1 /* 2 * Copyright (c) 2006, 2023, 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/mulnode.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/connode.hpp" 30 #include "opto/convertnode.hpp" 31 #include "opto/loopnode.hpp" 32 #include "opto/opaquenode.hpp" 33 #include "opto/predicates.hpp" 34 #include "opto/rootnode.hpp" 35 36 //================= Loop Unswitching ===================== 37 // 38 // orig: transformed: 39 // if (invariant-test) then 40 // predicates predicates 41 // loop loop 42 // stmt1 stmt1 43 // if (invariant-test) then stmt2 44 // stmt2 stmt4 45 // else endloop 46 // stmt3 else 47 // endif predicates [clone] 48 // stmt4 loop [clone] 49 // endloop stmt1 [clone] 50 // stmt3 51 // stmt4 [clone] 52 // endloop 53 // endif 54 // 55 // Note: the "else" clause may be empty 56 57 58 //------------------------------policy_unswitching----------------------------- 59 // Return TRUE or FALSE if the loop should be unswitched 60 // (ie. clone loop with an invariant test that does not exit the loop) 61 bool IdealLoopTree::policy_unswitching( PhaseIdealLoop *phase ) const { 62 if (!LoopUnswitching) { 63 return false; 64 } 65 if (!_head->is_Loop()) { 66 return false; 67 } 68 69 // If nodes are depleted, some transform has miscalculated its needs. 70 assert(!phase->exceeding_node_budget(), "sanity"); 71 72 // check for vectorized loops, any unswitching was already applied 73 if (_head->is_CountedLoop() && _head->as_CountedLoop()->is_unroll_only()) { 74 return false; 75 } 76 77 LoopNode* head = _head->as_Loop(); 78 if (head->unswitch_count() + 1 > head->unswitch_max()) { 79 return false; 80 } 81 82 if (head->is_flat_arrays()) { 83 return false; 84 } 85 86 Node_List unswitch_iffs; 87 if (phase->find_unswitching_candidate(this, unswitch_iffs) == nullptr) { 88 return false; 89 } 90 91 // Too speculative if running low on nodes. 92 return phase->may_require_nodes(est_loop_clone_sz(2)); 93 } 94 95 //------------------------------find_unswitching_candidate----------------------------- 96 // Find candidate "if" for unswitching 97 IfNode* PhaseIdealLoop::find_unswitching_candidate(const IdealLoopTree *loop, Node_List& unswitch_iffs) const { 98 99 // Find first invariant test that doesn't exit the loop 100 LoopNode *head = loop->_head->as_Loop(); 101 IfNode* unswitch_iff = nullptr; 102 Node* n = head->in(LoopNode::LoopBackControl); 103 while (n != head) { 104 Node* n_dom = idom(n); 105 if (n->is_Region()) { 106 if (n_dom->is_If()) { 107 IfNode* iff = n_dom->as_If(); 108 if (iff->in(1)->is_Bool()) { 109 BoolNode* bol = iff->in(1)->as_Bool(); 110 if (bol->in(1)->is_Cmp()) { 111 // If condition is invariant and not a loop exit, 112 // then found reason to unswitch. 113 if (loop->is_invariant(bol) && !loop->is_loop_exit(iff)) { 114 unswitch_iff = iff; 115 } 116 } 117 } 118 } 119 } 120 n = n_dom; 121 } 122 if (unswitch_iff != nullptr) { 123 unswitch_iffs.push(unswitch_iff); 124 } 125 126 // Collect all non-flat array checks for unswitching to create a fast loop 127 // without checks (only non-flat array accesses) and a slow loop with checks. 128 if (unswitch_iff == nullptr || unswitch_iff->is_flat_array_check(&_igvn)) { 129 for (uint i = 0; i < loop->_body.size(); i++) { 130 IfNode* n = loop->_body.at(i)->isa_If(); 131 if (n != nullptr && n != unswitch_iff && n->is_flat_array_check(&_igvn) && 132 loop->is_invariant(n->in(1)) && !loop->is_loop_exit(n)) { 133 unswitch_iffs.push(n); 134 if (unswitch_iff == nullptr) { 135 unswitch_iff = n; 136 } 137 } 138 } 139 } 140 return unswitch_iff; 141 } 142 143 //------------------------------do_unswitching----------------------------- 144 // Clone loop with an invariant test (that does not exit) and 145 // insert a clone of the test that selects which version to 146 // execute. 147 void PhaseIdealLoop::do_unswitching(IdealLoopTree *loop, Node_List &old_new) { 148 LoopNode* head = loop->_head->as_Loop(); 149 if (has_control_dependencies_from_predicates(head)) { 150 return; 151 } 152 153 // Find first invariant test that doesn't exit the loop 154 Node_List unswitch_iffs; 155 IfNode* unswitch_iff = find_unswitching_candidate((const IdealLoopTree *)loop, unswitch_iffs); 156 assert(unswitch_iff != nullptr && unswitch_iff == unswitch_iffs.at(0), "should be at least one"); 157 bool flat_array_checks = unswitch_iffs.size() > 1; 158 159 #ifndef PRODUCT 160 if (TraceLoopOpts) { 161 tty->print("Unswitch %d ", head->unswitch_count()+1); 162 loop->dump_head(); 163 for (uint i = 0; i < unswitch_iffs.size(); i++) { 164 unswitch_iffs.at(i)->dump(3); 165 tty->cr(); 166 } 167 } 168 #endif 169 170 // Need to revert back to normal loop 171 if (head->is_CountedLoop() && !head->as_CountedLoop()->is_normal_loop()) { 172 head->as_CountedLoop()->set_normal_loop(); 173 } 174 175 IfNode* invar_iff = create_slow_version_of_loop(loop, old_new, unswitch_iffs, CloneIncludesStripMined); 176 ProjNode* proj_true = invar_iff->proj_out(1); 177 verify_fast_loop(head, proj_true); 178 179 // Increment unswitch count 180 LoopNode* head_clone = old_new[head->_idx]->as_Loop(); 181 int nct = head->unswitch_count() + 1; 182 head->set_unswitch_count(nct); 183 head_clone->set_unswitch_count(nct); 184 185 // Hoist invariant casts out of each loop to the appropriate control projection. 186 Node_List worklist; 187 for (uint i = 0; i < unswitch_iffs.size(); i++) { 188 IfNode* iff = unswitch_iffs.at(i)->as_If(); 189 for (DUIterator_Fast imax, i = iff->fast_outs(imax); i < imax; i++) { 190 ProjNode* proj = iff->fast_out(i)->as_Proj(); 191 // Copy to a worklist for easier manipulation 192 for (DUIterator_Fast jmax, j = proj->fast_outs(jmax); j < jmax; j++) { 193 Node* use = proj->fast_out(j); 194 if (use->Opcode() == Op_CheckCastPP && loop->is_invariant(use->in(1))) { 195 worklist.push(use); 196 } 197 } 198 ProjNode* invar_proj = invar_iff->proj_out(proj->_con)->as_Proj(); 199 while (worklist.size() > 0) { 200 Node* use = worklist.pop(); 201 Node* nuse = use->clone(); 202 nuse->set_req(0, invar_proj); 203 _igvn.replace_input_of(use, 1, nuse); 204 register_new_node(nuse, invar_proj); 205 // Same for the clone if we are removing checks from the slow loop 206 if (!flat_array_checks) { 207 Node* use_clone = old_new[use->_idx]; 208 _igvn.replace_input_of(use_clone, 1, nuse); 209 } 210 } 211 } 212 } 213 214 // Hardwire the control paths in the loops into if(true) and if(false) 215 for (uint i = 0; i < unswitch_iffs.size(); i++) { 216 IfNode* iff = unswitch_iffs.at(i)->as_If(); 217 _igvn.rehash_node_delayed(iff); 218 dominated_by(proj_true->as_IfProj(), iff); 219 } 220 IfNode* unswitch_iff_clone = old_new[unswitch_iff->_idx]->as_If(); 221 if (!flat_array_checks) { 222 ProjNode* proj_false = invar_iff->proj_out(0)->as_Proj(); 223 _igvn.rehash_node_delayed(unswitch_iff_clone); 224 dominated_by(proj_false->as_IfProj(), unswitch_iff_clone); 225 } else { 226 // Leave the flat array checks in the slow loop and 227 // prevent it from being unswitched again based on these checks. 228 head_clone->mark_flat_arrays(); 229 } 230 231 // Reoptimize loops 232 loop->record_for_igvn(); 233 for(int i = loop->_body.size() - 1; i >= 0 ; i--) { 234 Node *n = loop->_body[i]; 235 Node *n_clone = old_new[n->_idx]; 236 _igvn._worklist.push(n_clone); 237 } 238 239 #ifndef PRODUCT 240 if (TraceLoopUnswitching) { 241 for (uint i = 0; i < unswitch_iffs.size(); i++) { 242 tty->print_cr("Loop unswitching orig: %d @ %d new: %d @ %d", 243 head->_idx, unswitch_iffs.at(i)->_idx, 244 old_new[head->_idx]->_idx, old_new[unswitch_iffs.at(i)->_idx]->_idx); 245 } 246 } 247 #endif 248 249 C->set_major_progress(); 250 } 251 252 bool PhaseIdealLoop::has_control_dependencies_from_predicates(LoopNode* head) const { 253 Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl); 254 Predicates predicates(entry); 255 if (predicates.has_any()) { 256 assert(entry->is_IfProj(), "sanity - must be ifProj since there is at least one predicate"); 257 if (entry->outcnt() > 1) { 258 // Bailout if there are predicates from which there are additional control dependencies (i.e. from loop 259 // entry 'entry') to previously partially peeled statements since this case is not handled and can lead 260 // to a wrong execution. Remove this bailout, once this is fixed. 261 return true; 262 } 263 } 264 return false; 265 } 266 267 //-------------------------create_slow_version_of_loop------------------------ 268 // Create a slow version of the loop by cloning the loop 269 // and inserting an if to select fast-slow versions. 270 // Return the inserted if. 271 IfNode* PhaseIdealLoop::create_slow_version_of_loop(IdealLoopTree *loop, 272 Node_List &old_new, 273 Node_List &unswitch_iffs, 274 CloneLoopMode mode) { 275 LoopNode* head = loop->_head->as_Loop(); 276 bool counted_loop = head->is_CountedLoop(); 277 Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl); 278 _igvn.rehash_node_delayed(entry); 279 IdealLoopTree* outer_loop = loop->_parent; 280 281 head->verify_strip_mined(1); 282 283 // Add test to new "if" outside of loop 284 IfNode* unswitch_iff = unswitch_iffs.at(0)->as_If(); 285 BoolNode* bol = unswitch_iff->in(1)->as_Bool(); 286 if (unswitch_iffs.size() > 1) { 287 // Flat array checks are used on array access to switch between 288 // a legacy object array access and a flat inline type array 289 // access. We want the performance impact on legacy accesses to be 290 // as small as possible so we make two copies of the loop: a fast 291 // one where all accesses are known to be legacy, a slow one where 292 // some accesses are to flat arrays. Flat array checks 293 // can be removed from the fast loop (true proj) but not from the 294 // slow loop (false proj) as it can have a mix of flat/legacy accesses. 295 assert(bol->_test._test == BoolTest::ne, "IfTrue proj must point to flat array"); 296 bol = bol->clone()->as_Bool(); 297 register_new_node(bol, entry); 298 FlatArrayCheckNode* cmp = bol->in(1)->clone()->as_FlatArrayCheck(); 299 register_new_node(cmp, entry); 300 bol->set_req(1, cmp); 301 // Combine all checks into a single one that fails if one array is a flat array 302 assert(cmp->req() == 3, "unexpected number of inputs for FlatArrayCheck"); 303 cmp->add_req_batch(C->top(), unswitch_iffs.size() - 1); 304 for (uint i = 0; i < unswitch_iffs.size(); i++) { 305 Node* array = unswitch_iffs.at(i)->in(1)->in(1)->in(FlatArrayCheckNode::ArrayOrKlass); 306 cmp->set_req(FlatArrayCheckNode::ArrayOrKlass + i, array); 307 } 308 } 309 310 IfNode* iff = (unswitch_iff->Opcode() == Op_RangeCheck) ? new RangeCheckNode(entry, bol, unswitch_iff->_prob, unswitch_iff->_fcnt) : 311 new IfNode(entry, bol, unswitch_iff->_prob, unswitch_iff->_fcnt); 312 register_node(iff, outer_loop, entry, dom_depth(entry)); 313 IfProjNode* iffast = new IfTrueNode(iff); 314 register_node(iffast, outer_loop, iff, dom_depth(iff)); 315 IfProjNode* ifslow = new IfFalseNode(iff); 316 register_node(ifslow, outer_loop, iff, dom_depth(iff)); 317 318 // Clone the loop body. The clone becomes the slow loop. The 319 // original pre-header will (illegally) have 3 control users 320 // (old & new loops & new if). 321 clone_loop(loop, old_new, dom_depth(head->skip_strip_mined()), mode, iff); 322 assert(old_new[head->_idx]->is_Loop(), "" ); 323 324 // Fast (true) and Slow (false) control 325 IfProjNode* iffast_pred = iffast; 326 IfProjNode* ifslow_pred = ifslow; 327 clone_parse_and_assertion_predicates_to_unswitched_loop(loop, old_new, iffast_pred, ifslow_pred); 328 329 Node* l = head->skip_strip_mined(); 330 _igvn.replace_input_of(l, LoopNode::EntryControl, iffast_pred); 331 set_idom(l, iffast_pred, dom_depth(l)); 332 LoopNode* slow_l = old_new[head->_idx]->as_Loop()->skip_strip_mined(); 333 _igvn.replace_input_of(slow_l, LoopNode::EntryControl, ifslow_pred); 334 set_idom(slow_l, ifslow_pred, dom_depth(l)); 335 336 recompute_dom_depth(); 337 338 return iff; 339 } 340 341 #ifdef ASSERT 342 void PhaseIdealLoop::verify_fast_loop(LoopNode* head, const ProjNode* proj_true) const { 343 assert(proj_true->is_IfTrue(), "must be true projection"); 344 Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl); 345 Predicates predicates(entry); 346 if (!predicates.has_any()) { 347 // No Parse Predicate. 348 Node* uniqc = proj_true->unique_ctrl_out(); 349 assert((uniqc == head && !head->is_strip_mined()) || (uniqc == head->in(LoopNode::EntryControl) 350 && head->is_strip_mined()), "must hold by construction if no predicates"); 351 } else { 352 // There is at least one Parse Predicate. When skipping all predicates/predicate blocks, we should end up 353 // at 'proj_true'. 354 assert(proj_true == predicates.entry(), "must hold by construction if at least one Parse Predicate"); 355 } 356 } 357 #endif // ASSERT 358