Mercurial > hg > truffle
annotate src/share/vm/c1/c1_IR.cpp @ 1972:f95d63e2154a
6989984: Use standard include model for Hospot
Summary: Replaced MakeDeps and the includeDB files with more standardized solutions.
Reviewed-by: coleenp, kvn, kamg
| author | stefank |
|---|---|
| date | Tue, 23 Nov 2010 13:22:55 -0800 |
| parents | 42a10fc37986 |
| children | 5ddfcf4b079e |
| rev | line source |
|---|---|
| 0 | 1 /* |
|
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1295
diff
changeset
|
2 * Copyright (c) 1999, 2010, Oracle and/or its affiliates. All rights reserved. |
| 0 | 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
| 4 * | |
| 5 * This code is free software; you can redistribute it and/or modify it | |
| 6 * under the terms of the GNU General Public License version 2 only, as | |
| 7 * published by the Free Software Foundation. | |
| 8 * | |
| 9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
| 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
| 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
| 12 * version 2 for more details (a copy is included in the LICENSE file that | |
| 13 * accompanied this code). | |
| 14 * | |
| 15 * You should have received a copy of the GNU General Public License version | |
| 16 * 2 along with this work; if not, write to the Free Software Foundation, | |
| 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
| 18 * | |
|
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1295
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
|
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1295
diff
changeset
|
20 * or visit www.oracle.com if you need additional information or have any |
|
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1295
diff
changeset
|
21 * questions. |
| 0 | 22 * |
| 23 */ | |
| 24 | |
| 1972 | 25 #include "precompiled.hpp" |
| 26 #include "c1/c1_Compilation.hpp" | |
| 27 #include "c1/c1_FrameMap.hpp" | |
| 28 #include "c1/c1_GraphBuilder.hpp" | |
| 29 #include "c1/c1_IR.hpp" | |
| 30 #include "c1/c1_InstructionPrinter.hpp" | |
| 31 #include "c1/c1_Optimizer.hpp" | |
| 32 #include "utilities/bitMap.inline.hpp" | |
| 0 | 33 |
| 34 | |
| 35 // Implementation of XHandlers | |
| 36 // | |
| 37 // Note: This code could eventually go away if we are | |
| 38 // just using the ciExceptionHandlerStream. | |
| 39 | |
| 40 XHandlers::XHandlers(ciMethod* method) : _list(method->exception_table_length()) { | |
| 41 ciExceptionHandlerStream s(method); | |
| 42 while (!s.is_done()) { | |
| 43 _list.append(new XHandler(s.handler())); | |
| 44 s.next(); | |
| 45 } | |
| 46 assert(s.count() == method->exception_table_length(), "exception table lengths inconsistent"); | |
| 47 } | |
| 48 | |
| 49 // deep copy of all XHandler contained in list | |
| 50 XHandlers::XHandlers(XHandlers* other) : | |
| 51 _list(other->length()) | |
| 52 { | |
| 53 for (int i = 0; i < other->length(); i++) { | |
| 54 _list.append(new XHandler(other->handler_at(i))); | |
| 55 } | |
| 56 } | |
| 57 | |
| 58 // Returns whether a particular exception type can be caught. Also | |
| 59 // returns true if klass is unloaded or any exception handler | |
| 60 // classes are unloaded. type_is_exact indicates whether the throw | |
| 61 // is known to be exactly that class or it might throw a subtype. | |
| 62 bool XHandlers::could_catch(ciInstanceKlass* klass, bool type_is_exact) const { | |
| 63 // the type is unknown so be conservative | |
| 64 if (!klass->is_loaded()) { | |
| 65 return true; | |
| 66 } | |
| 67 | |
| 68 for (int i = 0; i < length(); i++) { | |
| 69 XHandler* handler = handler_at(i); | |
| 70 if (handler->is_catch_all()) { | |
| 71 // catch of ANY | |
| 72 return true; | |
| 73 } | |
| 74 ciInstanceKlass* handler_klass = handler->catch_klass(); | |
| 75 // if it's unknown it might be catchable | |
| 76 if (!handler_klass->is_loaded()) { | |
| 77 return true; | |
| 78 } | |
| 79 // if the throw type is definitely a subtype of the catch type | |
| 80 // then it can be caught. | |
| 81 if (klass->is_subtype_of(handler_klass)) { | |
| 82 return true; | |
| 83 } | |
| 84 if (!type_is_exact) { | |
| 85 // If the type isn't exactly known then it can also be caught by | |
| 86 // catch statements where the inexact type is a subtype of the | |
| 87 // catch type. | |
| 88 // given: foo extends bar extends Exception | |
| 89 // throw bar can be caught by catch foo, catch bar, and catch | |
| 90 // Exception, however it can't be caught by any handlers without | |
| 91 // bar in its type hierarchy. | |
| 92 if (handler_klass->is_subtype_of(klass)) { | |
| 93 return true; | |
| 94 } | |
| 95 } | |
| 96 } | |
| 97 | |
| 98 return false; | |
| 99 } | |
| 100 | |
| 101 | |
| 102 bool XHandlers::equals(XHandlers* others) const { | |
| 103 if (others == NULL) return false; | |
| 104 if (length() != others->length()) return false; | |
| 105 | |
| 106 for (int i = 0; i < length(); i++) { | |
| 107 if (!handler_at(i)->equals(others->handler_at(i))) return false; | |
| 108 } | |
| 109 return true; | |
| 110 } | |
| 111 | |
| 112 bool XHandler::equals(XHandler* other) const { | |
| 113 assert(entry_pco() != -1 && other->entry_pco() != -1, "must have entry_pco"); | |
| 114 | |
| 115 if (entry_pco() != other->entry_pco()) return false; | |
| 116 if (scope_count() != other->scope_count()) return false; | |
| 117 if (_desc != other->_desc) return false; | |
| 118 | |
| 119 assert(entry_block() == other->entry_block(), "entry_block must be equal when entry_pco is equal"); | |
| 120 return true; | |
| 121 } | |
| 122 | |
| 123 | |
| 124 // Implementation of IRScope | |
| 125 BlockBegin* IRScope::build_graph(Compilation* compilation, int osr_bci) { | |
| 126 GraphBuilder gm(compilation, this); | |
| 127 NOT_PRODUCT(if (PrintValueNumbering && Verbose) gm.print_stats()); | |
| 128 if (compilation->bailed_out()) return NULL; | |
| 129 return gm.start(); | |
| 130 } | |
| 131 | |
| 132 | |
| 133 IRScope::IRScope(Compilation* compilation, IRScope* caller, int caller_bci, ciMethod* method, int osr_bci, bool create_graph) | |
| 134 : _callees(2) | |
| 135 , _compilation(compilation) | |
| 136 , _requires_phi_function(method->max_locals()) | |
| 137 { | |
| 138 _caller = caller; | |
| 139 _level = caller == NULL ? 0 : caller->level() + 1; | |
| 140 _method = method; | |
| 141 _xhandlers = new XHandlers(method); | |
| 142 _number_of_locks = 0; | |
| 143 _monitor_pairing_ok = method->has_balanced_monitors(); | |
| 144 _start = NULL; | |
| 145 | |
| 146 if (osr_bci == -1) { | |
| 147 _requires_phi_function.clear(); | |
| 148 } else { | |
| 149 // selective creation of phi functions is not possibel in osr-methods | |
| 150 _requires_phi_function.set_range(0, method->max_locals()); | |
| 151 } | |
| 152 | |
| 153 assert(method->holder()->is_loaded() , "method holder must be loaded"); | |
| 154 | |
| 155 // build graph if monitor pairing is ok | |
| 156 if (create_graph && monitor_pairing_ok()) _start = build_graph(compilation, osr_bci); | |
| 157 } | |
| 158 | |
| 159 | |
| 160 int IRScope::max_stack() const { | |
| 161 int my_max = method()->max_stack(); | |
| 162 int callee_max = 0; | |
| 163 for (int i = 0; i < number_of_callees(); i++) { | |
| 164 callee_max = MAX2(callee_max, callee_no(i)->max_stack()); | |
| 165 } | |
| 166 return my_max + callee_max; | |
| 167 } | |
| 168 | |
| 169 | |
|
900
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
170 bool IRScopeDebugInfo::should_reexecute() { |
|
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
171 ciMethod* cur_method = scope()->method(); |
|
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
172 int cur_bci = bci(); |
|
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
173 if (cur_method != NULL && cur_bci != SynchronizationEntryBCI) { |
|
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
174 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); |
|
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
175 return Interpreter::bytecode_should_reexecute(code); |
|
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
176 } else |
|
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
177 return false; |
|
9987d9d5eb0e
6833129: specjvm98 fails with NullPointerException in the compiler with -XX:DeoptimizeALot
cfang
parents:
470
diff
changeset
|
178 } |
| 0 | 179 |
| 180 | |
| 181 // Implementation of CodeEmitInfo | |
| 182 | |
| 183 // Stack must be NON-null | |
| 1819 | 184 CodeEmitInfo::CodeEmitInfo(ValueStack* stack, XHandlers* exception_handlers) |
| 0 | 185 : _scope(stack->scope()) |
| 186 , _scope_debug_info(NULL) | |
| 187 , _oop_map(NULL) | |
| 188 , _stack(stack) | |
| 189 , _exception_handlers(exception_handlers) | |
| 1564 | 190 , _is_method_handle_invoke(false) { |
| 0 | 191 assert(_stack != NULL, "must be non null"); |
| 192 } | |
| 193 | |
| 194 | |
| 1819 | 195 CodeEmitInfo::CodeEmitInfo(CodeEmitInfo* info, ValueStack* stack) |
| 0 | 196 : _scope(info->_scope) |
| 197 , _exception_handlers(NULL) | |
| 198 , _scope_debug_info(NULL) | |
| 1564 | 199 , _oop_map(NULL) |
| 1819 | 200 , _stack(stack == NULL ? info->_stack : stack) |
| 1564 | 201 , _is_method_handle_invoke(info->_is_method_handle_invoke) { |
| 0 | 202 |
| 203 // deep copy of exception handlers | |
| 204 if (info->_exception_handlers != NULL) { | |
| 205 _exception_handlers = new XHandlers(info->_exception_handlers); | |
| 206 } | |
| 207 } | |
| 208 | |
| 209 | |
| 1564 | 210 void CodeEmitInfo::record_debug_info(DebugInformationRecorder* recorder, int pc_offset) { |
| 0 | 211 // record the safepoint before recording the debug info for enclosing scopes |
| 212 recorder->add_safepoint(pc_offset, _oop_map->deep_copy()); | |
| 1564 | 213 _scope_debug_info->record_debug_info(recorder, pc_offset, true/*topmost*/, _is_method_handle_invoke); |
| 0 | 214 recorder->end_safepoint(pc_offset); |
| 215 } | |
| 216 | |
| 217 | |
| 218 void CodeEmitInfo::add_register_oop(LIR_Opr opr) { | |
| 219 assert(_oop_map != NULL, "oop map must already exist"); | |
| 220 assert(opr->is_single_cpu(), "should not call otherwise"); | |
| 221 | |
| 222 VMReg name = frame_map()->regname(opr); | |
| 223 _oop_map->set_oop(name); | |
| 224 } | |
| 225 | |
| 226 | |
| 227 | |
| 228 | |
| 229 // Implementation of IR | |
| 230 | |
| 231 IR::IR(Compilation* compilation, ciMethod* method, int osr_bci) : | |
| 232 _locals_size(in_WordSize(-1)) | |
| 233 , _num_loops(0) { | |
| 234 // setup IR fields | |
| 235 _compilation = compilation; | |
| 236 _top_scope = new IRScope(compilation, NULL, -1, method, osr_bci, true); | |
| 237 _code = NULL; | |
| 238 } | |
| 239 | |
| 240 | |
| 241 void IR::optimize() { | |
| 242 Optimizer opt(this); | |
| 1783 | 243 if (!compilation()->profile_branches()) { |
| 244 if (DoCEE) { | |
| 245 opt.eliminate_conditional_expressions(); | |
| 0 | 246 #ifndef PRODUCT |
| 1783 | 247 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after CEE"); print(true); } |
| 248 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after CEE"); print(false); } | |
| 0 | 249 #endif |
| 1783 | 250 } |
| 251 if (EliminateBlocks) { | |
| 252 opt.eliminate_blocks(); | |
| 0 | 253 #ifndef PRODUCT |
| 1783 | 254 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after block elimination"); print(true); } |
| 255 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after block elimination"); print(false); } | |
| 0 | 256 #endif |
| 1783 | 257 } |
| 0 | 258 } |
| 259 if (EliminateNullChecks) { | |
| 260 opt.eliminate_null_checks(); | |
| 261 #ifndef PRODUCT | |
| 262 if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after null check elimination"); print(true); } | |
| 263 if (PrintIR || PrintIR1 ) { tty->print_cr("IR after null check elimination"); print(false); } | |
| 264 #endif | |
| 265 } | |
| 266 } | |
| 267 | |
| 268 | |
| 269 static int sort_pairs(BlockPair** a, BlockPair** b) { | |
| 270 if ((*a)->from() == (*b)->from()) { | |
| 271 return (*a)->to()->block_id() - (*b)->to()->block_id(); | |
| 272 } else { | |
| 273 return (*a)->from()->block_id() - (*b)->from()->block_id(); | |
| 274 } | |
| 275 } | |
| 276 | |
| 277 | |
| 278 class CriticalEdgeFinder: public BlockClosure { | |
| 279 BlockPairList blocks; | |
| 280 IR* _ir; | |
| 281 | |
| 282 public: | |
| 283 CriticalEdgeFinder(IR* ir): _ir(ir) {} | |
| 284 void block_do(BlockBegin* bb) { | |
| 285 BlockEnd* be = bb->end(); | |
| 286 int nos = be->number_of_sux(); | |
| 287 if (nos >= 2) { | |
| 288 for (int i = 0; i < nos; i++) { | |
| 289 BlockBegin* sux = be->sux_at(i); | |
| 290 if (sux->number_of_preds() >= 2) { | |
| 291 blocks.append(new BlockPair(bb, sux)); | |
| 292 } | |
| 293 } | |
| 294 } | |
| 295 } | |
| 296 | |
| 297 void split_edges() { | |
| 298 BlockPair* last_pair = NULL; | |
| 299 blocks.sort(sort_pairs); | |
| 300 for (int i = 0; i < blocks.length(); i++) { | |
| 301 BlockPair* pair = blocks.at(i); | |
| 302 if (last_pair != NULL && pair->is_same(last_pair)) continue; | |
| 303 BlockBegin* from = pair->from(); | |
| 304 BlockBegin* to = pair->to(); | |
| 305 BlockBegin* split = from->insert_block_between(to); | |
| 306 #ifndef PRODUCT | |
| 307 if ((PrintIR || PrintIR1) && Verbose) { | |
| 308 tty->print_cr("Split critical edge B%d -> B%d (new block B%d)", | |
| 309 from->block_id(), to->block_id(), split->block_id()); | |
| 310 } | |
| 311 #endif | |
| 312 last_pair = pair; | |
| 313 } | |
| 314 } | |
| 315 }; | |
| 316 | |
| 317 void IR::split_critical_edges() { | |
| 318 CriticalEdgeFinder cef(this); | |
| 319 | |
| 320 iterate_preorder(&cef); | |
| 321 cef.split_edges(); | |
| 322 } | |
| 323 | |
| 324 | |
| 1584 | 325 class UseCountComputer: public ValueVisitor, BlockClosure { |
| 0 | 326 private: |
| 1584 | 327 void visit(Value* n) { |
| 0 | 328 // Local instructions and Phis for expression stack values at the |
| 329 // start of basic blocks are not added to the instruction list | |
| 1899 | 330 if (!(*n)->is_linked() && (*n)->can_be_linked()) { |
| 0 | 331 assert(false, "a node was not appended to the graph"); |
| 1584 | 332 Compilation::current()->bailout("a node was not appended to the graph"); |
| 0 | 333 } |
| 334 // use n's input if not visited before | |
| 335 if (!(*n)->is_pinned() && !(*n)->has_uses()) { | |
| 336 // note: a) if the instruction is pinned, it will be handled by compute_use_count | |
| 337 // b) if the instruction has uses, it was touched before | |
| 338 // => in both cases we don't need to update n's values | |
| 339 uses_do(n); | |
| 340 } | |
| 341 // use n | |
| 342 (*n)->_use_count++; | |
| 343 } | |
| 344 | |
| 1584 | 345 Values* worklist; |
| 346 int depth; | |
| 0 | 347 enum { |
| 348 max_recurse_depth = 20 | |
| 349 }; | |
| 350 | |
| 1584 | 351 void uses_do(Value* n) { |
| 0 | 352 depth++; |
| 353 if (depth > max_recurse_depth) { | |
| 354 // don't allow the traversal to recurse too deeply | |
| 355 worklist->push(*n); | |
| 356 } else { | |
| 1584 | 357 (*n)->input_values_do(this); |
| 0 | 358 // special handling for some instructions |
| 359 if ((*n)->as_BlockEnd() != NULL) { | |
| 360 // note on BlockEnd: | |
| 361 // must 'use' the stack only if the method doesn't | |
| 362 // terminate, however, in those cases stack is empty | |
| 1584 | 363 (*n)->state_values_do(this); |
| 0 | 364 } |
| 365 } | |
| 366 depth--; | |
| 367 } | |
| 368 | |
| 1584 | 369 void block_do(BlockBegin* b) { |
| 0 | 370 depth = 0; |
| 371 // process all pinned nodes as the roots of expression trees | |
| 372 for (Instruction* n = b; n != NULL; n = n->next()) { | |
| 373 if (n->is_pinned()) uses_do(&n); | |
| 374 } | |
| 375 assert(depth == 0, "should have counted back down"); | |
| 376 | |
| 377 // now process any unpinned nodes which recursed too deeply | |
| 378 while (worklist->length() > 0) { | |
| 379 Value t = worklist->pop(); | |
| 380 if (!t->is_pinned()) { | |
| 381 // compute the use count | |
| 382 uses_do(&t); | |
| 383 | |
| 384 // pin the instruction so that LIRGenerator doesn't recurse | |
| 385 // too deeply during it's evaluation. | |
| 386 t->pin(); | |
| 387 } | |
| 388 } | |
| 389 assert(depth == 0, "should have counted back down"); | |
| 390 } | |
| 391 | |
| 1584 | 392 UseCountComputer() { |
| 393 worklist = new Values(); | |
| 394 depth = 0; | |
| 395 } | |
| 396 | |
| 0 | 397 public: |
| 398 static void compute(BlockList* blocks) { | |
| 1584 | 399 UseCountComputer ucc; |
| 400 blocks->iterate_backward(&ucc); | |
| 0 | 401 } |
| 402 }; | |
| 403 | |
| 404 | |
| 405 // helper macro for short definition of trace-output inside code | |
| 406 #ifndef PRODUCT | |
| 407 #define TRACE_LINEAR_SCAN(level, code) \ | |
| 408 if (TraceLinearScanLevel >= level) { \ | |
| 409 code; \ | |
| 410 } | |
| 411 #else | |
| 412 #define TRACE_LINEAR_SCAN(level, code) | |
| 413 #endif | |
| 414 | |
| 415 class ComputeLinearScanOrder : public StackObj { | |
| 416 private: | |
| 417 int _max_block_id; // the highest block_id of a block | |
| 418 int _num_blocks; // total number of blocks (smaller than _max_block_id) | |
| 419 int _num_loops; // total number of loops | |
| 420 bool _iterative_dominators;// method requires iterative computation of dominatiors | |
| 421 | |
| 422 BlockList* _linear_scan_order; // the resulting list of blocks in correct order | |
| 423 | |
| 424 BitMap _visited_blocks; // used for recursive processing of blocks | |
| 425 BitMap _active_blocks; // used for recursive processing of blocks | |
| 426 BitMap _dominator_blocks; // temproary BitMap used for computation of dominator | |
| 427 intArray _forward_branches; // number of incoming forward branches for each block | |
| 428 BlockList _loop_end_blocks; // list of all loop end blocks collected during count_edges | |
| 429 BitMap2D _loop_map; // two-dimensional bit set: a bit is set if a block is contained in a loop | |
| 430 BlockList _work_list; // temporary list (used in mark_loops and compute_order) | |
| 431 | |
| 1783 | 432 Compilation* _compilation; |
| 433 | |
| 0 | 434 // accessors for _visited_blocks and _active_blocks |
| 435 void init_visited() { _active_blocks.clear(); _visited_blocks.clear(); } | |
| 436 bool is_visited(BlockBegin* b) const { return _visited_blocks.at(b->block_id()); } | |
| 437 bool is_active(BlockBegin* b) const { return _active_blocks.at(b->block_id()); } | |
| 438 void set_visited(BlockBegin* b) { assert(!is_visited(b), "already set"); _visited_blocks.set_bit(b->block_id()); } | |
| 439 void set_active(BlockBegin* b) { assert(!is_active(b), "already set"); _active_blocks.set_bit(b->block_id()); } | |
| 440 void clear_active(BlockBegin* b) { assert(is_active(b), "not already"); _active_blocks.clear_bit(b->block_id()); } | |
| 441 | |
| 442 // accessors for _forward_branches | |
| 443 void inc_forward_branches(BlockBegin* b) { _forward_branches.at_put(b->block_id(), _forward_branches.at(b->block_id()) + 1); } | |
| 444 int dec_forward_branches(BlockBegin* b) { _forward_branches.at_put(b->block_id(), _forward_branches.at(b->block_id()) - 1); return _forward_branches.at(b->block_id()); } | |
| 445 | |
| 446 // accessors for _loop_map | |
| 447 bool is_block_in_loop (int loop_idx, BlockBegin* b) const { return _loop_map.at(loop_idx, b->block_id()); } | |
| 448 void set_block_in_loop (int loop_idx, BlockBegin* b) { _loop_map.set_bit(loop_idx, b->block_id()); } | |
| 449 void clear_block_in_loop(int loop_idx, int block_id) { _loop_map.clear_bit(loop_idx, block_id); } | |
| 450 | |
| 451 // count edges between blocks | |
| 452 void count_edges(BlockBegin* cur, BlockBegin* parent); | |
| 453 | |
| 454 // loop detection | |
| 455 void mark_loops(); | |
| 456 void clear_non_natural_loops(BlockBegin* start_block); | |
| 457 void assign_loop_depth(BlockBegin* start_block); | |
| 458 | |
| 459 // computation of final block order | |
| 460 BlockBegin* common_dominator(BlockBegin* a, BlockBegin* b); | |
| 461 void compute_dominator(BlockBegin* cur, BlockBegin* parent); | |
| 462 int compute_weight(BlockBegin* cur); | |
| 463 bool ready_for_processing(BlockBegin* cur); | |
| 464 void sort_into_work_list(BlockBegin* b); | |
| 465 void append_block(BlockBegin* cur); | |
| 466 void compute_order(BlockBegin* start_block); | |
| 467 | |
| 468 // fixup of dominators for non-natural loops | |
| 469 bool compute_dominators_iter(); | |
| 470 void compute_dominators(); | |
| 471 | |
| 472 // debug functions | |
| 473 NOT_PRODUCT(void print_blocks();) | |
| 474 DEBUG_ONLY(void verify();) | |
| 475 | |
| 1783 | 476 Compilation* compilation() const { return _compilation; } |
| 0 | 477 public: |
| 1783 | 478 ComputeLinearScanOrder(Compilation* c, BlockBegin* start_block); |
| 0 | 479 |
| 480 // accessors for final result | |
| 481 BlockList* linear_scan_order() const { return _linear_scan_order; } | |
| 482 int num_loops() const { return _num_loops; } | |
| 483 }; | |
| 484 | |
| 485 | |
| 1783 | 486 ComputeLinearScanOrder::ComputeLinearScanOrder(Compilation* c, BlockBegin* start_block) : |
| 0 | 487 _max_block_id(BlockBegin::number_of_blocks()), |
| 488 _num_blocks(0), | |
| 489 _num_loops(0), | |
| 490 _iterative_dominators(false), | |
| 491 _visited_blocks(_max_block_id), | |
| 492 _active_blocks(_max_block_id), | |
| 493 _dominator_blocks(_max_block_id), | |
| 494 _forward_branches(_max_block_id, 0), | |
| 495 _loop_end_blocks(8), | |
| 496 _work_list(8), | |
| 497 _linear_scan_order(NULL), // initialized later with correct size | |
| 1783 | 498 _loop_map(0, 0), // initialized later with correct size |
| 499 _compilation(c) | |
| 0 | 500 { |
| 501 TRACE_LINEAR_SCAN(2, "***** computing linear-scan block order"); | |
| 502 | |
| 503 init_visited(); | |
| 504 count_edges(start_block, NULL); | |
| 505 | |
| 1783 | 506 if (compilation()->is_profiling()) { |
| 507 compilation()->method()->method_data()->set_compilation_stats(_num_loops, _num_blocks); | |
| 508 } | |
| 509 | |
| 0 | 510 if (_num_loops > 0) { |
| 511 mark_loops(); | |
| 512 clear_non_natural_loops(start_block); | |
| 513 assign_loop_depth(start_block); | |
| 514 } | |
| 515 | |
| 516 compute_order(start_block); | |
| 517 compute_dominators(); | |
| 518 | |
| 519 NOT_PRODUCT(print_blocks()); | |
| 520 DEBUG_ONLY(verify()); | |
| 521 } | |
| 522 | |
| 523 | |
| 524 // Traverse the CFG: | |
| 525 // * count total number of blocks | |
| 526 // * count all incoming edges and backward incoming edges | |
| 527 // * number loop header blocks | |
| 528 // * create a list with all loop end blocks | |
| 529 void ComputeLinearScanOrder::count_edges(BlockBegin* cur, BlockBegin* parent) { | |
| 530 TRACE_LINEAR_SCAN(3, tty->print_cr("Enter count_edges for block B%d coming from B%d", cur->block_id(), parent != NULL ? parent->block_id() : -1)); | |
| 531 assert(cur->dominator() == NULL, "dominator already initialized"); | |
| 532 | |
| 533 if (is_active(cur)) { | |
| 534 TRACE_LINEAR_SCAN(3, tty->print_cr("backward branch")); | |
| 535 assert(is_visited(cur), "block must be visisted when block is active"); | |
| 536 assert(parent != NULL, "must have parent"); | |
| 537 | |
| 538 cur->set(BlockBegin::linear_scan_loop_header_flag); | |
| 539 cur->set(BlockBegin::backward_branch_target_flag); | |
| 540 | |
| 541 parent->set(BlockBegin::linear_scan_loop_end_flag); | |
|
428
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
542 |
|
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
543 // When a loop header is also the start of an exception handler, then the backward branch is |
|
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
544 // an exception edge. Because such edges are usually critical edges which cannot be split, the |
|
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
545 // loop must be excluded here from processing. |
|
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
546 if (cur->is_set(BlockBegin::exception_entry_flag)) { |
|
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
547 // Make sure that dominators are correct in this weird situation |
|
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
548 _iterative_dominators = true; |
|
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
549 return; |
|
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
550 } |
|
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
551 assert(parent->number_of_sux() == 1 && parent->sux_at(0) == cur, |
|
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
552 "loop end blocks must have one successor (critical edges are split)"); |
|
334969144810
6758445: loop heads that are exception entry points can crash during count_edges/mark_loops
never
parents:
0
diff
changeset
|
553 |
| 0 | 554 _loop_end_blocks.append(parent); |
| 555 return; | |
| 556 } | |
| 557 | |
| 558 // increment number of incoming forward branches | |
| 559 inc_forward_branches(cur); | |
| 560 | |
| 561 if (is_visited(cur)) { | |
| 562 TRACE_LINEAR_SCAN(3, tty->print_cr("block already visited")); | |
| 563 return; | |
| 564 } | |
| 565 | |
| 566 _num_blocks++; | |
| 567 set_visited(cur); | |
| 568 set_active(cur); | |
| 569 | |
| 570 // recursive call for all successors | |
| 571 int i; | |
| 572 for (i = cur->number_of_sux() - 1; i >= 0; i--) { | |
| 573 count_edges(cur->sux_at(i), cur); | |
| 574 } | |
| 575 for (i = cur->number_of_exception_handlers() - 1; i >= 0; i--) { | |
| 576 count_edges(cur->exception_handler_at(i), cur); | |
| 577 } | |
| 578 | |
| 579 clear_active(cur); | |
| 580 | |
| 581 // Each loop has a unique number. | |
| 582 // When multiple loops are nested, assign_loop_depth assumes that the | |
| 583 // innermost loop has the lowest number. This is guaranteed by setting | |
| 584 // the loop number after the recursive calls for the successors above | |
| 585 // have returned. | |
| 586 if (cur->is_set(BlockBegin::linear_scan_loop_header_flag)) { | |
| 587 assert(cur->loop_index() == -1, "cannot set loop-index twice"); | |
| 588 TRACE_LINEAR_SCAN(3, tty->print_cr("Block B%d is loop header of loop %d", cur->block_id(), _num_loops)); | |
| 589 | |
| 590 cur->set_loop_index(_num_loops); | |
| 591 _num_loops++; | |
| 592 } | |
| 593 | |
| 594 TRACE_LINEAR_SCAN(3, tty->print_cr("Finished count_edges for block B%d", cur->block_id())); | |
| 595 } | |
| 596 | |
| 597 | |
| 598 void ComputeLinearScanOrder::mark_loops() { | |
| 599 TRACE_LINEAR_SCAN(3, tty->print_cr("----- marking loops")); | |
| 600 | |
| 601 _loop_map = BitMap2D(_num_loops, _max_block_id); | |
| 602 _loop_map.clear(); | |
| 603 | |
| 604 for (int i = _loop_end_blocks.length() - 1; i >= 0; i--) { | |
| 605 BlockBegin* loop_end = _loop_end_blocks.at(i); | |
| 606 BlockBegin* loop_start = loop_end->sux_at(0); | |
| 607 int loop_idx = loop_start->loop_index(); | |
| 608 | |
| 609 TRACE_LINEAR_SCAN(3, tty->print_cr("Processing loop from B%d to B%d (loop %d):", loop_start->block_id(), loop_end->block_id(), loop_idx)); | |
| 610 assert(loop_end->is_set(BlockBegin::linear_scan_loop_end_flag), "loop end flag must be set"); | |
| 611 assert(loop_end->number_of_sux() == 1, "incorrect number of successors"); | |
| 612 assert(loop_start->is_set(BlockBegin::linear_scan_loop_header_flag), "loop header flag must be set"); | |
| 613 assert(loop_idx >= 0 && loop_idx < _num_loops, "loop index not set"); | |
| 614 assert(_work_list.is_empty(), "work list must be empty before processing"); | |
| 615 | |
| 616 // add the end-block of the loop to the working list | |
| 617 _work_list.push(loop_end); | |
| 618 set_block_in_loop(loop_idx, loop_end); | |
| 619 do { | |
| 620 BlockBegin* cur = _work_list.pop(); | |
| 621 | |
| 622 TRACE_LINEAR_SCAN(3, tty->print_cr(" processing B%d", cur->block_id())); | |
| 623 assert(is_block_in_loop(loop_idx, cur), "bit in loop map must be set when block is in work list"); | |
| 624 | |
| 625 // recursive processing of all predecessors ends when start block of loop is reached | |
| 626 if (cur != loop_start && !cur->is_set(BlockBegin::osr_entry_flag)) { | |
| 627 for (int j = cur->number_of_preds() - 1; j >= 0; j--) { | |
| 628 BlockBegin* pred = cur->pred_at(j); | |
| 629 | |
| 630 if (!is_block_in_loop(loop_idx, pred) /*&& !pred->is_set(BlockBeginosr_entry_flag)*/) { | |
| 631 // this predecessor has not been processed yet, so add it to work list | |
| 632 TRACE_LINEAR_SCAN(3, tty->print_cr(" pushing B%d", pred->block_id())); | |
| 633 _work_list.push(pred); | |
| 634 set_block_in_loop(loop_idx, pred); | |
| 635 } | |
| 636 } | |
| 637 } | |
| 638 } while (!_work_list.is_empty()); | |
| 639 } | |
| 640 } | |
| 641 | |
| 642 | |
| 643 // check for non-natural loops (loops where the loop header does not dominate | |
| 644 // all other loop blocks = loops with mulitple entries). | |
| 645 // such loops are ignored | |
| 646 void ComputeLinearScanOrder::clear_non_natural_loops(BlockBegin* start_block) { | |
| 647 for (int i = _num_loops - 1; i >= 0; i--) { | |
| 648 if (is_block_in_loop(i, start_block)) { | |
| 649 // loop i contains the entry block of the method | |
| 650 // -> this is not a natural loop, so ignore it | |
| 651 TRACE_LINEAR_SCAN(2, tty->print_cr("Loop %d is non-natural, so it is ignored", i)); | |
| 652 | |
| 653 for (int block_id = _max_block_id - 1; block_id >= 0; block_id--) { | |
| 654 clear_block_in_loop(i, block_id); | |
| 655 } | |
| 656 _iterative_dominators = true; | |
| 657 } | |
| 658 } | |
| 659 } | |
| 660 | |
| 661 void ComputeLinearScanOrder::assign_loop_depth(BlockBegin* start_block) { | |
| 662 TRACE_LINEAR_SCAN(3, "----- computing loop-depth and weight"); | |
| 663 init_visited(); | |
| 664 | |
| 665 assert(_work_list.is_empty(), "work list must be empty before processing"); | |
| 666 _work_list.append(start_block); | |
| 667 | |
| 668 do { | |
| 669 BlockBegin* cur = _work_list.pop(); | |
| 670 | |
| 671 if (!is_visited(cur)) { | |
| 672 set_visited(cur); | |
| 673 TRACE_LINEAR_SCAN(4, tty->print_cr("Computing loop depth for block B%d", cur->block_id())); | |
| 674 | |
| 675 // compute loop-depth and loop-index for the block | |
| 676 assert(cur->loop_depth() == 0, "cannot set loop-depth twice"); | |
| 677 int i; | |
| 678 int loop_depth = 0; | |
| 679 int min_loop_idx = -1; | |
| 680 for (i = _num_loops - 1; i >= 0; i--) { | |
| 681 if (is_block_in_loop(i, cur)) { | |
| 682 loop_depth++; | |
| 683 min_loop_idx = i; | |
| 684 } | |
| 685 } | |
| 686 cur->set_loop_depth(loop_depth); | |
| 687 cur->set_loop_index(min_loop_idx); | |
| 688 | |
| 689 // append all unvisited successors to work list | |
| 690 for (i = cur->number_of_sux() - 1; i >= 0; i--) { | |
| 691 _work_list.append(cur->sux_at(i)); | |
| 692 } | |
| 693 for (i = cur->number_of_exception_handlers() - 1; i >= 0; i--) { | |
| 694 _work_list.append(cur->exception_handler_at(i)); | |
| 695 } | |
| 696 } | |
| 697 } while (!_work_list.is_empty()); | |
| 698 } | |
| 699 | |
| 700 | |
| 701 BlockBegin* ComputeLinearScanOrder::common_dominator(BlockBegin* a, BlockBegin* b) { | |
| 702 assert(a != NULL && b != NULL, "must have input blocks"); | |
| 703 | |
| 704 _dominator_blocks.clear(); | |
| 705 while (a != NULL) { | |
| 706 _dominator_blocks.set_bit(a->block_id()); | |
| 707 assert(a->dominator() != NULL || a == _linear_scan_order->at(0), "dominator must be initialized"); | |
| 708 a = a->dominator(); | |
| 709 } | |
| 710 while (b != NULL && !_dominator_blocks.at(b->block_id())) { | |
| 711 assert(b->dominator() != NULL || b == _linear_scan_order->at(0), "dominator must be initialized"); | |
| 712 b = b->dominator(); | |
| 713 } | |
| 714 | |
| 715 assert(b != NULL, "could not find dominator"); | |
| 716 return b; | |
| 717 } | |
| 718 | |
| 719 void ComputeLinearScanOrder::compute_dominator(BlockBegin* cur, BlockBegin* parent) { | |
| 720 if (cur->dominator() == NULL) { | |
| 721 TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: initializing dominator of B%d to B%d", cur->block_id(), parent->block_id())); | |
| 722 cur->set_dominator(parent); | |
| 723 | |
| 724 } else if (!(cur->is_set(BlockBegin::linear_scan_loop_header_flag) && parent->is_set(BlockBegin::linear_scan_loop_end_flag))) { | |
| 725 TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: computing dominator of B%d: common dominator of B%d and B%d is B%d", cur->block_id(), parent->block_id(), cur->dominator()->block_id(), common_dominator(cur->dominator(), parent)->block_id())); | |
| 726 assert(cur->number_of_preds() > 1, ""); | |
| 727 cur->set_dominator(common_dominator(cur->dominator(), parent)); | |
| 728 } | |
| 729 } | |
| 730 | |
| 731 | |
| 732 int ComputeLinearScanOrder::compute_weight(BlockBegin* cur) { | |
| 733 BlockBegin* single_sux = NULL; | |
| 734 if (cur->number_of_sux() == 1) { | |
| 735 single_sux = cur->sux_at(0); | |
| 736 } | |
| 737 | |
| 738 // limit loop-depth to 15 bit (only for security reason, it will never be so big) | |
| 739 int weight = (cur->loop_depth() & 0x7FFF) << 16; | |
| 740 | |
| 741 // general macro for short definition of weight flags | |
| 742 // the first instance of INC_WEIGHT_IF has the highest priority | |
| 743 int cur_bit = 15; | |
| 744 #define INC_WEIGHT_IF(condition) if ((condition)) { weight |= (1 << cur_bit); } cur_bit--; | |
| 745 | |
| 746 // this is necessery for the (very rare) case that two successing blocks have | |
| 747 // the same loop depth, but a different loop index (can happen for endless loops | |
| 748 // with exception handlers) | |
| 749 INC_WEIGHT_IF(!cur->is_set(BlockBegin::linear_scan_loop_header_flag)); | |
| 750 | |
| 751 // loop end blocks (blocks that end with a backward branch) are added | |
| 752 // after all other blocks of the loop. | |
| 753 INC_WEIGHT_IF(!cur->is_set(BlockBegin::linear_scan_loop_end_flag)); | |
| 754 | |
| 755 // critical edge split blocks are prefered because than they have a bigger | |
| 756 // proability to be completely empty | |
| 757 INC_WEIGHT_IF(cur->is_set(BlockBegin::critical_edge_split_flag)); | |
| 758 | |
| 759 // exceptions should not be thrown in normal control flow, so these blocks | |
| 760 // are added as late as possible | |
| 761 INC_WEIGHT_IF(cur->end()->as_Throw() == NULL && (single_sux == NULL || single_sux->end()->as_Throw() == NULL)); | |
| 762 INC_WEIGHT_IF(cur->end()->as_Return() == NULL && (single_sux == NULL || single_sux->end()->as_Return() == NULL)); | |
| 763 | |
| 764 // exceptions handlers are added as late as possible | |
| 765 INC_WEIGHT_IF(!cur->is_set(BlockBegin::exception_entry_flag)); | |
| 766 | |
| 767 // guarantee that weight is > 0 | |
| 768 weight |= 1; | |
| 769 | |
| 770 #undef INC_WEIGHT_IF | |
| 771 assert(cur_bit >= 0, "too many flags"); | |
| 772 assert(weight > 0, "weight cannot become negative"); | |
| 773 | |
| 774 return weight; | |
| 775 } | |
| 776 | |
| 777 bool ComputeLinearScanOrder::ready_for_processing(BlockBegin* cur) { | |
| 778 // Discount the edge just traveled. | |
| 779 // When the number drops to zero, all forward branches were processed | |
| 780 if (dec_forward_branches(cur) != 0) { | |
| 781 return false; | |
| 782 } | |
| 783 | |
| 784 assert(_linear_scan_order->index_of(cur) == -1, "block already processed (block can be ready only once)"); | |
| 785 assert(_work_list.index_of(cur) == -1, "block already in work-list (block can be ready only once)"); | |
| 786 return true; | |
| 787 } | |
| 788 | |
| 789 void ComputeLinearScanOrder::sort_into_work_list(BlockBegin* cur) { | |
| 790 assert(_work_list.index_of(cur) == -1, "block already in work list"); | |
| 791 | |
| 792 int cur_weight = compute_weight(cur); | |
| 793 | |
| 794 // the linear_scan_number is used to cache the weight of a block | |
| 795 cur->set_linear_scan_number(cur_weight); | |
| 796 | |
| 797 #ifndef PRODUCT | |
| 798 if (StressLinearScan) { | |
| 799 _work_list.insert_before(0, cur); | |
| 800 return; | |
| 801 } | |
| 802 #endif | |
| 803 | |
| 804 _work_list.append(NULL); // provide space for new element | |
| 805 | |
| 806 int insert_idx = _work_list.length() - 1; | |
| 807 while (insert_idx > 0 && _work_list.at(insert_idx - 1)->linear_scan_number() > cur_weight) { | |
| 808 _work_list.at_put(insert_idx, _work_list.at(insert_idx - 1)); | |
| 809 insert_idx--; | |
| 810 } | |
| 811 _work_list.at_put(insert_idx, cur); | |
| 812 | |
| 813 TRACE_LINEAR_SCAN(3, tty->print_cr("Sorted B%d into worklist. new worklist:", cur->block_id())); | |
| 814 TRACE_LINEAR_SCAN(3, for (int i = 0; i < _work_list.length(); i++) tty->print_cr("%8d B%2d weight:%6x", i, _work_list.at(i)->block_id(), _work_list.at(i)->linear_scan_number())); | |
| 815 | |
| 816 #ifdef ASSERT | |
| 817 for (int i = 0; i < _work_list.length(); i++) { | |
| 818 assert(_work_list.at(i)->linear_scan_number() > 0, "weight not set"); | |
| 819 assert(i == 0 || _work_list.at(i - 1)->linear_scan_number() <= _work_list.at(i)->linear_scan_number(), "incorrect order in worklist"); | |
| 820 } | |
| 821 #endif | |
| 822 } | |
| 823 | |
| 824 void ComputeLinearScanOrder::append_block(BlockBegin* cur) { | |
| 825 TRACE_LINEAR_SCAN(3, tty->print_cr("appending block B%d (weight 0x%6x) to linear-scan order", cur->block_id(), cur->linear_scan_number())); | |
| 826 assert(_linear_scan_order->index_of(cur) == -1, "cannot add the same block twice"); | |
| 827 | |
| 828 // currently, the linear scan order and code emit order are equal. | |
| 829 // therefore the linear_scan_number and the weight of a block must also | |
| 830 // be equal. | |
| 831 cur->set_linear_scan_number(_linear_scan_order->length()); | |
| 832 _linear_scan_order->append(cur); | |
| 833 } | |
| 834 | |
| 835 void ComputeLinearScanOrder::compute_order(BlockBegin* start_block) { | |
| 836 TRACE_LINEAR_SCAN(3, "----- computing final block order"); | |
| 837 | |
| 838 // the start block is always the first block in the linear scan order | |
| 839 _linear_scan_order = new BlockList(_num_blocks); | |
| 840 append_block(start_block); | |
| 841 | |
| 842 assert(start_block->end()->as_Base() != NULL, "start block must end with Base-instruction"); | |
| 843 BlockBegin* std_entry = ((Base*)start_block->end())->std_entry(); | |
| 844 BlockBegin* osr_entry = ((Base*)start_block->end())->osr_entry(); | |
| 845 | |
| 846 BlockBegin* sux_of_osr_entry = NULL; | |
| 847 if (osr_entry != NULL) { | |
| 848 // special handling for osr entry: | |
| 849 // ignore the edge between the osr entry and its successor for processing | |
| 850 // the osr entry block is added manually below | |
| 851 assert(osr_entry->number_of_sux() == 1, "osr entry must have exactly one successor"); | |
| 852 assert(osr_entry->sux_at(0)->number_of_preds() >= 2, "sucessor of osr entry must have two predecessors (otherwise it is not present in normal control flow"); | |
| 853 | |
| 854 sux_of_osr_entry = osr_entry->sux_at(0); | |
| 855 dec_forward_branches(sux_of_osr_entry); | |
| 856 | |
| 857 compute_dominator(osr_entry, start_block); | |
| 858 _iterative_dominators = true; | |
| 859 } | |
| 860 compute_dominator(std_entry, start_block); | |
| 861 | |
| 862 // start processing with standard entry block | |
| 863 assert(_work_list.is_empty(), "list must be empty before processing"); | |
| 864 | |
| 865 if (ready_for_processing(std_entry)) { | |
| 866 sort_into_work_list(std_entry); | |
| 867 } else { | |
| 868 assert(false, "the std_entry must be ready for processing (otherwise, the method has no start block)"); | |
| 869 } | |
| 870 | |
| 871 do { | |
| 872 BlockBegin* cur = _work_list.pop(); | |
| 873 | |
| 874 if (cur == sux_of_osr_entry) { | |
| 875 // the osr entry block is ignored in normal processing, it is never added to the | |
| 876 // work list. Instead, it is added as late as possible manually here. | |
| 877 append_block(osr_entry); | |
| 878 compute_dominator(cur, osr_entry); | |
| 879 } | |
| 880 append_block(cur); | |
| 881 | |
| 882 int i; | |
| 883 int num_sux = cur->number_of_sux(); | |
| 884 // changed loop order to get "intuitive" order of if- and else-blocks | |
| 885 for (i = 0; i < num_sux; i++) { | |
| 886 BlockBegin* sux = cur->sux_at(i); | |
| 887 compute_dominator(sux, cur); | |
| 888 if (ready_for_processing(sux)) { | |
| 889 sort_into_work_list(sux); | |
| 890 } | |
| 891 } | |
| 892 num_sux = cur->number_of_exception_handlers(); | |
| 893 for (i = 0; i < num_sux; i++) { | |
| 894 BlockBegin* sux = cur->exception_handler_at(i); | |
| 895 compute_dominator(sux, cur); | |
| 896 if (ready_for_processing(sux)) { | |
| 897 sort_into_work_list(sux); | |
| 898 } | |
| 899 } | |
| 900 } while (_work_list.length() > 0); | |
| 901 } | |
| 902 | |
| 903 | |
| 904 bool ComputeLinearScanOrder::compute_dominators_iter() { | |
| 905 bool changed = false; | |
| 906 int num_blocks = _linear_scan_order->length(); | |
| 907 | |
| 908 assert(_linear_scan_order->at(0)->dominator() == NULL, "must not have dominator"); | |
| 909 assert(_linear_scan_order->at(0)->number_of_preds() == 0, "must not have predecessors"); | |
| 910 for (int i = 1; i < num_blocks; i++) { | |
| 911 BlockBegin* block = _linear_scan_order->at(i); | |
| 912 | |
| 913 BlockBegin* dominator = block->pred_at(0); | |
| 914 int num_preds = block->number_of_preds(); | |
| 915 for (int i = 1; i < num_preds; i++) { | |
| 916 dominator = common_dominator(dominator, block->pred_at(i)); | |
| 917 } | |
| 918 | |
| 919 if (dominator != block->dominator()) { | |
| 920 TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: updating dominator of B%d from B%d to B%d", block->block_id(), block->dominator()->block_id(), dominator->block_id())); | |
| 921 | |
| 922 block->set_dominator(dominator); | |
| 923 changed = true; | |
| 924 } | |
| 925 } | |
| 926 return changed; | |
| 927 } | |
| 928 | |
| 929 void ComputeLinearScanOrder::compute_dominators() { | |
| 930 TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing dominators (iterative computation reqired: %d)", _iterative_dominators)); | |
| 931 | |
| 932 // iterative computation of dominators is only required for methods with non-natural loops | |
| 933 // and OSR-methods. For all other methods, the dominators computed when generating the | |
| 934 // linear scan block order are correct. | |
| 935 if (_iterative_dominators) { | |
| 936 do { | |
| 937 TRACE_LINEAR_SCAN(1, tty->print_cr("DOM: next iteration of fix-point calculation")); | |
| 938 } while (compute_dominators_iter()); | |
| 939 } | |
| 940 | |
| 941 // check that dominators are correct | |
| 942 assert(!compute_dominators_iter(), "fix point not reached"); | |
| 943 } | |
| 944 | |
| 945 | |
| 946 #ifndef PRODUCT | |
| 947 void ComputeLinearScanOrder::print_blocks() { | |
| 948 if (TraceLinearScanLevel >= 2) { | |
| 949 tty->print_cr("----- loop information:"); | |
| 950 for (int block_idx = 0; block_idx < _linear_scan_order->length(); block_idx++) { | |
| 951 BlockBegin* cur = _linear_scan_order->at(block_idx); | |
| 952 | |
| 953 tty->print("%4d: B%2d: ", cur->linear_scan_number(), cur->block_id()); | |
| 954 for (int loop_idx = 0; loop_idx < _num_loops; loop_idx++) { | |
| 955 tty->print ("%d ", is_block_in_loop(loop_idx, cur)); | |
| 956 } | |
| 957 tty->print_cr(" -> loop_index: %2d, loop_depth: %2d", cur->loop_index(), cur->loop_depth()); | |
| 958 } | |
| 959 } | |
| 960 | |
| 961 if (TraceLinearScanLevel >= 1) { | |
| 962 tty->print_cr("----- linear-scan block order:"); | |
| 963 for (int block_idx = 0; block_idx < _linear_scan_order->length(); block_idx++) { | |
| 964 BlockBegin* cur = _linear_scan_order->at(block_idx); | |
| 965 tty->print("%4d: B%2d loop: %2d depth: %2d", cur->linear_scan_number(), cur->block_id(), cur->loop_index(), cur->loop_depth()); | |
| 966 | |
| 967 tty->print(cur->is_set(BlockBegin::exception_entry_flag) ? " ex" : " "); | |
| 968 tty->print(cur->is_set(BlockBegin::critical_edge_split_flag) ? " ce" : " "); | |
| 969 tty->print(cur->is_set(BlockBegin::linear_scan_loop_header_flag) ? " lh" : " "); | |
| 970 tty->print(cur->is_set(BlockBegin::linear_scan_loop_end_flag) ? " le" : " "); | |
| 971 | |
| 972 if (cur->dominator() != NULL) { | |
| 973 tty->print(" dom: B%d ", cur->dominator()->block_id()); | |
| 974 } else { | |
| 975 tty->print(" dom: NULL "); | |
| 976 } | |
| 977 | |
| 978 if (cur->number_of_preds() > 0) { | |
| 979 tty->print(" preds: "); | |
| 980 for (int j = 0; j < cur->number_of_preds(); j++) { | |
| 981 BlockBegin* pred = cur->pred_at(j); | |
| 982 tty->print("B%d ", pred->block_id()); | |
| 983 } | |
| 984 } | |
| 985 if (cur->number_of_sux() > 0) { | |
| 986 tty->print(" sux: "); | |
| 987 for (int j = 0; j < cur->number_of_sux(); j++) { | |
| 988 BlockBegin* sux = cur->sux_at(j); | |
| 989 tty->print("B%d ", sux->block_id()); | |
| 990 } | |
| 991 } | |
| 992 if (cur->number_of_exception_handlers() > 0) { | |
| 993 tty->print(" ex: "); | |
| 994 for (int j = 0; j < cur->number_of_exception_handlers(); j++) { | |
| 995 BlockBegin* ex = cur->exception_handler_at(j); | |
| 996 tty->print("B%d ", ex->block_id()); | |
| 997 } | |
| 998 } | |
| 999 tty->cr(); | |
| 1000 } | |
| 1001 } | |
| 1002 } | |
| 1003 #endif | |
| 1004 | |
| 1005 #ifdef ASSERT | |
| 1006 void ComputeLinearScanOrder::verify() { | |
| 1007 assert(_linear_scan_order->length() == _num_blocks, "wrong number of blocks in list"); | |
| 1008 | |
| 1009 if (StressLinearScan) { | |
| 1010 // blocks are scrambled when StressLinearScan is used | |
| 1011 return; | |
| 1012 } | |
| 1013 | |
| 1014 // check that all successors of a block have a higher linear-scan-number | |
| 1015 // and that all predecessors of a block have a lower linear-scan-number | |
| 1016 // (only backward branches of loops are ignored) | |
| 1017 int i; | |
| 1018 for (i = 0; i < _linear_scan_order->length(); i++) { | |
| 1019 BlockBegin* cur = _linear_scan_order->at(i); | |
| 1020 | |
| 1021 assert(cur->linear_scan_number() == i, "incorrect linear_scan_number"); | |
| 1022 assert(cur->linear_scan_number() >= 0 && cur->linear_scan_number() == _linear_scan_order->index_of(cur), "incorrect linear_scan_number"); | |
| 1023 | |
| 1024 int j; | |
| 1025 for (j = cur->number_of_sux() - 1; j >= 0; j--) { | |
| 1026 BlockBegin* sux = cur->sux_at(j); | |
| 1027 | |
| 1028 assert(sux->linear_scan_number() >= 0 && sux->linear_scan_number() == _linear_scan_order->index_of(sux), "incorrect linear_scan_number"); | |
| 1029 if (!cur->is_set(BlockBegin::linear_scan_loop_end_flag)) { | |
| 1030 assert(cur->linear_scan_number() < sux->linear_scan_number(), "invalid order"); | |
| 1031 } | |
| 1032 if (cur->loop_depth() == sux->loop_depth()) { | |
| 1033 assert(cur->loop_index() == sux->loop_index() || sux->is_set(BlockBegin::linear_scan_loop_header_flag), "successing blocks with same loop depth must have same loop index"); | |
| 1034 } | |
| 1035 } | |
| 1036 | |
| 1037 for (j = cur->number_of_preds() - 1; j >= 0; j--) { | |
| 1038 BlockBegin* pred = cur->pred_at(j); | |
| 1039 | |
| 1040 assert(pred->linear_scan_number() >= 0 && pred->linear_scan_number() == _linear_scan_order->index_of(pred), "incorrect linear_scan_number"); | |
| 1041 if (!cur->is_set(BlockBegin::linear_scan_loop_header_flag)) { | |
| 1042 assert(cur->linear_scan_number() > pred->linear_scan_number(), "invalid order"); | |
| 1043 } | |
| 1044 if (cur->loop_depth() == pred->loop_depth()) { | |
| 1045 assert(cur->loop_index() == pred->loop_index() || cur->is_set(BlockBegin::linear_scan_loop_header_flag), "successing blocks with same loop depth must have same loop index"); | |
| 1046 } | |
| 1047 | |
| 1048 assert(cur->dominator()->linear_scan_number() <= cur->pred_at(j)->linear_scan_number(), "dominator must be before predecessors"); | |
| 1049 } | |
| 1050 | |
| 1051 // check dominator | |
| 1052 if (i == 0) { | |
| 1053 assert(cur->dominator() == NULL, "first block has no dominator"); | |
| 1054 } else { | |
| 1055 assert(cur->dominator() != NULL, "all but first block must have dominator"); | |
| 1056 } | |
| 1057 assert(cur->number_of_preds() != 1 || cur->dominator() == cur->pred_at(0), "Single predecessor must also be dominator"); | |
| 1058 } | |
| 1059 | |
| 1060 // check that all loops are continuous | |
| 1061 for (int loop_idx = 0; loop_idx < _num_loops; loop_idx++) { | |
| 1062 int block_idx = 0; | |
| 1063 assert(!is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx)), "the first block must not be present in any loop"); | |
| 1064 | |
| 1065 // skip blocks before the loop | |
| 1066 while (block_idx < _num_blocks && !is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx))) { | |
| 1067 block_idx++; | |
| 1068 } | |
| 1069 // skip blocks of loop | |
| 1070 while (block_idx < _num_blocks && is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx))) { | |
| 1071 block_idx++; | |
| 1072 } | |
| 1073 // after the first non-loop block, there must not be another loop-block | |
| 1074 while (block_idx < _num_blocks) { | |
| 1075 assert(!is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx)), "loop not continuous in linear-scan order"); | |
| 1076 block_idx++; | |
| 1077 } | |
| 1078 } | |
| 1079 } | |
| 1080 #endif | |
| 1081 | |
| 1082 | |
| 1083 void IR::compute_code() { | |
| 1084 assert(is_valid(), "IR must be valid"); | |
| 1085 | |
| 1783 | 1086 ComputeLinearScanOrder compute_order(compilation(), start()); |
| 0 | 1087 _num_loops = compute_order.num_loops(); |
| 1088 _code = compute_order.linear_scan_order(); | |
| 1089 } | |
| 1090 | |
| 1091 | |
| 1092 void IR::compute_use_counts() { | |
| 1093 // make sure all values coming out of this block get evaluated. | |
| 1094 int num_blocks = _code->length(); | |
| 1095 for (int i = 0; i < num_blocks; i++) { | |
| 1096 _code->at(i)->end()->state()->pin_stack_for_linear_scan(); | |
| 1097 } | |
| 1098 | |
| 1099 // compute use counts | |
| 1100 UseCountComputer::compute(_code); | |
| 1101 } | |
| 1102 | |
| 1103 | |
| 1104 void IR::iterate_preorder(BlockClosure* closure) { | |
| 1105 assert(is_valid(), "IR must be valid"); | |
| 1106 start()->iterate_preorder(closure); | |
| 1107 } | |
| 1108 | |
| 1109 | |
| 1110 void IR::iterate_postorder(BlockClosure* closure) { | |
| 1111 assert(is_valid(), "IR must be valid"); | |
| 1112 start()->iterate_postorder(closure); | |
| 1113 } | |
| 1114 | |
| 1115 void IR::iterate_linear_scan_order(BlockClosure* closure) { | |
| 1116 linear_scan_order()->iterate_forward(closure); | |
| 1117 } | |
| 1118 | |
| 1119 | |
| 1120 #ifndef PRODUCT | |
| 1121 class BlockPrinter: public BlockClosure { | |
| 1122 private: | |
| 1123 InstructionPrinter* _ip; | |
| 1124 bool _cfg_only; | |
| 1125 bool _live_only; | |
| 1126 | |
| 1127 public: | |
| 1128 BlockPrinter(InstructionPrinter* ip, bool cfg_only, bool live_only = false) { | |
| 1129 _ip = ip; | |
| 1130 _cfg_only = cfg_only; | |
| 1131 _live_only = live_only; | |
| 1132 } | |
| 1133 | |
| 1134 virtual void block_do(BlockBegin* block) { | |
| 1135 if (_cfg_only) { | |
| 1136 _ip->print_instr(block); tty->cr(); | |
| 1137 } else { | |
| 1138 block->print_block(*_ip, _live_only); | |
| 1139 } | |
| 1140 } | |
| 1141 }; | |
| 1142 | |
| 1143 | |
| 1144 void IR::print(BlockBegin* start, bool cfg_only, bool live_only) { | |
| 1145 ttyLocker ttyl; | |
| 1146 InstructionPrinter ip(!cfg_only); | |
| 1147 BlockPrinter bp(&ip, cfg_only, live_only); | |
| 1148 start->iterate_preorder(&bp); | |
| 1149 tty->cr(); | |
| 1150 } | |
| 1151 | |
| 1152 void IR::print(bool cfg_only, bool live_only) { | |
| 1153 if (is_valid()) { | |
| 1154 print(start(), cfg_only, live_only); | |
| 1155 } else { | |
| 1156 tty->print_cr("invalid IR"); | |
| 1157 } | |
| 1158 } | |
| 1159 | |
| 1160 | |
| 1161 define_array(BlockListArray, BlockList*) | |
| 1162 define_stack(BlockListList, BlockListArray) | |
| 1163 | |
| 1164 class PredecessorValidator : public BlockClosure { | |
| 1165 private: | |
| 1166 BlockListList* _predecessors; | |
| 1167 BlockList* _blocks; | |
| 1168 | |
| 1169 static int cmp(BlockBegin** a, BlockBegin** b) { | |
| 1170 return (*a)->block_id() - (*b)->block_id(); | |
| 1171 } | |
| 1172 | |
| 1173 public: | |
| 1174 PredecessorValidator(IR* hir) { | |
| 1175 ResourceMark rm; | |
| 1176 _predecessors = new BlockListList(BlockBegin::number_of_blocks(), NULL); | |
| 1177 _blocks = new BlockList(); | |
| 1178 | |
| 1179 int i; | |
| 1180 hir->start()->iterate_preorder(this); | |
| 1181 if (hir->code() != NULL) { | |
| 1182 assert(hir->code()->length() == _blocks->length(), "must match"); | |
| 1183 for (i = 0; i < _blocks->length(); i++) { | |
| 1184 assert(hir->code()->contains(_blocks->at(i)), "should be in both lists"); | |
| 1185 } | |
| 1186 } | |
| 1187 | |
| 1188 for (i = 0; i < _blocks->length(); i++) { | |
| 1189 BlockBegin* block = _blocks->at(i); | |
| 1190 BlockList* preds = _predecessors->at(block->block_id()); | |
| 1191 if (preds == NULL) { | |
| 1192 assert(block->number_of_preds() == 0, "should be the same"); | |
| 1193 continue; | |
| 1194 } | |
| 1195 | |
| 1196 // clone the pred list so we can mutate it | |
| 1197 BlockList* pred_copy = new BlockList(); | |
| 1198 int j; | |
| 1199 for (j = 0; j < block->number_of_preds(); j++) { | |
| 1200 pred_copy->append(block->pred_at(j)); | |
| 1201 } | |
| 1202 // sort them in the same order | |
| 1203 preds->sort(cmp); | |
| 1204 pred_copy->sort(cmp); | |
| 1205 int length = MIN2(preds->length(), block->number_of_preds()); | |
| 1206 for (j = 0; j < block->number_of_preds(); j++) { | |
| 1207 assert(preds->at(j) == pred_copy->at(j), "must match"); | |
| 1208 } | |
| 1209 | |
| 1210 assert(preds->length() == block->number_of_preds(), "should be the same"); | |
| 1211 } | |
| 1212 } | |
| 1213 | |
| 1214 virtual void block_do(BlockBegin* block) { | |
| 1215 _blocks->append(block); | |
| 1216 BlockEnd* be = block->end(); | |
| 1217 int n = be->number_of_sux(); | |
| 1218 int i; | |
| 1219 for (i = 0; i < n; i++) { | |
| 1220 BlockBegin* sux = be->sux_at(i); | |
| 1221 assert(!sux->is_set(BlockBegin::exception_entry_flag), "must not be xhandler"); | |
| 1222 | |
| 1223 BlockList* preds = _predecessors->at_grow(sux->block_id(), NULL); | |
| 1224 if (preds == NULL) { | |
| 1225 preds = new BlockList(); | |
| 1226 _predecessors->at_put(sux->block_id(), preds); | |
| 1227 } | |
| 1228 preds->append(block); | |
| 1229 } | |
| 1230 | |
| 1231 n = block->number_of_exception_handlers(); | |
| 1232 for (i = 0; i < n; i++) { | |
| 1233 BlockBegin* sux = block->exception_handler_at(i); | |
| 1234 assert(sux->is_set(BlockBegin::exception_entry_flag), "must be xhandler"); | |
| 1235 | |
| 1236 BlockList* preds = _predecessors->at_grow(sux->block_id(), NULL); | |
| 1237 if (preds == NULL) { | |
| 1238 preds = new BlockList(); | |
| 1239 _predecessors->at_put(sux->block_id(), preds); | |
| 1240 } | |
| 1241 preds->append(block); | |
| 1242 } | |
| 1243 } | |
| 1244 }; | |
| 1245 | |
| 1246 void IR::verify() { | |
| 1247 #ifdef ASSERT | |
| 1248 PredecessorValidator pv(this); | |
| 1249 #endif | |
| 1250 } | |
| 1251 | |
| 1252 #endif // PRODUCT | |
| 1253 | |
| 1584 | 1254 void SubstitutionResolver::visit(Value* v) { |
| 0 | 1255 Value v0 = *v; |
| 1256 if (v0) { | |
| 1257 Value vs = v0->subst(); | |
| 1258 if (vs != v0) { | |
| 1259 *v = v0->subst(); | |
| 1260 } | |
| 1261 } | |
| 1262 } | |
| 1263 | |
| 1264 #ifdef ASSERT | |
| 1584 | 1265 class SubstitutionChecker: public ValueVisitor { |
| 1266 void visit(Value* v) { | |
| 1267 Value v0 = *v; | |
| 1268 if (v0) { | |
| 1269 Value vs = v0->subst(); | |
| 1270 assert(vs == v0, "missed substitution"); | |
| 1271 } | |
| 0 | 1272 } |
| 1584 | 1273 }; |
| 0 | 1274 #endif |
| 1275 | |
| 1276 | |
| 1277 void SubstitutionResolver::block_do(BlockBegin* block) { | |
| 1278 Instruction* last = NULL; | |
| 1279 for (Instruction* n = block; n != NULL;) { | |
| 1584 | 1280 n->values_do(this); |
| 0 | 1281 // need to remove this instruction from the instruction stream |
| 1282 if (n->subst() != n) { | |
| 1283 assert(last != NULL, "must have last"); | |
| 1819 | 1284 last->set_next(n->next()); |
| 0 | 1285 } else { |
| 1286 last = n; | |
| 1287 } | |
| 1288 n = last->next(); | |
| 1289 } | |
| 1290 | |
| 1291 #ifdef ASSERT | |
| 1584 | 1292 SubstitutionChecker check_substitute; |
| 1293 if (block->state()) block->state()->values_do(&check_substitute); | |
| 1294 block->block_values_do(&check_substitute); | |
| 1295 if (block->end() && block->end()->state()) block->end()->state()->values_do(&check_substitute); | |
| 0 | 1296 #endif |
| 1297 } |