1 /*
   2  * Copyright (c) 2006, 2012, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "memory/allocation.inline.hpp"
  27 #include "opto/connode.hpp"
  28 #include "opto/convertnode.hpp"
  29 #include "opto/loopnode.hpp"
  30 #include "opto/opaquenode.hpp"
  31 #include "opto/rootnode.hpp"
  32 
  33 //================= Loop Unswitching =====================
  34 //
  35 // orig:                       transformed:
  36 //                               if (invariant-test) then
  37 //  predicate                      predicate
  38 //  loop                           loop
  39 //    stmt1                          stmt1
  40 //    if (invariant-test) then       stmt2
  41 //      stmt2                        stmt4
  42 //    else                         endloop
  43 //      stmt3                    else
  44 //    endif                        predicate [clone]
  45 //    stmt4                        loop [clone]
  46 //  endloop                          stmt1 [clone]
  47 //                                   stmt3
  48 //                                   stmt4 [clone]
  49 //                                 endloop
  50 //                               endif
  51 //
  52 // Note: the "else" clause may be empty
  53 
  54 //------------------------------policy_unswitching-----------------------------
  55 // Return TRUE or FALSE if the loop should be unswitched
  56 // (ie. clone loop with an invariant test that does not exit the loop)
  57 bool IdealLoopTree::policy_unswitching( PhaseIdealLoop *phase ) const {
  58   if (!LoopUnswitching) {
  59     return false;
  60   }
  61   if (!_head->is_Loop()) {
  62     return false;
  63   }
  64 
  65   // If nodes are depleted, some transform has miscalculated its needs.
  66   assert(!phase->exceeding_node_budget(), "sanity");
  67 
  68   // check for vectorized loops, any unswitching was already applied
  69   if (_head->is_CountedLoop() && _head->as_CountedLoop()->is_unroll_only()) {
  70     return false;
  71   }
  72 
  73   LoopNode* head = _head->as_Loop();
  74   if (head->unswitch_count() + 1 > head->unswitch_max()) {
  75     return false;
  76   }
  77   if (phase->find_unswitching_candidate(this) == NULL) {
  78     return false;
  79   }
  80 
  81   // Too speculative if running low on nodes.
  82   return phase->may_require_nodes(est_loop_clone_sz(3, _body.size()));
  83 }
  84 
  85 //------------------------------find_unswitching_candidate-----------------------------
  86 // Find candidate "if" for unswitching
  87 IfNode* PhaseIdealLoop::find_unswitching_candidate(const IdealLoopTree *loop) const {
  88 
  89   // Find first invariant test that doesn't exit the loop
  90   LoopNode *head = loop->_head->as_Loop();
  91   IfNode* unswitch_iff = NULL;
  92   Node* n = head->in(LoopNode::LoopBackControl);
  93   while (n != head) {
  94     Node* n_dom = idom(n);
  95     if (n->is_Region()) {
  96       if (n_dom->is_If()) {
  97         IfNode* iff = n_dom->as_If();
  98         if (iff->in(1)->is_Bool()) {
  99           BoolNode* bol = iff->in(1)->as_Bool();
 100           if (bol->in(1)->is_Cmp()) {
 101             // If condition is invariant and not a loop exit,
 102             // then found reason to unswitch.
 103             if (loop->is_invariant(bol) && !loop->is_loop_exit(iff)) {
 104               unswitch_iff = iff;
 105             }
 106           }
 107         }
 108       }
 109     }
 110     n = n_dom;
 111   }
 112   return unswitch_iff;
 113 }
 114 
 115 //------------------------------do_unswitching-----------------------------
 116 // Clone loop with an invariant test (that does not exit) and
 117 // insert a clone of the test that selects which version to
 118 // execute.
 119 void PhaseIdealLoop::do_unswitching (IdealLoopTree *loop, Node_List &old_new) {
 120 
 121   // Find first invariant test that doesn't exit the loop
 122   LoopNode *head = loop->_head->as_Loop();
 123 
 124   IfNode* unswitch_iff = find_unswitching_candidate((const IdealLoopTree *)loop);
 125   assert(unswitch_iff != NULL, "should be at least one");
 126 
 127 #ifndef PRODUCT
 128   if (TraceLoopOpts) {
 129     tty->print("Unswitch   %d ", head->unswitch_count()+1);
 130     loop->dump_head();
 131   }
 132 #endif
 133 
 134   // Need to revert back to normal loop
 135   if (head->is_CountedLoop() && !head->as_CountedLoop()->is_normal_loop()) {
 136     head->as_CountedLoop()->set_normal_loop();
 137   }
 138 
 139   ProjNode* proj_true = create_slow_version_of_loop(loop, old_new, unswitch_iff->Opcode(), CloneIncludesStripMined);
 140 
 141 #ifdef ASSERT
 142   Node* uniqc = proj_true->unique_ctrl_out();
 143   Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
 144   Node* predicate = find_predicate(entry);
 145   if (predicate != NULL) {
 146     entry = skip_loop_predicates(entry);
 147   }
 148   if (predicate != NULL && UseLoopPredicate) {
 149     // We may have two predicates, find first.
 150     Node* n = find_predicate(entry);
 151     if (n != NULL) {
 152       predicate = n;
 153       entry = skip_loop_predicates(entry);
 154     }
 155   }
 156   if (predicate != NULL && UseProfiledLoopPredicate) {
 157     entry = find_predicate(entry);
 158     if (entry != NULL) predicate = entry;
 159   }
 160   if (predicate != NULL) predicate = predicate->in(0);
 161   assert(proj_true->is_IfTrue() &&
 162          (predicate == NULL && uniqc == head && !head->is_strip_mined() ||
 163           predicate == NULL && uniqc == head->in(LoopNode::EntryControl) && head->is_strip_mined() ||
 164           predicate != NULL && uniqc == predicate), "by construction");
 165 #endif
 166   // Increment unswitch count
 167   LoopNode* head_clone = old_new[head->_idx]->as_Loop();
 168   int nct = head->unswitch_count() + 1;
 169   head->set_unswitch_count(nct);
 170   head_clone->set_unswitch_count(nct);
 171 
 172   // Add test to new "if" outside of loop
 173   IfNode* invar_iff   = proj_true->in(0)->as_If();
 174   Node* invar_iff_c   = invar_iff->in(0);
 175   BoolNode* bol       = unswitch_iff->in(1)->as_Bool();
 176   invar_iff->set_req(1, bol);
 177   invar_iff->_prob    = unswitch_iff->_prob;
 178 
 179   ProjNode* proj_false = invar_iff->proj_out(0)->as_Proj();
 180 
 181   // Hoist invariant casts out of each loop to the appropriate
 182   // control projection.
 183 
 184   Node_List worklist;
 185 
 186   for (DUIterator_Fast imax, i = unswitch_iff->fast_outs(imax); i < imax; i++) {
 187     ProjNode* proj= unswitch_iff->fast_out(i)->as_Proj();
 188     // Copy to a worklist for easier manipulation
 189     for (DUIterator_Fast jmax, j = proj->fast_outs(jmax); j < jmax; j++) {
 190       Node* use = proj->fast_out(j);
 191       if (use->Opcode() == Op_CheckCastPP && loop->is_invariant(use->in(1))) {
 192         worklist.push(use);
 193       }
 194     }
 195     ProjNode* invar_proj = invar_iff->proj_out(proj->_con)->as_Proj();
 196     while (worklist.size() > 0) {
 197       Node* use = worklist.pop();
 198       Node* nuse = use->clone();
 199       nuse->set_req(0, invar_proj);
 200       _igvn.replace_input_of(use, 1, nuse);
 201       register_new_node(nuse, invar_proj);
 202       // Same for the clone
 203       Node* use_clone = old_new[use->_idx];
 204       _igvn.replace_input_of(use_clone, 1, nuse);
 205     }
 206   }
 207 
 208   // Hardwire the control paths in the loops into if(true) and if(false)
 209   _igvn.rehash_node_delayed(unswitch_iff);
 210   short_circuit_if(unswitch_iff, proj_true);
 211 
 212   IfNode* unswitch_iff_clone = old_new[unswitch_iff->_idx]->as_If();
 213   _igvn.rehash_node_delayed(unswitch_iff_clone);
 214   short_circuit_if(unswitch_iff_clone, proj_false);
 215 
 216   // Reoptimize loops
 217   loop->record_for_igvn();
 218   for(int i = loop->_body.size() - 1; i >= 0 ; i--) {
 219     Node *n = loop->_body[i];
 220     Node *n_clone = old_new[n->_idx];
 221     _igvn._worklist.push(n_clone);
 222   }
 223 
 224 #ifndef PRODUCT
 225   if (TraceLoopUnswitching) {
 226     tty->print_cr("Loop unswitching orig: %d @ %d  new: %d @ %d",
 227                   head->_idx,                unswitch_iff->_idx,
 228                   old_new[head->_idx]->_idx, unswitch_iff_clone->_idx);
 229   }
 230 #endif
 231 
 232   C->set_major_progress();
 233 }
 234 
 235 //-------------------------create_slow_version_of_loop------------------------
 236 // Create a slow version of the loop by cloning the loop
 237 // and inserting an if to select fast-slow versions.
 238 // Return control projection of the entry to the fast version.
 239 ProjNode* PhaseIdealLoop::create_slow_version_of_loop(IdealLoopTree *loop,
 240                                                       Node_List &old_new,
 241                                                       int opcode,
 242                                                       CloneLoopMode mode) {
 243   LoopNode* head  = loop->_head->as_Loop();
 244   bool counted_loop = head->is_CountedLoop();
 245   Node*     entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
 246   _igvn.rehash_node_delayed(entry);
 247   IdealLoopTree* outer_loop = loop->_parent;
 248 
 249   head->verify_strip_mined(1);
 250 
 251   Node *cont      = _igvn.intcon(1);
 252   set_ctrl(cont, C->root());
 253   Node* opq       = new Opaque1Node(C, cont);
 254   register_node(opq, outer_loop, entry, dom_depth(entry));
 255   Node *bol       = new Conv2BNode(opq);
 256   register_node(bol, outer_loop, entry, dom_depth(entry));
 257   IfNode* iff = (opcode == Op_RangeCheck) ? new RangeCheckNode(entry, bol, PROB_MAX, COUNT_UNKNOWN) :
 258     new IfNode(entry, bol, PROB_MAX, COUNT_UNKNOWN);
 259   register_node(iff, outer_loop, entry, dom_depth(entry));
 260   ProjNode* iffast = new IfTrueNode(iff);
 261   register_node(iffast, outer_loop, iff, dom_depth(iff));
 262   ProjNode* ifslow = new IfFalseNode(iff);
 263   register_node(ifslow, outer_loop, iff, dom_depth(iff));
 264 
 265   // Clone the loop body.  The clone becomes the fast loop.  The
 266   // original pre-header will (illegally) have 3 control users
 267   // (old & new loops & new if).
 268   clone_loop(loop, old_new, dom_depth(head->skip_strip_mined()), mode, iff);
 269   assert(old_new[head->_idx]->is_Loop(), "" );
 270 
 271   // Fast (true) control
 272   Node* iffast_pred = clone_loop_predicates(entry, iffast, !counted_loop);
 273 
 274   // Slow (false) control
 275   Node* ifslow_pred = clone_loop_predicates(entry, ifslow, !counted_loop);
 276 
 277   Node* l = head->skip_strip_mined();
 278   _igvn.replace_input_of(l, LoopNode::EntryControl, iffast_pred);
 279   set_idom(l, iffast_pred, dom_depth(l));
 280   LoopNode* slow_l = old_new[head->_idx]->as_Loop()->skip_strip_mined();
 281   _igvn.replace_input_of(slow_l, LoopNode::EntryControl, ifslow_pred);
 282   set_idom(slow_l, ifslow_pred, dom_depth(l));
 283 
 284   recompute_dom_depth();
 285 
 286   return iffast;
 287 }
 288 
 289 LoopNode* PhaseIdealLoop::create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk) {
 290   Node_List old_new;
 291   LoopNode* head  = loop->_head->as_Loop();
 292   bool counted_loop = head->is_CountedLoop();
 293   Node*     entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
 294   _igvn.rehash_node_delayed(entry);
 295   IdealLoopTree* outer_loop = head->is_strip_mined() ? loop->_parent->_parent : loop->_parent;
 296 
 297   ConINode* const_1 = _igvn.intcon(1);
 298   set_ctrl(const_1, C->root());
 299   IfNode* iff = new IfNode(entry, const_1, PROB_MAX, COUNT_UNKNOWN);
 300   register_node(iff, outer_loop, entry, dom_depth(entry));
 301   ProjNode* iffast = new IfTrueNode(iff);
 302   register_node(iffast, outer_loop, iff, dom_depth(iff));
 303   ProjNode* ifslow = new IfFalseNode(iff);
 304   register_node(ifslow, outer_loop, iff, dom_depth(iff));
 305 
 306   // Clone the loop body.  The clone becomes the fast loop.  The
 307   // original pre-header will (illegally) have 3 control users
 308   // (old & new loops & new if).
 309   clone_loop(loop, old_new, dom_depth(head), CloneIncludesStripMined, iff);
 310   assert(old_new[head->_idx]->is_Loop(), "" );
 311 
 312   LoopNode* slow_head = old_new[head->_idx]->as_Loop();
 313 
 314 #ifndef PRODUCT
 315   if (TraceLoopOpts) {
 316     tty->print_cr("PhaseIdealLoop::create_reserve_version_of_loop:");
 317     tty->print("\t iff = %d, ", iff->_idx); iff->dump();
 318     tty->print("\t iffast = %d, ", iffast->_idx); iffast->dump();
 319     tty->print("\t ifslow = %d, ", ifslow->_idx); ifslow->dump();
 320     tty->print("\t before replace_input_of: head = %d, ", head->_idx); head->dump();
 321     tty->print("\t before replace_input_of: slow_head = %d, ", slow_head->_idx); slow_head->dump();
 322   }
 323 #endif
 324 
 325   // Fast (true) control
 326   _igvn.replace_input_of(head->skip_strip_mined(), LoopNode::EntryControl, iffast);
 327   // Slow (false) control
 328   _igvn.replace_input_of(slow_head->skip_strip_mined(), LoopNode::EntryControl, ifslow);
 329 
 330   recompute_dom_depth();
 331 
 332   lk->set_iff(iff);
 333 
 334 #ifndef PRODUCT
 335   if (TraceLoopOpts ) {
 336     tty->print("\t after  replace_input_of: head = %d, ", head->_idx); head->dump();
 337     tty->print("\t after  replace_input_of: slow_head = %d, ", slow_head->_idx); slow_head->dump();
 338   }
 339 #endif
 340 
 341   return slow_head->as_Loop();
 342 }
 343 
 344 CountedLoopReserveKit::CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active = true) :
 345   _phase(phase),
 346   _lpt(loop),
 347   _lp(NULL),
 348   _iff(NULL),
 349   _lp_reserved(NULL),
 350   _has_reserved(false),
 351   _use_new(false),
 352   _active(active)
 353   {
 354     create_reserve();
 355   };
 356 
 357 CountedLoopReserveKit::~CountedLoopReserveKit() {
 358   if (!_active) {
 359     return;
 360   }
 361 
 362   if (_has_reserved && !_use_new) {
 363     // intcon(0)->iff-node reverts CF to the reserved copy
 364     ConINode* const_0 = _phase->_igvn.intcon(0);
 365     _phase->set_ctrl(const_0, _phase->C->root());
 366     _iff->set_req(1, const_0);
 367 
 368     #ifndef PRODUCT
 369       if (TraceLoopOpts) {
 370         tty->print_cr("CountedLoopReserveKit::~CountedLoopReserveKit()");
 371         tty->print("\t discard loop %d and revert to the reserved loop clone %d: ", _lp->_idx, _lp_reserved->_idx);
 372         _lp_reserved->dump();
 373       }
 374     #endif
 375   }
 376 }
 377 
 378 bool CountedLoopReserveKit::create_reserve() {
 379   if (!_active) {
 380     return false;
 381   }
 382 
 383   if(!_lpt->_head->is_CountedLoop()) {
 384     if (TraceLoopOpts) {
 385       tty->print_cr("CountedLoopReserveKit::create_reserve: %d not counted loop", _lpt->_head->_idx);
 386     }
 387     return false;
 388   }
 389   CountedLoopNode *cl = _lpt->_head->as_CountedLoop();
 390   if (!cl->is_valid_counted_loop()) {
 391     if (TraceLoopOpts) {
 392       tty->print_cr("CountedLoopReserveKit::create_reserve: %d not valid counted loop", cl->_idx);
 393     }
 394     return false; // skip malformed counted loop
 395   }
 396   if (!cl->is_main_loop()) {
 397     bool loop_not_canonical = true;
 398     if (cl->is_post_loop() && (cl->slp_max_unroll() > 0)) {
 399       loop_not_canonical = false;
 400     }
 401     // only reject some loop forms
 402     if (loop_not_canonical) {
 403       if (TraceLoopOpts) {
 404         tty->print_cr("CountedLoopReserveKit::create_reserve: %d not canonical loop", cl->_idx);
 405       }
 406       return false; // skip normal, pre, and post (conditionally) loops
 407     }
 408   }
 409 
 410   _lp = _lpt->_head->as_Loop();
 411   _lp_reserved = _phase->create_reserve_version_of_loop(_lpt, this);
 412 
 413   if (!_lp_reserved->is_CountedLoop()) {
 414     return false;
 415   }
 416 
 417   Node* ifslow_pred = _lp_reserved->skip_strip_mined()->in(LoopNode::EntryControl);
 418 
 419   if (!ifslow_pred->is_IfFalse()) {
 420     return false;
 421   }
 422 
 423   Node* iff = ifslow_pred->in(0);
 424   if (!iff->is_If() || iff != _iff) {
 425     return false;
 426   }
 427 
 428   if (iff->in(1)->Opcode() != Op_ConI) {
 429     return false;
 430   }
 431 
 432   return _has_reserved = true;
 433 }