final_int = (int)final_con; |
kvn@2877 | 800 | // The final value should be in integer range since the loop |
kvn@2877 | 801 | // is counted and the limit was checked for overflow. |
kvn@2877 | 802 | assert(final_con == (long)final_int, "final value should be integer"); |
kvn@2877 | 803 | return TypeInt::make(final_int); |
kvn@2877 | 804 | } |
kvn@2877 | 805 | |
kvn@2877 | 806 | return bottom_type(); // TypeInt::INT |
kvn@2877 | 807 | } |
kvn@2877 | 808 | |
kvn@2877 | 809 | //------------------------------Ideal------------------------------------------ |
kvn@2877 | 810 | // Return a node which is more "ideal" than the current node. |
kvn@2877 | 811 | Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
kvn@2877 | 812 | if (phase->type(in(Init)) == Type::TOP || |
kvn@2877 | 813 | phase->type(in(Limit)) == Type::TOP || |
kvn@2877 | 814 | phase->type(in(Stride)) == Type::TOP) |
kvn@2877 | 815 | return NULL; // Dead |
kvn@2877 | 816 | |
kvn@2877 | 817 | int stride_con = phase->type(in(Stride))->is_int()->get_con(); |
kvn@2877 | 818 | if (stride_con == 1) |
kvn@2877 | 819 | return NULL; // Identity |
kvn@2877 | 820 | |
kvn@2877 | 821 | if (in(Init)->is_Con() && in(Limit)->is_Con()) |
kvn@2877 | 822 | return NULL; // Value |
kvn@2877 | 823 | |
kvn@2877 | 824 | // Delay following optimizations until all loop optimizations |
kvn@2877 | 825 | // done to keep Ideal graph simple. |
kvn@2877 | 826 | if (!can_reshape || phase->C->major_progress()) |
kvn@2877 | 827 | return NULL; |
kvn@2877 | 828 | |
kvn@2877 | 829 | const TypeInt* init_t = phase->type(in(Init) )->is_int(); |
kvn@2877 | 830 | const TypeInt* limit_t = phase->type(in(Limit))->is_int(); |
kvn@2877 | 831 | int stride_p; |
kvn@2877 | 832 | long lim, ini; |
kvn@2877 | 833 | julong max; |
kvn@2877 | 834 | if (stride_con > 0) { |
kvn@2877 | 835 | stride_p = stride_con; |
kvn@2877 | 836 | lim = limit_t->_hi; |
kvn@2877 | 837 | ini = init_t->_lo; |
kvn@2877 | 838 | max = (julong)max_jint; |
kvn@2877 | 839 | } else { |
kvn@2877 | 840 | stride_p = -stride_con; |
kvn@2877 | 841 | lim = init_t->_hi; |
kvn@2877 | 842 | ini = limit_t->_lo; |
kvn@2877 | 843 | max = (julong)min_jint; |
kvn@2877 | 844 | } |
kvn@2877 | 845 | julong range = lim - ini + stride_p; |
kvn@2877 | 846 | if (range <= max) { |
kvn@2877 | 847 | // Convert to integer expression if it is not overflow. |
kvn@2877 | 848 | Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1)); |
kvn@2877 | 849 | Node *range = phase->transform(new (phase->C, 3) SubINode(in(Limit), in(Init))); |
kvn@2877 | 850 | Node *bias = phase->transform(new (phase->C, 3) AddINode(range, stride_m)); |
kvn@2877 | 851 | Node *trip = phase->transform(new (phase->C, 3) DivINode(0, bias, in(Stride))); |
kvn@2877 | 852 | Node *span = phase->transform(new (phase->C, 3) MulINode(trip, in(Stride))); |
kvn@2877 | 853 | return new (phase->C, 3) AddINode(span, in(Init)); // exact limit |
kvn@2877 | 854 | } |
kvn@2877 | 855 | |
kvn@2877 | 856 | if (is_power_of_2(stride_p) || // divisor is 2^n |
kvn@2877 | 857 | !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node? |
kvn@2877 | 858 | // Convert to long expression to avoid integer overflow |
kvn@2877 | 859 | // and let igvn optimizer convert this division. |
kvn@2877 | 860 | // |
kvn@2877 | 861 | Node* init = phase->transform( new (phase->C, 2) ConvI2LNode(in(Init))); |
kvn@2877 | 862 | Node* limit = phase->transform( new (phase->C, 2) ConvI2LNode(in(Limit))); |
kvn@2877 | 863 | Node* stride = phase->longcon(stride_con); |
kvn@2877 | 864 | Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1)); |
kvn@2877 | 865 | |
kvn@2877 | 866 | Node *range = phase->transform(new (phase->C, 3) SubLNode(limit, init)); |
kvn@2877 | 867 | Node *bias = phase->transform(new (phase->C, 3) AddLNode(range, stride_m)); |
kvn@2877 | 868 | Node *span; |
kvn@2877 | 869 | if (stride_con > 0 && is_power_of_2(stride_p)) { |
kvn@2877 | 870 | // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride) |
kvn@2877 | 871 | // and avoid generating rounding for division. Zero trip guard should |
kvn@2877 | 872 | // guarantee that init < limit but sometimes the guard is missing and |
kvn@2877 | 873 | // we can get situation when init > limit. Note, for the empty loop |
kvn@2877 | 874 | // optimization zero trip guard is generated explicitly which leaves |
kvn@2877 | 875 | // only RCE predicate where exact limit is used and the predicate |
kvn@2877 | 876 | // will simply fail forcing recompilation. |
kvn@2877 | 877 | Node* neg_stride = phase->longcon(-stride_con); |
kvn@2877 | 878 | span = phase->transform(new (phase->C, 3) AndLNode(bias, neg_stride)); |
kvn@2877 | 879 | } else { |
kvn@2877 | 880 | Node *trip = phase->transform(new (phase->C, 3) DivLNode(0, bias, stride)); |
kvn@2877 | 881 | span = phase->transform(new (phase->C, 3) MulLNode(trip, stride)); |
kvn@2877 | 882 | } |
kvn@2877 | 883 | // Convert back to int |
kvn@2877 | 884 | Node *span_int = phase->transform(new (phase->C, 2) ConvL2INode(span)); |
kvn@2877 | 885 | return new (phase->C, 3) AddINode(span_int, in(Init)); // exact limit |
kvn@2877 | 886 | } |
kvn@2877 | 887 | |
kvn@2877 | 888 | return NULL; // No progress |
kvn@2877 | 889 | } |
kvn@2877 | 890 | |
kvn@2877 | 891 | //------------------------------Identity--------------------------------------- |
kvn@2877 | 892 | // If stride == 1 return limit node. |
kvn@2877 | 893 | Node *LoopLimitNode::Identity( PhaseTransform *phase ) { |
kvn@2877 | 894 | int stride_con = phase->type(in(Stride))->is_int()->get_con(); |
kvn@2877 | 895 | if (stride_con == 1 || stride_con == -1) |
kvn@2877 | 896 | return in(Limit); |
kvn@2877 | 897 | return this; |
kvn@2877 | 898 | } |
kvn@2877 | 899 | |
kvn@2877 | 900 | //============================================================================= |
duke@435 | 901 | //----------------------match_incr_with_optional_truncation-------------------- |
duke@435 | 902 | // Match increment with optional truncation: |
duke@435 | 903 | // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16 |
duke@435 | 904 | // Return NULL for failure. Success returns the increment node. |
duke@435 | 905 | Node* CountedLoopNode::match_incr_with_optional_truncation( |
duke@435 | 906 | Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) { |
duke@435 | 907 | // Quick cutouts: |
brutisso@3489 | 908 | if (expr == NULL || expr->req() != 3) return NULL; |
duke@435 | 909 | |
duke@435 | 910 | Node *t1 = NULL; |
duke@435 | 911 | Node *t2 = NULL; |
duke@435 | 912 | const TypeInt* trunc_t = TypeInt::INT; |
duke@435 | 913 | Node* n1 = expr; |
duke@435 | 914 | int n1op = n1->Opcode(); |
duke@435 | 915 | |
duke@435 | 916 | // Try to strip (n1 & M) or (n1 << N >> N) from n1. |
duke@435 | 917 | if (n1op == Op_AndI && |
duke@435 | 918 | n1->in(2)->is_Con() && |
duke@435 | 919 | n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) { |
duke@435 | 920 | // %%% This check should match any mask of 2**K-1. |
duke@435 | 921 | t1 = n1; |
duke@435 | 922 | n1 = t1->in(1); |
duke@435 | 923 | n1op = n1->Opcode(); |
duke@435 | 924 | trunc_t = TypeInt::CHAR; |
duke@435 | 925 | } else if (n1op == Op_RShiftI && |
duke@435 | 926 | n1->in(1) != NULL && |
duke@435 | 927 | n1->in(1)->Opcode() == Op_LShiftI && |
duke@435 | 928 | n1->in(2) == n1->in(1)->in(2) && |
duke@435 | 929 | n1->in(2)->is_Con()) { |
duke@435 | 930 | jint shift = n1->in(2)->bottom_type()->is_int()->get_con(); |
duke@435 | 931 | // %%% This check should match any shift in [1..31]. |
duke@435 | 932 | if (shift == 16 || shift == 8) { |
duke@435 | 933 | t1 = n1; |
duke@435 | 934 | t2 = t1->in(1); |
duke@435 | 935 | n1 = t2->in(1); |
duke@435 | 936 | n1op = n1->Opcode(); |
duke@435 | 937 | if (shift == 16) { |
duke@435 | 938 | trunc_t = TypeInt::SHORT; |
duke@435 | 939 | } else if (shift == 8) { |
duke@435 | 940 | trunc_t = TypeInt::BYTE; |
duke@435 | 941 | } |
duke@435 | 942 | } |
duke@435 | 943 | } |
duke@435 | 944 | |
duke@435 | 945 | // If (maybe after stripping) it is an AddI, we won: |
duke@435 | 946 | if (n1op == Op_AddI) { |
duke@435 | 947 | *trunc1 = t1; |
duke@435 | 948 | *trunc2 = t2; |
duke@435 | 949 | *trunc_type = trunc_t; |
duke@435 | 950 | return n1; |
duke@435 | 951 | } |
duke@435 | 952 | |
duke@435 | 953 | // failed |
duke@435 | 954 | return NULL; |
duke@435 | 955 | } |
duke@435 | 956 | |
duke@435 | 957 | |
duke@435 | 958 | //------------------------------filtered_type-------------------------------- |
duke@435 | 959 | // Return a type based on condition control flow |
duke@435 | 960 | // A successful return will be a type that is restricted due |
duke@435 | 961 | // to a series of dominating if-tests, such as: |
duke@435 | 962 | // if (i < 10) { |
duke@435 | 963 | // if (i > 0) { |
duke@435 | 964 | // here: "i" type is [1..10) |
duke@435 | 965 | // } |
duke@435 | 966 | // } |
duke@435 | 967 | // or a control flow merge |
duke@435 | 968 | // if (i < 10) { |
duke@435 | 969 | // do { |
duke@435 | 970 | // phi( , ) -- at top of loop type is [min_int..10) |
duke@435 | 971 | // i = ? |
duke@435 | 972 | // } while ( i < 10) |
duke@435 | 973 | // |
duke@435 | 974 | const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) { |
duke@435 | 975 | assert(n && n->bottom_type()->is_int(), "must be int"); |
duke@435 | 976 | const TypeInt* filtered_t = NULL; |
duke@435 | 977 | if (!n->is_Phi()) { |
duke@435 | 978 | assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control"); |
duke@435 | 979 | filtered_t = filtered_type_from_dominators(n, n_ctrl); |
duke@435 | 980 | |
duke@435 | 981 | } else { |
duke@435 | 982 | Node* phi = n->as_Phi(); |
duke@435 | 983 | Node* region = phi->in(0); |
duke@435 | 984 | assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region"); |
duke@435 | 985 | if (region && region != C->top()) { |
duke@435 | 986 | for (uint i = 1; i < phi->req(); i++) { |
duke@435 | 987 | Node* val = phi->in(i); |
duke@435 | 988 | Node* use_c = region->in(i); |
duke@435 | 989 | const TypeInt* val_t = filtered_type_from_dominators(val, use_c); |
duke@435 | 990 | if (val_t != NULL) { |
duke@435 | 991 | if (filtered_t == NULL) { |
duke@435 | 992 | filtered_t = val_t; |
duke@435 | 993 | } else { |
duke@435 | 994 | filtered_t = filtered_t->meet(val_t)->is_int(); |
duke@435 | 995 | } |
duke@435 | 996 | } |
duke@435 | 997 | } |
duke@435 | 998 | } |
duke@435 | 999 | } |
duke@435 | 1000 | const TypeInt* n_t = _igvn.type(n)->is_int(); |
duke@435 | 1001 | if (filtered_t != NULL) { |
duke@435 | 1002 | n_t = n_t->join(filtered_t)->is_int(); |
duke@435 | 1003 | } |
duke@435 | 1004 | return n_t; |
duke@435 | 1005 | } |
duke@435 | 1006 | |
duke@435 | 1007 | |
duke@435 | 1008 | //------------------------------filtered_type_from_dominators-------------------------------- |
duke@435 | 1009 | // Return a possibly more restrictive type for val based on condition control flow of dominators |
duke@435 | 1010 | const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) { |
duke@435 | 1011 | if (val->is_Con()) { |
duke@435 | 1012 | return val->bottom_type()->is_int(); |
duke@435 | 1013 | } |
duke@435 | 1014 | uint if_limit = 10; // Max number of dominating if's visited |
duke@435 | 1015 | const TypeInt* rtn_t = NULL; |
duke@435 | 1016 | |
duke@435 | 1017 | if (use_ctrl && use_ctrl != C->top()) { |
duke@435 | 1018 | Node* val_ctrl = get_ctrl(val); |
duke@435 | 1019 | uint val_dom_depth = dom_depth(val_ctrl); |
duke@435 | 1020 | Node* pred = use_ctrl; |
duke@435 | 1021 | uint if_cnt = 0; |
duke@435 | 1022 | while (if_cnt < if_limit) { |
duke@435 | 1023 | if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) { |
duke@435 | 1024 | if_cnt++; |
never@452 | 1025 | const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred); |
duke@435 | 1026 | if (if_t != NULL) { |
duke@435 | 1027 | if (rtn_t == NULL) { |
duke@435 | 1028 | rtn_t = if_t; |
duke@435 | 1029 | } else { |
duke@435 | 1030 | rtn_t = rtn_t->join(if_t)->is_int(); |
duke@435 | 1031 | } |
duke@435 | 1032 | } |
duke@435 | 1033 | } |
duke@435 | 1034 | pred = idom(pred); |
duke@435 | 1035 | if (pred == NULL || pred == C->top()) { |
duke@435 | 1036 | break; |
duke@435 | 1037 | } |
duke@435 | 1038 | // Stop if going beyond definition block of val |
duke@435 | 1039 | if (dom_depth(pred) < val_dom_depth) { |
duke@435 | 1040 | break; |
duke@435 | 1041 | } |
duke@435 | 1042 | } |
duke@435 | 1043 | } |
duke@435 | 1044 | return rtn_t; |
duke@435 | 1045 | } |
duke@435 | 1046 | |
duke@435 | 1047 | |
duke@435 | 1048 | //------------------------------dump_spec-------------------------------------- |
duke@435 | 1049 | // Dump special per-node info |
duke@435 | 1050 | #ifndef PRODUCT |
duke@435 | 1051 | void CountedLoopEndNode::dump_spec(outputStream *st) const { |
duke@435 | 1052 | if( in(TestValue)->is_Bool() ) { |
duke@435 | 1053 | BoolTest bt( test_trip()); // Added this for g++. |
duke@435 | 1054 | |
duke@435 | 1055 | st->print("["); |
duke@435 | 1056 | bt.dump_on(st); |
duke@435 | 1057 | st->print("]"); |
duke@435 | 1058 | } |
duke@435 | 1059 | st->print(" "); |
duke@435 | 1060 | IfNode::dump_spec(st); |
duke@435 | 1061 | } |
duke@435 | 1062 | #endif |
duke@435 | 1063 | |
duke@435 | 1064 | //============================================================================= |
duke@435 | 1065 | //------------------------------is_member-------------------------------------- |
duke@435 | 1066 | // Is 'l' a member of 'this'? |
duke@435 | 1067 | int IdealLoopTree::is_member( const IdealLoopTree *l ) const { |
duke@435 | 1068 | while( l->_nest > _nest ) l = l->_parent; |
duke@435 | 1069 | return l == this; |
duke@435 | 1070 | } |
duke@435 | 1071 | |
duke@435 | 1072 | //------------------------------set_nest--------------------------------------- |
duke@435 | 1073 | // Set loop tree nesting depth. Accumulate _has_call bits. |
duke@435 | 1074 | int IdealLoopTree::set_nest( uint depth ) { |
duke@435 | 1075 | _nest = depth; |
duke@435 | 1076 | int bits = _has_call; |
duke@435 | 1077 | if( _child ) bits |= _child->set_nest(depth+1); |
duke@435 | 1078 | if( bits ) _has_call = 1; |
duke@435 | 1079 | if( _next ) bits |= _next ->set_nest(depth ); |
duke@435 | 1080 | return bits; |
duke@435 | 1081 | } |
duke@435 | 1082 | |
duke@435 | 1083 | //------------------------------split_fall_in---------------------------------- |
duke@435 | 1084 | // Split out multiple fall-in edges from the loop header. Move them to a |
duke@435 | 1085 | // private RegionNode before the loop. This becomes the loop landing pad. |
duke@435 | 1086 | void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) { |
duke@435 | 1087 | PhaseIterGVN &igvn = phase->_igvn; |
duke@435 | 1088 | uint i; |
duke@435 | 1089 | |
duke@435 | 1090 | // Make a new RegionNode to be the landing pad. |
duke@435 | 1091 | Node *landing_pad = new (phase->C, fall_in_cnt+1) RegionNode( fall_in_cnt+1 ); |
duke@435 | 1092 | phase->set_loop(landing_pad,_parent); |
duke@435 | 1093 | // Gather all the fall-in control paths into the landing pad |
duke@435 | 1094 | uint icnt = fall_in_cnt; |
duke@435 | 1095 | uint oreq = _head->req(); |
duke@435 | 1096 | for( i = oreq-1; i>0; i-- ) |
duke@435 | 1097 | if( !phase->is_member( this, _head->in(i) ) ) |
duke@435 | 1098 | landing_pad->set_req(icnt--,_head->in(i)); |
duke@435 | 1099 | |
duke@435 | 1100 | // Peel off PhiNode edges as well |
duke@435 | 1101 | for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 1102 | Node *oj = _head->fast_out(j); |
duke@435 | 1103 | if( oj->is_Phi() ) { |
duke@435 | 1104 | PhiNode* old_phi = oj->as_Phi(); |
duke@435 | 1105 | assert( old_phi->region() == _head, "" ); |
duke@435 | 1106 | igvn.hash_delete(old_phi); // Yank from hash before hacking edges |
duke@435 | 1107 | Node *p = PhiNode::make_blank(landing_pad, old_phi); |
duke@435 | 1108 | uint icnt = fall_in_cnt; |
duke@435 | 1109 | for( i = oreq-1; i>0; i-- ) { |
duke@435 | 1110 | if( !phase->is_member( this, _head->in(i) ) ) { |
duke@435 | 1111 | p->init_req(icnt--, old_phi->in(i)); |
duke@435 | 1112 | // Go ahead and clean out old edges from old phi |
duke@435 | 1113 | old_phi->del_req(i); |
duke@435 | 1114 | } |
duke@435 | 1115 | } |
duke@435 | 1116 | // Search for CSE's here, because ZKM.jar does a lot of |
duke@435 | 1117 | // loop hackery and we need to be a little incremental |
duke@435 | 1118 | // with the CSE to avoid O(N^2) node blow-up. |
duke@435 | 1119 | Node *p2 = igvn.hash_find_insert(p); // Look for a CSE |
duke@435 | 1120 | if( p2 ) { // Found CSE |
duke@435 | 1121 | p->destruct(); // Recover useless new node |
duke@435 | 1122 | p = p2; // Use old node |
duke@435 | 1123 | } else { |
duke@435 | 1124 | igvn.register_new_node_with_optimizer(p, old_phi); |
duke@435 | 1125 | } |
duke@435 | 1126 | // Make old Phi refer to new Phi. |
duke@435 | 1127 | old_phi->add_req(p); |
duke@435 | 1128 | // Check for the special case of making the old phi useless and |
duke@435 | 1129 | // disappear it. In JavaGrande I have a case where this useless |
duke@435 | 1130 | // Phi is the loop limit and prevents recognizing a CountedLoop |
duke@435 | 1131 | // which in turn prevents removing an empty loop. |
duke@435 | 1132 | Node *id_old_phi = old_phi->Identity( &igvn ); |
duke@435 | 1133 | if( id_old_phi != old_phi ) { // Found a simple identity? |
kvn@1976 | 1134 | // Note that I cannot call 'replace_node' here, because |
duke@435 | 1135 | // that will yank the edge from old_phi to the Region and |
duke@435 | 1136 | // I'm mid-iteration over the Region's uses. |
duke@435 | 1137 | for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) { |
duke@435 | 1138 | Node* use = old_phi->last_out(i); |
kvn@3847 | 1139 | igvn.rehash_node_delayed(use); |
duke@435 | 1140 | uint uses_found = 0; |
duke@435 | 1141 | for (uint j = 0; j < use->len(); j++) { |
duke@435 | 1142 | if (use->in(j) == old_phi) { |
duke@435 | 1143 | if (j < use->req()) use->set_req (j, id_old_phi); |
duke@435 | 1144 | else use->set_prec(j, id_old_phi); |
duke@435 | 1145 | uses_found++; |
duke@435 | 1146 | } |
duke@435 | 1147 | } |
duke@435 | 1148 | i -= uses_found; // we deleted 1 or more copies of this edge |
duke@435 | 1149 | } |
duke@435 | 1150 | } |
duke@435 | 1151 | igvn._worklist.push(old_phi); |
duke@435 | 1152 | } |
duke@435 | 1153 | } |
duke@435 | 1154 | // Finally clean out the fall-in edges from the RegionNode |
duke@435 | 1155 | for( i = oreq-1; i>0; i-- ) { |
duke@435 | 1156 | if( !phase->is_member( this, _head->in(i) ) ) { |
duke@435 | 1157 | _head->del_req(i); |
duke@435 | 1158 | } |
duke@435 | 1159 | } |
duke@435 | 1160 | // Transform landing pad |
duke@435 | 1161 | igvn.register_new_node_with_optimizer(landing_pad, _head); |
duke@435 | 1162 | // Insert landing pad into the header |
duke@435 | 1163 | _head->add_req(landing_pad); |
duke@435 | 1164 | } |
duke@435 | 1165 | |
duke@435 | 1166 | //------------------------------split_outer_loop------------------------------- |
duke@435 | 1167 | // Split out the outermost loop from this shared header. |
duke@435 | 1168 | void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) { |
duke@435 | 1169 | PhaseIterGVN &igvn = phase->_igvn; |
duke@435 | 1170 | |
duke@435 | 1171 | // Find index of outermost loop; it should also be my tail. |
duke@435 | 1172 | uint outer_idx = 1; |
duke@435 | 1173 | while( _head->in(outer_idx) != _tail ) outer_idx++; |
duke@435 | 1174 | |
duke@435 | 1175 | // Make a LoopNode for the outermost loop. |
duke@435 | 1176 | Node *ctl = _head->in(LoopNode::EntryControl); |
duke@435 | 1177 | Node *outer = new (phase->C, 3) LoopNode( ctl, _head->in(outer_idx) ); |
duke@435 | 1178 | outer = igvn.register_new_node_with_optimizer(outer, _head); |
duke@435 | 1179 | phase->set_created_loop_node(); |
kvn@2727 | 1180 | |
duke@435 | 1181 | // Outermost loop falls into '_head' loop |
kvn@3043 | 1182 | _head->set_req(LoopNode::EntryControl, outer); |
duke@435 | 1183 | _head->del_req(outer_idx); |
duke@435 | 1184 | // Split all the Phis up between '_head' loop and 'outer' loop. |
duke@435 | 1185 | for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 1186 | Node *out = _head->fast_out(j); |
duke@435 | 1187 | if( out->is_Phi() ) { |
duke@435 | 1188 | PhiNode *old_phi = out->as_Phi(); |
duke@435 | 1189 | assert( old_phi->region() == _head, "" ); |
duke@435 | 1190 | Node *phi = PhiNode::make_blank(outer, old_phi); |
duke@435 | 1191 | phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl)); |
duke@435 | 1192 | phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx)); |
duke@435 | 1193 | phi = igvn.register_new_node_with_optimizer(phi, old_phi); |
duke@435 | 1194 | // Make old Phi point to new Phi on the fall-in path |
kvn@3847 | 1195 | igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi); |
duke@435 | 1196 | old_phi->del_req(outer_idx); |
duke@435 | 1197 | } |
duke@435 | 1198 | } |
duke@435 | 1199 | |
duke@435 | 1200 | // Use the new loop head instead of the old shared one |
duke@435 | 1201 | _head = outer; |
duke@435 | 1202 | phase->set_loop(_head, this); |
duke@435 | 1203 | } |
duke@435 | 1204 | |
duke@435 | 1205 | //------------------------------fix_parent------------------------------------- |
duke@435 | 1206 | static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) { |
duke@435 | 1207 | loop->_parent = parent; |
duke@435 | 1208 | if( loop->_child ) fix_parent( loop->_child, loop ); |
duke@435 | 1209 | if( loop->_next ) fix_parent( loop->_next , parent ); |
duke@435 | 1210 | } |
duke@435 | 1211 | |
duke@435 | 1212 | //------------------------------estimate_path_freq----------------------------- |
duke@435 | 1213 | static float estimate_path_freq( Node *n ) { |
duke@435 | 1214 | // Try to extract some path frequency info |
duke@435 | 1215 | IfNode *iff; |
duke@435 | 1216 | for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests |
duke@435 | 1217 | uint nop = n->Opcode(); |
duke@435 | 1218 | if( nop == Op_SafePoint ) { // Skip any safepoint |
duke@435 | 1219 | n = n->in(0); |
duke@435 | 1220 | continue; |
duke@435 | 1221 | } |
duke@435 | 1222 | if( nop == Op_CatchProj ) { // Get count from a prior call |
duke@435 | 1223 | // Assume call does not always throw exceptions: means the call-site |
duke@435 | 1224 | // count is also the frequency of the fall-through path. |
duke@435 | 1225 | assert( n->is_CatchProj(), "" ); |
duke@435 | 1226 | if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index ) |
duke@435 | 1227 | return 0.0f; // Assume call exception path is rare |
duke@435 | 1228 | Node *call = n->in(0)->in(0)->in(0); |
duke@435 | 1229 | assert( call->is_Call(), "expect a call here" ); |
duke@435 | 1230 | const JVMState *jvms = ((CallNode*)call)->jvms(); |
duke@435 | 1231 | ciMethodData* methodData = jvms->method()->method_data(); |
duke@435 | 1232 | if (!methodData->is_mature()) return 0.0f; // No call-site data |
duke@435 | 1233 | ciProfileData* data = methodData->bci_to_data(jvms->bci()); |
duke@435 | 1234 | if ((data == NULL) || !data->is_CounterData()) { |
duke@435 | 1235 | // no call profile available, try call's control input |
duke@435 | 1236 | n = n->in(0); |
duke@435 | 1237 | continue; |
duke@435 | 1238 | } |
duke@435 | 1239 | return data->as_CounterData()->count()/FreqCountInvocations; |
duke@435 | 1240 | } |
duke@435 | 1241 | // See if there's a gating IF test |
duke@435 | 1242 | Node *n_c = n->in(0); |
duke@435 | 1243 | if( !n_c->is_If() ) break; // No estimate available |
duke@435 | 1244 | iff = n_c->as_If(); |
duke@435 | 1245 | if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count? |
duke@435 | 1246 | // Compute how much count comes on this path |
duke@435 | 1247 | return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt; |
duke@435 | 1248 | // Have no count info. Skip dull uncommon-trap like branches. |
duke@435 | 1249 | if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) || |
duke@435 | 1250 | (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) ) |
duke@435 | 1251 | break; |
duke@435 | 1252 | // Skip through never-taken branch; look for a real loop exit. |
duke@435 | 1253 | n = iff->in(0); |
duke@435 | 1254 | } |
duke@435 | 1255 | return 0.0f; // No estimate available |
duke@435 | 1256 | } |
duke@435 | 1257 | |
duke@435 | 1258 | //------------------------------merge_many_backedges--------------------------- |
duke@435 | 1259 | // Merge all the backedges from the shared header into a private Region. |
duke@435 | 1260 | // Feed that region as the one backedge to this loop. |
duke@435 | 1261 | void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) { |
duke@435 | 1262 | uint i; |
duke@435 | 1263 | |
duke@435 | 1264 | // Scan for the top 2 hottest backedges |
duke@435 | 1265 | float hotcnt = 0.0f; |
duke@435 | 1266 | float warmcnt = 0.0f; |
duke@435 | 1267 | uint hot_idx = 0; |
duke@435 | 1268 | // Loop starts at 2 because slot 1 is the fall-in path |
duke@435 | 1269 | for( i = 2; i < _head->req(); i++ ) { |
duke@435 | 1270 | float cnt = estimate_path_freq(_head->in(i)); |
duke@435 | 1271 | if( cnt > hotcnt ) { // Grab hottest path |
duke@435 | 1272 | warmcnt = hotcnt; |
duke@435 | 1273 | hotcnt = cnt; |
duke@435 | 1274 | hot_idx = i; |
duke@435 | 1275 | } else if( cnt > warmcnt ) { // And 2nd hottest path |
duke@435 | 1276 | warmcnt = cnt; |
duke@435 | 1277 | } |
duke@435 | 1278 | } |
duke@435 | 1279 | |
duke@435 | 1280 | // See if the hottest backedge is worthy of being an inner loop |
duke@435 | 1281 | // by being much hotter than the next hottest backedge. |
duke@435 | 1282 | if( hotcnt <= 0.0001 || |
duke@435 | 1283 | hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge |
duke@435 | 1284 | |
duke@435 | 1285 | // Peel out the backedges into a private merge point; peel |
duke@435 | 1286 | // them all except optionally hot_idx. |
duke@435 | 1287 | PhaseIterGVN &igvn = phase->_igvn; |
duke@435 | 1288 | |
duke@435 | 1289 | Node *hot_tail = NULL; |
duke@435 | 1290 | // Make a Region for the merge point |
duke@435 | 1291 | Node *r = new (phase->C, 1) RegionNode(1); |
duke@435 | 1292 | for( i = 2; i < _head->req(); i++ ) { |
duke@435 | 1293 | if( i != hot_idx ) |
duke@435 | 1294 | r->add_req( _head->in(i) ); |
duke@435 | 1295 | else hot_tail = _head->in(i); |
duke@435 | 1296 | } |
duke@435 | 1297 | igvn.register_new_node_with_optimizer(r, _head); |
duke@435 | 1298 | // Plug region into end of loop _head, followed by hot_tail |
duke@435 | 1299 | while( _head->req() > 3 ) _head->del_req( _head->req()-1 ); |
duke@435 | 1300 | _head->set_req(2, r); |
duke@435 | 1301 | if( hot_idx ) _head->add_req(hot_tail); |
duke@435 | 1302 | |
duke@435 | 1303 | // Split all the Phis up between '_head' loop and the Region 'r' |
duke@435 | 1304 | for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 1305 | Node *out = _head->fast_out(j); |
duke@435 | 1306 | if( out->is_Phi() ) { |
duke@435 | 1307 | PhiNode* n = out->as_Phi(); |
duke@435 | 1308 | igvn.hash_delete(n); // Delete from hash before hacking edges |
duke@435 | 1309 | Node *hot_phi = NULL; |
duke@435 | 1310 | Node *phi = new (phase->C, r->req()) PhiNode(r, n->type(), n->adr_type()); |
duke@435 | 1311 | // Check all inputs for the ones to peel out |
duke@435 | 1312 | uint j = 1; |
duke@435 | 1313 | for( uint i = 2; i < n->req(); i++ ) { |
duke@435 | 1314 | if( i != hot_idx ) |
duke@435 | 1315 | phi->set_req( j++, n->in(i) ); |
duke@435 | 1316 | else hot_phi = n->in(i); |
duke@435 | 1317 | } |
duke@435 | 1318 | // Register the phi but do not transform until whole place transforms |
duke@435 | 1319 | igvn.register_new_node_with_optimizer(phi, n); |
duke@435 | 1320 | // Add the merge phi to the old Phi |
duke@435 | 1321 | while( n->req() > 3 ) n->del_req( n->req()-1 ); |
duke@435 | 1322 | n->set_req(2, phi); |
duke@435 | 1323 | if( hot_idx ) n->add_req(hot_phi); |
duke@435 | 1324 | } |
duke@435 | 1325 | } |
duke@435 | 1326 | |
duke@435 | 1327 | |
duke@435 | 1328 | // Insert a new IdealLoopTree inserted below me. Turn it into a clone |
duke@435 | 1329 | // of self loop tree. Turn self into a loop headed by _head and with |
duke@435 | 1330 | // tail being the new merge point. |
duke@435 | 1331 | IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail ); |
duke@435 | 1332 | phase->set_loop(_tail,ilt); // Adjust tail |
duke@435 | 1333 | _tail = r; // Self's tail is new merge point |
duke@435 | 1334 | phase->set_loop(r,this); |
duke@435 | 1335 | ilt->_child = _child; // New guy has my children |
duke@435 | 1336 | _child = ilt; // Self has new guy as only child |
duke@435 | 1337 | ilt->_parent = this; // new guy has self for parent |
duke@435 | 1338 | ilt->_nest = _nest; // Same nesting depth (for now) |
duke@435 | 1339 | |
duke@435 | 1340 | // Starting with 'ilt', look for child loop trees using the same shared |
duke@435 | 1341 | // header. Flatten these out; they will no longer be loops in the end. |
duke@435 | 1342 | IdealLoopTree **pilt = &_child; |
duke@435 | 1343 | while( ilt ) { |
duke@435 | 1344 | if( ilt->_head == _head ) { |
duke@435 | 1345 | uint i; |
duke@435 | 1346 | for( i = 2; i < _head->req(); i++ ) |
duke@435 | 1347 | if( _head->in(i) == ilt->_tail ) |
duke@435 | 1348 | break; // Still a loop |
duke@435 | 1349 | if( i == _head->req() ) { // No longer a loop |
duke@435 | 1350 | // Flatten ilt. Hang ilt's "_next" list from the end of |
duke@435 | 1351 | // ilt's '_child' list. Move the ilt's _child up to replace ilt. |
duke@435 | 1352 | IdealLoopTree **cp = &ilt->_child; |
duke@435 | 1353 | while( *cp ) cp = &(*cp)->_next; // Find end of child list |
duke@435 | 1354 | *cp = ilt->_next; // Hang next list at end of child list |
duke@435 | 1355 | *pilt = ilt->_child; // Move child up to replace ilt |
duke@435 | 1356 | ilt->_head = NULL; // Flag as a loop UNIONED into parent |
duke@435 | 1357 | ilt = ilt->_child; // Repeat using new ilt |
duke@435 | 1358 | continue; // do not advance over ilt->_child |
duke@435 | 1359 | } |
duke@435 | 1360 | assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" ); |
duke@435 | 1361 | phase->set_loop(_head,ilt); |
duke@435 | 1362 | } |
duke@435 | 1363 | pilt = &ilt->_child; // Advance to next |
duke@435 | 1364 | ilt = *pilt; |
duke@435 | 1365 | } |
duke@435 | 1366 | |
duke@435 | 1367 | if( _child ) fix_parent( _child, this ); |
duke@435 | 1368 | } |
duke@435 | 1369 | |
duke@435 | 1370 | //------------------------------beautify_loops--------------------------------- |
duke@435 | 1371 | // Split shared headers and insert loop landing pads. |
duke@435 | 1372 | // Insert a LoopNode to replace the RegionNode. |
duke@435 | 1373 | // Return TRUE if loop tree is structurally changed. |
duke@435 | 1374 | bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) { |
duke@435 | 1375 | bool result = false; |
duke@435 | 1376 | // Cache parts in locals for easy |
duke@435 | 1377 | PhaseIterGVN &igvn = phase->_igvn; |
duke@435 | 1378 | |
duke@435 | 1379 | igvn.hash_delete(_head); // Yank from hash before hacking edges |
duke@435 | 1380 | |
duke@435 | 1381 | // Check for multiple fall-in paths. Peel off a landing pad if need be. |
duke@435 | 1382 | int fall_in_cnt = 0; |
duke@435 | 1383 | for( uint i = 1; i < _head->req(); i++ ) |
duke@435 | 1384 | if( !phase->is_member( this, _head->in(i) ) ) |
duke@435 | 1385 | fall_in_cnt++; |
duke@435 | 1386 | assert( fall_in_cnt, "at least 1 fall-in path" ); |
duke@435 | 1387 | if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins |
duke@435 | 1388 | split_fall_in( phase, fall_in_cnt ); |
duke@435 | 1389 | |
duke@435 | 1390 | // Swap inputs to the _head and all Phis to move the fall-in edge to |
duke@435 | 1391 | // the left. |
duke@435 | 1392 | fall_in_cnt = 1; |
duke@435 | 1393 | while( phase->is_member( this, _head->in(fall_in_cnt) ) ) |
duke@435 | 1394 | fall_in_cnt++; |
duke@435 | 1395 | if( fall_in_cnt > 1 ) { |
duke@435 | 1396 | // Since I am just swapping inputs I do not need to update def-use info |
duke@435 | 1397 | Node *tmp = _head->in(1); |
duke@435 | 1398 | _head->set_req( 1, _head->in(fall_in_cnt) ); |
duke@435 | 1399 | _head->set_req( fall_in_cnt, tmp ); |
duke@435 | 1400 | // Swap also all Phis |
duke@435 | 1401 | for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) { |
duke@435 | 1402 | Node* phi = _head->fast_out(i); |
duke@435 | 1403 | if( phi->is_Phi() ) { |
duke@435 | 1404 | igvn.hash_delete(phi); // Yank from hash before hacking edges |
duke@435 | 1405 | tmp = phi->in(1); |
duke@435 | 1406 | phi->set_req( 1, phi->in(fall_in_cnt) ); |
duke@435 | 1407 | phi->set_req( fall_in_cnt, tmp ); |
duke@435 | 1408 | } |
duke@435 | 1409 | } |
duke@435 | 1410 | } |
duke@435 | 1411 | assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" ); |
duke@435 | 1412 | assert( phase->is_member( this, _head->in(2) ), "right edge is loop" ); |
duke@435 | 1413 | |
duke@435 | 1414 | // If I am a shared header (multiple backedges), peel off the many |
duke@435 | 1415 | // backedges into a private merge point and use the merge point as |
duke@435 | 1416 | // the one true backedge. |
duke@435 | 1417 | if( _head->req() > 3 ) { |
kvn@2727 | 1418 | // Merge the many backedges into a single backedge but leave |
kvn@2727 | 1419 | // the hottest backedge as separate edge for the following peel. |
duke@435 | 1420 | merge_many_backedges( phase ); |
duke@435 | 1421 | result = true; |
duke@435 | 1422 | } |
duke@435 | 1423 | |
kvn@2727 | 1424 | // If I have one hot backedge, peel off myself loop. |
duke@435 | 1425 | // I better be the outermost loop. |
duke@435 | 1426 | if( _head->req() > 3 ) { |
duke@435 | 1427 | split_outer_loop( phase ); |
duke@435 | 1428 | result = true; |
duke@435 | 1429 | |
duke@435 | 1430 | } else if( !_head->is_Loop() && !_irreducible ) { |
duke@435 | 1431 | // Make a new LoopNode to replace the old loop head |
duke@435 | 1432 | Node *l = new (phase->C, 3) LoopNode( _head->in(1), _head->in(2) ); |
duke@435 | 1433 | l = igvn.register_new_node_with_optimizer(l, _head); |
duke@435 | 1434 | phase->set_created_loop_node(); |
duke@435 | 1435 | // Go ahead and replace _head |
kvn@1976 | 1436 | phase->_igvn.replace_node( _head, l ); |
duke@435 | 1437 | _head = l; |
duke@435 | 1438 | phase->set_loop(_head, this); |
duke@435 | 1439 | } |
duke@435 | 1440 | |
duke@435 | 1441 | // Now recursively beautify nested loops |
duke@435 | 1442 | if( _child ) result |= _child->beautify_loops( phase ); |
duke@435 | 1443 | if( _next ) result |= _next ->beautify_loops( phase ); |
duke@435 | 1444 | return result; |
duke@435 | 1445 | } |
duke@435 | 1446 | |
duke@435 | 1447 | //------------------------------allpaths_check_safepts---------------------------- |
duke@435 | 1448 | // Allpaths backwards scan from loop tail, terminating each path at first safepoint |
duke@435 | 1449 | // encountered. Helper for check_safepts. |
duke@435 | 1450 | void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) { |
duke@435 | 1451 | assert(stack.size() == 0, "empty stack"); |
duke@435 | 1452 | stack.push(_tail); |
duke@435 | 1453 | visited.Clear(); |
duke@435 | 1454 | visited.set(_tail->_idx); |
duke@435 | 1455 | while (stack.size() > 0) { |
duke@435 | 1456 | Node* n = stack.pop(); |
duke@435 | 1457 | if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { |
duke@435 | 1458 | // Terminate this path |
duke@435 | 1459 | } else if (n->Opcode() == Op_SafePoint) { |
duke@435 | 1460 | if (_phase->get_loop(n) != this) { |
duke@435 | 1461 | if (_required_safept == NULL) _required_safept = new Node_List(); |
duke@435 | 1462 | _required_safept->push(n); // save the one closest to the tail |
duke@435 | 1463 | } |
duke@435 | 1464 | // Terminate this path |
duke@435 | 1465 | } else { |
duke@435 | 1466 | uint start = n->is_Region() ? 1 : 0; |
duke@435 | 1467 | uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1; |
duke@435 | 1468 | for (uint i = start; i < end; i++) { |
duke@435 | 1469 | Node* in = n->in(i); |
duke@435 | 1470 | assert(in->is_CFG(), "must be"); |
duke@435 | 1471 | if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) { |
duke@435 | 1472 | stack.push(in); |
duke@435 | 1473 | } |
duke@435 | 1474 | } |
duke@435 | 1475 | } |
duke@435 | 1476 | } |
duke@435 | 1477 | } |
duke@435 | 1478 | |
duke@435 | 1479 | //------------------------------check_safepts---------------------------- |
duke@435 | 1480 | // Given dominators, try to find loops with calls that must always be |
duke@435 | 1481 | // executed (call dominates loop tail). These loops do not need non-call |
duke@435 | 1482 | // safepoints (ncsfpt). |
duke@435 | 1483 | // |
duke@435 | 1484 | // A complication is that a safepoint in a inner loop may be needed |
duke@435 | 1485 | // by an outer loop. In the following, the inner loop sees it has a |
duke@435 | 1486 | // call (block 3) on every path from the head (block 2) to the |
duke@435 | 1487 | // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint) |
duke@435 | 1488 | // in block 2, _but_ this leaves the outer loop without a safepoint. |
duke@435 | 1489 | // |
duke@435 | 1490 | // entry 0 |
duke@435 | 1491 | // | |
duke@435 | 1492 | // v |
duke@435 | 1493 | // outer 1,2 +->1 |
duke@435 | 1494 | // | | |
duke@435 | 1495 | // | v |
duke@435 | 1496 | // | 2<---+ ncsfpt in 2 |
duke@435 | 1497 | // |_/|\ | |
duke@435 | 1498 | // | v | |
duke@435 | 1499 | // inner 2,3 / 3 | call in 3 |
duke@435 | 1500 | // / | | |
duke@435 | 1501 | // v +--+ |
duke@435 | 1502 | // exit 4 |
duke@435 | 1503 | // |
duke@435 | 1504 | // |
duke@435 | 1505 | // This method creates a list (_required_safept) of ncsfpt nodes that must |
duke@435 | 1506 | // be protected is created for each loop. When a ncsfpt maybe deleted, it |
duke@435 | 1507 | // is first looked for in the lists for the outer loops of the current loop. |
duke@435 | 1508 | // |
duke@435 | 1509 | // The insights into the problem: |
duke@435 | 1510 | // A) counted loops are okay |
duke@435 | 1511 | // B) innermost loops are okay (only an inner loop can delete |
duke@435 | 1512 | // a ncsfpt needed by an outer loop) |
duke@435 | 1513 | // C) a loop is immune from an inner loop deleting a safepoint |
duke@435 | 1514 | // if the loop has a call on the idom-path |
duke@435 | 1515 | // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the |
duke@435 | 1516 | // idom-path that is not in a nested loop |
duke@435 | 1517 | // E) otherwise, an ncsfpt on the idom-path that is nested in an inner |
duke@435 | 1518 | // loop needs to be prevented from deletion by an inner loop |
duke@435 | 1519 | // |
duke@435 | 1520 | // There are two analyses: |
duke@435 | 1521 | // 1) The first, and cheaper one, scans the loop body from |
duke@435 | 1522 | // tail to head following the idom (immediate dominator) |
duke@435 | 1523 | // chain, looking for the cases (C,D,E) above. |
duke@435 | 1524 | // Since inner loops are scanned before outer loops, there is summary |
duke@435 | 1525 | // information about inner loops. Inner loops can be skipped over |
duke@435 | 1526 | // when the tail of an inner loop is encountered. |
duke@435 | 1527 | // |
duke@435 | 1528 | // 2) The second, invoked if the first fails to find a call or ncsfpt on |
duke@435 | 1529 | // the idom path (which is rare), scans all predecessor control paths |
duke@435 | 1530 | // from the tail to the head, terminating a path when a call or sfpt |
duke@435 | 1531 | // is encountered, to find the ncsfpt's that are closest to the tail. |
duke@435 | 1532 | // |
duke@435 | 1533 | void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) { |
duke@435 | 1534 | // Bottom up traversal |
duke@435 | 1535 | IdealLoopTree* ch = _child; |
kvn@4023 | 1536 | if (_child) _child->check_safepts(visited, stack); |
kvn@4023 | 1537 | if (_next) _next ->check_safepts(visited, stack); |
duke@435 | 1538 | |
duke@435 | 1539 | if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) { |
duke@435 | 1540 | bool has_call = false; // call on dom-path |
duke@435 | 1541 | bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth |
duke@435 | 1542 | Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth |
duke@435 | 1543 | // Scan the dom-path nodes from tail to head |
duke@435 | 1544 | for (Node* n = tail(); n != _head; n = _phase->idom(n)) { |
duke@435 | 1545 | if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { |
duke@435 | 1546 | has_call = true; |
duke@435 | 1547 | _has_sfpt = 1; // Then no need for a safept! |
duke@435 | 1548 | break; |
duke@435 | 1549 | } else if (n->Opcode() == Op_SafePoint) { |
duke@435 | 1550 | if (_phase->get_loop(n) == this) { |
duke@435 | 1551 | has_local_ncsfpt = true; |
duke@435 | 1552 | break; |
duke@435 | 1553 | } |
duke@435 | 1554 | if (nonlocal_ncsfpt == NULL) { |
duke@435 | 1555 | nonlocal_ncsfpt = n; // save the one closest to the tail |
duke@435 | 1556 | } |
duke@435 | 1557 | } else { |
duke@435 | 1558 | IdealLoopTree* nlpt = _phase->get_loop(n); |
duke@435 | 1559 | if (this != nlpt) { |
duke@435 | 1560 | // If at an inner loop tail, see if the inner loop has already |
duke@435 | 1561 | // recorded seeing a call on the dom-path (and stop.) If not, |
duke@435 | 1562 | // jump to the head of the inner loop. |
duke@435 | 1563 | assert(is_member(nlpt), "nested loop"); |
duke@435 | 1564 | Node* tail = nlpt->_tail; |
duke@435 | 1565 | if (tail->in(0)->is_If()) tail = tail->in(0); |
duke@435 | 1566 | if (n == tail) { |
duke@435 | 1567 | // If inner loop has call on dom-path, so does outer loop |
duke@435 | 1568 | if (nlpt->_has_sfpt) { |
duke@435 | 1569 | has_call = true; |
duke@435 | 1570 | _has_sfpt = 1; |
duke@435 | 1571 | break; |
duke@435 | 1572 | } |
duke@435 | 1573 | // Skip to head of inner loop |
duke@435 | 1574 | assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head"); |
duke@435 | 1575 | n = nlpt->_head; |
duke@435 | 1576 | } |
duke@435 | 1577 | } |
duke@435 | 1578 | } |
duke@435 | 1579 | } |
duke@435 | 1580 | // Record safept's that this loop needs preserved when an |
duke@435 | 1581 | // inner loop attempts to delete it's safepoints. |
duke@435 | 1582 | if (_child != NULL && !has_call && !has_local_ncsfpt) { |
duke@435 | 1583 | if (nonlocal_ncsfpt != NULL) { |
duke@435 | 1584 | if (_required_safept == NULL) _required_safept = new Node_List(); |
duke@435 | 1585 | _required_safept->push(nonlocal_ncsfpt); |
duke@435 | 1586 | } else { |
duke@435 | 1587 | // Failed to find a suitable safept on the dom-path. Now use |
duke@435 | 1588 | // an all paths walk from tail to head, looking for safepoints to preserve. |
duke@435 | 1589 | allpaths_check_safepts(visited, stack); |
duke@435 | 1590 | } |
duke@435 | 1591 | } |
duke@435 | 1592 | } |
duke@435 | 1593 | } |
duke@435 | 1594 | |
duke@435 | 1595 | //---------------------------is_deleteable_safept---------------------------- |
duke@435 | 1596 | // Is safept not required by an outer loop? |
duke@435 | 1597 | bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) { |
duke@435 | 1598 | assert(sfpt->Opcode() == Op_SafePoint, ""); |
duke@435 | 1599 | IdealLoopTree* lp = get_loop(sfpt)->_parent; |
duke@435 | 1600 | while (lp != NULL) { |
duke@435 | 1601 | Node_List* sfpts = lp->_required_safept; |
duke@435 | 1602 | if (sfpts != NULL) { |
duke@435 | 1603 | for (uint i = 0; i < sfpts->size(); i++) { |
duke@435 | 1604 | if (sfpt == sfpts->at(i)) |
duke@435 | 1605 | return false; |
duke@435 | 1606 | } |
duke@435 | 1607 | } |
duke@435 | 1608 | lp = lp->_parent; |
duke@435 | 1609 | } |
duke@435 | 1610 | return true; |
duke@435 | 1611 | } |
duke@435 | 1612 | |
kvn@2665 | 1613 | //---------------------------replace_parallel_iv------------------------------- |
kvn@2665 | 1614 | // Replace parallel induction variable (parallel to trip counter) |
kvn@2665 | 1615 | void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) { |
kvn@2665 | 1616 | assert(loop->_head->is_CountedLoop(), ""); |
kvn@2665 | 1617 | CountedLoopNode *cl = loop->_head->as_CountedLoop(); |
kvn@3048 | 1618 | if (!cl->is_valid_counted_loop()) |
kvn@3048 | 1619 | return; // skip malformed counted loop |
kvn@2665 | 1620 | Node *incr = cl->incr(); |
kvn@2665 | 1621 | if (incr == NULL) |
kvn@2665 | 1622 | return; // Dead loop? |
kvn@2665 | 1623 | Node *init = cl->init_trip(); |
kvn@2665 | 1624 | Node *phi = cl->phi(); |
kvn@2665 | 1625 | int stride_con = cl->stride_con(); |
kvn@2665 | 1626 | |
kvn@2665 | 1627 | // Visit all children, looking for Phis |
kvn@2665 | 1628 | for (DUIterator i = cl->outs(); cl->has_out(i); i++) { |
kvn@2665 | 1629 | Node *out = cl->out(i); |
kvn@2665 | 1630 | // Look for other phis (secondary IVs). Skip dead ones |
kvn@2665 | 1631 | if (!out->is_Phi() || out == phi || !has_node(out)) |
kvn@2665 | 1632 | continue; |
kvn@2665 | 1633 | PhiNode* phi2 = out->as_Phi(); |
kvn@2665 | 1634 | Node *incr2 = phi2->in( LoopNode::LoopBackControl ); |
kvn@2665 | 1635 | // Look for induction variables of the form: X += constant |
kvn@2665 | 1636 | if (phi2->region() != loop->_head || |
kvn@2665 | 1637 | incr2->req() != 3 || |
kvn@2665 | 1638 | incr2->in(1) != phi2 || |
kvn@2665 | 1639 | incr2 == incr || |
kvn@2665 | 1640 | incr2->Opcode() != Op_AddI || |
kvn@2665 | 1641 | !incr2->in(2)->is_Con()) |
kvn@2665 | 1642 | continue; |
kvn@2665 | 1643 | |
kvn@2665 | 1644 | // Check for parallel induction variable (parallel to trip counter) |
kvn@2665 | 1645 | // via an affine function. In particular, count-down loops with |
kvn@2665 | 1646 | // count-up array indices are common. We only RCE references off |
kvn@2665 | 1647 | // the trip-counter, so we need to convert all these to trip-counter |
kvn@2665 | 1648 | // expressions. |
kvn@2665 | 1649 | Node *init2 = phi2->in( LoopNode::EntryControl ); |
kvn@2665 | 1650 | int stride_con2 = incr2->in(2)->get_int(); |
kvn@2665 | 1651 | |
kvn@2665 | 1652 | // The general case here gets a little tricky. We want to find the |
kvn@2665 | 1653 | // GCD of all possible parallel IV's and make a new IV using this |
kvn@2665 | 1654 | // GCD for the loop. Then all possible IVs are simple multiples of |
kvn@2665 | 1655 | // the GCD. In practice, this will cover very few extra loops. |
kvn@2665 | 1656 | // Instead we require 'stride_con2' to be a multiple of 'stride_con', |
kvn@2665 | 1657 | // where +/-1 is the common case, but other integer multiples are |
kvn@2665 | 1658 | // also easy to handle. |
kvn@2665 | 1659 | int ratio_con = stride_con2/stride_con; |
kvn@2665 | 1660 | |
kvn@2665 | 1661 | if ((ratio_con * stride_con) == stride_con2) { // Check for exact |
kvn@3135 | 1662 | #ifndef PRODUCT |
kvn@3135 | 1663 | if (TraceLoopOpts) { |
kvn@3135 | 1664 | tty->print("Parallel IV: %d ", phi2->_idx); |
kvn@3135 | 1665 | loop->dump_head(); |
kvn@3135 | 1666 | } |
kvn@3135 | 1667 | #endif |
kvn@2665 | 1668 | // Convert to using the trip counter. The parallel induction |
kvn@2665 | 1669 | // variable differs from the trip counter by a loop-invariant |
kvn@2665 | 1670 | // amount, the difference between their respective initial values. |
kvn@2665 | 1671 | // It is scaled by the 'ratio_con'. |
kvn@2665 | 1672 | Node* ratio = _igvn.intcon(ratio_con); |
kvn@2665 | 1673 | set_ctrl(ratio, C->root()); |
kvn@3135 | 1674 | Node* ratio_init = new (C, 3) MulINode(init, ratio); |
kvn@3135 | 1675 | _igvn.register_new_node_with_optimizer(ratio_init, init); |
kvn@3135 | 1676 | set_early_ctrl(ratio_init); |
kvn@3135 | 1677 | Node* diff = new (C, 3) SubINode(init2, ratio_init); |
kvn@3135 | 1678 | _igvn.register_new_node_with_optimizer(diff, init2); |
kvn@3135 | 1679 | set_early_ctrl(diff); |
kvn@3135 | 1680 | Node* ratio_idx = new (C, 3) MulINode(phi, ratio); |
kvn@3135 | 1681 | _igvn.register_new_node_with_optimizer(ratio_idx, phi); |
kvn@3135 | 1682 | set_ctrl(ratio_idx, cl); |
kvn@3135 | 1683 | Node* add = new (C, 3) AddINode(ratio_idx, diff); |
kvn@3135 | 1684 | _igvn.register_new_node_with_optimizer(add); |
kvn@3135 | 1685 | set_ctrl(add, cl); |
kvn@2665 | 1686 | _igvn.replace_node( phi2, add ); |
kvn@2665 | 1687 | // Sometimes an induction variable is unused |
kvn@2665 | 1688 | if (add->outcnt() == 0) { |
kvn@2665 | 1689 | _igvn.remove_dead_node(add); |
kvn@2665 | 1690 | } |
kvn@2665 | 1691 | --i; // deleted this phi; rescan starting with next position |
kvn@2665 | 1692 | continue; |
kvn@2665 | 1693 | } |
kvn@2665 | 1694 | } |
kvn@2665 | 1695 | } |
kvn@2665 | 1696 | |
duke@435 | 1697 | //------------------------------counted_loop----------------------------------- |
duke@435 | 1698 | // Convert to counted loops where possible |
duke@435 | 1699 | void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) { |
duke@435 | 1700 | |
duke@435 | 1701 | // For grins, set the inner-loop flag here |
kvn@2665 | 1702 | if (!_child) { |
kvn@2665 | 1703 | if (_head->is_Loop()) _head->as_Loop()->set_inner_loop(); |
duke@435 | 1704 | } |
duke@435 | 1705 | |
kvn@2665 | 1706 | if (_head->is_CountedLoop() || |
kvn@2665 | 1707 | phase->is_counted_loop(_head, this)) { |
duke@435 | 1708 | _has_sfpt = 1; // Indicate we do not need a safepoint here |
duke@435 | 1709 | |
kvn@4023 | 1710 | // Look for safepoints to remove. |
kvn@4023 | 1711 | Node_List* sfpts = _safepts; |
kvn@4023 | 1712 | if (sfpts != NULL) { |
kvn@4023 | 1713 | for (uint i = 0; i < sfpts->size(); i++) { |
kvn@4023 | 1714 | Node* n = sfpts->at(i); |
kvn@4023 | 1715 | assert(phase->get_loop(n) == this, ""); |
kvn@4023 | 1716 | if (phase->is_deleteable_safept(n)) { |
kvn@4023 | 1717 | phase->lazy_replace(n, n->in(TypeFunc::Control)); |
kvn@4023 | 1718 | } |
kvn@4023 | 1719 | } |
kvn@4023 | 1720 | } |
duke@435 | 1721 | |
kvn@2665 | 1722 | // Look for induction variables |
kvn@2665 | 1723 | phase->replace_parallel_iv(this); |
duke@435 | 1724 | |
duke@435 | 1725 | } else if (_parent != NULL && !_irreducible) { |
duke@435 | 1726 | // Not a counted loop. |
kvn@4023 | 1727 | // Look for a safepoint on the idom-path. |
kvn@4023 | 1728 | Node* sfpt = tail(); |
kvn@4023 | 1729 | for (; sfpt != _head; sfpt = phase->idom(sfpt)) { |
kvn@4023 | 1730 | if (sfpt->Opcode() == Op_SafePoint && phase->get_loop(sfpt) == this) |
kvn@4023 | 1731 | break; // Found one |
duke@435 | 1732 | } |
kvn@4023 | 1733 | // Delete other safepoints in this loop. |
kvn@4023 | 1734 | Node_List* sfpts = _safepts; |
kvn@4023 | 1735 | if (sfpts != NULL && sfpt != _head && sfpt->Opcode() == Op_SafePoint) { |
kvn@4023 | 1736 | for (uint i = 0; i < sfpts->size(); i++) { |
kvn@4023 | 1737 | Node* n = sfpts->at(i); |
kvn@4023 | 1738 | assert(phase->get_loop(n) == this, ""); |
kvn@4023 | 1739 | if (n != sfpt && phase->is_deleteable_safept(n)) { |
kvn@4023 | 1740 | phase->lazy_replace(n, n->in(TypeFunc::Control)); |
kvn@4023 | 1741 | } |
kvn@4023 | 1742 | } |
duke@435 | 1743 | } |
duke@435 | 1744 | } |
duke@435 | 1745 | |
duke@435 | 1746 | // Recursively |
kvn@2665 | 1747 | if (_child) _child->counted_loop( phase ); |
kvn@2665 | 1748 | if (_next) _next ->counted_loop( phase ); |
duke@435 | 1749 | } |
duke@435 | 1750 | |
duke@435 | 1751 | #ifndef PRODUCT |
duke@435 | 1752 | //------------------------------dump_head-------------------------------------- |
duke@435 | 1753 | // Dump 1 liner for loop header info |
duke@435 | 1754 | void IdealLoopTree::dump_head( ) const { |
kvn@2665 | 1755 | for (uint i=0; i<_nest; i++) |
duke@435 | 1756 | tty->print(" "); |
duke@435 | 1757 | tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx); |
kvn@2665 | 1758 | if (_irreducible) tty->print(" IRREDUCIBLE"); |
kvn@2877 | 1759 | Node* entry = _head->in(LoopNode::EntryControl); |
kvn@2877 | 1760 | if (LoopLimitCheck) { |
kvn@2877 | 1761 | Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); |
kvn@2877 | 1762 | if (predicate != NULL ) { |
kvn@2877 | 1763 | tty->print(" limit_check"); |
kvn@2877 | 1764 | entry = entry->in(0)->in(0); |
kvn@2877 | 1765 | } |
kvn@2877 | 1766 | } |
kvn@2665 | 1767 | if (UseLoopPredicate) { |
kvn@2877 | 1768 | entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); |
kvn@2727 | 1769 | if (entry != NULL) { |
kvn@2665 | 1770 | tty->print(" predicated"); |
kvn@2665 | 1771 | } |
kvn@2665 | 1772 | } |
kvn@2665 | 1773 | if (_head->is_CountedLoop()) { |
duke@435 | 1774 | CountedLoopNode *cl = _head->as_CountedLoop(); |
duke@435 | 1775 | tty->print(" counted"); |
kvn@2747 | 1776 | |
kvn@2747 | 1777 | Node* init_n = cl->init_trip(); |
kvn@2747 | 1778 | if (init_n != NULL && init_n->is_Con()) |
kvn@2747 | 1779 | tty->print(" [%d,", cl->init_trip()->get_int()); |
kvn@2747 | 1780 | else |
kvn@2747 | 1781 | tty->print(" [int,"); |
kvn@2747 | 1782 | Node* limit_n = cl->limit(); |
kvn@2747 | 1783 | if (limit_n != NULL && limit_n->is_Con()) |
kvn@2747 | 1784 | tty->print("%d),", cl->limit()->get_int()); |
kvn@2747 | 1785 | else |
kvn@2747 | 1786 | tty->print("int),"); |
kvn@2747 | 1787 | int stride_con = cl->stride_con(); |
kvn@2747 | 1788 | if (stride_con > 0) tty->print("+"); |
kvn@2747 | 1789 | tty->print("%d", stride_con); |
kvn@2747 | 1790 | |
kvn@4001 | 1791 | tty->print(" (%d iters) ", (int)cl->profile_trip_cnt()); |
kvn@4001 | 1792 | |
kvn@2665 | 1793 | if (cl->is_pre_loop ()) tty->print(" pre" ); |
kvn@2665 | 1794 | if (cl->is_main_loop()) tty->print(" main"); |
kvn@2665 | 1795 | if (cl->is_post_loop()) tty->print(" post"); |
duke@435 | 1796 | } |
duke@435 | 1797 | tty->cr(); |
duke@435 | 1798 | } |
duke@435 | 1799 | |
duke@435 | 1800 | //------------------------------dump------------------------------------------- |
duke@435 | 1801 | // Dump loops by loop tree |
duke@435 | 1802 | void IdealLoopTree::dump( ) const { |
duke@435 | 1803 | dump_head(); |
kvn@2665 | 1804 | if (_child) _child->dump(); |
kvn@2665 | 1805 | if (_next) _next ->dump(); |
duke@435 | 1806 | } |
duke@435 | 1807 | |
duke@435 | 1808 | #endif |
duke@435 | 1809 | |
never@802 | 1810 | static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) { |
never@802 | 1811 | if (loop == root) { |
never@802 | 1812 | if (loop->_child != NULL) { |
never@802 | 1813 | log->begin_head("loop_tree"); |
never@802 | 1814 | log->end_head(); |
never@802 | 1815 | if( loop->_child ) log_loop_tree(root, loop->_child, log); |
never@802 | 1816 | log->tail("loop_tree"); |
never@802 | 1817 | assert(loop->_next == NULL, "what?"); |
never@802 | 1818 | } |
never@802 | 1819 | } else { |
never@802 | 1820 | Node* head = loop->_head; |
never@802 | 1821 | log->begin_head("loop"); |
never@802 | 1822 | log->print(" idx='%d' ", head->_idx); |
never@802 | 1823 | if (loop->_irreducible) log->print("irreducible='1' "); |
never@802 | 1824 | if (head->is_Loop()) { |
never@802 | 1825 | if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' "); |
never@802 | 1826 | if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' "); |
never@802 | 1827 | } |
never@802 | 1828 | if (head->is_CountedLoop()) { |
never@802 | 1829 | CountedLoopNode* cl = head->as_CountedLoop(); |
never@802 | 1830 | if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx()); |
never@802 | 1831 | if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx); |
never@802 | 1832 | if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx()); |
never@802 | 1833 | } |
never@802 | 1834 | log->end_head(); |
never@802 | 1835 | if( loop->_child ) log_loop_tree(root, loop->_child, log); |
never@802 | 1836 | log->tail("loop"); |
never@802 | 1837 | if( loop->_next ) log_loop_tree(root, loop->_next, log); |
never@802 | 1838 | } |
never@802 | 1839 | } |
never@802 | 1840 | |
cfang@1607 | 1841 | //---------------------collect_potentially_useful_predicates----------------------- |
cfang@1607 | 1842 | // Helper function to collect potentially useful predicates to prevent them from |
cfang@1607 | 1843 | // being eliminated by PhaseIdealLoop::eliminate_useless_predicates |
cfang@1607 | 1844 | void PhaseIdealLoop::collect_potentially_useful_predicates( |
cfang@1607 | 1845 | IdealLoopTree * loop, Unique_Node_List &useful_predicates) { |
cfang@1607 | 1846 | if (loop->_child) { // child |
cfang@1607 | 1847 | collect_potentially_useful_predicates(loop->_child, useful_predicates); |
cfang@1607 | 1848 | } |
cfang@1607 | 1849 | |
cfang@1607 | 1850 | // self (only loops that we can apply loop predication may use their predicates) |
kvn@2665 | 1851 | if (loop->_head->is_Loop() && |
kvn@2665 | 1852 | !loop->_irreducible && |
cfang@1607 | 1853 | !loop->tail()->is_top()) { |
kvn@2665 | 1854 | LoopNode* lpn = loop->_head->as_Loop(); |
cfang@1607 | 1855 | Node* entry = lpn->in(LoopNode::EntryControl); |
kvn@2877 | 1856 | Node* predicate_proj = find_predicate(entry); // loop_limit_check first |
cfang@1607 | 1857 | if (predicate_proj != NULL ) { // right pattern that can be used by loop predication |
kvn@2665 | 1858 | assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be"); |
cfang@1607 | 1859 | useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one |
kvn@2877 | 1860 | entry = entry->in(0)->in(0); |
kvn@2877 | 1861 | } |
kvn@2877 | 1862 | predicate_proj = find_predicate(entry); // Predicate |
kvn@2877 | 1863 | if (predicate_proj != NULL ) { |
kvn@2877 | 1864 | useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one |
cfang@1607 | 1865 | } |
cfang@1607 | 1866 | } |
cfang@1607 | 1867 | |
kvn@2665 | 1868 | if (loop->_next) { // sibling |
cfang@1607 | 1869 | collect_potentially_useful_predicates(loop->_next, useful_predicates); |
cfang@1607 | 1870 | } |
cfang@1607 | 1871 | } |
cfang@1607 | 1872 | |
cfang@1607 | 1873 | //------------------------eliminate_useless_predicates----------------------------- |
cfang@1607 | 1874 | // Eliminate all inserted predicates if they could not be used by loop predication. |
kvn@2877 | 1875 | // Note: it will also eliminates loop limits check predicate since it also uses |
kvn@2877 | 1876 | // Opaque1 node (see Parse::add_predicate()). |
cfang@1607 | 1877 | void PhaseIdealLoop::eliminate_useless_predicates() { |
kvn@2665 | 1878 | if (C->predicate_count() == 0) |
kvn@2665 | 1879 | return; // no predicate left |
cfang@1607 | 1880 | |
cfang@1607 | 1881 | Unique_Node_List useful_predicates; // to store useful predicates |
cfang@1607 | 1882 | if (C->has_loops()) { |
cfang@1607 | 1883 | collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates); |
cfang@1607 | 1884 | } |
cfang@1607 | 1885 | |
cfang@1607 | 1886 | for (int i = C->predicate_count(); i > 0; i--) { |
cfang@1607 | 1887 | Node * n = C->predicate_opaque1_node(i-1); |
cfang@1607 | 1888 | assert(n->Opcode() == Op_Opaque1, "must be"); |
cfang@1607 | 1889 | if (!useful_predicates.member(n)) { // not in the useful list |
cfang@1607 | 1890 | _igvn.replace_node(n, n->in(1)); |
cfang@1607 | 1891 | } |
cfang@1607 | 1892 | } |
cfang@1607 | 1893 | } |
cfang@1607 | 1894 | |
duke@435 | 1895 | //============================================================================= |
never@1356 | 1896 | //----------------------------build_and_optimize------------------------------- |
duke@435 | 1897 | // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to |
duke@435 | 1898 | // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups. |
kvn@3260 | 1899 | void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool skip_loop_opts) { |
kvn@2555 | 1900 | ResourceMark rm; |
kvn@2555 | 1901 | |
never@1356 | 1902 | int old_progress = C->major_progress(); |
never@2685 | 1903 | uint orig_worklist_size = _igvn._worklist.size(); |
never@1356 | 1904 | |
duke@435 | 1905 | // Reset major-progress flag for the driver's heuristics |
duke@435 | 1906 | C->clear_major_progress(); |
duke@435 | 1907 | |
duke@435 | 1908 | #ifndef PRODUCT |
duke@435 | 1909 | // Capture for later assert |
duke@435 | 1910 | uint unique = C->unique(); |
duke@435 | 1911 | _loop_invokes++; |
duke@435 | 1912 | _loop_work += unique; |
duke@435 | 1913 | #endif |
duke@435 | 1914 | |
duke@435 | 1915 | // True if the method has at least 1 irreducible loop |
duke@435 | 1916 | _has_irreducible_loops = false; |
duke@435 | 1917 | |
duke@435 | 1918 | _created_loop_node = false; |
duke@435 | 1919 | |
duke@435 | 1920 | Arena *a = Thread::current()->resource_area(); |
duke@435 | 1921 | VectorSet visited(a); |
duke@435 | 1922 | // Pre-grow the mapping from Nodes to IdealLoopTrees. |
duke@435 | 1923 | _nodes.map(C->unique(), NULL); |
duke@435 | 1924 | memset(_nodes.adr(), 0, wordSize * C->unique()); |
duke@435 | 1925 | |
duke@435 | 1926 | // Pre-build the top-level outermost loop tree entry |
duke@435 | 1927 | _ltree_root = new IdealLoopTree( this, C->root(), C->root() ); |
duke@435 | 1928 | // Do not need a safepoint at the top level |
duke@435 | 1929 | _ltree_root->_has_sfpt = 1; |
duke@435 | 1930 | |
kvn@2727 | 1931 | // Initialize Dominators. |
kvn@2727 | 1932 | // Checked in clone_loop_predicate() during beautify_loops(). |
kvn@2727 | 1933 | _idom_size = 0; |
kvn@2727 | 1934 | _idom = NULL; |
kvn@2727 | 1935 | _dom_depth = NULL; |
kvn@2727 | 1936 | _dom_stk = NULL; |
kvn@2727 | 1937 | |
duke@435 | 1938 | // Empty pre-order array |
duke@435 | 1939 | allocate_preorders(); |
duke@435 | 1940 | |
duke@435 | 1941 | // Build a loop tree on the fly. Build a mapping from CFG nodes to |
duke@435 | 1942 | // IdealLoopTree entries. Data nodes are NOT walked. |
duke@435 | 1943 | build_loop_tree(); |
duke@435 | 1944 | // Check for bailout, and return |
duke@435 | 1945 | if (C->failing()) { |
duke@435 | 1946 | return; |
duke@435 | 1947 | } |
duke@435 | 1948 | |
duke@435 | 1949 | // No loops after all |
never@1356 | 1950 | if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false); |
duke@435 | 1951 | |
duke@435 | 1952 | // There should always be an outer loop containing the Root and Return nodes. |
duke@435 | 1953 | // If not, we have a degenerate empty program. Bail out in this case. |
duke@435 | 1954 | if (!has_node(C->root())) { |
never@1356 | 1955 | if (!_verify_only) { |
never@1356 | 1956 | C->clear_major_progress(); |
never@1356 | 1957 | C->record_method_not_compilable("empty program detected during loop optimization"); |
never@1356 | 1958 | } |
duke@435 | 1959 | return; |
duke@435 | 1960 | } |
duke@435 | 1961 | |
duke@435 | 1962 | // Nothing to do, so get out |
kvn@3311 | 1963 | if( !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only ) { |
duke@435 | 1964 | _igvn.optimize(); // Cleanup NeverBranches |
duke@435 | 1965 | return; |
duke@435 | 1966 | } |
duke@435 | 1967 | |
duke@435 | 1968 | // Set loop nesting depth |
duke@435 | 1969 | _ltree_root->set_nest( 0 ); |
duke@435 | 1970 | |
duke@435 | 1971 | // Split shared headers and insert loop landing pads. |
duke@435 | 1972 | // Do not bother doing this on the Root loop of course. |
never@1356 | 1973 | if( !_verify_me && !_verify_only && _ltree_root->_child ) { |
never@2685 | 1974 | C->print_method("Before beautify loops", 3); |
duke@435 | 1975 | if( _ltree_root->_child->beautify_loops( this ) ) { |
duke@435 | 1976 | // Re-build loop tree! |
duke@435 | 1977 | _ltree_root->_child = NULL; |
duke@435 | 1978 | _nodes.clear(); |
duke@435 | 1979 | reallocate_preorders(); |
duke@435 | 1980 | build_loop_tree(); |
duke@435 | 1981 | // Check for bailout, and return |
duke@435 | 1982 | if (C->failing()) { |
duke@435 | 1983 | return; |
duke@435 | 1984 | } |
duke@435 | 1985 | // Reset loop nesting depth |
duke@435 | 1986 | _ltree_root->set_nest( 0 ); |
never@657 | 1987 | |
never@657 | 1988 | C->print_method("After beautify loops", 3); |
duke@435 | 1989 | } |
duke@435 | 1990 | } |
duke@435 | 1991 | |
duke@435 | 1992 | // Build Dominators for elision of NULL checks & loop finding. |
duke@435 | 1993 | // Since nodes do not have a slot for immediate dominator, make |
twisti@1040 | 1994 | // a persistent side array for that info indexed on node->_idx. |
duke@435 | 1995 | _idom_size = C->unique(); |
duke@435 | 1996 | _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size ); |
duke@435 | 1997 | _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size ); |
duke@435 | 1998 | _dom_stk = NULL; // Allocated on demand in recompute_dom_depth |
duke@435 | 1999 | memset( _dom_depth, 0, _idom_size * sizeof(uint) ); |
duke@435 | 2000 | |
duke@435 | 2001 | Dominators(); |
duke@435 | 2002 | |
never@1356 | 2003 | if (!_verify_only) { |
never@1356 | 2004 | // As a side effect, Dominators removed any unreachable CFG paths |
never@1356 | 2005 | // into RegionNodes. It doesn't do this test against Root, so |
never@1356 | 2006 | // we do it here. |
never@1356 | 2007 | for( uint i = 1; i < C->root()->req(); i++ ) { |
never@1356 | 2008 | if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root? |
kvn@3847 | 2009 | _igvn.delete_input_of(C->root(), i); |
never@1356 | 2010 | i--; // Rerun same iteration on compressed edges |
never@1356 | 2011 | } |
duke@435 | 2012 | } |
never@1356 | 2013 | |
never@1356 | 2014 | // Given dominators, try to find inner loops with calls that must |
never@1356 | 2015 | // always be executed (call dominates loop tail). These loops do |
never@1356 | 2016 | // not need a separate safepoint. |
never@1356 | 2017 | Node_List cisstack(a); |
never@1356 | 2018 | _ltree_root->check_safepts(visited, cisstack); |
duke@435 | 2019 | } |
duke@435 | 2020 | |
duke@435 | 2021 | // Walk the DATA nodes and place into loops. Find earliest control |
duke@435 | 2022 | // node. For CFG nodes, the _nodes array starts out and remains |
duke@435 | 2023 | // holding the associated IdealLoopTree pointer. For DATA nodes, the |
duke@435 | 2024 | // _nodes array holds the earliest legal controlling CFG node. |
duke@435 | 2025 | |
duke@435 | 2026 | // Allocate stack with enough space to avoid frequent realloc |
duke@435 | 2027 | int stack_size = (C->unique() >> 1) + 16; // (unique>>1)+16 from Java2D stats |
duke@435 | 2028 | Node_Stack nstack( a, stack_size ); |
duke@435 | 2029 | |
duke@435 | 2030 | visited.Clear(); |
duke@435 | 2031 | Node_List worklist(a); |
duke@435 | 2032 | // Don't need C->root() on worklist since |
duke@435 | 2033 | // it will be processed among C->top() inputs |
duke@435 | 2034 | worklist.push( C->top() ); |
duke@435 | 2035 | visited.set( C->top()->_idx ); // Set C->top() as visited now |
never@1356 | 2036 | build_loop_early( visited, worklist, nstack ); |
duke@435 | 2037 | |
duke@435 | 2038 | // Given early legal placement, try finding counted loops. This placement |
duke@435 | 2039 | // is good enough to discover most loop invariants. |
never@1356 | 2040 | if( !_verify_me && !_verify_only ) |
duke@435 | 2041 | _ltree_root->counted_loop( this ); |
duke@435 | 2042 | |
duke@435 | 2043 | // Find latest loop placement. Find ideal loop placement. |
duke@435 | 2044 | visited.Clear(); |
duke@435 | 2045 | init_dom_lca_tags(); |
duke@435 | 2046 | // Need C->root() on worklist when processing outs |
duke@435 | 2047 | worklist.push( C->root() ); |
duke@435 | 2048 | NOT_PRODUCT( C->verify_graph_edges(); ) |
duke@435 | 2049 | worklist.push( C->top() ); |
never@1356 | 2050 | build_loop_late( visited, worklist, nstack ); |
never@1356 | 2051 | |
never@1356 | 2052 | if (_verify_only) { |
never@1356 | 2053 | // restore major progress flag |
never@1356 | 2054 | for (int i = 0; i < old_progress; i++) |
never@1356 | 2055 | C->set_major_progress(); |
never@1356 | 2056 | assert(C->unique() == unique, "verification mode made Nodes? ? ?"); |
never@2685 | 2057 | assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything"); |
never@1356 | 2058 | return; |
never@1356 | 2059 | } |
duke@435 | 2060 | |
kvn@2727 | 2061 | // Some parser-inserted loop predicates could never be used by loop |
kvn@2727 | 2062 | // predication or they were moved away from loop during some optimizations. |
kvn@2727 | 2063 | // For example, peeling. Eliminate them before next loop optimizations. |
kvn@2877 | 2064 | if (UseLoopPredicate || LoopLimitCheck) { |
cfang@1607 | 2065 | eliminate_useless_predicates(); |
cfang@1607 | 2066 | } |
cfang@1607 | 2067 | |
duke@435 | 2068 | // clear out the dead code |
duke@435 | 2069 | while(_deadlist.size()) { |
never@1356 | 2070 | _igvn.remove_globally_dead_node(_deadlist.pop()); |
duke@435 | 2071 | } |
duke@435 | 2072 | |
duke@435 | 2073 | #ifndef PRODUCT |
duke@435 | 2074 | C->verify_graph_edges(); |
kvn@2665 | 2075 | if (_verify_me) { // Nested verify pass? |
duke@435 | 2076 | // Check to see if the verify mode is broken |
duke@435 | 2077 | assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?"); |
duke@435 | 2078 | return; |
duke@435 | 2079 | } |
kvn@2665 | 2080 | if(VerifyLoopOptimizations) verify(); |
kvn@2665 | 2081 | if(TraceLoopOpts && C->has_loops()) { |
kvn@2665 | 2082 | _ltree_root->dump(); |
kvn@2665 | 2083 | } |
duke@435 | 2084 | #endif |
duke@435 | 2085 | |
kvn@3260 | 2086 | if (skip_loop_opts) { |
kvn@3260 | 2087 | // Cleanup any modified bits |
kvn@3260 | 2088 | _igvn.optimize(); |
kvn@3260 | 2089 | |
kvn@3260 | 2090 | if (C->log() != NULL) { |
kvn@3260 | 2091 | log_loop_tree(_ltree_root, _ltree_root, C->log()); |
kvn@3260 | 2092 | } |
kvn@3260 | 2093 | return; |
kvn@3260 | 2094 | } |
kvn@3260 | 2095 | |
duke@435 | 2096 | if (ReassociateInvariants) { |
duke@435 | 2097 | // Reassociate invariants and prep for split_thru_phi |
duke@435 | 2098 | for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { |
duke@435 | 2099 | IdealLoopTree* lpt = iter.current(); |
duke@435 | 2100 | if (!lpt->is_counted() || !lpt->is_inner()) continue; |
duke@435 | 2101 | |
duke@435 | 2102 | lpt->reassociate_invariants(this); |
duke@435 | 2103 | |
duke@435 | 2104 | // Because RCE opportunities can be masked by split_thru_phi, |
duke@435 | 2105 | // look for RCE candidates and inhibit split_thru_phi |
duke@435 | 2106 | // on just their loop-phi's for this pass of loop opts |
cfang@1607 | 2107 | if (SplitIfBlocks && do_split_ifs) { |
duke@435 | 2108 | if (lpt->policy_range_check(this)) { |
kvn@474 | 2109 | lpt->_rce_candidate = 1; // = true |
duke@435 | 2110 | } |
duke@435 | 2111 | } |
duke@435 | 2112 | } |
duke@435 | 2113 | } |
duke@435 | 2114 | |
duke@435 | 2115 | // Check for aggressive application of split-if and other transforms |
duke@435 | 2116 | // that require basic-block info (like cloning through Phi's) |
duke@435 | 2117 | if( SplitIfBlocks && do_split_ifs ) { |
duke@435 | 2118 | visited.Clear(); |
duke@435 | 2119 | split_if_with_blocks( visited, nstack ); |
duke@435 | 2120 | NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); ); |
duke@435 | 2121 | } |
duke@435 | 2122 | |
cfang@1607 | 2123 | // Perform loop predication before iteration splitting |
kvn@2727 | 2124 | if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) { |
cfang@1607 | 2125 | _ltree_root->_child->loop_predication(this); |
cfang@1607 | 2126 | } |
cfang@1607 | 2127 | |
never@2118 | 2128 | if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) { |
never@2118 | 2129 | if (do_intrinsify_fill()) { |
never@2118 | 2130 | C->set_major_progress(); |
never@2118 | 2131 | } |
never@2118 | 2132 | } |
never@2118 | 2133 | |
duke@435 | 2134 | // Perform iteration-splitting on inner loops. Split iterations to avoid |
duke@435 | 2135 | // range checks or one-shot null checks. |
duke@435 | 2136 | |
duke@435 | 2137 | // If split-if's didn't hack the graph too bad (no CFG changes) |
duke@435 | 2138 | // then do loop opts. |
cfang@1607 | 2139 | if (C->has_loops() && !C->major_progress()) { |
duke@435 | 2140 | memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) ); |
duke@435 | 2141 | _ltree_root->_child->iteration_split( this, worklist ); |
duke@435 | 2142 | // No verify after peeling! GCM has hoisted code out of the loop. |
duke@435 | 2143 | // After peeling, the hoisted code could sink inside the peeled area. |
duke@435 | 2144 | // The peeling code does not try to recompute the best location for |
duke@435 | 2145 | // all the code before the peeled area, so the verify pass will always |
duke@435 | 2146 | // complain about it. |
duke@435 | 2147 | } |
duke@435 | 2148 | // Do verify graph edges in any case |
duke@435 | 2149 | NOT_PRODUCT( C->verify_graph_edges(); ); |
duke@435 | 2150 | |
cfang@1607 | 2151 | if (!do_split_ifs) { |
duke@435 | 2152 | // We saw major progress in Split-If to get here. We forced a |
duke@435 | 2153 | // pass with unrolling and not split-if, however more split-if's |
duke@435 | 2154 | // might make progress. If the unrolling didn't make progress |
duke@435 | 2155 | // then the major-progress flag got cleared and we won't try |
duke@435 | 2156 | // another round of Split-If. In particular the ever-common |
duke@435 | 2157 | // instance-of/check-cast pattern requires at least 2 rounds of |
duke@435 | 2158 | // Split-If to clear out. |
duke@435 | 2159 | C->set_major_progress(); |
duke@435 | 2160 | } |
duke@435 | 2161 | |
duke@435 | 2162 | // Repeat loop optimizations if new loops were seen |
duke@435 | 2163 | if (created_loop_node()) { |
duke@435 | 2164 | C->set_major_progress(); |
duke@435 | 2165 | } |
duke@435 | 2166 | |
kvn@2727 | 2167 | // Keep loop predicates and perform optimizations with them |
kvn@2727 | 2168 | // until no more loop optimizations could be done. |
kvn@2727 | 2169 | // After that switch predicates off and do more loop optimizations. |
kvn@2727 | 2170 | if (!C->major_progress() && (C->predicate_count() > 0)) { |
kvn@2727 | 2171 | C->cleanup_loop_predicates(_igvn); |
kvn@2727 | 2172 | #ifndef PRODUCT |
kvn@2727 | 2173 | if (TraceLoopOpts) { |
kvn@2727 | 2174 | tty->print_cr("PredicatesOff"); |
kvn@2727 | 2175 | } |
kvn@2727 | 2176 | #endif |
kvn@2727 | 2177 | C->set_major_progress(); |
kvn@2727 | 2178 | } |
duke@435 | 2179 | |
kvn@2727 | 2180 | // Convert scalar to superword operations at the end of all loop opts. |
duke@435 | 2181 | if (UseSuperWord && C->has_loops() && !C->major_progress()) { |
duke@435 | 2182 | // SuperWord transform |
duke@435 | 2183 | SuperWord sw(this); |
duke@435 | 2184 | for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { |
duke@435 | 2185 | IdealLoopTree* lpt = iter.current(); |
duke@435 | 2186 | if (lpt->is_counted()) { |
duke@435 | 2187 | sw.transform_loop(lpt); |
duke@435 | 2188 | } |
duke@435 | 2189 | } |
duke@435 | 2190 | } |
duke@435 | 2191 | |
duke@435 | 2192 | // Cleanup any modified bits |
duke@435 | 2193 | _igvn.optimize(); |
duke@435 | 2194 | |
never@802 | 2195 | // disable assert until issue with split_flow_path is resolved (6742111) |
never@802 | 2196 | // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(), |
never@802 | 2197 | // "shouldn't introduce irreducible loops"); |
never@802 | 2198 | |
never@802 | 2199 | if (C->log() != NULL) { |
never@802 | 2200 | log_loop_tree(_ltree_root, _ltree_root, C->log()); |
never@802 | 2201 | } |
duke@435 | 2202 | } |
duke@435 | 2203 | |
duke@435 | 2204 | #ifndef PRODUCT |
duke@435 | 2205 | //------------------------------print_statistics------------------------------- |
duke@435 | 2206 | int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes |
duke@435 | 2207 | int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique |
duke@435 | 2208 | void PhaseIdealLoop::print_statistics() { |
duke@435 | 2209 | tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work); |
duke@435 | 2210 | } |
duke@435 | 2211 | |
duke@435 | 2212 | //------------------------------verify----------------------------------------- |
duke@435 | 2213 | // Build a verify-only PhaseIdealLoop, and see that it agrees with me. |
duke@435 | 2214 | static int fail; // debug only, so its multi-thread dont care |
duke@435 | 2215 | void PhaseIdealLoop::verify() const { |
duke@435 | 2216 | int old_progress = C->major_progress(); |
duke@435 | 2217 | ResourceMark rm; |
never@1356 | 2218 | PhaseIdealLoop loop_verify( _igvn, this ); |
duke@435 | 2219 | VectorSet visited(Thread::current()->resource_area()); |
duke@435 | 2220 | |
duke@435 | 2221 | fail = 0; |
duke@435 | 2222 | verify_compare( C->root(), &loop_verify, visited ); |
duke@435 | 2223 | assert( fail == 0, "verify loops failed" ); |
duke@435 | 2224 | // Verify loop structure is the same |
duke@435 | 2225 | _ltree_root->verify_tree(loop_verify._ltree_root, NULL); |
duke@435 | 2226 | // Reset major-progress. It was cleared by creating a verify version of |
duke@435 | 2227 | // PhaseIdealLoop. |
duke@435 | 2228 | for( int i=0; i<old_progress; i++ ) |
duke@435 | 2229 | C->set_major_progress(); |
duke@435 | 2230 | } |
duke@435 | 2231 | |
duke@435 | 2232 | //------------------------------verify_compare--------------------------------- |
duke@435 | 2233 | // Make sure me and the given PhaseIdealLoop agree on key data structures |
duke@435 | 2234 | void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const { |
duke@435 | 2235 | if( !n ) return; |
duke@435 | 2236 | if( visited.test_set( n->_idx ) ) return; |
duke@435 | 2237 | if( !_nodes[n->_idx] ) { // Unreachable |
duke@435 | 2238 | assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" ); |
duke@435 | 2239 | return; |
duke@435 | 2240 | } |
duke@435 | 2241 | |
duke@435 | 2242 | uint i; |
duke@435 | 2243 | for( i = 0; i < n->req(); i++ ) |
duke@435 | 2244 | verify_compare( n->in(i), loop_verify, visited ); |
duke@435 | 2245 | |
duke@435 | 2246 | // Check the '_nodes' block/loop structure |
duke@435 | 2247 | i = n->_idx; |
duke@435 | 2248 | if( has_ctrl(n) ) { // We have control; verify has loop or ctrl |
duke@435 | 2249 | if( _nodes[i] != loop_verify->_nodes[i] && |
duke@435 | 2250 | get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) { |
duke@435 | 2251 | tty->print("Mismatched control setting for: "); |
duke@435 | 2252 | n->dump(); |
duke@435 | 2253 | if( fail++ > 10 ) return; |
duke@435 | 2254 | Node *c = get_ctrl_no_update(n); |
duke@435 | 2255 | tty->print("We have it as: "); |
duke@435 | 2256 | if( c->in(0) ) c->dump(); |
duke@435 | 2257 | else tty->print_cr("N%d",c->_idx); |
duke@435 | 2258 | tty->print("Verify thinks: "); |
duke@435 | 2259 | if( loop_verify->has_ctrl(n) ) |
duke@435 | 2260 | loop_verify->get_ctrl_no_update(n)->dump(); |
duke@435 | 2261 | else |
duke@435 | 2262 | loop_verify->get_loop_idx(n)->dump(); |
duke@435 | 2263 | tty->cr(); |
duke@435 | 2264 | } |
duke@435 | 2265 | } else { // We have a loop |
duke@435 | 2266 | IdealLoopTree *us = get_loop_idx(n); |
duke@435 | 2267 | if( loop_verify->has_ctrl(n) ) { |
duke@435 | 2268 | tty->print("Mismatched loop setting for: "); |
duke@435 | 2269 | n->dump(); |
duke@435 | 2270 | if( fail++ > 10 ) return; |
duke@435 | 2271 | tty->print("We have it as: "); |
duke@435 | 2272 | us->dump(); |
duke@435 | 2273 | tty->print("Verify thinks: "); |
duke@435 | 2274 | loop_verify->get_ctrl_no_update(n)->dump(); |
duke@435 | 2275 | tty->cr(); |
duke@435 | 2276 | } else if (!C->major_progress()) { |
duke@435 | 2277 | // Loop selection can be messed up if we did a major progress |
duke@435 | 2278 | // operation, like split-if. Do not verify in that case. |
duke@435 | 2279 | IdealLoopTree *them = loop_verify->get_loop_idx(n); |
duke@435 | 2280 | if( us->_head != them->_head || us->_tail != them->_tail ) { |
duke@435 | 2281 | tty->print("Unequals loops for: "); |
duke@435 | 2282 | n->dump(); |
duke@435 | 2283 | if( fail++ > 10 ) return; |
duke@435 | 2284 | tty->print("We have it as: "); |
duke@435 | 2285 | us->dump(); |
duke@435 | 2286 | tty->print("Verify thinks: "); |
duke@435 | 2287 | them->dump(); |
duke@435 | 2288 | tty->cr(); |
duke@435 | 2289 | } |
duke@435 | 2290 | } |
duke@435 | 2291 | } |
duke@435 | 2292 | |
duke@435 | 2293 | // Check for immediate dominators being equal |
duke@435 | 2294 | if( i >= _idom_size ) { |
duke@435 | 2295 | if( !n->is_CFG() ) return; |
duke@435 | 2296 | tty->print("CFG Node with no idom: "); |
duke@435 | 2297 | n->dump(); |
duke@435 | 2298 | return; |
duke@435 | 2299 | } |
duke@435 | 2300 | if( !n->is_CFG() ) return; |
duke@435 | 2301 | if( n == C->root() ) return; // No IDOM here |
duke@435 | 2302 | |
duke@435 | 2303 | assert(n->_idx == i, "sanity"); |
duke@435 | 2304 | Node *id = idom_no_update(n); |
duke@435 | 2305 | if( id != loop_verify->idom_no_update(n) ) { |
duke@435 | 2306 | tty->print("Unequals idoms for: "); |
duke@435 | 2307 | n->dump(); |
duke@435 | 2308 | if( fail++ > 10 ) return; |
duke@435 | 2309 | tty->print("We have it as: "); |
duke@435 | 2310 | id->dump(); |
duke@435 | 2311 | tty->print("Verify thinks: "); |
duke@435 | 2312 | loop_verify->idom_no_update(n)->dump(); |
duke@435 | 2313 | tty->cr(); |
duke@435 | 2314 | } |
duke@435 | 2315 | |
duke@435 | 2316 | } |
duke@435 | 2317 | |
duke@435 | 2318 | //------------------------------verify_tree------------------------------------ |
duke@435 | 2319 | // Verify that tree structures match. Because the CFG can change, siblings |
duke@435 | 2320 | // within the loop tree can be reordered. We attempt to deal with that by |
duke@435 | 2321 | // reordering the verify's loop tree if possible. |
duke@435 | 2322 | void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const { |
duke@435 | 2323 | assert( _parent == parent, "Badly formed loop tree" ); |
duke@435 | 2324 | |
duke@435 | 2325 | // Siblings not in same order? Attempt to re-order. |
duke@435 | 2326 | if( _head != loop->_head ) { |
duke@435 | 2327 | // Find _next pointer to update |
duke@435 | 2328 | IdealLoopTree **pp = &loop->_parent->_child; |
duke@435 | 2329 | while( *pp != loop ) |
duke@435 | 2330 | pp = &((*pp)->_next); |
duke@435 | 2331 | // Find proper sibling to be next |
duke@435 | 2332 | IdealLoopTree **nn = &loop->_next; |
duke@435 | 2333 | while( (*nn) && (*nn)->_head != _head ) |
duke@435 | 2334 | nn = &((*nn)->_next); |
duke@435 | 2335 | |
duke@435 | 2336 | // Check for no match. |
duke@435 | 2337 | if( !(*nn) ) { |
duke@435 | 2338 | // Annoyingly, irreducible loops can pick different headers |
duke@435 | 2339 | // after a major_progress operation, so the rest of the loop |
duke@435 | 2340 | // tree cannot be matched. |
duke@435 | 2341 | if (_irreducible && Compile::current()->major_progress()) return; |
duke@435 | 2342 | assert( 0, "failed to match loop tree" ); |
duke@435 | 2343 | } |
duke@435 | 2344 | |
duke@435 | 2345 | // Move (*nn) to (*pp) |
duke@435 | 2346 | IdealLoopTree *hit = *nn; |
duke@435 | 2347 | *nn = hit->_next; |
duke@435 | 2348 | hit->_next = loop; |
duke@435 | 2349 | *pp = loop; |
duke@435 | 2350 | loop = hit; |
duke@435 | 2351 | // Now try again to verify |
duke@435 | 2352 | } |
duke@435 | 2353 | |
duke@435 | 2354 | assert( _head == loop->_head , "mismatched loop head" ); |
duke@435 | 2355 | Node *tail = _tail; // Inline a non-updating version of |
duke@435 | 2356 | while( !tail->in(0) ) // the 'tail()' call. |
duke@435 | 2357 | tail = tail->in(1); |
duke@435 | 2358 | assert( tail == loop->_tail, "mismatched loop tail" ); |
duke@435 | 2359 | |
duke@435 | 2360 | // Counted loops that are guarded should be able to find their guards |
duke@435 | 2361 | if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) { |
duke@435 | 2362 | CountedLoopNode *cl = _head->as_CountedLoop(); |
duke@435 | 2363 | Node *init = cl->init_trip(); |
duke@435 | 2364 | Node *ctrl = cl->in(LoopNode::EntryControl); |
duke@435 | 2365 | assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); |
duke@435 | 2366 | Node *iff = ctrl->in(0); |
duke@435 | 2367 | assert( iff->Opcode() == Op_If, "" ); |
duke@435 | 2368 | Node *bol = iff->in(1); |
duke@435 | 2369 | assert( bol->Opcode() == Op_Bool, "" ); |
duke@435 | 2370 | Node *cmp = bol->in(1); |
duke@435 | 2371 | assert( cmp->Opcode() == Op_CmpI, "" ); |
duke@435 | 2372 | Node *add = cmp->in(1); |
duke@435 | 2373 | Node *opaq; |
duke@435 | 2374 | if( add->Opcode() == Op_Opaque1 ) { |
duke@435 | 2375 | opaq = add; |
duke@435 | 2376 | } else { |
duke@435 | 2377 | assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" ); |
duke@435 | 2378 | assert( add == init, "" ); |
duke@435 | 2379 | opaq = cmp->in(2); |
duke@435 | 2380 | } |
duke@435 | 2381 | assert( opaq->Opcode() == Op_Opaque1, "" ); |
duke@435 | 2382 | |
duke@435 | 2383 | } |
duke@435 | 2384 | |
duke@435 | 2385 | if (_child != NULL) _child->verify_tree(loop->_child, this); |
duke@435 | 2386 | if (_next != NULL) _next ->verify_tree(loop->_next, parent); |
duke@435 | 2387 | // Innermost loops need to verify loop bodies, |
duke@435 | 2388 | // but only if no 'major_progress' |
duke@435 | 2389 | int fail = 0; |
duke@435 | 2390 | if (!Compile::current()->major_progress() && _child == NULL) { |
duke@435 | 2391 | for( uint i = 0; i < _body.size(); i++ ) { |
duke@435 | 2392 | Node *n = _body.at(i); |
duke@435 | 2393 | if (n->outcnt() == 0) continue; // Ignore dead |
duke@435 | 2394 | uint j; |
duke@435 | 2395 | for( j = 0; j < loop->_body.size(); j++ ) |
duke@435 | 2396 | if( loop->_body.at(j) == n ) |
duke@435 | 2397 | break; |
duke@435 | 2398 | if( j == loop->_body.size() ) { // Not found in loop body |
duke@435 | 2399 | // Last ditch effort to avoid assertion: Its possible that we |
duke@435 | 2400 | // have some users (so outcnt not zero) but are still dead. |
duke@435 | 2401 | // Try to find from root. |
duke@435 | 2402 | if (Compile::current()->root()->find(n->_idx)) { |
duke@435 | 2403 | fail++; |
duke@435 | 2404 | tty->print("We have that verify does not: "); |
duke@435 | 2405 | n->dump(); |
duke@435 | 2406 | } |
duke@435 | 2407 | } |
duke@435 | 2408 | } |
duke@435 | 2409 | for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) { |
duke@435 | 2410 | Node *n = loop->_body.at(i2); |
duke@435 | 2411 | if (n->outcnt() == 0) continue; // Ignore dead |
duke@435 | 2412 | uint j; |
duke@435 | 2413 | for( j = 0; j < _body.size(); j++ ) |
duke@435 | 2414 | if( _body.at(j) == n ) |
duke@435 | 2415 | break; |
duke@435 | 2416 | if( j == _body.size() ) { // Not found in loop body |
duke@435 | 2417 | // Last ditch effort to avoid assertion: Its possible that we |
duke@435 | 2418 | // have some users (so outcnt not zero) but are still dead. |
duke@435 | 2419 | // Try to find from root. |
duke@435 | 2420 | if (Compile::current()->root()->find(n->_idx)) { |
duke@435 | 2421 | fail++; |
duke@435 | 2422 | tty->print("Verify has that we do not: "); |
duke@435 | 2423 | n->dump(); |
duke@435 | 2424 | } |
duke@435 | 2425 | } |
duke@435 | 2426 | } |
duke@435 | 2427 | assert( !fail, "loop body mismatch" ); |
duke@435 | 2428 | } |
duke@435 | 2429 | } |
duke@435 | 2430 | |
duke@435 | 2431 | #endif |
duke@435 | 2432 | |
duke@435 | 2433 | //------------------------------set_idom--------------------------------------- |
duke@435 | 2434 | void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) { |
duke@435 | 2435 | uint idx = d->_idx; |
duke@435 | 2436 | if (idx >= _idom_size) { |
duke@435 | 2437 | uint newsize = _idom_size<<1; |
duke@435 | 2438 | while( idx >= newsize ) { |
duke@435 | 2439 | newsize <<= 1; |
duke@435 | 2440 | } |
duke@435 | 2441 | _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize); |
duke@435 | 2442 | _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize); |
duke@435 | 2443 | memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) ); |
duke@435 | 2444 | _idom_size = newsize; |
duke@435 | 2445 | } |
duke@435 | 2446 | _idom[idx] = n; |
duke@435 | 2447 | _dom_depth[idx] = dom_depth; |
duke@435 | 2448 | } |
duke@435 | 2449 | |
duke@435 | 2450 | //------------------------------recompute_dom_depth--------------------------------------- |
duke@435 | 2451 | // The dominator tree is constructed with only parent pointers. |
duke@435 | 2452 | // This recomputes the depth in the tree by first tagging all |
duke@435 | 2453 | // nodes as "no depth yet" marker. The next pass then runs up |
duke@435 | 2454 | // the dom tree from each node marked "no depth yet", and computes |
duke@435 | 2455 | // the depth on the way back down. |
duke@435 | 2456 | void PhaseIdealLoop::recompute_dom_depth() { |
duke@435 | 2457 | uint no_depth_marker = C->unique(); |
duke@435 | 2458 | uint i; |
duke@435 | 2459 | // Initialize depth to "no depth yet" |
duke@435 | 2460 | for (i = 0; i < _idom_size; i++) { |
duke@435 | 2461 | if (_dom_depth[i] > 0 && _idom[i] != NULL) { |
duke@435 | 2462 | _dom_depth[i] = no_depth_marker; |
duke@435 | 2463 | } |
duke@435 | 2464 | } |
duke@435 | 2465 | if (_dom_stk == NULL) { |
duke@435 | 2466 | uint init_size = C->unique() / 100; // Guess that 1/100 is a reasonable initial size. |
duke@435 | 2467 | if (init_size < 10) init_size = 10; |
kvn@2555 | 2468 | _dom_stk = new GrowableArray<uint>(init_size); |
duke@435 | 2469 | } |
duke@435 | 2470 | // Compute new depth for each node. |
duke@435 | 2471 | for (i = 0; i < _idom_size; i++) { |
duke@435 | 2472 | uint j = i; |
duke@435 | 2473 | // Run up the dom tree to find a node with a depth |
duke@435 | 2474 | while (_dom_depth[j] == no_depth_marker) { |
duke@435 | 2475 | _dom_stk->push(j); |
duke@435 | 2476 | j = _idom[j]->_idx; |
duke@435 | 2477 | } |
duke@435 | 2478 | // Compute the depth on the way back down this tree branch |
duke@435 | 2479 | uint dd = _dom_depth[j] + 1; |
duke@435 | 2480 | while (_dom_stk->length() > 0) { |
duke@435 | 2481 | uint j = _dom_stk->pop(); |
duke@435 | 2482 | _dom_depth[j] = dd; |
duke@435 | 2483 | dd++; |
duke@435 | 2484 | } |
duke@435 | 2485 | } |
duke@435 | 2486 | } |
duke@435 | 2487 | |
duke@435 | 2488 | //------------------------------sort------------------------------------------- |
duke@435 | 2489 | // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the |
duke@435 | 2490 | // loop tree, not the root. |
duke@435 | 2491 | IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) { |
duke@435 | 2492 | if( !innermost ) return loop; // New innermost loop |
duke@435 | 2493 | |
duke@435 | 2494 | int loop_preorder = get_preorder(loop->_head); // Cache pre-order number |
duke@435 | 2495 | assert( loop_preorder, "not yet post-walked loop" ); |
duke@435 | 2496 | IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer |
duke@435 | 2497 | IdealLoopTree *l = *pp; // Do I go before or after 'l'? |
duke@435 | 2498 | |
duke@435 | 2499 | // Insert at start of list |
duke@435 | 2500 | while( l ) { // Insertion sort based on pre-order |
duke@435 | 2501 | if( l == loop ) return innermost; // Already on list! |
duke@435 | 2502 | int l_preorder = get_preorder(l->_head); // Cache pre-order number |
duke@435 | 2503 | assert( l_preorder, "not yet post-walked l" ); |
duke@435 | 2504 | // Check header pre-order number to figure proper nesting |
duke@435 | 2505 | if( loop_preorder > l_preorder ) |
duke@435 | 2506 | break; // End of insertion |
duke@435 | 2507 | // If headers tie (e.g., shared headers) check tail pre-order numbers. |
duke@435 | 2508 | // Since I split shared headers, you'd think this could not happen. |
duke@435 | 2509 | // BUT: I must first do the preorder numbering before I can discover I |
duke@435 | 2510 | // have shared headers, so the split headers all get the same preorder |
duke@435 | 2511 | // number as the RegionNode they split from. |
duke@435 | 2512 | if( loop_preorder == l_preorder && |
duke@435 | 2513 | get_preorder(loop->_tail) < get_preorder(l->_tail) ) |
duke@435 | 2514 | break; // Also check for shared headers (same pre#) |
duke@435 | 2515 | pp = &l->_parent; // Chain up list |
duke@435 | 2516 | l = *pp; |
duke@435 | 2517 | } |
duke@435 | 2518 | // Link into list |
duke@435 | 2519 | // Point predecessor to me |
duke@435 | 2520 | *pp = loop; |
duke@435 | 2521 | // Point me to successor |
duke@435 | 2522 | IdealLoopTree *p = loop->_parent; |
duke@435 | 2523 | loop->_parent = l; // Point me to successor |
duke@435 | 2524 | if( p ) sort( p, innermost ); // Insert my parents into list as well |
duke@435 | 2525 | return innermost; |
duke@435 | 2526 | } |
duke@435 | 2527 | |
duke@435 | 2528 | //------------------------------build_loop_tree-------------------------------- |
duke@435 | 2529 | // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit |
duke@435 | 2530 | // bits. The _nodes[] array is mapped by Node index and holds a NULL for |
duke@435 | 2531 | // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the |
duke@435 | 2532 | // tightest enclosing IdealLoopTree for post-walked. |
duke@435 | 2533 | // |
duke@435 | 2534 | // During my forward walk I do a short 1-layer lookahead to see if I can find |
duke@435 | 2535 | // a loop backedge with that doesn't have any work on the backedge. This |
duke@435 | 2536 | // helps me construct nested loops with shared headers better. |
duke@435 | 2537 | // |
duke@435 | 2538 | // Once I've done the forward recursion, I do the post-work. For each child |
duke@435 | 2539 | // I check to see if there is a backedge. Backedges define a loop! I |
duke@435 | 2540 | // insert an IdealLoopTree at the target of the backedge. |
duke@435 | 2541 | // |
duke@435 | 2542 | // During the post-work I also check to see if I have several children |
duke@435 | 2543 | // belonging to different loops. If so, then this Node is a decision point |
duke@435 | 2544 | // where control flow can choose to change loop nests. It is at this |
duke@435 | 2545 | // decision point where I can figure out how loops are nested. At this |
duke@435 | 2546 | // time I can properly order the different loop nests from my children. |
duke@435 | 2547 | // Note that there may not be any backedges at the decision point! |
duke@435 | 2548 | // |
duke@435 | 2549 | // Since the decision point can be far removed from the backedges, I can't |
duke@435 | 2550 | // order my loops at the time I discover them. Thus at the decision point |
duke@435 | 2551 | // I need to inspect loop header pre-order numbers to properly nest my |
duke@435 | 2552 | // loops. This means I need to sort my childrens' loops by pre-order. |
duke@435 | 2553 | // The sort is of size number-of-control-children, which generally limits |
duke@435 | 2554 | // it to size 2 (i.e., I just choose between my 2 target loops). |
duke@435 | 2555 | void PhaseIdealLoop::build_loop_tree() { |
duke@435 | 2556 | // Allocate stack of size C->unique()/2 to avoid frequent realloc |
duke@435 | 2557 | GrowableArray <Node *> bltstack(C->unique() >> 1); |
duke@435 | 2558 | Node *n = C->root(); |
duke@435 | 2559 | bltstack.push(n); |
duke@435 | 2560 | int pre_order = 1; |
duke@435 | 2561 | int stack_size; |
duke@435 | 2562 | |
duke@435 | 2563 | while ( ( stack_size = bltstack.length() ) != 0 ) { |
duke@435 | 2564 | n = bltstack.top(); // Leave node on stack |
duke@435 | 2565 | if ( !is_visited(n) ) { |
duke@435 | 2566 | // ---- Pre-pass Work ---- |
duke@435 | 2567 | // Pre-walked but not post-walked nodes need a pre_order number. |
duke@435 | 2568 | |
duke@435 | 2569 | set_preorder_visited( n, pre_order ); // set as visited |
duke@435 | 2570 | |
duke@435 | 2571 | // ---- Scan over children ---- |
duke@435 | 2572 | // Scan first over control projections that lead to loop headers. |
duke@435 | 2573 | // This helps us find inner-to-outer loops with shared headers better. |
duke@435 | 2574 | |
duke@435 | 2575 | // Scan children's children for loop headers. |
duke@435 | 2576 | for ( int i = n->outcnt() - 1; i >= 0; --i ) { |
duke@435 | 2577 | Node* m = n->raw_out(i); // Child |
duke@435 | 2578 | if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children |
duke@435 | 2579 | // Scan over children's children to find loop |
duke@435 | 2580 | for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 2581 | Node* l = m->fast_out(j); |
duke@435 | 2582 | if( is_visited(l) && // Been visited? |
duke@435 | 2583 | !is_postvisited(l) && // But not post-visited |
duke@435 | 2584 | get_preorder(l) < pre_order ) { // And smaller pre-order |
duke@435 | 2585 | // Found! Scan the DFS down this path before doing other paths |
duke@435 | 2586 | bltstack.push(m); |
duke@435 | 2587 | break; |
duke@435 | 2588 | } |
duke@435 | 2589 | } |
duke@435 | 2590 | } |
duke@435 | 2591 | } |
duke@435 | 2592 | pre_order++; |
duke@435 | 2593 | } |
duke@435 | 2594 | else if ( !is_postvisited(n) ) { |
duke@435 | 2595 | // Note: build_loop_tree_impl() adds out edges on rare occasions, |
duke@435 | 2596 | // such as com.sun.rsasign.am::a. |
duke@435 | 2597 | // For non-recursive version, first, process current children. |
duke@435 | 2598 | // On next iteration, check if additional children were added. |
duke@435 | 2599 | for ( int k = n->outcnt() - 1; k >= 0; --k ) { |
duke@435 | 2600 | Node* u = n->raw_out(k); |
duke@435 | 2601 | if ( u->is_CFG() && !is_visited(u) ) { |
duke@435 | 2602 | bltstack.push(u); |
duke@435 | 2603 | } |
duke@435 | 2604 | } |
duke@435 | 2605 | if ( bltstack.length() == stack_size ) { |
duke@435 | 2606 | // There were no additional children, post visit node now |
duke@435 | 2607 | (void)bltstack.pop(); // Remove node from stack |
duke@435 | 2608 | pre_order = build_loop_tree_impl( n, pre_order ); |
duke@435 | 2609 | // Check for bailout |
duke@435 | 2610 | if (C->failing()) { |
duke@435 | 2611 | return; |
duke@435 | 2612 | } |
duke@435 | 2613 | // Check to grow _preorders[] array for the case when |
duke@435 | 2614 | // build_loop_tree_impl() adds new nodes. |
duke@435 | 2615 | check_grow_preorders(); |
duke@435 | 2616 | } |
duke@435 | 2617 | } |
duke@435 | 2618 | else { |
duke@435 | 2619 | (void)bltstack.pop(); // Remove post-visited node from stack |
duke@435 | 2620 | } |
duke@435 | 2621 | } |
duke@435 | 2622 | } |
duke@435 | 2623 | |
duke@435 | 2624 | //------------------------------build_loop_tree_impl--------------------------- |
duke@435 | 2625 | int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) { |
duke@435 | 2626 | // ---- Post-pass Work ---- |
duke@435 | 2627 | // Pre-walked but not post-walked nodes need a pre_order number. |
duke@435 | 2628 | |
duke@435 | 2629 | // Tightest enclosing loop for this Node |
duke@435 | 2630 | IdealLoopTree *innermost = NULL; |
duke@435 | 2631 | |
duke@435 | 2632 | // For all children, see if any edge is a backedge. If so, make a loop |
duke@435 | 2633 | // for it. Then find the tightest enclosing loop for the self Node. |
duke@435 | 2634 | for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { |
duke@435 | 2635 | Node* m = n->fast_out(i); // Child |
duke@435 | 2636 | if( n == m ) continue; // Ignore control self-cycles |
duke@435 | 2637 | if( !m->is_CFG() ) continue;// Ignore non-CFG edges |
duke@435 | 2638 | |
duke@435 | 2639 | IdealLoopTree *l; // Child's loop |
duke@435 | 2640 | if( !is_postvisited(m) ) { // Child visited but not post-visited? |
duke@435 | 2641 | // Found a backedge |
duke@435 | 2642 | assert( get_preorder(m) < pre_order, "should be backedge" ); |
duke@435 | 2643 | // Check for the RootNode, which is already a LoopNode and is allowed |
duke@435 | 2644 | // to have multiple "backedges". |
duke@435 | 2645 | if( m == C->root()) { // Found the root? |
duke@435 | 2646 | l = _ltree_root; // Root is the outermost LoopNode |
duke@435 | 2647 | } else { // Else found a nested loop |
duke@435 | 2648 | // Insert a LoopNode to mark this loop. |
duke@435 | 2649 | l = new IdealLoopTree(this, m, n); |
duke@435 | 2650 | } // End of Else found a nested loop |
duke@435 | 2651 | if( !has_loop(m) ) // If 'm' does not already have a loop set |
duke@435 | 2652 | set_loop(m, l); // Set loop header to loop now |
duke@435 | 2653 | |
duke@435 | 2654 | } else { // Else not a nested loop |
duke@435 | 2655 | if( !_nodes[m->_idx] ) continue; // Dead code has no loop |
duke@435 | 2656 | l = get_loop(m); // Get previously determined loop |
duke@435 | 2657 | // If successor is header of a loop (nest), move up-loop till it |
duke@435 | 2658 | // is a member of some outer enclosing loop. Since there are no |
duke@435 | 2659 | // shared headers (I've split them already) I only need to go up |
duke@435 | 2660 | // at most 1 level. |
duke@435 | 2661 | while( l && l->_head == m ) // Successor heads loop? |
duke@435 | 2662 | l = l->_parent; // Move up 1 for me |
duke@435 | 2663 | // If this loop is not properly parented, then this loop |
duke@435 | 2664 | // has no exit path out, i.e. its an infinite loop. |
duke@435 | 2665 | if( !l ) { |
duke@435 | 2666 | // Make loop "reachable" from root so the CFG is reachable. Basically |
duke@435 | 2667 | // insert a bogus loop exit that is never taken. 'm', the loop head, |
duke@435 | 2668 | // points to 'n', one (of possibly many) fall-in paths. There may be |
duke@435 | 2669 | // many backedges as well. |
duke@435 | 2670 | |
duke@435 | 2671 | // Here I set the loop to be the root loop. I could have, after |
duke@435 | 2672 | // inserting a bogus loop exit, restarted the recursion and found my |
duke@435 | 2673 | // new loop exit. This would make the infinite loop a first-class |
duke@435 | 2674 | // loop and it would then get properly optimized. What's the use of |
duke@435 | 2675 | // optimizing an infinite loop? |
duke@435 | 2676 | l = _ltree_root; // Oops, found infinite loop |
duke@435 | 2677 | |
never@1356 | 2678 | if (!_verify_only) { |
never@1356 | 2679 | // Insert the NeverBranch between 'm' and it's control user. |
never@1356 | 2680 | NeverBranchNode *iff = new (C, 1) NeverBranchNode( m ); |
never@1356 | 2681 | _igvn.register_new_node_with_optimizer(iff); |
never@1356 | 2682 | set_loop(iff, l); |
never@1356 | 2683 | Node *if_t = new (C, 1) CProjNode( iff, 0 ); |
never@1356 | 2684 | _igvn.register_new_node_with_optimizer(if_t); |
never@1356 | 2685 | set_loop(if_t, l); |
duke@435 | 2686 | |
never@1356 | 2687 | Node* cfg = NULL; // Find the One True Control User of m |
never@1356 | 2688 | for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { |
never@1356 | 2689 | Node* x = m->fast_out(j); |
never@1356 | 2690 | if (x->is_CFG() && x != m && x != iff) |
never@1356 | 2691 | { cfg = x; break; } |
never@1356 | 2692 | } |
never@1356 | 2693 | assert(cfg != NULL, "must find the control user of m"); |
never@1356 | 2694 | uint k = 0; // Probably cfg->in(0) |
never@1356 | 2695 | while( cfg->in(k) != m ) k++; // But check incase cfg is a Region |
never@1356 | 2696 | cfg->set_req( k, if_t ); // Now point to NeverBranch |
never@1356 | 2697 | |
never@1356 | 2698 | // Now create the never-taken loop exit |
never@1356 | 2699 | Node *if_f = new (C, 1) CProjNode( iff, 1 ); |
never@1356 | 2700 | _igvn.register_new_node_with_optimizer(if_f); |
never@1356 | 2701 | set_loop(if_f, l); |
never@1356 | 2702 | // Find frame ptr for Halt. Relies on the optimizer |
never@1356 | 2703 | // V-N'ing. Easier and quicker than searching through |
never@1356 | 2704 | // the program structure. |
never@1356 | 2705 | Node *frame = new (C, 1) ParmNode( C->start(), TypeFunc::FramePtr ); |
never@1356 | 2706 | _igvn.register_new_node_with_optimizer(frame); |
never@1356 | 2707 | // Halt & Catch Fire |
never@1356 | 2708 | Node *halt = new (C, TypeFunc::Parms) HaltNode( if_f, frame ); |
never@1356 | 2709 | _igvn.register_new_node_with_optimizer(halt); |
never@1356 | 2710 | set_loop(halt, l); |
never@1356 | 2711 | C->root()->add_req(halt); |
duke@435 | 2712 | } |
duke@435 | 2713 | set_loop(C->root(), _ltree_root); |
duke@435 | 2714 | } |
duke@435 | 2715 | } |
duke@435 | 2716 | // Weeny check for irreducible. This child was already visited (this |
duke@435 | 2717 | // IS the post-work phase). Is this child's loop header post-visited |
duke@435 | 2718 | // as well? If so, then I found another entry into the loop. |
never@1356 | 2719 | if (!_verify_only) { |
never@1356 | 2720 | while( is_postvisited(l->_head) ) { |
never@1356 | 2721 | // found irreducible |
never@1356 | 2722 | l->_irreducible = 1; // = true |
never@1356 | 2723 | l = l->_parent; |
never@1356 | 2724 | _has_irreducible_loops = true; |
never@1356 | 2725 | // Check for bad CFG here to prevent crash, and bailout of compile |
never@1356 | 2726 | if (l == NULL) { |
never@1356 | 2727 | C->record_method_not_compilable("unhandled CFG detected during loop optimization"); |
never@1356 | 2728 | return pre_order; |
never@1356 | 2729 | } |
duke@435 | 2730 | } |
duke@435 | 2731 | } |
duke@435 | 2732 | |
duke@435 | 2733 | // This Node might be a decision point for loops. It is only if |
duke@435 | 2734 | // it's children belong to several different loops. The sort call |
duke@435 | 2735 | // does a trivial amount of work if there is only 1 child or all |
duke@435 | 2736 | // children belong to the same loop. If however, the children |
duke@435 | 2737 | // belong to different loops, the sort call will properly set the |
duke@435 | 2738 | // _parent pointers to show how the loops nest. |
duke@435 | 2739 | // |
duke@435 | 2740 | // In any case, it returns the tightest enclosing loop. |
duke@435 | 2741 | innermost = sort( l, innermost ); |
duke@435 | 2742 | } |
duke@435 | 2743 | |
duke@435 | 2744 | // Def-use info will have some dead stuff; dead stuff will have no |
duke@435 | 2745 | // loop decided on. |
duke@435 | 2746 | |
duke@435 | 2747 | // Am I a loop header? If so fix up my parent's child and next ptrs. |
duke@435 | 2748 | if( innermost && innermost->_head == n ) { |
duke@435 | 2749 | assert( get_loop(n) == innermost, "" ); |
duke@435 | 2750 | IdealLoopTree *p = innermost->_parent; |
duke@435 | 2751 | IdealLoopTree *l = innermost; |
duke@435 | 2752 | while( p && l->_head == n ) { |
duke@435 | 2753 | l->_next = p->_child; // Put self on parents 'next child' |
duke@435 | 2754 | p->_child = l; // Make self as first child of parent |
duke@435 | 2755 | l = p; // Now walk up the parent chain |
duke@435 | 2756 | p = l->_parent; |
duke@435 | 2757 | } |
duke@435 | 2758 | } else { |
duke@435 | 2759 | // Note that it is possible for a LoopNode to reach here, if the |
duke@435 | 2760 | // backedge has been made unreachable (hence the LoopNode no longer |
duke@435 | 2761 | // denotes a Loop, and will eventually be removed). |
duke@435 | 2762 | |
duke@435 | 2763 | // Record tightest enclosing loop for self. Mark as post-visited. |
duke@435 | 2764 | set_loop(n, innermost); |
duke@435 | 2765 | // Also record has_call flag early on |
duke@435 | 2766 | if( innermost ) { |
duke@435 | 2767 | if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) { |
duke@435 | 2768 | // Do not count uncommon calls |
duke@435 | 2769 | if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) { |
duke@435 | 2770 | Node *iff = n->in(0)->in(0); |
kvn@3882 | 2771 | // No any calls for vectorized loops. |
kvn@3882 | 2772 | if( UseSuperWord || !iff->is_If() || |
duke@435 | 2773 | (n->in(0)->Opcode() == Op_IfFalse && |
duke@435 | 2774 | (1.0 - iff->as_If()->_prob) >= 0.01) || |
duke@435 | 2775 | (iff->as_If()->_prob >= 0.01) ) |
duke@435 | 2776 | innermost->_has_call = 1; |
duke@435 | 2777 | } |
kvn@474 | 2778 | } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) { |
kvn@474 | 2779 | // Disable loop optimizations if the loop has a scalar replaceable |
kvn@474 | 2780 | // allocation. This disabling may cause a potential performance lost |
kvn@474 | 2781 | // if the allocation is not eliminated for some reason. |
kvn@474 | 2782 | innermost->_allow_optimizations = false; |
kvn@474 | 2783 | innermost->_has_call = 1; // = true |
kvn@4023 | 2784 | } else if (n->Opcode() == Op_SafePoint) { |
kvn@4023 | 2785 | // Record all safepoints in this loop. |
kvn@4023 | 2786 | if (innermost->_safepts == NULL) innermost->_safepts = new Node_List(); |
kvn@4023 | 2787 | innermost->_safepts->push(n); |
duke@435 | 2788 | } |
duke@435 | 2789 | } |
duke@435 | 2790 | } |
duke@435 | 2791 | |
duke@435 | 2792 | // Flag as post-visited now |
duke@435 | 2793 | set_postvisited(n); |
duke@435 | 2794 | return pre_order; |
duke@435 | 2795 | } |
duke@435 | 2796 | |
duke@435 | 2797 | |
duke@435 | 2798 | //------------------------------build_loop_early------------------------------- |
duke@435 | 2799 | // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. |
duke@435 | 2800 | // First pass computes the earliest controlling node possible. This is the |
duke@435 | 2801 | // controlling input with the deepest dominating depth. |
never@1356 | 2802 | void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { |
duke@435 | 2803 | while (worklist.size() != 0) { |
duke@435 | 2804 | // Use local variables nstack_top_n & nstack_top_i to cache values |
duke@435 | 2805 | // on nstack's top. |
duke@435 | 2806 | Node *nstack_top_n = worklist.pop(); |
duke@435 | 2807 | uint nstack_top_i = 0; |
duke@435 | 2808 | //while_nstack_nonempty: |
duke@435 | 2809 | while (true) { |
duke@435 | 2810 | // Get parent node and next input's index from stack's top. |
duke@435 | 2811 | Node *n = nstack_top_n; |
duke@435 | 2812 | uint i = nstack_top_i; |
duke@435 | 2813 | uint cnt = n->req(); // Count of inputs |
duke@435 | 2814 | if (i == 0) { // Pre-process the node. |
duke@435 | 2815 | if( has_node(n) && // Have either loop or control already? |
duke@435 | 2816 | !has_ctrl(n) ) { // Have loop picked out already? |
duke@435 | 2817 | // During "merge_many_backedges" we fold up several nested loops |
duke@435 | 2818 | // into a single loop. This makes the members of the original |
duke@435 | 2819 | // loop bodies pointing to dead loops; they need to move up |
duke@435 | 2820 | // to the new UNION'd larger loop. I set the _head field of these |
duke@435 | 2821 | // dead loops to NULL and the _parent field points to the owning |
duke@435 | 2822 | // loop. Shades of UNION-FIND algorithm. |
duke@435 | 2823 | IdealLoopTree *ilt; |
duke@435 | 2824 | while( !(ilt = get_loop(n))->_head ) { |
duke@435 | 2825 | // Normally I would use a set_loop here. But in this one special |
duke@435 | 2826 | // case, it is legal (and expected) to change what loop a Node |
duke@435 | 2827 | // belongs to. |
duke@435 | 2828 | _nodes.map(n->_idx, (Node*)(ilt->_parent) ); |
duke@435 | 2829 | } |
duke@435 | 2830 | // Remove safepoints ONLY if I've already seen I don't need one. |
duke@435 | 2831 | // (the old code here would yank a 2nd safepoint after seeing a |
duke@435 | 2832 | // first one, even though the 1st did not dominate in the loop body |
duke@435 | 2833 | // and thus could be avoided indefinitely) |
never@1356 | 2834 | if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint && |
duke@435 | 2835 | is_deleteable_safept(n)) { |
duke@435 | 2836 | Node *in = n->in(TypeFunc::Control); |
duke@435 | 2837 | lazy_replace(n,in); // Pull safepoint now |
kvn@4023 | 2838 | if (ilt->_safepts != NULL) { |
kvn@4023 | 2839 | ilt->_safepts->yank(n); |
kvn@4023 | 2840 | } |
duke@435 | 2841 | // Carry on with the recursion "as if" we are walking |
duke@435 | 2842 | // only the control input |
duke@435 | 2843 | if( !visited.test_set( in->_idx ) ) { |
duke@435 | 2844 | worklist.push(in); // Visit this guy later, using worklist |
duke@435 | 2845 | } |
duke@435 | 2846 | // Get next node from nstack: |
duke@435 | 2847 | // - skip n's inputs processing by setting i > cnt; |
duke@435 | 2848 | // - we also will not call set_early_ctrl(n) since |
duke@435 | 2849 | // has_node(n) == true (see the condition above). |
duke@435 | 2850 | i = cnt + 1; |
duke@435 | 2851 | } |
duke@435 | 2852 | } |
duke@435 | 2853 | } // if (i == 0) |
duke@435 | 2854 | |
duke@435 | 2855 | // Visit all inputs |
duke@435 | 2856 | bool done = true; // Assume all n's inputs will be processed |
duke@435 | 2857 | while (i < cnt) { |
duke@435 | 2858 | Node *in = n->in(i); |
duke@435 | 2859 | ++i; |
duke@435 | 2860 | if (in == NULL) continue; |
duke@435 | 2861 | if (in->pinned() && !in->is_CFG()) |
duke@435 | 2862 | set_ctrl(in, in->in(0)); |
duke@435 | 2863 | int is_visited = visited.test_set( in->_idx ); |
duke@435 | 2864 | if (!has_node(in)) { // No controlling input yet? |
duke@435 | 2865 | assert( !in->is_CFG(), "CFG Node with no controlling input?" ); |
duke@435 | 2866 | assert( !is_visited, "visit only once" ); |
duke@435 | 2867 | nstack.push(n, i); // Save parent node and next input's index. |
duke@435 | 2868 | nstack_top_n = in; // Process current input now. |
duke@435 | 2869 | nstack_top_i = 0; |
duke@435 | 2870 | done = false; // Not all n's inputs processed. |
duke@435 | 2871 | break; // continue while_nstack_nonempty; |
duke@435 | 2872 | } else if (!is_visited) { |
duke@435 | 2873 | // This guy has a location picked out for him, but has not yet |
duke@435 | 2874 | // been visited. Happens to all CFG nodes, for instance. |
duke@435 | 2875 | // Visit him using the worklist instead of recursion, to break |
duke@435 | 2876 | // cycles. Since he has a location already we do not need to |
duke@435 | 2877 | // find his location before proceeding with the current Node. |
duke@435 | 2878 | worklist.push(in); // Visit this guy later, using worklist |
duke@435 | 2879 | } |
duke@435 | 2880 | } |
duke@435 | 2881 | if (done) { |
duke@435 | 2882 | // All of n's inputs have been processed, complete post-processing. |
duke@435 | 2883 | |
twisti@1040 | 2884 | // Compute earliest point this Node can go. |
duke@435 | 2885 | // CFG, Phi, pinned nodes already know their controlling input. |
duke@435 | 2886 | if (!has_node(n)) { |
duke@435 | 2887 | // Record earliest legal location |
duke@435 | 2888 | set_early_ctrl( n ); |
duke@435 | 2889 | } |
duke@435 | 2890 | if (nstack.is_empty()) { |
duke@435 | 2891 | // Finished all nodes on stack. |
duke@435 | 2892 | // Process next node on the worklist. |
duke@435 | 2893 | break; |
duke@435 | 2894 | } |
duke@435 | 2895 | // Get saved parent node and next input's index. |
duke@435 | 2896 | nstack_top_n = nstack.node(); |
duke@435 | 2897 | nstack_top_i = nstack.index(); |
duke@435 | 2898 | nstack.pop(); |
duke@435 | 2899 | } |
duke@435 | 2900 | } // while (true) |
duke@435 | 2901 | } |
duke@435 | 2902 | } |
duke@435 | 2903 | |
duke@435 | 2904 | //------------------------------dom_lca_internal-------------------------------- |
duke@435 | 2905 | // Pair-wise LCA |
duke@435 | 2906 | Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const { |
duke@435 | 2907 | if( !n1 ) return n2; // Handle NULL original LCA |
duke@435 | 2908 | assert( n1->is_CFG(), "" ); |
duke@435 | 2909 | assert( n2->is_CFG(), "" ); |
duke@435 | 2910 | // find LCA of all uses |
duke@435 | 2911 | uint d1 = dom_depth(n1); |
duke@435 | 2912 | uint d2 = dom_depth(n2); |
duke@435 | 2913 | while (n1 != n2) { |
duke@435 | 2914 | if (d1 > d2) { |
duke@435 | 2915 | n1 = idom(n1); |
duke@435 | 2916 | d1 = dom_depth(n1); |
duke@435 | 2917 | } else if (d1 < d2) { |
duke@435 | 2918 | n2 = idom(n2); |
duke@435 | 2919 | d2 = dom_depth(n2); |
duke@435 | 2920 | } else { |
duke@435 | 2921 | // Here d1 == d2. Due to edits of the dominator-tree, sections |
duke@435 | 2922 | // of the tree might have the same depth. These sections have |
duke@435 | 2923 | // to be searched more carefully. |
duke@435 | 2924 | |
duke@435 | 2925 | // Scan up all the n1's with equal depth, looking for n2. |
duke@435 | 2926 | Node *t1 = idom(n1); |
duke@435 | 2927 | while (dom_depth(t1) == d1) { |
duke@435 | 2928 | if (t1 == n2) return n2; |
duke@435 | 2929 | t1 = idom(t1); |
duke@435 | 2930 | } |
duke@435 | 2931 | // Scan up all the n2's with equal depth, looking for n1. |
duke@435 | 2932 | Node *t2 = idom(n2); |
duke@435 | 2933 | while (dom_depth(t2) == d2) { |
duke@435 | 2934 | if (t2 == n1) return n1; |
duke@435 | 2935 | t2 = idom(t2); |
duke@435 | 2936 | } |
duke@435 | 2937 | // Move up to a new dominator-depth value as well as up the dom-tree. |
duke@435 | 2938 | n1 = t1; |
duke@435 | 2939 | n2 = t2; |
duke@435 | 2940 | d1 = dom_depth(n1); |
duke@435 | 2941 | d2 = dom_depth(n2); |
duke@435 | 2942 | } |
duke@435 | 2943 | } |
duke@435 | 2944 | return n1; |
duke@435 | 2945 | } |
duke@435 | 2946 | |
duke@435 | 2947 | //------------------------------compute_idom----------------------------------- |
duke@435 | 2948 | // Locally compute IDOM using dom_lca call. Correct only if the incoming |
duke@435 | 2949 | // IDOMs are correct. |
duke@435 | 2950 | Node *PhaseIdealLoop::compute_idom( Node *region ) const { |
duke@435 | 2951 | assert( region->is_Region(), "" ); |
duke@435 | 2952 | Node *LCA = NULL; |
duke@435 | 2953 | for( uint i = 1; i < region->req(); i++ ) { |
duke@435 | 2954 | if( region->in(i) != C->top() ) |
duke@435 | 2955 | LCA = dom_lca( LCA, region->in(i) ); |
duke@435 | 2956 | } |
duke@435 | 2957 | return LCA; |
duke@435 | 2958 | } |
duke@435 | 2959 | |
never@1356 | 2960 | bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) { |
never@1356 | 2961 | bool had_error = false; |
never@1356 | 2962 | #ifdef ASSERT |
never@1356 | 2963 | if (early != C->root()) { |
never@1356 | 2964 | // Make sure that there's a dominance path from use to LCA |
never@1356 | 2965 | Node* d = use; |
never@1356 | 2966 | while (d != LCA) { |
never@1356 | 2967 | d = idom(d); |
never@1356 | 2968 | if (d == C->root()) { |
never@1356 | 2969 | tty->print_cr("*** Use %d isn't dominated by def %s", use->_idx, n->_idx); |
never@1356 | 2970 | n->dump(); |
never@1356 | 2971 | use->dump(); |
never@1356 | 2972 | had_error = true; |
never@1356 | 2973 | break; |
never@1356 | 2974 | } |
never@1356 | 2975 | } |
never@1356 | 2976 | } |
never@1356 | 2977 | #endif |
never@1356 | 2978 | return had_error; |
never@1356 | 2979 | } |
duke@435 | 2980 | |
never@1356 | 2981 | |
never@1356 | 2982 | Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) { |
duke@435 | 2983 | // Compute LCA over list of uses |
never@1356 | 2984 | bool had_error = false; |
duke@435 | 2985 | Node *LCA = NULL; |
duke@435 | 2986 | for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) { |
duke@435 | 2987 | Node* c = n->fast_out(i); |
duke@435 | 2988 | if (_nodes[c->_idx] == NULL) |
duke@435 | 2989 | continue; // Skip the occasional dead node |
duke@435 | 2990 | if( c->is_Phi() ) { // For Phis, we must land above on the path |
duke@435 | 2991 | for( uint j=1; j<c->req(); j++ ) {// For all inputs |
duke@435 | 2992 | if( c->in(j) == n ) { // Found matching input? |
duke@435 | 2993 | Node *use = c->in(0)->in(j); |
never@1356 | 2994 | if (_verify_only && use->is_top()) continue; |
duke@435 | 2995 | LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); |
never@1356 | 2996 | if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error; |
duke@435 | 2997 | } |
duke@435 | 2998 | } |
duke@435 | 2999 | } else { |
duke@435 | 3000 | // For CFG data-users, use is in the block just prior |
duke@435 | 3001 | Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0); |
duke@435 | 3002 | LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); |
never@1356 | 3003 | if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error; |
duke@435 | 3004 | } |
duke@435 | 3005 | } |
never@1356 | 3006 | assert(!had_error, "bad dominance"); |
never@1356 | 3007 | return LCA; |
never@1356 | 3008 | } |
never@1356 | 3009 | |
never@1356 | 3010 | //------------------------------get_late_ctrl---------------------------------- |
never@1356 | 3011 | // Compute latest legal control. |
never@1356 | 3012 | Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) { |
never@1356 | 3013 | assert(early != NULL, "early control should not be NULL"); |
never@1356 | 3014 | |
never@1356 | 3015 | Node* LCA = compute_lca_of_uses(n, early); |
never@1356 | 3016 | #ifdef ASSERT |
never@1356 | 3017 | if (LCA == C->root() && LCA != early) { |
never@1356 | 3018 | // def doesn't dominate uses so print some useful debugging output |
never@1356 | 3019 | compute_lca_of_uses(n, early, true); |
never@1356 | 3020 | } |
never@1356 | 3021 | #endif |
duke@435 | 3022 | |
duke@435 | 3023 | // if this is a load, check for anti-dependent stores |
duke@435 | 3024 | // We use a conservative algorithm to identify potential interfering |
duke@435 | 3025 | // instructions and for rescheduling the load. The users of the memory |
duke@435 | 3026 | // input of this load are examined. Any use which is not a load and is |
duke@435 | 3027 | // dominated by early is considered a potentially interfering store. |
duke@435 | 3028 | // This can produce false positives. |
duke@435 | 3029 | if (n->is_Load() && LCA != early) { |
duke@435 | 3030 | Node_List worklist; |
duke@435 | 3031 | |
duke@435 | 3032 | Node *mem = n->in(MemNode::Memory); |
duke@435 | 3033 | for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { |
duke@435 | 3034 | Node* s = mem->fast_out(i); |
duke@435 | 3035 | worklist.push(s); |
duke@435 | 3036 | } |
duke@435 | 3037 | while(worklist.size() != 0 && LCA != early) { |
duke@435 | 3038 | Node* s = worklist.pop(); |
duke@435 | 3039 | if (s->is_Load()) { |
duke@435 | 3040 | continue; |
duke@435 | 3041 | } else if (s->is_MergeMem()) { |
duke@435 | 3042 | for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) { |
duke@435 | 3043 | Node* s1 = s->fast_out(i); |
duke@435 | 3044 | worklist.push(s1); |
duke@435 | 3045 | } |
duke@435 | 3046 | } else { |
duke@435 | 3047 | Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0); |
duke@435 | 3048 | assert(sctrl != NULL || s->outcnt() == 0, "must have control"); |
duke@435 | 3049 | if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) { |
duke@435 | 3050 | LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n); |
duke@435 | 3051 | } |
duke@435 | 3052 | } |
duke@435 | 3053 | } |
duke@435 | 3054 | } |
duke@435 | 3055 | |
duke@435 | 3056 | assert(LCA == find_non_split_ctrl(LCA), "unexpected late control"); |
duke@435 | 3057 | return LCA; |
duke@435 | 3058 | } |
duke@435 | 3059 | |
duke@435 | 3060 | // true if CFG node d dominates CFG node n |
duke@435 | 3061 | bool PhaseIdealLoop::is_dominator(Node *d, Node *n) { |
duke@435 | 3062 | if (d == n) |
duke@435 | 3063 | return true; |
duke@435 | 3064 | assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); |
duke@435 | 3065 | uint dd = dom_depth(d); |
duke@435 | 3066 | while (dom_depth(n) >= dd) { |
duke@435 | 3067 | if (n == d) |
duke@435 | 3068 | return true; |
duke@435 | 3069 | n = idom(n); |
duke@435 | 3070 | } |
duke@435 | 3071 | return false; |
duke@435 | 3072 | } |
duke@435 | 3073 | |
duke@435 | 3074 | //------------------------------dom_lca_for_get_late_ctrl_internal------------- |
duke@435 | 3075 | // Pair-wise LCA with tags. |
duke@435 | 3076 | // Tag each index with the node 'tag' currently being processed |
duke@435 | 3077 | // before advancing up the dominator chain using idom(). |
duke@435 | 3078 | // Later calls that find a match to 'tag' know that this path has already |
duke@435 | 3079 | // been considered in the current LCA (which is input 'n1' by convention). |
duke@435 | 3080 | // Since get_late_ctrl() is only called once for each node, the tag array |
duke@435 | 3081 | // does not need to be cleared between calls to get_late_ctrl(). |
duke@435 | 3082 | // Algorithm trades a larger constant factor for better asymptotic behavior |
duke@435 | 3083 | // |
duke@435 | 3084 | Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) { |
duke@435 | 3085 | uint d1 = dom_depth(n1); |
duke@435 | 3086 | uint d2 = dom_depth(n2); |
duke@435 | 3087 | |
duke@435 | 3088 | do { |
duke@435 | 3089 | if (d1 > d2) { |
duke@435 | 3090 | // current lca is deeper than n2 |
duke@435 | 3091 | _dom_lca_tags.map(n1->_idx, tag); |
duke@435 | 3092 | n1 = idom(n1); |
duke@435 | 3093 | d1 = dom_depth(n1); |
duke@435 | 3094 | } else if (d1 < d2) { |
duke@435 | 3095 | // n2 is deeper than current lca |
duke@435 | 3096 | Node *memo = _dom_lca_tags[n2->_idx]; |
duke@435 | 3097 | if( memo == tag ) { |
duke@435 | 3098 | return n1; // Return the current LCA |
duke@435 | 3099 | } |
duke@435 | 3100 | _dom_lca_tags.map(n2->_idx, tag); |
duke@435 | 3101 | n2 = idom(n2); |
duke@435 | 3102 | d2 = dom_depth(n2); |
duke@435 | 3103 | } else { |
duke@435 | 3104 | // Here d1 == d2. Due to edits of the dominator-tree, sections |
duke@435 | 3105 | // of the tree might have the same depth. These sections have |
duke@435 | 3106 | // to be searched more carefully. |
duke@435 | 3107 | |
duke@435 | 3108 | // Scan up all the n1's with equal depth, looking for n2. |
duke@435 | 3109 | _dom_lca_tags.map(n1->_idx, tag); |
duke@435 | 3110 | Node *t1 = idom(n1); |
duke@435 | 3111 | while (dom_depth(t1) == d1) { |
duke@435 | 3112 | if (t1 == n2) return n2; |
duke@435 | 3113 | _dom_lca_tags.map(t1->_idx, tag); |
duke@435 | 3114 | t1 = idom(t1); |
duke@435 | 3115 | } |
duke@435 | 3116 | // Scan up all the n2's with equal depth, looking for n1. |
duke@435 | 3117 | _dom_lca_tags.map(n2->_idx, tag); |
duke@435 | 3118 | Node *t2 = idom(n2); |
duke@435 | 3119 | while (dom_depth(t2) == d2) { |
duke@435 | 3120 | if (t2 == n1) return n1; |
duke@435 | 3121 | _dom_lca_tags.map(t2->_idx, tag); |
duke@435 | 3122 | t2 = idom(t2); |
duke@435 | 3123 | } |
duke@435 | 3124 | // Move up to a new dominator-depth value as well as up the dom-tree. |
duke@435 | 3125 | n1 = t1; |
duke@435 | 3126 | n2 = t2; |
duke@435 | 3127 | d1 = dom_depth(n1); |
duke@435 | 3128 | d2 = dom_depth(n2); |
duke@435 | 3129 | } |
duke@435 | 3130 | } while (n1 != n2); |
duke@435 | 3131 | return n1; |
duke@435 | 3132 | } |
duke@435 | 3133 | |
duke@435 | 3134 | //------------------------------init_dom_lca_tags------------------------------ |
duke@435 | 3135 | // Tag could be a node's integer index, 32bits instead of 64bits in some cases |
duke@435 | 3136 | // Intended use does not involve any growth for the array, so it could |
duke@435 | 3137 | // be of fixed size. |
duke@435 | 3138 | void PhaseIdealLoop::init_dom_lca_tags() { |
duke@435 | 3139 | uint limit = C->unique() + 1; |
duke@435 | 3140 | _dom_lca_tags.map( limit, NULL ); |
duke@435 | 3141 | #ifdef ASSERT |
duke@435 | 3142 | for( uint i = 0; i < limit; ++i ) { |
duke@435 | 3143 | assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); |
duke@435 | 3144 | } |
duke@435 | 3145 | #endif // ASSERT |
duke@435 | 3146 | } |
duke@435 | 3147 | |
duke@435 | 3148 | //------------------------------clear_dom_lca_tags------------------------------ |
duke@435 | 3149 | // Tag could be a node's integer index, 32bits instead of 64bits in some cases |
duke@435 | 3150 | // Intended use does not involve any growth for the array, so it could |
duke@435 | 3151 | // be of fixed size. |
duke@435 | 3152 | void PhaseIdealLoop::clear_dom_lca_tags() { |
duke@435 | 3153 | uint limit = C->unique() + 1; |
duke@435 | 3154 | _dom_lca_tags.map( limit, NULL ); |
duke@435 | 3155 | _dom_lca_tags.clear(); |
duke@435 | 3156 | #ifdef ASSERT |
duke@435 | 3157 | for( uint i = 0; i < limit; ++i ) { |
duke@435 | 3158 | assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); |
duke@435 | 3159 | } |
duke@435 | 3160 | #endif // ASSERT |
duke@435 | 3161 | } |
duke@435 | 3162 | |
duke@435 | 3163 | //------------------------------build_loop_late-------------------------------- |
duke@435 | 3164 | // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. |
duke@435 | 3165 | // Second pass finds latest legal placement, and ideal loop placement. |
never@1356 | 3166 | void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { |
duke@435 | 3167 | while (worklist.size() != 0) { |
duke@435 | 3168 | Node *n = worklist.pop(); |
duke@435 | 3169 | // Only visit once |
duke@435 | 3170 | if (visited.test_set(n->_idx)) continue; |
duke@435 | 3171 | uint cnt = n->outcnt(); |
duke@435 | 3172 | uint i = 0; |
duke@435 | 3173 | while (true) { |
duke@435 | 3174 | assert( _nodes[n->_idx], "no dead nodes" ); |
duke@435 | 3175 | // Visit all children |
duke@435 | 3176 | if (i < cnt) { |
duke@435 | 3177 | Node* use = n->raw_out(i); |
duke@435 | 3178 | ++i; |
duke@435 | 3179 | // Check for dead uses. Aggressively prune such junk. It might be |
duke@435 | 3180 | // dead in the global sense, but still have local uses so I cannot |
duke@435 | 3181 | // easily call 'remove_dead_node'. |
duke@435 | 3182 | if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead? |
duke@435 | 3183 | // Due to cycles, we might not hit the same fixed point in the verify |
duke@435 | 3184 | // pass as we do in the regular pass. Instead, visit such phis as |
duke@435 | 3185 | // simple uses of the loop head. |
duke@435 | 3186 | if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) { |
duke@435 | 3187 | if( !visited.test(use->_idx) ) |
duke@435 | 3188 | worklist.push(use); |
duke@435 | 3189 | } else if( !visited.test_set(use->_idx) ) { |
duke@435 | 3190 | nstack.push(n, i); // Save parent and next use's index. |
duke@435 | 3191 | n = use; // Process all children of current use. |
duke@435 | 3192 | cnt = use->outcnt(); |
duke@435 | 3193 | i = 0; |
duke@435 | 3194 | } |
duke@435 | 3195 | } else { |
duke@435 | 3196 | // Do not visit around the backedge of loops via data edges. |
duke@435 | 3197 | // push dead code onto a worklist |
duke@435 | 3198 | _deadlist.push(use); |
duke@435 | 3199 | } |
duke@435 | 3200 | } else { |
duke@435 | 3201 | // All of n's children have been processed, complete post-processing. |
never@1356 | 3202 | build_loop_late_post(n); |
duke@435 | 3203 | if (nstack.is_empty()) { |
duke@435 | 3204 | // Finished all nodes on stack. |
duke@435 | 3205 | // Process next node on the worklist. |
duke@435 | 3206 | break; |
duke@435 | 3207 | } |
duke@435 | 3208 | // Get saved parent node and next use's index. Visit the rest of uses. |
duke@435 | 3209 | n = nstack.node(); |
duke@435 | 3210 | cnt = n->outcnt(); |
duke@435 | 3211 | i = nstack.index(); |
duke@435 | 3212 | nstack.pop(); |
duke@435 | 3213 | } |
duke@435 | 3214 | } |
duke@435 | 3215 | } |
duke@435 | 3216 | } |
duke@435 | 3217 | |
duke@435 | 3218 | //------------------------------build_loop_late_post--------------------------- |
duke@435 | 3219 | // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. |
duke@435 | 3220 | // Second pass finds latest legal placement, and ideal loop placement. |
never@1356 | 3221 | void PhaseIdealLoop::build_loop_late_post( Node *n ) { |
duke@435 | 3222 | |
never@1356 | 3223 | if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress() && !_verify_only) { |
duke@435 | 3224 | _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops. |
duke@435 | 3225 | } |
duke@435 | 3226 | |
duke@435 | 3227 | // CFG and pinned nodes already handled |
duke@435 | 3228 | if( n->in(0) ) { |
duke@435 | 3229 | if( n->in(0)->is_top() ) return; // Dead? |
duke@435 | 3230 | |
duke@435 | 3231 | // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads |
duke@435 | 3232 | // _must_ be pinned (they have to observe their control edge of course). |
duke@435 | 3233 | // Unlike Stores (which modify an unallocable resource, the memory |
duke@435 | 3234 | // state), Mods/Loads can float around. So free them up. |
duke@435 | 3235 | bool pinned = true; |
duke@435 | 3236 | switch( n->Opcode() ) { |
duke@435 | 3237 | case Op_DivI: |
duke@435 | 3238 | case Op_DivF: |
duke@435 | 3239 | case Op_DivD: |
duke@435 | 3240 | case Op_ModI: |
duke@435 | 3241 | case Op_ModF: |
duke@435 | 3242 | case Op_ModD: |
duke@435 | 3243 | case Op_LoadB: // Same with Loads; they can sink |
kvn@3882 | 3244 | case Op_LoadUB: // during loop optimizations. |
kvn@3882 | 3245 | case Op_LoadUS: |
duke@435 | 3246 | case Op_LoadD: |
duke@435 | 3247 | case Op_LoadF: |
duke@435 | 3248 | case Op_LoadI: |
duke@435 | 3249 | case Op_LoadKlass: |
kvn@728 | 3250 | case Op_LoadNKlass: |
duke@435 | 3251 | case Op_LoadL: |
duke@435 | 3252 | case Op_LoadS: |
duke@435 | 3253 | case Op_LoadP: |
kvn@728 | 3254 | case Op_LoadN: |
duke@435 | 3255 | case Op_LoadRange: |
duke@435 | 3256 | case Op_LoadD_unaligned: |
duke@435 | 3257 | case Op_LoadL_unaligned: |
duke@435 | 3258 | case Op_StrComp: // Does a bunch of load-like effects |
cfang@1116 | 3259 | case Op_StrEquals: |
cfang@1116 | 3260 | case Op_StrIndexOf: |
rasbold@604 | 3261 | case Op_AryEq: |
duke@435 | 3262 | pinned = false; |
duke@435 | 3263 | } |
duke@435 | 3264 | if( pinned ) { |
twisti@1040 | 3265 | IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n)); |
twisti@1040 | 3266 | if( !chosen_loop->_child ) // Inner loop? |
twisti@1040 | 3267 | chosen_loop->_body.push(n); // Collect inner loops |
duke@435 | 3268 | return; |
duke@435 | 3269 | } |
duke@435 | 3270 | } else { // No slot zero |
duke@435 | 3271 | if( n->is_CFG() ) { // CFG with no slot 0 is dead |
duke@435 | 3272 | _nodes.map(n->_idx,0); // No block setting, it's globally dead |
duke@435 | 3273 | return; |
duke@435 | 3274 | } |
duke@435 | 3275 | assert(!n->is_CFG() || n->outcnt() == 0, ""); |
duke@435 | 3276 | } |
duke@435 | 3277 | |
duke@435 | 3278 | // Do I have a "safe range" I can select over? |
duke@435 | 3279 | Node *early = get_ctrl(n);// Early location already computed |
duke@435 | 3280 | |
duke@435 | 3281 | // Compute latest point this Node can go |
duke@435 | 3282 | Node *LCA = get_late_ctrl( n, early ); |
duke@435 | 3283 | // LCA is NULL due to uses being dead |
duke@435 | 3284 | if( LCA == NULL ) { |
duke@435 | 3285 | #ifdef ASSERT |
duke@435 | 3286 | for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) { |
duke@435 | 3287 | assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead"); |
duke@435 | 3288 | } |
duke@435 | 3289 | #endif |
duke@435 | 3290 | _nodes.map(n->_idx, 0); // This node is useless |
duke@435 | 3291 | _deadlist.push(n); |
duke@435 | 3292 | return; |
duke@435 | 3293 | } |
duke@435 | 3294 | assert(LCA != NULL && !LCA->is_top(), "no dead nodes"); |
duke@435 | 3295 | |
duke@435 | 3296 | Node *legal = LCA; // Walk 'legal' up the IDOM chain |
duke@435 | 3297 | Node *least = legal; // Best legal position so far |
duke@435 | 3298 | while( early != legal ) { // While not at earliest legal |
cfang@1607 | 3299 | #ifdef ASSERT |
cfang@1607 | 3300 | if (legal->is_Start() && !early->is_Root()) { |
cfang@1607 | 3301 | // Bad graph. Print idom path and fail. |
kvn@3408 | 3302 | dump_bad_graph(n, early, LCA); |
cfang@1607 | 3303 | assert(false, "Bad graph detected in build_loop_late"); |
cfang@1607 | 3304 | } |
cfang@1607 | 3305 | #endif |
duke@435 | 3306 | // Find least loop nesting depth |
duke@435 | 3307 | legal = idom(legal); // Bump up the IDOM tree |
duke@435 | 3308 | // Check for lower nesting depth |
duke@435 | 3309 | if( get_loop(legal)->_nest < get_loop(least)->_nest ) |
duke@435 | 3310 | least = legal; |
duke@435 | 3311 | } |
never@1356 | 3312 | assert(early == legal || legal != C->root(), "bad dominance of inputs"); |
duke@435 | 3313 | |
duke@435 | 3314 | // Try not to place code on a loop entry projection |
duke@435 | 3315 | // which can inhibit range check elimination. |
duke@435 | 3316 | if (least != early) { |
duke@435 | 3317 | Node* ctrl_out = least->unique_ctrl_out(); |
duke@435 | 3318 | if (ctrl_out && ctrl_out->is_CountedLoop() && |
duke@435 | 3319 | least == ctrl_out->in(LoopNode::EntryControl)) { |
duke@435 | 3320 | Node* least_dom = idom(least); |
duke@435 | 3321 | if (get_loop(least_dom)->is_member(get_loop(least))) { |
duke@435 | 3322 | least = least_dom; |
duke@435 | 3323 | } |
duke@435 | 3324 | } |
duke@435 | 3325 | } |
duke@435 | 3326 | |
duke@435 | 3327 | #ifdef ASSERT |
duke@435 | 3328 | // If verifying, verify that 'verify_me' has a legal location |
duke@435 | 3329 | // and choose it as our location. |
never@1356 | 3330 | if( _verify_me ) { |
never@1356 | 3331 | Node *v_ctrl = _verify_me->get_ctrl_no_update(n); |
duke@435 | 3332 | Node *legal = LCA; |
duke@435 | 3333 | while( early != legal ) { // While not at earliest legal |
duke@435 | 3334 | if( legal == v_ctrl ) break; // Check for prior good location |
duke@435 | 3335 | legal = idom(legal) ;// Bump up the IDOM tree |
duke@435 | 3336 | } |
duke@435 | 3337 | // Check for prior good location |
duke@435 | 3338 | if( legal == v_ctrl ) least = legal; // Keep prior if found |
duke@435 | 3339 | } |
duke@435 | 3340 | #endif |
duke@435 | 3341 | |
duke@435 | 3342 | // Assign discovered "here or above" point |
duke@435 | 3343 | least = find_non_split_ctrl(least); |
duke@435 | 3344 | set_ctrl(n, least); |
duke@435 | 3345 | |
duke@435 | 3346 | // Collect inner loop bodies |
twisti@1040 | 3347 | IdealLoopTree *chosen_loop = get_loop(least); |
twisti@1040 | 3348 | if( !chosen_loop->_child ) // Inner loop? |
twisti@1040 | 3349 | chosen_loop->_body.push(n);// Collect inner loops |
duke@435 | 3350 | } |
duke@435 | 3351 | |
kvn@3408 | 3352 | #ifdef ASSERT |
kvn@3408 | 3353 | void PhaseIdealLoop::dump_bad_graph(Node* n, Node* early, Node* LCA) { |
kvn@3408 | 3354 | tty->print_cr( "Bad graph detected in build_loop_late"); |
kvn@3408 | 3355 | tty->print("n: "); n->dump(); |
kvn@3408 | 3356 | tty->print("early(n): "); early->dump(); |
kvn@3408 | 3357 | if (n->in(0) != NULL && !n->in(0)->is_top() && |
kvn@3408 | 3358 | n->in(0) != early && !n->in(0)->is_Root()) { |
kvn@3408 | 3359 | tty->print("n->in(0): "); n->in(0)->dump(); |
kvn@3408 | 3360 | } |
kvn@3408 | 3361 | for (uint i = 1; i < n->req(); i++) { |
kvn@3408 | 3362 | Node* in1 = n->in(i); |
kvn@3408 | 3363 | if (in1 != NULL && in1 != n && !in1->is_top()) { |
kvn@3408 | 3364 | tty->print("n->in(%d): ", i); in1->dump(); |
kvn@3408 | 3365 | Node* in1_early = get_ctrl(in1); |
kvn@3408 | 3366 | tty->print("early(n->in(%d)): ", i); in1_early->dump(); |
kvn@3408 | 3367 | if (in1->in(0) != NULL && !in1->in(0)->is_top() && |
kvn@3408 | 3368 | in1->in(0) != in1_early && !in1->in(0)->is_Root()) { |
kvn@3408 | 3369 | tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump(); |
kvn@3408 | 3370 | } |
kvn@3408 | 3371 | for (uint j = 1; j < in1->req(); j++) { |
kvn@3408 | 3372 | Node* in2 = in1->in(j); |
kvn@3408 | 3373 | if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) { |
kvn@3408 | 3374 | tty->print("n->in(%d)->in(%d): ", i, j); in2->dump(); |
kvn@3408 | 3375 | Node* in2_early = get_ctrl(in2); |
kvn@3408 | 3376 | tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump(); |
kvn@3408 | 3377 | if (in2->in(0) != NULL && !in2->in(0)->is_top() && |
kvn@3408 | 3378 | in2->in(0) != in2_early && !in2->in(0)->is_Root()) { |
kvn@3408 | 3379 | tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump(); |
kvn@3408 | 3380 | } |
kvn@3408 | 3381 | } |
kvn@3408 | 3382 | } |
kvn@3408 | 3383 | } |
kvn@3408 | 3384 | } |
kvn@3408 | 3385 | tty->cr(); |
kvn@3408 | 3386 | tty->print("LCA(n): "); LCA->dump(); |
kvn@3408 | 3387 | for (uint i = 0; i < n->outcnt(); i++) { |
kvn@3408 | 3388 | Node* u1 = n->raw_out(i); |
kvn@3408 | 3389 | if (u1 == n) |
kvn@3408 | 3390 | continue; |
kvn@3408 | 3391 | tty->print("n->out(%d): ", i); u1->dump(); |
kvn@3408 | 3392 | if (u1->is_CFG()) { |
kvn@3408 | 3393 | for (uint j = 0; j < u1->outcnt(); j++) { |
kvn@3408 | 3394 | Node* u2 = u1->raw_out(j); |
kvn@3408 | 3395 | if (u2 != u1 && u2 != n && u2->is_CFG()) { |
kvn@3408 | 3396 | tty->print("n->out(%d)->out(%d): ", i, j); u2->dump(); |
kvn@3408 | 3397 | } |
kvn@3408 | 3398 | } |
kvn@3408 | 3399 | } else { |
kvn@3408 | 3400 | Node* u1_later = get_ctrl(u1); |
kvn@3408 | 3401 | tty->print("later(n->out(%d)): ", i); u1_later->dump(); |
kvn@3408 | 3402 | if (u1->in(0) != NULL && !u1->in(0)->is_top() && |
kvn@3408 | 3403 | u1->in(0) != u1_later && !u1->in(0)->is_Root()) { |
kvn@3408 | 3404 | tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump(); |
kvn@3408 | 3405 | } |
kvn@3408 | 3406 | for (uint j = 0; j < u1->outcnt(); j++) { |
kvn@3408 | 3407 | Node* u2 = u1->raw_out(j); |
kvn@3408 | 3408 | if (u2 == n || u2 == u1) |
kvn@3408 | 3409 | continue; |
kvn@3408 | 3410 | tty->print("n->out(%d)->out(%d): ", i, j); u2->dump(); |
kvn@3408 | 3411 | if (!u2->is_CFG()) { |
kvn@3408 | 3412 | Node* u2_later = get_ctrl(u2); |
kvn@3408 | 3413 | tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump(); |
kvn@3408 | 3414 | if (u2->in(0) != NULL && !u2->in(0)->is_top() && |
kvn@3408 | 3415 | u2->in(0) != u2_later && !u2->in(0)->is_Root()) { |
kvn@3408 | 3416 | tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump(); |
kvn@3408 | 3417 | } |
kvn@3408 | 3418 | } |
kvn@3408 | 3419 | } |
kvn@3408 | 3420 | } |
kvn@3408 | 3421 | } |
kvn@3408 | 3422 | tty->cr(); |
kvn@3408 | 3423 | int ct = 0; |
kvn@3408 | 3424 | Node *dbg_legal = LCA; |
kvn@3408 | 3425 | while(!dbg_legal->is_Start() && ct < 100) { |
kvn@3408 | 3426 | tty->print("idom[%d] ",ct); dbg_legal->dump(); |
kvn@3408 | 3427 | ct++; |
kvn@3408 | 3428 | dbg_legal = idom(dbg_legal); |
kvn@3408 | 3429 | } |
kvn@3408 | 3430 | tty->cr(); |
kvn@3408 | 3431 | } |
kvn@3408 | 3432 | #endif |
kvn@3408 | 3433 | |
duke@435 | 3434 | #ifndef PRODUCT |
duke@435 | 3435 | //------------------------------dump------------------------------------------- |
duke@435 | 3436 | void PhaseIdealLoop::dump( ) const { |
duke@435 | 3437 | ResourceMark rm; |
duke@435 | 3438 | Arena* arena = Thread::current()->resource_area(); |
duke@435 | 3439 | Node_Stack stack(arena, C->unique() >> 2); |
duke@435 | 3440 | Node_List rpo_list; |
duke@435 | 3441 | VectorSet visited(arena); |
duke@435 | 3442 | visited.set(C->top()->_idx); |
duke@435 | 3443 | rpo( C->root(), stack, visited, rpo_list ); |
duke@435 | 3444 | // Dump root loop indexed by last element in PO order |
duke@435 | 3445 | dump( _ltree_root, rpo_list.size(), rpo_list ); |
duke@435 | 3446 | } |
duke@435 | 3447 | |
duke@435 | 3448 | void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const { |
never@802 | 3449 | loop->dump_head(); |
duke@435 | 3450 | |
duke@435 | 3451 | // Now scan for CFG nodes in the same loop |
duke@435 | 3452 | for( uint j=idx; j > 0; j-- ) { |
duke@435 | 3453 | Node *n = rpo_list[j-1]; |
duke@435 | 3454 | if( !_nodes[n->_idx] ) // Skip dead nodes |
duke@435 | 3455 | continue; |
duke@435 | 3456 | if( get_loop(n) != loop ) { // Wrong loop nest |
duke@435 | 3457 | if( get_loop(n)->_head == n && // Found nested loop? |
duke@435 | 3458 | get_loop(n)->_parent == loop ) |
duke@435 | 3459 | dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly |
duke@435 | 3460 | continue; |
duke@435 | 3461 | } |
duke@435 | 3462 | |
duke@435 | 3463 | // Dump controlling node |
duke@435 | 3464 | for( uint x = 0; x < loop->_nest; x++ ) |
duke@435 | 3465 | tty->print(" "); |
duke@435 | 3466 | tty->print("C"); |
duke@435 | 3467 | if( n == C->root() ) { |
duke@435 | 3468 | n->dump(); |
duke@435 | 3469 | } else { |
duke@435 | 3470 | Node* cached_idom = idom_no_update(n); |
duke@435 | 3471 | Node *computed_idom = n->in(0); |
duke@435 | 3472 | if( n->is_Region() ) { |
duke@435 | 3473 | computed_idom = compute_idom(n); |
duke@435 | 3474 | // computed_idom() will return n->in(0) when idom(n) is an IfNode (or |
duke@435 | 3475 | // any MultiBranch ctrl node), so apply a similar transform to |
duke@435 | 3476 | // the cached idom returned from idom_no_update. |
duke@435 | 3477 | cached_idom = find_non_split_ctrl(cached_idom); |
duke@435 | 3478 | } |
duke@435 | 3479 | tty->print(" ID:%d",computed_idom->_idx); |
duke@435 | 3480 | n->dump(); |
duke@435 | 3481 | if( cached_idom != computed_idom ) { |
duke@435 | 3482 | tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d", |
duke@435 | 3483 | computed_idom->_idx, cached_idom->_idx); |
duke@435 | 3484 | } |
duke@435 | 3485 | } |
duke@435 | 3486 | // Dump nodes it controls |
duke@435 | 3487 | for( uint k = 0; k < _nodes.Size(); k++ ) { |
duke@435 | 3488 | // (k < C->unique() && get_ctrl(find(k)) == n) |
duke@435 | 3489 | if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) { |
duke@435 | 3490 | Node *m = C->root()->find(k); |
duke@435 | 3491 | if( m && m->outcnt() > 0 ) { |
duke@435 | 3492 | if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) { |
duke@435 | 3493 | tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p", |
duke@435 | 3494 | _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL); |
duke@435 | 3495 | } |
duke@435 | 3496 | for( uint j = 0; j < loop->_nest; j++ ) |
duke@435 | 3497 | tty->print(" "); |
duke@435 | 3498 | tty->print(" "); |
duke@435 | 3499 | m->dump(); |
duke@435 | 3500 | } |
duke@435 | 3501 | } |
duke@435 | 3502 | } |
duke@435 | 3503 | } |
duke@435 | 3504 | } |
duke@435 | 3505 | |
duke@435 | 3506 | // Collect a R-P-O for the whole CFG. |
duke@435 | 3507 | // Result list is in post-order (scan backwards for RPO) |
duke@435 | 3508 | void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const { |
duke@435 | 3509 | stk.push(start, 0); |
duke@435 | 3510 | visited.set(start->_idx); |
duke@435 | 3511 | |
duke@435 | 3512 | while (stk.is_nonempty()) { |
duke@435 | 3513 | Node* m = stk.node(); |
duke@435 | 3514 | uint idx = stk.index(); |
duke@435 | 3515 | if (idx < m->outcnt()) { |
duke@435 | 3516 | stk.set_index(idx + 1); |
duke@435 | 3517 | Node* n = m->raw_out(idx); |
duke@435 | 3518 | if (n->is_CFG() && !visited.test_set(n->_idx)) { |
duke@435 | 3519 | stk.push(n, 0); |
duke@435 | 3520 | } |
duke@435 | 3521 | } else { |
duke@435 | 3522 | rpo_list.push(m); |
duke@435 | 3523 | stk.pop(); |
duke@435 | 3524 | } |
duke@435 | 3525 | } |
duke@435 | 3526 | } |
duke@435 | 3527 | #endif |
duke@435 | 3528 | |
duke@435 | 3529 | |
duke@435 | 3530 | //============================================================================= |
duke@435 | 3531 | //------------------------------LoopTreeIterator----------------------------------- |
duke@435 | 3532 | |
duke@435 | 3533 | // Advance to next loop tree using a preorder, left-to-right traversal. |
duke@435 | 3534 | void LoopTreeIterator::next() { |
duke@435 | 3535 | assert(!done(), "must not be done."); |
duke@435 | 3536 | if (_curnt->_child != NULL) { |
duke@435 | 3537 | _curnt = _curnt->_child; |
duke@435 | 3538 | } else if (_curnt->_next != NULL) { |
duke@435 | 3539 | _curnt = _curnt->_next; |
duke@435 | 3540 | } else { |
duke@435 | 3541 | while (_curnt != _root && _curnt->_next == NULL) { |
duke@435 | 3542 | _curnt = _curnt->_parent; |
duke@435 | 3543 | } |
duke@435 | 3544 | if (_curnt == _root) { |
duke@435 | 3545 | _curnt = NULL; |
duke@435 | 3546 | assert(done(), "must be done."); |
duke@435 | 3547 | } else { |
duke@435 | 3548 | assert(_curnt->_next != NULL, "must be more to do"); |
duke@435 | 3549 | _curnt = _curnt->_next; |
duke@435 | 3550 | } |
duke@435 | 3551 | } |
duke@435 | 3552 | } |