1 /* 2 * Copyright (c) 1998, 2010, 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 "asm/assembler.inline.hpp" 27 #include "code/debugInfo.hpp" 28 #include "code/debugInfoRec.hpp" 29 #include "compiler/compileBroker.hpp" 30 #include "compiler/oopMap.hpp" 31 #include "memory/allocation.inline.hpp" 32 #include "opto/callnode.hpp" 33 #include "opto/cfgnode.hpp" 34 #include "opto/locknode.hpp" 35 #include "opto/machnode.hpp" 36 #include "opto/output.hpp" 37 #include "opto/regalloc.hpp" 38 #include "opto/runtime.hpp" 39 #include "opto/subnode.hpp" 40 #include "opto/type.hpp" 41 #include "runtime/handles.inline.hpp" 42 #include "utilities/xmlstream.hpp" 43 44 extern uint size_java_to_interp(); 45 extern uint reloc_java_to_interp(); 46 extern uint size_exception_handler(); 47 extern uint size_deopt_handler(); 48 49 #ifndef PRODUCT 50 #define DEBUG_ARG(x) , x 51 #else 52 #define DEBUG_ARG(x) 53 #endif 54 55 extern int emit_exception_handler(CodeBuffer &cbuf); 56 extern int emit_deopt_handler(CodeBuffer &cbuf); 57 58 //------------------------------Output----------------------------------------- 59 // Convert Nodes to instruction bits and pass off to the VM 60 void Compile::Output() { 61 // RootNode goes 62 assert( _cfg->_broot->_nodes.size() == 0, "" ); 63 64 // The number of new nodes (mostly MachNop) is proportional to 65 // the number of java calls and inner loops which are aligned. 66 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 + 67 C->inner_loops()*(OptoLoopAlignment-1)), 68 "out of nodes before code generation" ) ) { 69 return; 70 } 71 // Make sure I can find the Start Node 72 Block_Array& bbs = _cfg->_bbs; 73 Block *entry = _cfg->_blocks[1]; 74 Block *broot = _cfg->_broot; 75 76 const StartNode *start = entry->_nodes[0]->as_Start(); 77 78 // Replace StartNode with prolog 79 MachPrologNode *prolog = new (this) MachPrologNode(); 80 entry->_nodes.map( 0, prolog ); 81 bbs.map( prolog->_idx, entry ); 82 bbs.map( start->_idx, NULL ); // start is no longer in any block 83 84 // Virtual methods need an unverified entry point 85 86 if( is_osr_compilation() ) { 87 if( PoisonOSREntry ) { 88 // TODO: Should use a ShouldNotReachHereNode... 89 _cfg->insert( broot, 0, new (this) MachBreakpointNode() ); 90 } 91 } else { 92 if( _method && !_method->flags().is_static() ) { 93 // Insert unvalidated entry point 94 _cfg->insert( broot, 0, new (this) MachUEPNode() ); 95 } 96 97 } 98 99 100 // Break before main entry point 101 if( (_method && _method->break_at_execute()) 102 #ifndef PRODUCT 103 ||(OptoBreakpoint && is_method_compilation()) 104 ||(OptoBreakpointOSR && is_osr_compilation()) 105 ||(OptoBreakpointC2R && !_method) 106 #endif 107 ) { 108 // checking for _method means that OptoBreakpoint does not apply to 109 // runtime stubs or frame converters 110 _cfg->insert( entry, 1, new (this) MachBreakpointNode() ); 111 } 112 113 // Insert epilogs before every return 114 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 115 Block *b = _cfg->_blocks[i]; 116 if( !b->is_connector() && b->non_connector_successor(0) == _cfg->_broot ) { // Found a program exit point? 117 Node *m = b->end(); 118 if( m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt ) { 119 MachEpilogNode *epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); 120 b->add_inst( epilog ); 121 bbs.map(epilog->_idx, b); 122 //_regalloc->set_bad(epilog->_idx); // Already initialized this way. 123 } 124 } 125 } 126 127 # ifdef ENABLE_ZAP_DEAD_LOCALS 128 if ( ZapDeadCompiledLocals ) Insert_zap_nodes(); 129 # endif 130 131 ScheduleAndBundle(); 132 133 #ifndef PRODUCT 134 if (trace_opto_output()) { 135 tty->print("\n---- After ScheduleAndBundle ----\n"); 136 for (uint i = 0; i < _cfg->_num_blocks; i++) { 137 tty->print("\nBB#%03d:\n", i); 138 Block *bb = _cfg->_blocks[i]; 139 for (uint j = 0; j < bb->_nodes.size(); j++) { 140 Node *n = bb->_nodes[j]; 141 OptoReg::Name reg = _regalloc->get_reg_first(n); 142 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : ""); 143 n->dump(); 144 } 145 } 146 } 147 #endif 148 149 if (failing()) return; 150 151 BuildOopMaps(); 152 153 if (failing()) return; 154 155 Fill_buffer(); 156 } 157 158 bool Compile::need_stack_bang(int frame_size_in_bytes) const { 159 // Determine if we need to generate a stack overflow check. 160 // Do it if the method is not a stub function and 161 // has java calls or has frame size > vm_page_size/8. 162 return (stub_function() == NULL && 163 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3)); 164 } 165 166 bool Compile::need_register_stack_bang() const { 167 // Determine if we need to generate a register stack overflow check. 168 // This is only used on architectures which have split register 169 // and memory stacks (ie. IA64). 170 // Bang if the method is not a stub function and has java calls 171 return (stub_function() == NULL && has_java_calls()); 172 } 173 174 # ifdef ENABLE_ZAP_DEAD_LOCALS 175 176 177 // In order to catch compiler oop-map bugs, we have implemented 178 // a debugging mode called ZapDeadCompilerLocals. 179 // This mode causes the compiler to insert a call to a runtime routine, 180 // "zap_dead_locals", right before each place in compiled code 181 // that could potentially be a gc-point (i.e., a safepoint or oop map point). 182 // The runtime routine checks that locations mapped as oops are really 183 // oops, that locations mapped as values do not look like oops, 184 // and that locations mapped as dead are not used later 185 // (by zapping them to an invalid address). 186 187 int Compile::_CompiledZap_count = 0; 188 189 void Compile::Insert_zap_nodes() { 190 bool skip = false; 191 192 193 // Dink with static counts because code code without the extra 194 // runtime calls is MUCH faster for debugging purposes 195 196 if ( CompileZapFirst == 0 ) ; // nothing special 197 else if ( CompileZapFirst > CompiledZap_count() ) skip = true; 198 else if ( CompileZapFirst == CompiledZap_count() ) 199 warning("starting zap compilation after skipping"); 200 201 if ( CompileZapLast == -1 ) ; // nothing special 202 else if ( CompileZapLast < CompiledZap_count() ) skip = true; 203 else if ( CompileZapLast == CompiledZap_count() ) 204 warning("about to compile last zap"); 205 206 ++_CompiledZap_count; // counts skipped zaps, too 207 208 if ( skip ) return; 209 210 211 if ( _method == NULL ) 212 return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care 213 214 // Insert call to zap runtime stub before every node with an oop map 215 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 216 Block *b = _cfg->_blocks[i]; 217 for ( uint j = 0; j < b->_nodes.size(); ++j ) { 218 Node *n = b->_nodes[j]; 219 220 // Determining if we should insert a zap-a-lot node in output. 221 // We do that for all nodes that has oopmap info, except for calls 222 // to allocation. Calls to allocation passes in the old top-of-eden pointer 223 // and expect the C code to reset it. Hence, there can be no safepoints between 224 // the inlined-allocation and the call to new_Java, etc. 225 // We also cannot zap monitor calls, as they must hold the microlock 226 // during the call to Zap, which also wants to grab the microlock. 227 bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL); 228 if ( insert ) { // it is MachSafePoint 229 if ( !n->is_MachCall() ) { 230 insert = false; 231 } else if ( n->is_MachCall() ) { 232 MachCallNode* call = n->as_MachCall(); 233 if (call->entry_point() == OptoRuntime::new_instance_Java() || 234 call->entry_point() == OptoRuntime::new_array_Java() || 235 call->entry_point() == OptoRuntime::multianewarray2_Java() || 236 call->entry_point() == OptoRuntime::multianewarray3_Java() || 237 call->entry_point() == OptoRuntime::multianewarray4_Java() || 238 call->entry_point() == OptoRuntime::multianewarray5_Java() || 239 call->entry_point() == OptoRuntime::slow_arraycopy_Java() || 240 call->entry_point() == OptoRuntime::complete_monitor_locking_Java() 241 ) { 242 insert = false; 243 } 244 } 245 if (insert) { 246 Node *zap = call_zap_node(n->as_MachSafePoint(), i); 247 b->_nodes.insert( j, zap ); 248 _cfg->_bbs.map( zap->_idx, b ); 249 ++j; 250 } 251 } 252 } 253 } 254 } 255 256 257 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) { 258 const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type(); 259 CallStaticJavaNode* ideal_node = 260 new (this, tf->domain()->cnt()) CallStaticJavaNode( tf, 261 OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()), 262 "call zap dead locals stub", 0, TypePtr::BOTTOM); 263 // We need to copy the OopMap from the site we're zapping at. 264 // We have to make a copy, because the zap site might not be 265 // a call site, and zap_dead is a call site. 266 OopMap* clone = node_to_check->oop_map()->deep_copy(); 267 268 // Add the cloned OopMap to the zap node 269 ideal_node->set_oop_map(clone); 270 return _matcher->match_sfpt(ideal_node); 271 } 272 273 //------------------------------is_node_getting_a_safepoint-------------------- 274 bool Compile::is_node_getting_a_safepoint( Node* n) { 275 // This code duplicates the logic prior to the call of add_safepoint 276 // below in this file. 277 if( n->is_MachSafePoint() ) return true; 278 return false; 279 } 280 281 # endif // ENABLE_ZAP_DEAD_LOCALS 282 283 //------------------------------compute_loop_first_inst_sizes------------------ 284 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top 285 // of a loop. When aligning a loop we need to provide enough instructions 286 // in cpu's fetch buffer to feed decoders. The loop alignment could be 287 // avoided if we have enough instructions in fetch buffer at the head of a loop. 288 // By default, the size is set to 999999 by Block's constructor so that 289 // a loop will be aligned if the size is not reset here. 290 // 291 // Note: Mach instructions could contain several HW instructions 292 // so the size is estimated only. 293 // 294 void Compile::compute_loop_first_inst_sizes() { 295 // The next condition is used to gate the loop alignment optimization. 296 // Don't aligned a loop if there are enough instructions at the head of a loop 297 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad 298 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is 299 // equal to 11 bytes which is the largest address NOP instruction. 300 if( MaxLoopPad < OptoLoopAlignment-1 ) { 301 uint last_block = _cfg->_num_blocks-1; 302 for( uint i=1; i <= last_block; i++ ) { 303 Block *b = _cfg->_blocks[i]; 304 // Check the first loop's block which requires an alignment. 305 if( b->loop_alignment() > (uint)relocInfo::addr_unit() ) { 306 uint sum_size = 0; 307 uint inst_cnt = NumberOfLoopInstrToAlign; 308 inst_cnt = b->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 309 310 // Check subsequent fallthrough blocks if the loop's first 311 // block(s) does not have enough instructions. 312 Block *nb = b; 313 while( inst_cnt > 0 && 314 i < last_block && 315 !_cfg->_blocks[i+1]->has_loop_alignment() && 316 !nb->has_successor(b) ) { 317 i++; 318 nb = _cfg->_blocks[i]; 319 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 320 } // while( inst_cnt > 0 && i < last_block ) 321 322 b->set_first_inst_size(sum_size); 323 } // f( b->head()->is_Loop() ) 324 } // for( i <= last_block ) 325 } // if( MaxLoopPad < OptoLoopAlignment-1 ) 326 } 327 328 //----------------------Shorten_branches--------------------------------------- 329 // The architecture description provides short branch variants for some long 330 // branch instructions. Replace eligible long branches with short branches. 331 void Compile::Shorten_branches(Label *labels, int& code_size, int& reloc_size, int& stub_size) { 332 333 // fill in the nop array for bundling computations 334 MachNode *_nop_list[Bundle::_nop_count]; 335 Bundle::initialize_nops(_nop_list, this); 336 337 // ------------------ 338 // Compute size of each block, method size, and relocation information size 339 uint *jmp_end = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks); 340 uint *blk_starts = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks+1); 341 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks); ) 342 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks); ) 343 blk_starts[0] = 0; 344 345 // Initialize the sizes to 0 346 code_size = 0; // Size in bytes of generated code 347 stub_size = 0; // Size in bytes of all stub entries 348 // Size in bytes of all relocation entries, including those in local stubs. 349 // Start with 2-bytes of reloc info for the unvalidated entry point 350 reloc_size = 1; // Number of relocation entries 351 352 // Make three passes. The first computes pessimistic blk_starts, 353 // relative jmp_end and reloc_size information. The second performs 354 // short branch substitution using the pessimistic sizing. The 355 // third inserts nops where needed. 356 357 Node *nj; // tmp 358 359 // Step one, perform a pessimistic sizing pass. 360 uint i; 361 uint min_offset_from_last_call = 1; // init to a positive value 362 uint nop_size = (new (this) MachNopNode())->size(_regalloc); 363 for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks 364 Block *b = _cfg->_blocks[i]; 365 366 // Sum all instruction sizes to compute block size 367 uint last_inst = b->_nodes.size(); 368 uint blk_size = 0; 369 for( uint j = 0; j<last_inst; j++ ) { 370 nj = b->_nodes[j]; 371 uint inst_size = nj->size(_regalloc); 372 blk_size += inst_size; 373 // Handle machine instruction nodes 374 if( nj->is_Mach() ) { 375 MachNode *mach = nj->as_Mach(); 376 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 377 reloc_size += mach->reloc(); 378 if( mach->is_MachCall() ) { 379 MachCallNode *mcall = mach->as_MachCall(); 380 // This destination address is NOT PC-relative 381 382 mcall->method_set((intptr_t)mcall->entry_point()); 383 384 if( mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method ) { 385 stub_size += size_java_to_interp(); 386 reloc_size += reloc_java_to_interp(); 387 } 388 } else if (mach->is_MachSafePoint()) { 389 // If call/safepoint are adjacent, account for possible 390 // nop to disambiguate the two safepoints. 391 if (min_offset_from_last_call == 0) { 392 blk_size += nop_size; 393 } 394 } 395 } 396 min_offset_from_last_call += inst_size; 397 // Remember end of call offset 398 if (nj->is_MachCall() && nj->as_MachCall()->is_safepoint_node()) { 399 min_offset_from_last_call = 0; 400 } 401 } 402 403 // During short branch replacement, we store the relative (to blk_starts) 404 // end of jump in jmp_end, rather than the absolute end of jump. This 405 // is so that we do not need to recompute sizes of all nodes when we compute 406 // correct blk_starts in our next sizing pass. 407 jmp_end[i] = blk_size; 408 DEBUG_ONLY( jmp_target[i] = 0; ) 409 410 // When the next block starts a loop, we may insert pad NOP 411 // instructions. Since we cannot know our future alignment, 412 // assume the worst. 413 if( i<_cfg->_num_blocks-1 ) { 414 Block *nb = _cfg->_blocks[i+1]; 415 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 416 if( max_loop_pad > 0 ) { 417 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 418 blk_size += max_loop_pad; 419 } 420 } 421 422 // Save block size; update total method size 423 blk_starts[i+1] = blk_starts[i]+blk_size; 424 } 425 426 // Step two, replace eligible long jumps. 427 428 // Note: this will only get the long branches within short branch 429 // range. Another pass might detect more branches that became 430 // candidates because the shortening in the first pass exposed 431 // more opportunities. Unfortunately, this would require 432 // recomputing the starting and ending positions for the blocks 433 for( i=0; i<_cfg->_num_blocks; i++ ) { 434 Block *b = _cfg->_blocks[i]; 435 436 int j; 437 // Find the branch; ignore trailing NOPs. 438 for( j = b->_nodes.size()-1; j>=0; j-- ) { 439 nj = b->_nodes[j]; 440 if( !nj->is_Mach() || nj->as_Mach()->ideal_Opcode() != Op_Con ) 441 break; 442 } 443 444 if (j >= 0) { 445 if( nj->is_Mach() && nj->as_Mach()->may_be_short_branch() ) { 446 MachNode *mach = nj->as_Mach(); 447 // This requires the TRUE branch target be in succs[0] 448 uint bnum = b->non_connector_successor(0)->_pre_order; 449 uintptr_t target = blk_starts[bnum]; 450 if( mach->is_pc_relative() ) { 451 int offset = target-(blk_starts[i] + jmp_end[i]); 452 if (_matcher->is_short_branch_offset(mach->rule(), offset)) { 453 // We've got a winner. Replace this branch. 454 MachNode* replacement = mach->short_branch_version(this); 455 b->_nodes.map(j, replacement); 456 mach->subsume_by(replacement); 457 458 // Update the jmp_end size to save time in our 459 // next pass. 460 jmp_end[i] -= (mach->size(_regalloc) - replacement->size(_regalloc)); 461 DEBUG_ONLY( jmp_target[i] = bnum; ); 462 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 463 } 464 } else { 465 #ifndef PRODUCT 466 mach->dump(3); 467 #endif 468 Unimplemented(); 469 } 470 } 471 } 472 } 473 474 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 475 // of a loop. It is used to determine the padding for loop alignment. 476 compute_loop_first_inst_sizes(); 477 478 // Step 3, compute the offsets of all the labels 479 uint last_call_adr = max_uint; 480 for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks 481 // copy the offset of the beginning to the corresponding label 482 assert(labels[i].is_unused(), "cannot patch at this point"); 483 labels[i].bind_loc(blk_starts[i], CodeBuffer::SECT_INSTS); 484 485 // insert padding for any instructions that need it 486 Block *b = _cfg->_blocks[i]; 487 uint last_inst = b->_nodes.size(); 488 uint adr = blk_starts[i]; 489 for( uint j = 0; j<last_inst; j++ ) { 490 nj = b->_nodes[j]; 491 if( nj->is_Mach() ) { 492 int padding = nj->as_Mach()->compute_padding(adr); 493 // If call/safepoint are adjacent insert a nop (5010568) 494 if (padding == 0 && nj->is_MachSafePoint() && !nj->is_MachCall() && 495 adr == last_call_adr ) { 496 padding = nop_size; 497 } 498 if(padding > 0) { 499 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 500 int nops_cnt = padding / nop_size; 501 MachNode *nop = new (this) MachNopNode(nops_cnt); 502 b->_nodes.insert(j++, nop); 503 _cfg->_bbs.map( nop->_idx, b ); 504 adr += padding; 505 last_inst++; 506 } 507 } 508 adr += nj->size(_regalloc); 509 510 // Remember end of call offset 511 if (nj->is_MachCall() && nj->as_MachCall()->is_safepoint_node()) { 512 last_call_adr = adr; 513 } 514 } 515 516 if ( i != _cfg->_num_blocks-1) { 517 // Get the size of the block 518 uint blk_size = adr - blk_starts[i]; 519 520 // When the next block is the top of a loop, we may insert pad NOP 521 // instructions. 522 Block *nb = _cfg->_blocks[i+1]; 523 int current_offset = blk_starts[i] + blk_size; 524 current_offset += nb->alignment_padding(current_offset); 525 // Save block size; update total method size 526 blk_starts[i+1] = current_offset; 527 } 528 } 529 530 #ifdef ASSERT 531 for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks 532 if( jmp_target[i] != 0 ) { 533 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_end[i]); 534 if (!_matcher->is_short_branch_offset(jmp_rule[i], offset)) { 535 tty->print_cr("target (%d) - jmp_end(%d) = offset (%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_end[i], offset, i, jmp_target[i]); 536 } 537 assert(_matcher->is_short_branch_offset(jmp_rule[i], offset), "Displacement too large for short jmp"); 538 } 539 } 540 #endif 541 542 // ------------------ 543 // Compute size for code buffer 544 code_size = blk_starts[i-1] + jmp_end[i-1]; 545 546 // Relocation records 547 reloc_size += 1; // Relo entry for exception handler 548 549 // Adjust reloc_size to number of record of relocation info 550 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 551 // a relocation index. 552 // The CodeBuffer will expand the locs array if this estimate is too low. 553 reloc_size *= 10 / sizeof(relocInfo); 554 } 555 556 //------------------------------FillLocArray----------------------------------- 557 // Create a bit of debug info and append it to the array. The mapping is from 558 // Java local or expression stack to constant, register or stack-slot. For 559 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 560 // entry has been taken care of and caller should skip it). 561 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 562 // This should never have accepted Bad before 563 assert(OptoReg::is_valid(regnum), "location must be valid"); 564 return (OptoReg::is_reg(regnum)) 565 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 566 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 567 } 568 569 570 ObjectValue* 571 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 572 for (int i = 0; i < objs->length(); i++) { 573 assert(objs->at(i)->is_object(), "corrupt object cache"); 574 ObjectValue* sv = (ObjectValue*) objs->at(i); 575 if (sv->id() == id) { 576 return sv; 577 } 578 } 579 // Otherwise.. 580 return NULL; 581 } 582 583 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 584 ObjectValue* sv ) { 585 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition"); 586 objs->append(sv); 587 } 588 589 590 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 591 GrowableArray<ScopeValue*> *array, 592 GrowableArray<ScopeValue*> *objs ) { 593 assert( local, "use _top instead of null" ); 594 if (array->length() != idx) { 595 assert(array->length() == idx + 1, "Unexpected array count"); 596 // Old functionality: 597 // return 598 // New functionality: 599 // Assert if the local is not top. In product mode let the new node 600 // override the old entry. 601 assert(local == top(), "LocArray collision"); 602 if (local == top()) { 603 return; 604 } 605 array->pop(); 606 } 607 const Type *t = local->bottom_type(); 608 609 // Is it a safepoint scalar object node? 610 if (local->is_SafePointScalarObject()) { 611 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 612 613 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx); 614 if (sv == NULL) { 615 ciKlass* cik = t->is_oopptr()->klass(); 616 assert(cik->is_instance_klass() || 617 cik->is_array_klass(), "Not supported allocation."); 618 sv = new ObjectValue(spobj->_idx, 619 new ConstantOopWriteValue(cik->constant_encoding())); 620 Compile::set_sv_for_object_node(objs, sv); 621 622 uint first_ind = spobj->first_index(); 623 for (uint i = 0; i < spobj->n_fields(); i++) { 624 Node* fld_node = sfpt->in(first_ind+i); 625 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 626 } 627 } 628 array->append(sv); 629 return; 630 } 631 632 // Grab the register number for the local 633 OptoReg::Name regnum = _regalloc->get_reg_first(local); 634 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 635 // Record the double as two float registers. 636 // The register mask for such a value always specifies two adjacent 637 // float registers, with the lower register number even. 638 // Normally, the allocation of high and low words to these registers 639 // is irrelevant, because nearly all operations on register pairs 640 // (e.g., StoreD) treat them as a single unit. 641 // Here, we assume in addition that the words in these two registers 642 // stored "naturally" (by operations like StoreD and double stores 643 // within the interpreter) such that the lower-numbered register 644 // is written to the lower memory address. This may seem like 645 // a machine dependency, but it is not--it is a requirement on 646 // the author of the <arch>.ad file to ensure that, for every 647 // even/odd double-register pair to which a double may be allocated, 648 // the word in the even single-register is stored to the first 649 // memory word. (Note that register numbers are completely 650 // arbitrary, and are not tied to any machine-level encodings.) 651 #ifdef _LP64 652 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 653 array->append(new ConstantIntValue(0)); 654 array->append(new_loc_value( _regalloc, regnum, Location::dbl )); 655 } else if ( t->base() == Type::Long ) { 656 array->append(new ConstantIntValue(0)); 657 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 658 } else if ( t->base() == Type::RawPtr ) { 659 // jsr/ret return address which must be restored into a the full 660 // width 64-bit stack slot. 661 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 662 } 663 #else //_LP64 664 #ifdef SPARC 665 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) { 666 // For SPARC we have to swap high and low words for 667 // long values stored in a single-register (g0-g7). 668 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 669 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 670 } else 671 #endif //SPARC 672 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 673 // Repack the double/long as two jints. 674 // The convention the interpreter uses is that the second local 675 // holds the first raw word of the native double representation. 676 // This is actually reasonable, since locals and stack arrays 677 // grow downwards in all implementations. 678 // (If, on some machine, the interpreter's Java locals or stack 679 // were to grow upwards, the embedded doubles would be word-swapped.) 680 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 681 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 682 } 683 #endif //_LP64 684 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 685 OptoReg::is_reg(regnum) ) { 686 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double() 687 ? Location::float_in_dbl : Location::normal )); 688 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 689 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long 690 ? Location::int_in_long : Location::normal )); 691 } else if( t->base() == Type::NarrowOop ) { 692 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop )); 693 } else { 694 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal )); 695 } 696 return; 697 } 698 699 // No register. It must be constant data. 700 switch (t->base()) { 701 case Type::Half: // Second half of a double 702 ShouldNotReachHere(); // Caller should skip 2nd halves 703 break; 704 case Type::AnyPtr: 705 array->append(new ConstantOopWriteValue(NULL)); 706 break; 707 case Type::AryPtr: 708 case Type::InstPtr: 709 case Type::KlassPtr: // fall through 710 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 711 break; 712 case Type::NarrowOop: 713 if (t == TypeNarrowOop::NULL_PTR) { 714 array->append(new ConstantOopWriteValue(NULL)); 715 } else { 716 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 717 } 718 break; 719 case Type::Int: 720 array->append(new ConstantIntValue(t->is_int()->get_con())); 721 break; 722 case Type::RawPtr: 723 // A return address (T_ADDRESS). 724 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 725 #ifdef _LP64 726 // Must be restored to the full-width 64-bit stack slot. 727 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 728 #else 729 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 730 #endif 731 break; 732 case Type::FloatCon: { 733 float f = t->is_float_constant()->getf(); 734 array->append(new ConstantIntValue(jint_cast(f))); 735 break; 736 } 737 case Type::DoubleCon: { 738 jdouble d = t->is_double_constant()->getd(); 739 #ifdef _LP64 740 array->append(new ConstantIntValue(0)); 741 array->append(new ConstantDoubleValue(d)); 742 #else 743 // Repack the double as two jints. 744 // The convention the interpreter uses is that the second local 745 // holds the first raw word of the native double representation. 746 // This is actually reasonable, since locals and stack arrays 747 // grow downwards in all implementations. 748 // (If, on some machine, the interpreter's Java locals or stack 749 // were to grow upwards, the embedded doubles would be word-swapped.) 750 jint *dp = (jint*)&d; 751 array->append(new ConstantIntValue(dp[1])); 752 array->append(new ConstantIntValue(dp[0])); 753 #endif 754 break; 755 } 756 case Type::Long: { 757 jlong d = t->is_long()->get_con(); 758 #ifdef _LP64 759 array->append(new ConstantIntValue(0)); 760 array->append(new ConstantLongValue(d)); 761 #else 762 // Repack the long as two jints. 763 // The convention the interpreter uses is that the second local 764 // holds the first raw word of the native double representation. 765 // This is actually reasonable, since locals and stack arrays 766 // grow downwards in all implementations. 767 // (If, on some machine, the interpreter's Java locals or stack 768 // were to grow upwards, the embedded doubles would be word-swapped.) 769 jint *dp = (jint*)&d; 770 array->append(new ConstantIntValue(dp[1])); 771 array->append(new ConstantIntValue(dp[0])); 772 #endif 773 break; 774 } 775 case Type::Top: // Add an illegal value here 776 array->append(new LocationValue(Location())); 777 break; 778 default: 779 ShouldNotReachHere(); 780 break; 781 } 782 } 783 784 // Determine if this node starts a bundle 785 bool Compile::starts_bundle(const Node *n) const { 786 return (_node_bundling_limit > n->_idx && 787 _node_bundling_base[n->_idx].starts_bundle()); 788 } 789 790 //--------------------------Process_OopMap_Node-------------------------------- 791 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) { 792 793 // Handle special safepoint nodes for synchronization 794 MachSafePointNode *sfn = mach->as_MachSafePoint(); 795 MachCallNode *mcall; 796 797 #ifdef ENABLE_ZAP_DEAD_LOCALS 798 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative"); 799 #endif 800 801 int safepoint_pc_offset = current_offset; 802 bool is_method_handle_invoke = false; 803 bool return_oop = false; 804 805 // Add the safepoint in the DebugInfoRecorder 806 if( !mach->is_MachCall() ) { 807 mcall = NULL; 808 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 809 } else { 810 mcall = mach->as_MachCall(); 811 812 // Is the call a MethodHandle call? 813 if (mcall->is_MachCallJava()) { 814 if (mcall->as_MachCallJava()->_method_handle_invoke) { 815 assert(has_method_handle_invokes(), "must have been set during call generation"); 816 is_method_handle_invoke = true; 817 } 818 } 819 820 // Check if a call returns an object. 821 if (mcall->return_value_is_used() && 822 mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) { 823 return_oop = true; 824 } 825 safepoint_pc_offset += mcall->ret_addr_offset(); 826 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 827 } 828 829 // Loop over the JVMState list to add scope information 830 // Do not skip safepoints with a NULL method, they need monitor info 831 JVMState* youngest_jvms = sfn->jvms(); 832 int max_depth = youngest_jvms->depth(); 833 834 // Allocate the object pool for scalar-replaced objects -- the map from 835 // small-integer keys (which can be recorded in the local and ostack 836 // arrays) to descriptions of the object state. 837 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 838 839 // Visit scopes from oldest to youngest. 840 for (int depth = 1; depth <= max_depth; depth++) { 841 JVMState* jvms = youngest_jvms->of_depth(depth); 842 int idx; 843 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 844 // Safepoints that do not have method() set only provide oop-map and monitor info 845 // to support GC; these do not support deoptimization. 846 int num_locs = (method == NULL) ? 0 : jvms->loc_size(); 847 int num_exps = (method == NULL) ? 0 : jvms->stk_size(); 848 int num_mon = jvms->nof_monitors(); 849 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), 850 "JVMS local count must match that of the method"); 851 852 // Add Local and Expression Stack Information 853 854 // Insert locals into the locarray 855 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 856 for( idx = 0; idx < num_locs; idx++ ) { 857 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 858 } 859 860 // Insert expression stack entries into the exparray 861 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 862 for( idx = 0; idx < num_exps; idx++ ) { 863 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 864 } 865 866 // Add in mappings of the monitors 867 assert( !method || 868 !method->is_synchronized() || 869 method->is_native() || 870 num_mon > 0 || 871 !GenerateSynchronizationCode, 872 "monitors must always exist for synchronized methods"); 873 874 // Build the growable array of ScopeValues for exp stack 875 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 876 877 // Loop over monitors and insert into array 878 for(idx = 0; idx < num_mon; idx++) { 879 // Grab the node that defines this monitor 880 Node* box_node = sfn->monitor_box(jvms, idx); 881 Node* obj_node = sfn->monitor_obj(jvms, idx); 882 883 // Create ScopeValue for object 884 ScopeValue *scval = NULL; 885 886 if( obj_node->is_SafePointScalarObject() ) { 887 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 888 scval = Compile::sv_for_node_id(objs, spobj->_idx); 889 if (scval == NULL) { 890 const Type *t = obj_node->bottom_type(); 891 ciKlass* cik = t->is_oopptr()->klass(); 892 assert(cik->is_instance_klass() || 893 cik->is_array_klass(), "Not supported allocation."); 894 ObjectValue* sv = new ObjectValue(spobj->_idx, 895 new ConstantOopWriteValue(cik->constant_encoding())); 896 Compile::set_sv_for_object_node(objs, sv); 897 898 uint first_ind = spobj->first_index(); 899 for (uint i = 0; i < spobj->n_fields(); i++) { 900 Node* fld_node = sfn->in(first_ind+i); 901 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 902 } 903 scval = sv; 904 } 905 } else if( !obj_node->is_Con() ) { 906 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); 907 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 908 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop ); 909 } else { 910 scval = new_loc_value( _regalloc, obj_reg, Location::oop ); 911 } 912 } else { 913 const TypePtr *tp = obj_node->bottom_type()->make_ptr(); 914 scval = new ConstantOopWriteValue(tp->is_instptr()->const_oop()->constant_encoding()); 915 } 916 917 OptoReg::Name box_reg = BoxLockNode::stack_slot(box_node); 918 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)); 919 while( !box_node->is_BoxLock() ) box_node = box_node->in(1); 920 monarray->append(new MonitorValue(scval, basic_lock, box_node->as_BoxLock()->is_eliminated())); 921 } 922 923 // We dump the object pool first, since deoptimization reads it in first. 924 debug_info()->dump_object_pool(objs); 925 926 // Build first class objects to pass to scope 927 DebugToken *locvals = debug_info()->create_scope_values(locarray); 928 DebugToken *expvals = debug_info()->create_scope_values(exparray); 929 DebugToken *monvals = debug_info()->create_monitor_values(monarray); 930 931 // Make method available for all Safepoints 932 ciMethod* scope_method = method ? method : _method; 933 // Describe the scope here 934 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 935 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 936 // Now we can describe the scope. 937 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals); 938 } // End jvms loop 939 940 // Mark the end of the scope set. 941 debug_info()->end_safepoint(safepoint_pc_offset); 942 } 943 944 945 946 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 947 class NonSafepointEmitter { 948 Compile* C; 949 JVMState* _pending_jvms; 950 int _pending_offset; 951 952 void emit_non_safepoint(); 953 954 public: 955 NonSafepointEmitter(Compile* compile) { 956 this->C = compile; 957 _pending_jvms = NULL; 958 _pending_offset = 0; 959 } 960 961 void observe_instruction(Node* n, int pc_offset) { 962 if (!C->debug_info()->recording_non_safepoints()) return; 963 964 Node_Notes* nn = C->node_notes_at(n->_idx); 965 if (nn == NULL || nn->jvms() == NULL) return; 966 if (_pending_jvms != NULL && 967 _pending_jvms->same_calls_as(nn->jvms())) { 968 // Repeated JVMS? Stretch it up here. 969 _pending_offset = pc_offset; 970 } else { 971 if (_pending_jvms != NULL && 972 _pending_offset < pc_offset) { 973 emit_non_safepoint(); 974 } 975 _pending_jvms = NULL; 976 if (pc_offset > C->debug_info()->last_pc_offset()) { 977 // This is the only way _pending_jvms can become non-NULL: 978 _pending_jvms = nn->jvms(); 979 _pending_offset = pc_offset; 980 } 981 } 982 } 983 984 // Stay out of the way of real safepoints: 985 void observe_safepoint(JVMState* jvms, int pc_offset) { 986 if (_pending_jvms != NULL && 987 !_pending_jvms->same_calls_as(jvms) && 988 _pending_offset < pc_offset) { 989 emit_non_safepoint(); 990 } 991 _pending_jvms = NULL; 992 } 993 994 void flush_at_end() { 995 if (_pending_jvms != NULL) { 996 emit_non_safepoint(); 997 } 998 _pending_jvms = NULL; 999 } 1000 }; 1001 1002 void NonSafepointEmitter::emit_non_safepoint() { 1003 JVMState* youngest_jvms = _pending_jvms; 1004 int pc_offset = _pending_offset; 1005 1006 // Clear it now: 1007 _pending_jvms = NULL; 1008 1009 DebugInformationRecorder* debug_info = C->debug_info(); 1010 assert(debug_info->recording_non_safepoints(), "sanity"); 1011 1012 debug_info->add_non_safepoint(pc_offset); 1013 int max_depth = youngest_jvms->depth(); 1014 1015 // Visit scopes from oldest to youngest. 1016 for (int depth = 1; depth <= max_depth; depth++) { 1017 JVMState* jvms = youngest_jvms->of_depth(depth); 1018 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 1019 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 1020 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute()); 1021 } 1022 1023 // Mark the end of the scope set. 1024 debug_info->end_non_safepoint(pc_offset); 1025 } 1026 1027 1028 1029 // helper for Fill_buffer bailout logic 1030 static void turn_off_compiler(Compile* C) { 1031 if (CodeCache::unallocated_capacity() >= CodeCacheMinimumFreeSpace*10) { 1032 // Do not turn off compilation if a single giant method has 1033 // blown the code cache size. 1034 C->record_failure("excessive request to CodeCache"); 1035 } else { 1036 // Let CompilerBroker disable further compilations. 1037 C->record_failure("CodeCache is full"); 1038 } 1039 } 1040 1041 1042 //------------------------------Fill_buffer------------------------------------ 1043 void Compile::Fill_buffer() { 1044 1045 // Set the initially allocated size 1046 int code_req = initial_code_capacity; 1047 int locs_req = initial_locs_capacity; 1048 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity; 1049 int const_req = initial_const_capacity; 1050 bool labels_not_set = true; 1051 1052 int pad_req = NativeCall::instruction_size; 1053 // The extra spacing after the code is necessary on some platforms. 1054 // Sometimes we need to patch in a jump after the last instruction, 1055 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 1056 1057 uint i; 1058 // Compute the byte offset where we can store the deopt pc. 1059 if (fixed_slots() != 0) { 1060 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 1061 } 1062 1063 // Compute prolog code size 1064 _method_size = 0; 1065 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; 1066 #ifdef IA64 1067 if (save_argument_registers()) { 1068 // 4815101: this is a stub with implicit and unknown precision fp args. 1069 // The usual spill mechanism can only generate stfd's in this case, which 1070 // doesn't work if the fp reg to spill contains a single-precision denorm. 1071 // Instead, we hack around the normal spill mechanism using stfspill's and 1072 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate 1073 // space here for the fp arg regs (f8-f15) we're going to thusly spill. 1074 // 1075 // If we ever implement 16-byte 'registers' == stack slots, we can 1076 // get rid of this hack and have SpillCopy generate stfspill/ldffill 1077 // instead of stfd/stfs/ldfd/ldfs. 1078 _frame_slots += 8*(16/BytesPerInt); 1079 } 1080 #endif 1081 assert( _frame_slots >= 0 && _frame_slots < 1000000, "sanity check" ); 1082 1083 // Create an array of unused labels, one for each basic block 1084 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, _cfg->_num_blocks+1); 1085 1086 for( i=0; i <= _cfg->_num_blocks; i++ ) { 1087 blk_labels[i].init(); 1088 } 1089 1090 if (has_mach_constant_base_node()) { 1091 // Fill the constant table. 1092 // Note: This must happen before Shorten_branches. 1093 for (i = 0; i < _cfg->_num_blocks; i++) { 1094 Block* b = _cfg->_blocks[i]; 1095 1096 for (uint j = 0; j < b->_nodes.size(); j++) { 1097 Node* n = b->_nodes[j]; 1098 1099 // If the node is a MachConstantNode evaluate the constant 1100 // value section. 1101 if (n->is_MachConstant()) { 1102 MachConstantNode* machcon = n->as_MachConstant(); 1103 machcon->eval_constant(C); 1104 } 1105 } 1106 } 1107 1108 // Calculate the offsets of the constants and the size of the 1109 // constant table (including the padding to the next section). 1110 constant_table().calculate_offsets_and_size(); 1111 const_req = constant_table().size(); 1112 } 1113 1114 // Initialize the space for the BufferBlob used to find and verify 1115 // instruction size in MachNode::emit_size() 1116 init_scratch_buffer_blob(const_req); 1117 if (failing()) return; // Out of memory 1118 1119 // If this machine supports different size branch offsets, then pre-compute 1120 // the length of the blocks 1121 if( _matcher->is_short_branch_offset(-1, 0) ) { 1122 Shorten_branches(blk_labels, code_req, locs_req, stub_req); 1123 labels_not_set = false; 1124 } 1125 1126 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1127 int exception_handler_req = size_exception_handler(); 1128 int deopt_handler_req = size_deopt_handler(); 1129 exception_handler_req += MAX_stubs_size; // add marginal slop for handler 1130 deopt_handler_req += MAX_stubs_size; // add marginal slop for handler 1131 stub_req += MAX_stubs_size; // ensure per-stub margin 1132 code_req += MAX_inst_size; // ensure per-instruction margin 1133 1134 if (StressCodeBuffers) 1135 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1136 1137 int total_req = 1138 const_req + 1139 code_req + 1140 pad_req + 1141 stub_req + 1142 exception_handler_req + 1143 deopt_handler_req; // deopt handler 1144 1145 if (has_method_handle_invokes()) 1146 total_req += deopt_handler_req; // deopt MH handler 1147 1148 CodeBuffer* cb = code_buffer(); 1149 cb->initialize(total_req, locs_req); 1150 1151 // Have we run out of code space? 1152 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1153 turn_off_compiler(this); 1154 return; 1155 } 1156 // Configure the code buffer. 1157 cb->initialize_consts_size(const_req); 1158 cb->initialize_stubs_size(stub_req); 1159 cb->initialize_oop_recorder(env()->oop_recorder()); 1160 1161 // fill in the nop array for bundling computations 1162 MachNode *_nop_list[Bundle::_nop_count]; 1163 Bundle::initialize_nops(_nop_list, this); 1164 1165 // Create oopmap set. 1166 _oop_map_set = new OopMapSet(); 1167 1168 // !!!!! This preserves old handling of oopmaps for now 1169 debug_info()->set_oopmaps(_oop_map_set); 1170 1171 // Count and start of implicit null check instructions 1172 uint inct_cnt = 0; 1173 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, _cfg->_num_blocks+1); 1174 1175 // Count and start of calls 1176 uint *call_returns = NEW_RESOURCE_ARRAY(uint, _cfg->_num_blocks+1); 1177 1178 uint return_offset = 0; 1179 int nop_size = (new (this) MachNopNode())->size(_regalloc); 1180 1181 int previous_offset = 0; 1182 int current_offset = 0; 1183 int last_call_offset = -1; 1184 1185 // Create an array of unused labels, one for each basic block, if printing is enabled 1186 #ifndef PRODUCT 1187 int *node_offsets = NULL; 1188 uint node_offset_limit = unique(); 1189 1190 1191 if ( print_assembly() ) 1192 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1193 #endif 1194 1195 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily 1196 1197 // Emit the constant table. 1198 if (has_mach_constant_base_node()) { 1199 constant_table().emit(*cb); 1200 } 1201 1202 // ------------------ 1203 // Now fill in the code buffer 1204 Node *delay_slot = NULL; 1205 1206 for( i=0; i < _cfg->_num_blocks; i++ ) { 1207 Block *b = _cfg->_blocks[i]; 1208 1209 Node *head = b->head(); 1210 1211 // If this block needs to start aligned (i.e, can be reached other 1212 // than by falling-thru from the previous block), then force the 1213 // start of a new bundle. 1214 if( Pipeline::requires_bundling() && starts_bundle(head) ) 1215 cb->flush_bundle(true); 1216 1217 // Define the label at the beginning of the basic block 1218 if (labels_not_set) { 1219 MacroAssembler(cb).bind(blk_labels[b->_pre_order]); 1220 } else { 1221 assert(blk_labels[b->_pre_order].loc_pos() == cb->insts_size(), 1222 err_msg("label position does not match code offset: %d != %d", 1223 blk_labels[b->_pre_order].loc_pos(), cb->insts_size())); 1224 } 1225 1226 uint last_inst = b->_nodes.size(); 1227 1228 // Emit block normally, except for last instruction. 1229 // Emit means "dump code bits into code buffer". 1230 for( uint j = 0; j<last_inst; j++ ) { 1231 1232 // Get the node 1233 Node* n = b->_nodes[j]; 1234 1235 // See if delay slots are supported 1236 if (valid_bundle_info(n) && 1237 node_bundling(n)->used_in_unconditional_delay()) { 1238 assert(delay_slot == NULL, "no use of delay slot node"); 1239 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1240 1241 delay_slot = n; 1242 continue; 1243 } 1244 1245 // If this starts a new instruction group, then flush the current one 1246 // (but allow split bundles) 1247 if( Pipeline::requires_bundling() && starts_bundle(n) ) 1248 cb->flush_bundle(false); 1249 1250 // The following logic is duplicated in the code ifdeffed for 1251 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It 1252 // should be factored out. Or maybe dispersed to the nodes? 1253 1254 // Special handling for SafePoint/Call Nodes 1255 bool is_mcall = false; 1256 if( n->is_Mach() ) { 1257 MachNode *mach = n->as_Mach(); 1258 is_mcall = n->is_MachCall(); 1259 bool is_sfn = n->is_MachSafePoint(); 1260 1261 // If this requires all previous instructions be flushed, then do so 1262 if( is_sfn || is_mcall || mach->alignment_required() != 1) { 1263 cb->flush_bundle(true); 1264 current_offset = cb->insts_size(); 1265 } 1266 1267 // align the instruction if necessary 1268 int padding = mach->compute_padding(current_offset); 1269 // Make sure safepoint node for polling is distinct from a call's 1270 // return by adding a nop if needed. 1271 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset ) { 1272 padding = nop_size; 1273 } 1274 assert( labels_not_set || padding == 0, "instruction should already be aligned"); 1275 1276 if(padding > 0) { 1277 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1278 int nops_cnt = padding / nop_size; 1279 MachNode *nop = new (this) MachNopNode(nops_cnt); 1280 b->_nodes.insert(j++, nop); 1281 last_inst++; 1282 _cfg->_bbs.map( nop->_idx, b ); 1283 nop->emit(*cb, _regalloc); 1284 cb->flush_bundle(true); 1285 current_offset = cb->insts_size(); 1286 } 1287 1288 // Remember the start of the last call in a basic block 1289 if (is_mcall) { 1290 MachCallNode *mcall = mach->as_MachCall(); 1291 1292 // This destination address is NOT PC-relative 1293 mcall->method_set((intptr_t)mcall->entry_point()); 1294 1295 // Save the return address 1296 call_returns[b->_pre_order] = current_offset + mcall->ret_addr_offset(); 1297 1298 if (!mcall->is_safepoint_node()) { 1299 is_mcall = false; 1300 is_sfn = false; 1301 } 1302 } 1303 1304 // sfn will be valid whenever mcall is valid now because of inheritance 1305 if( is_sfn || is_mcall ) { 1306 1307 // Handle special safepoint nodes for synchronization 1308 if( !is_mcall ) { 1309 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1310 // !!!!! Stubs only need an oopmap right now, so bail out 1311 if( sfn->jvms()->method() == NULL) { 1312 // Write the oopmap directly to the code blob??!! 1313 # ifdef ENABLE_ZAP_DEAD_LOCALS 1314 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive"); 1315 # endif 1316 continue; 1317 } 1318 } // End synchronization 1319 1320 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1321 current_offset); 1322 Process_OopMap_Node(mach, current_offset); 1323 } // End if safepoint 1324 1325 // If this is a null check, then add the start of the previous instruction to the list 1326 else if( mach->is_MachNullCheck() ) { 1327 inct_starts[inct_cnt++] = previous_offset; 1328 } 1329 1330 // If this is a branch, then fill in the label with the target BB's label 1331 else if ( mach->is_Branch() ) { 1332 1333 if ( mach->ideal_Opcode() == Op_Jump ) { 1334 for (uint h = 0; h < b->_num_succs; h++ ) { 1335 Block* succs_block = b->_succs[h]; 1336 for (uint j = 1; j < succs_block->num_preds(); j++) { 1337 Node* jpn = succs_block->pred(j); 1338 if ( jpn->is_JumpProj() && jpn->in(0) == mach ) { 1339 uint block_num = succs_block->non_connector()->_pre_order; 1340 Label *blkLabel = &blk_labels[block_num]; 1341 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1342 } 1343 } 1344 } 1345 } else { 1346 // For Branchs 1347 // This requires the TRUE branch target be in succs[0] 1348 uint block_num = b->non_connector_successor(0)->_pre_order; 1349 mach->label_set( blk_labels[block_num], block_num ); 1350 } 1351 } 1352 1353 #ifdef ASSERT 1354 // Check that oop-store precedes the card-mark 1355 else if( mach->ideal_Opcode() == Op_StoreCM ) { 1356 uint storeCM_idx = j; 1357 Node *oop_store = mach->in(mach->_cnt); // First precedence edge 1358 assert( oop_store != NULL, "storeCM expects a precedence edge"); 1359 uint i4; 1360 for( i4 = 0; i4 < last_inst; ++i4 ) { 1361 if( b->_nodes[i4] == oop_store ) break; 1362 } 1363 // Note: This test can provide a false failure if other precedence 1364 // edges have been added to the storeCMNode. 1365 assert( i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1366 } 1367 #endif 1368 1369 else if( !n->is_Proj() ) { 1370 // Remember the beginning of the previous instruction, in case 1371 // it's followed by a flag-kill and a null-check. Happens on 1372 // Intel all the time, with add-to-memory kind of opcodes. 1373 previous_offset = current_offset; 1374 } 1375 } 1376 1377 // Verify that there is sufficient space remaining 1378 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1379 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1380 turn_off_compiler(this); 1381 return; 1382 } 1383 1384 // Save the offset for the listing 1385 #ifndef PRODUCT 1386 if( node_offsets && n->_idx < node_offset_limit ) 1387 node_offsets[n->_idx] = cb->insts_size(); 1388 #endif 1389 1390 // "Normal" instruction case 1391 n->emit(*cb, _regalloc); 1392 current_offset = cb->insts_size(); 1393 non_safepoints.observe_instruction(n, current_offset); 1394 1395 // mcall is last "call" that can be a safepoint 1396 // record it so we can see if a poll will directly follow it 1397 // in which case we'll need a pad to make the PcDesc sites unique 1398 // see 5010568. This can be slightly inaccurate but conservative 1399 // in the case that return address is not actually at current_offset. 1400 // This is a small price to pay. 1401 1402 if (is_mcall) { 1403 last_call_offset = current_offset; 1404 } 1405 1406 // See if this instruction has a delay slot 1407 if ( valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1408 assert(delay_slot != NULL, "expecting delay slot node"); 1409 1410 // Back up 1 instruction 1411 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size()); 1412 1413 // Save the offset for the listing 1414 #ifndef PRODUCT 1415 if( node_offsets && delay_slot->_idx < node_offset_limit ) 1416 node_offsets[delay_slot->_idx] = cb->insts_size(); 1417 #endif 1418 1419 // Support a SafePoint in the delay slot 1420 if( delay_slot->is_MachSafePoint() ) { 1421 MachNode *mach = delay_slot->as_Mach(); 1422 // !!!!! Stubs only need an oopmap right now, so bail out 1423 if( !mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL ) { 1424 // Write the oopmap directly to the code blob??!! 1425 # ifdef ENABLE_ZAP_DEAD_LOCALS 1426 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive"); 1427 # endif 1428 delay_slot = NULL; 1429 continue; 1430 } 1431 1432 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1433 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1434 adjusted_offset); 1435 // Generate an OopMap entry 1436 Process_OopMap_Node(mach, adjusted_offset); 1437 } 1438 1439 // Insert the delay slot instruction 1440 delay_slot->emit(*cb, _regalloc); 1441 1442 // Don't reuse it 1443 delay_slot = NULL; 1444 } 1445 1446 } // End for all instructions in block 1447 1448 // If the next block is the top of a loop, pad this block out to align 1449 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1450 if( i<_cfg->_num_blocks-1 ) { 1451 Block *nb = _cfg->_blocks[i+1]; 1452 uint padding = nb->alignment_padding(current_offset); 1453 if( padding > 0 ) { 1454 MachNode *nop = new (this) MachNopNode(padding / nop_size); 1455 b->_nodes.insert( b->_nodes.size(), nop ); 1456 _cfg->_bbs.map( nop->_idx, b ); 1457 nop->emit(*cb, _regalloc); 1458 current_offset = cb->insts_size(); 1459 } 1460 } 1461 1462 } // End of for all blocks 1463 1464 non_safepoints.flush_at_end(); 1465 1466 // Offset too large? 1467 if (failing()) return; 1468 1469 // Define a pseudo-label at the end of the code 1470 MacroAssembler(cb).bind( blk_labels[_cfg->_num_blocks] ); 1471 1472 // Compute the size of the first block 1473 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1474 1475 assert(cb->insts_size() < 500000, "method is unreasonably large"); 1476 1477 // ------------------ 1478 1479 #ifndef PRODUCT 1480 // Information on the size of the method, without the extraneous code 1481 Scheduling::increment_method_size(cb->insts_size()); 1482 #endif 1483 1484 // ------------------ 1485 // Fill in exception table entries. 1486 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1487 1488 // Only java methods have exception handlers and deopt handlers 1489 if (_method) { 1490 // Emit the exception handler code. 1491 _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb)); 1492 // Emit the deopt handler code. 1493 _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb)); 1494 1495 // Emit the MethodHandle deopt handler code (if required). 1496 if (has_method_handle_invokes()) { 1497 // We can use the same code as for the normal deopt handler, we 1498 // just need a different entry point address. 1499 _code_offsets.set_value(CodeOffsets::DeoptMH, emit_deopt_handler(*cb)); 1500 } 1501 } 1502 1503 // One last check for failed CodeBuffer::expand: 1504 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1505 turn_off_compiler(this); 1506 return; 1507 } 1508 1509 #ifndef PRODUCT 1510 // Dump the assembly code, including basic-block numbers 1511 if (print_assembly()) { 1512 ttyLocker ttyl; // keep the following output all in one block 1513 if (!VMThread::should_terminate()) { // test this under the tty lock 1514 // This output goes directly to the tty, not the compiler log. 1515 // To enable tools to match it up with the compilation activity, 1516 // be sure to tag this tty output with the compile ID. 1517 if (xtty != NULL) { 1518 xtty->head("opto_assembly compile_id='%d'%s", compile_id(), 1519 is_osr_compilation() ? " compile_kind='osr'" : 1520 ""); 1521 } 1522 if (method() != NULL) { 1523 method()->print_oop(); 1524 print_codes(); 1525 } 1526 dump_asm(node_offsets, node_offset_limit); 1527 if (xtty != NULL) { 1528 xtty->tail("opto_assembly"); 1529 } 1530 } 1531 } 1532 #endif 1533 1534 } 1535 1536 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1537 _inc_table.set_size(cnt); 1538 1539 uint inct_cnt = 0; 1540 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1541 Block *b = _cfg->_blocks[i]; 1542 Node *n = NULL; 1543 int j; 1544 1545 // Find the branch; ignore trailing NOPs. 1546 for( j = b->_nodes.size()-1; j>=0; j-- ) { 1547 n = b->_nodes[j]; 1548 if( !n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con ) 1549 break; 1550 } 1551 1552 // If we didn't find anything, continue 1553 if( j < 0 ) continue; 1554 1555 // Compute ExceptionHandlerTable subtable entry and add it 1556 // (skip empty blocks) 1557 if( n->is_Catch() ) { 1558 1559 // Get the offset of the return from the call 1560 uint call_return = call_returns[b->_pre_order]; 1561 #ifdef ASSERT 1562 assert( call_return > 0, "no call seen for this basic block" ); 1563 while( b->_nodes[--j]->Opcode() == Op_MachProj ) ; 1564 assert( b->_nodes[j]->is_Call(), "CatchProj must follow call" ); 1565 #endif 1566 // last instruction is a CatchNode, find it's CatchProjNodes 1567 int nof_succs = b->_num_succs; 1568 // allocate space 1569 GrowableArray<intptr_t> handler_bcis(nof_succs); 1570 GrowableArray<intptr_t> handler_pcos(nof_succs); 1571 // iterate through all successors 1572 for (int j = 0; j < nof_succs; j++) { 1573 Block* s = b->_succs[j]; 1574 bool found_p = false; 1575 for( uint k = 1; k < s->num_preds(); k++ ) { 1576 Node *pk = s->pred(k); 1577 if( pk->is_CatchProj() && pk->in(0) == n ) { 1578 const CatchProjNode* p = pk->as_CatchProj(); 1579 found_p = true; 1580 // add the corresponding handler bci & pco information 1581 if( p->_con != CatchProjNode::fall_through_index ) { 1582 // p leads to an exception handler (and is not fall through) 1583 assert(s == _cfg->_blocks[s->_pre_order],"bad numbering"); 1584 // no duplicates, please 1585 if( !handler_bcis.contains(p->handler_bci()) ) { 1586 uint block_num = s->non_connector()->_pre_order; 1587 handler_bcis.append(p->handler_bci()); 1588 handler_pcos.append(blk_labels[block_num].loc_pos()); 1589 } 1590 } 1591 } 1592 } 1593 assert(found_p, "no matching predecessor found"); 1594 // Note: Due to empty block removal, one block may have 1595 // several CatchProj inputs, from the same Catch. 1596 } 1597 1598 // Set the offset of the return from the call 1599 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); 1600 continue; 1601 } 1602 1603 // Handle implicit null exception table updates 1604 if( n->is_MachNullCheck() ) { 1605 uint block_num = b->non_connector_successor(0)->_pre_order; 1606 _inc_table.append( inct_starts[inct_cnt++], blk_labels[block_num].loc_pos() ); 1607 continue; 1608 } 1609 } // End of for all blocks fill in exception table entries 1610 } 1611 1612 // Static Variables 1613 #ifndef PRODUCT 1614 uint Scheduling::_total_nop_size = 0; 1615 uint Scheduling::_total_method_size = 0; 1616 uint Scheduling::_total_branches = 0; 1617 uint Scheduling::_total_unconditional_delays = 0; 1618 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1619 #endif 1620 1621 // Initializer for class Scheduling 1622 1623 Scheduling::Scheduling(Arena *arena, Compile &compile) 1624 : _arena(arena), 1625 _cfg(compile.cfg()), 1626 _bbs(compile.cfg()->_bbs), 1627 _regalloc(compile.regalloc()), 1628 _reg_node(arena), 1629 _bundle_instr_count(0), 1630 _bundle_cycle_number(0), 1631 _scheduled(arena), 1632 _available(arena), 1633 _next_node(NULL), 1634 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]), 1635 _pinch_free_list(arena) 1636 #ifndef PRODUCT 1637 , _branches(0) 1638 , _unconditional_delays(0) 1639 #endif 1640 { 1641 // Create a MachNopNode 1642 _nop = new (&compile) MachNopNode(); 1643 1644 // Now that the nops are in the array, save the count 1645 // (but allow entries for the nops) 1646 _node_bundling_limit = compile.unique(); 1647 uint node_max = _regalloc->node_regs_max_index(); 1648 1649 compile.set_node_bundling_limit(_node_bundling_limit); 1650 1651 // This one is persistent within the Compile class 1652 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1653 1654 // Allocate space for fixed-size arrays 1655 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1656 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1657 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1658 1659 // Clear the arrays 1660 memset(_node_bundling_base, 0, node_max * sizeof(Bundle)); 1661 memset(_node_latency, 0, node_max * sizeof(unsigned short)); 1662 memset(_uses, 0, node_max * sizeof(short)); 1663 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 1664 1665 // Clear the bundling information 1666 memcpy(_bundle_use_elements, 1667 Pipeline_Use::elaborated_elements, 1668 sizeof(Pipeline_Use::elaborated_elements)); 1669 1670 // Get the last node 1671 Block *bb = _cfg->_blocks[_cfg->_blocks.size()-1]; 1672 1673 _next_node = bb->_nodes[bb->_nodes.size()-1]; 1674 } 1675 1676 #ifndef PRODUCT 1677 // Scheduling destructor 1678 Scheduling::~Scheduling() { 1679 _total_branches += _branches; 1680 _total_unconditional_delays += _unconditional_delays; 1681 } 1682 #endif 1683 1684 // Step ahead "i" cycles 1685 void Scheduling::step(uint i) { 1686 1687 Bundle *bundle = node_bundling(_next_node); 1688 bundle->set_starts_bundle(); 1689 1690 // Update the bundle record, but leave the flags information alone 1691 if (_bundle_instr_count > 0) { 1692 bundle->set_instr_count(_bundle_instr_count); 1693 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1694 } 1695 1696 // Update the state information 1697 _bundle_instr_count = 0; 1698 _bundle_cycle_number += i; 1699 _bundle_use.step(i); 1700 } 1701 1702 void Scheduling::step_and_clear() { 1703 Bundle *bundle = node_bundling(_next_node); 1704 bundle->set_starts_bundle(); 1705 1706 // Update the bundle record 1707 if (_bundle_instr_count > 0) { 1708 bundle->set_instr_count(_bundle_instr_count); 1709 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1710 1711 _bundle_cycle_number += 1; 1712 } 1713 1714 // Clear the bundling information 1715 _bundle_instr_count = 0; 1716 _bundle_use.reset(); 1717 1718 memcpy(_bundle_use_elements, 1719 Pipeline_Use::elaborated_elements, 1720 sizeof(Pipeline_Use::elaborated_elements)); 1721 } 1722 1723 //------------------------------ScheduleAndBundle------------------------------ 1724 // Perform instruction scheduling and bundling over the sequence of 1725 // instructions in backwards order. 1726 void Compile::ScheduleAndBundle() { 1727 1728 // Don't optimize this if it isn't a method 1729 if (!_method) 1730 return; 1731 1732 // Don't optimize this if scheduling is disabled 1733 if (!do_scheduling()) 1734 return; 1735 1736 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); ) 1737 1738 // Create a data structure for all the scheduling information 1739 Scheduling scheduling(Thread::current()->resource_area(), *this); 1740 1741 // Initialize the space for the BufferBlob used to find and verify 1742 // instruction size in MachNode::emit_size() 1743 init_scratch_buffer_blob(MAX_const_size); 1744 if (failing()) return; // Out of memory 1745 1746 // Walk backwards over each basic block, computing the needed alignment 1747 // Walk over all the basic blocks 1748 scheduling.DoScheduling(); 1749 } 1750 1751 //------------------------------ComputeLocalLatenciesForward------------------- 1752 // Compute the latency of all the instructions. This is fairly simple, 1753 // because we already have a legal ordering. Walk over the instructions 1754 // from first to last, and compute the latency of the instruction based 1755 // on the latency of the preceding instruction(s). 1756 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { 1757 #ifndef PRODUCT 1758 if (_cfg->C->trace_opto_output()) 1759 tty->print("# -> ComputeLocalLatenciesForward\n"); 1760 #endif 1761 1762 // Walk over all the schedulable instructions 1763 for( uint j=_bb_start; j < _bb_end; j++ ) { 1764 1765 // This is a kludge, forcing all latency calculations to start at 1. 1766 // Used to allow latency 0 to force an instruction to the beginning 1767 // of the bb 1768 uint latency = 1; 1769 Node *use = bb->_nodes[j]; 1770 uint nlen = use->len(); 1771 1772 // Walk over all the inputs 1773 for ( uint k=0; k < nlen; k++ ) { 1774 Node *def = use->in(k); 1775 if (!def) 1776 continue; 1777 1778 uint l = _node_latency[def->_idx] + use->latency(k); 1779 if (latency < l) 1780 latency = l; 1781 } 1782 1783 _node_latency[use->_idx] = latency; 1784 1785 #ifndef PRODUCT 1786 if (_cfg->C->trace_opto_output()) { 1787 tty->print("# latency %4d: ", latency); 1788 use->dump(); 1789 } 1790 #endif 1791 } 1792 1793 #ifndef PRODUCT 1794 if (_cfg->C->trace_opto_output()) 1795 tty->print("# <- ComputeLocalLatenciesForward\n"); 1796 #endif 1797 1798 } // end ComputeLocalLatenciesForward 1799 1800 // See if this node fits into the present instruction bundle 1801 bool Scheduling::NodeFitsInBundle(Node *n) { 1802 uint n_idx = n->_idx; 1803 1804 // If this is the unconditional delay instruction, then it fits 1805 if (n == _unconditional_delay_slot) { 1806 #ifndef PRODUCT 1807 if (_cfg->C->trace_opto_output()) 1808 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 1809 #endif 1810 return (true); 1811 } 1812 1813 // If the node cannot be scheduled this cycle, skip it 1814 if (_current_latency[n_idx] > _bundle_cycle_number) { 1815 #ifndef PRODUCT 1816 if (_cfg->C->trace_opto_output()) 1817 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 1818 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 1819 #endif 1820 return (false); 1821 } 1822 1823 const Pipeline *node_pipeline = n->pipeline(); 1824 1825 uint instruction_count = node_pipeline->instructionCount(); 1826 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1827 instruction_count = 0; 1828 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1829 instruction_count++; 1830 1831 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 1832 #ifndef PRODUCT 1833 if (_cfg->C->trace_opto_output()) 1834 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 1835 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 1836 #endif 1837 return (false); 1838 } 1839 1840 // Don't allow non-machine nodes to be handled this way 1841 if (!n->is_Mach() && instruction_count == 0) 1842 return (false); 1843 1844 // See if there is any overlap 1845 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 1846 1847 if (delay > 0) { 1848 #ifndef PRODUCT 1849 if (_cfg->C->trace_opto_output()) 1850 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 1851 #endif 1852 return false; 1853 } 1854 1855 #ifndef PRODUCT 1856 if (_cfg->C->trace_opto_output()) 1857 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 1858 #endif 1859 1860 return true; 1861 } 1862 1863 Node * Scheduling::ChooseNodeToBundle() { 1864 uint siz = _available.size(); 1865 1866 if (siz == 0) { 1867 1868 #ifndef PRODUCT 1869 if (_cfg->C->trace_opto_output()) 1870 tty->print("# ChooseNodeToBundle: NULL\n"); 1871 #endif 1872 return (NULL); 1873 } 1874 1875 // Fast path, if only 1 instruction in the bundle 1876 if (siz == 1) { 1877 #ifndef PRODUCT 1878 if (_cfg->C->trace_opto_output()) { 1879 tty->print("# ChooseNodeToBundle (only 1): "); 1880 _available[0]->dump(); 1881 } 1882 #endif 1883 return (_available[0]); 1884 } 1885 1886 // Don't bother, if the bundle is already full 1887 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 1888 for ( uint i = 0; i < siz; i++ ) { 1889 Node *n = _available[i]; 1890 1891 // Skip projections, we'll handle them another way 1892 if (n->is_Proj()) 1893 continue; 1894 1895 // This presupposed that instructions are inserted into the 1896 // available list in a legality order; i.e. instructions that 1897 // must be inserted first are at the head of the list 1898 if (NodeFitsInBundle(n)) { 1899 #ifndef PRODUCT 1900 if (_cfg->C->trace_opto_output()) { 1901 tty->print("# ChooseNodeToBundle: "); 1902 n->dump(); 1903 } 1904 #endif 1905 return (n); 1906 } 1907 } 1908 } 1909 1910 // Nothing fits in this bundle, choose the highest priority 1911 #ifndef PRODUCT 1912 if (_cfg->C->trace_opto_output()) { 1913 tty->print("# ChooseNodeToBundle: "); 1914 _available[0]->dump(); 1915 } 1916 #endif 1917 1918 return _available[0]; 1919 } 1920 1921 //------------------------------AddNodeToAvailableList------------------------- 1922 void Scheduling::AddNodeToAvailableList(Node *n) { 1923 assert( !n->is_Proj(), "projections never directly made available" ); 1924 #ifndef PRODUCT 1925 if (_cfg->C->trace_opto_output()) { 1926 tty->print("# AddNodeToAvailableList: "); 1927 n->dump(); 1928 } 1929 #endif 1930 1931 int latency = _current_latency[n->_idx]; 1932 1933 // Insert in latency order (insertion sort) 1934 uint i; 1935 for ( i=0; i < _available.size(); i++ ) 1936 if (_current_latency[_available[i]->_idx] > latency) 1937 break; 1938 1939 // Special Check for compares following branches 1940 if( n->is_Mach() && _scheduled.size() > 0 ) { 1941 int op = n->as_Mach()->ideal_Opcode(); 1942 Node *last = _scheduled[0]; 1943 if( last->is_MachIf() && last->in(1) == n && 1944 ( op == Op_CmpI || 1945 op == Op_CmpU || 1946 op == Op_CmpP || 1947 op == Op_CmpF || 1948 op == Op_CmpD || 1949 op == Op_CmpL ) ) { 1950 1951 // Recalculate position, moving to front of same latency 1952 for ( i=0 ; i < _available.size(); i++ ) 1953 if (_current_latency[_available[i]->_idx] >= latency) 1954 break; 1955 } 1956 } 1957 1958 // Insert the node in the available list 1959 _available.insert(i, n); 1960 1961 #ifndef PRODUCT 1962 if (_cfg->C->trace_opto_output()) 1963 dump_available(); 1964 #endif 1965 } 1966 1967 //------------------------------DecrementUseCounts----------------------------- 1968 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 1969 for ( uint i=0; i < n->len(); i++ ) { 1970 Node *def = n->in(i); 1971 if (!def) continue; 1972 if( def->is_Proj() ) // If this is a machine projection, then 1973 def = def->in(0); // propagate usage thru to the base instruction 1974 1975 if( _bbs[def->_idx] != bb ) // Ignore if not block-local 1976 continue; 1977 1978 // Compute the latency 1979 uint l = _bundle_cycle_number + n->latency(i); 1980 if (_current_latency[def->_idx] < l) 1981 _current_latency[def->_idx] = l; 1982 1983 // If this does not have uses then schedule it 1984 if ((--_uses[def->_idx]) == 0) 1985 AddNodeToAvailableList(def); 1986 } 1987 } 1988 1989 //------------------------------AddNodeToBundle-------------------------------- 1990 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 1991 #ifndef PRODUCT 1992 if (_cfg->C->trace_opto_output()) { 1993 tty->print("# AddNodeToBundle: "); 1994 n->dump(); 1995 } 1996 #endif 1997 1998 // Remove this from the available list 1999 uint i; 2000 for (i = 0; i < _available.size(); i++) 2001 if (_available[i] == n) 2002 break; 2003 assert(i < _available.size(), "entry in _available list not found"); 2004 _available.remove(i); 2005 2006 // See if this fits in the current bundle 2007 const Pipeline *node_pipeline = n->pipeline(); 2008 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 2009 2010 // Check for instructions to be placed in the delay slot. We 2011 // do this before we actually schedule the current instruction, 2012 // because the delay slot follows the current instruction. 2013 if (Pipeline::_branch_has_delay_slot && 2014 node_pipeline->hasBranchDelay() && 2015 !_unconditional_delay_slot) { 2016 2017 uint siz = _available.size(); 2018 2019 // Conditional branches can support an instruction that 2020 // is unconditionally executed and not dependent by the 2021 // branch, OR a conditionally executed instruction if 2022 // the branch is taken. In practice, this means that 2023 // the first instruction at the branch target is 2024 // copied to the delay slot, and the branch goes to 2025 // the instruction after that at the branch target 2026 if ( n->is_Mach() && n->is_Branch() ) { 2027 2028 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 2029 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 2030 2031 #ifndef PRODUCT 2032 _branches++; 2033 #endif 2034 2035 // At least 1 instruction is on the available list 2036 // that is not dependent on the branch 2037 for (uint i = 0; i < siz; i++) { 2038 Node *d = _available[i]; 2039 const Pipeline *avail_pipeline = d->pipeline(); 2040 2041 // Don't allow safepoints in the branch shadow, that will 2042 // cause a number of difficulties 2043 if ( avail_pipeline->instructionCount() == 1 && 2044 !avail_pipeline->hasMultipleBundles() && 2045 !avail_pipeline->hasBranchDelay() && 2046 Pipeline::instr_has_unit_size() && 2047 d->size(_regalloc) == Pipeline::instr_unit_size() && 2048 NodeFitsInBundle(d) && 2049 !node_bundling(d)->used_in_delay()) { 2050 2051 if (d->is_Mach() && !d->is_MachSafePoint()) { 2052 // A node that fits in the delay slot was found, so we need to 2053 // set the appropriate bits in the bundle pipeline information so 2054 // that it correctly indicates resource usage. Later, when we 2055 // attempt to add this instruction to the bundle, we will skip 2056 // setting the resource usage. 2057 _unconditional_delay_slot = d; 2058 node_bundling(n)->set_use_unconditional_delay(); 2059 node_bundling(d)->set_used_in_unconditional_delay(); 2060 _bundle_use.add_usage(avail_pipeline->resourceUse()); 2061 _current_latency[d->_idx] = _bundle_cycle_number; 2062 _next_node = d; 2063 ++_bundle_instr_count; 2064 #ifndef PRODUCT 2065 _unconditional_delays++; 2066 #endif 2067 break; 2068 } 2069 } 2070 } 2071 } 2072 2073 // No delay slot, add a nop to the usage 2074 if (!_unconditional_delay_slot) { 2075 // See if adding an instruction in the delay slot will overflow 2076 // the bundle. 2077 if (!NodeFitsInBundle(_nop)) { 2078 #ifndef PRODUCT 2079 if (_cfg->C->trace_opto_output()) 2080 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2081 #endif 2082 step(1); 2083 } 2084 2085 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2086 _next_node = _nop; 2087 ++_bundle_instr_count; 2088 } 2089 2090 // See if the instruction in the delay slot requires a 2091 // step of the bundles 2092 if (!NodeFitsInBundle(n)) { 2093 #ifndef PRODUCT 2094 if (_cfg->C->trace_opto_output()) 2095 tty->print("# *** STEP(branch won't fit) ***\n"); 2096 #endif 2097 // Update the state information 2098 _bundle_instr_count = 0; 2099 _bundle_cycle_number += 1; 2100 _bundle_use.step(1); 2101 } 2102 } 2103 2104 // Get the number of instructions 2105 uint instruction_count = node_pipeline->instructionCount(); 2106 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2107 instruction_count = 0; 2108 2109 // Compute the latency information 2110 uint delay = 0; 2111 2112 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2113 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2114 if (relative_latency < 0) 2115 relative_latency = 0; 2116 2117 delay = _bundle_use.full_latency(relative_latency, node_usage); 2118 2119 // Does not fit in this bundle, start a new one 2120 if (delay > 0) { 2121 step(delay); 2122 2123 #ifndef PRODUCT 2124 if (_cfg->C->trace_opto_output()) 2125 tty->print("# *** STEP(%d) ***\n", delay); 2126 #endif 2127 } 2128 } 2129 2130 // If this was placed in the delay slot, ignore it 2131 if (n != _unconditional_delay_slot) { 2132 2133 if (delay == 0) { 2134 if (node_pipeline->hasMultipleBundles()) { 2135 #ifndef PRODUCT 2136 if (_cfg->C->trace_opto_output()) 2137 tty->print("# *** STEP(multiple instructions) ***\n"); 2138 #endif 2139 step(1); 2140 } 2141 2142 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2143 #ifndef PRODUCT 2144 if (_cfg->C->trace_opto_output()) 2145 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2146 instruction_count + _bundle_instr_count, 2147 Pipeline::_max_instrs_per_cycle); 2148 #endif 2149 step(1); 2150 } 2151 } 2152 2153 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2154 _bundle_instr_count++; 2155 2156 // Set the node's latency 2157 _current_latency[n->_idx] = _bundle_cycle_number; 2158 2159 // Now merge the functional unit information 2160 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2161 _bundle_use.add_usage(node_usage); 2162 2163 // Increment the number of instructions in this bundle 2164 _bundle_instr_count += instruction_count; 2165 2166 // Remember this node for later 2167 if (n->is_Mach()) 2168 _next_node = n; 2169 } 2170 2171 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2172 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2173 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2174 // into the block. All other scheduled nodes get put in the schedule here. 2175 int op = n->Opcode(); 2176 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2177 (op != Op_Node && // Not an unused antidepedence node and 2178 // not an unallocated boxlock 2179 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2180 2181 // Push any trailing projections 2182 if( bb->_nodes[bb->_nodes.size()-1] != n ) { 2183 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2184 Node *foi = n->fast_out(i); 2185 if( foi->is_Proj() ) 2186 _scheduled.push(foi); 2187 } 2188 } 2189 2190 // Put the instruction in the schedule list 2191 _scheduled.push(n); 2192 } 2193 2194 #ifndef PRODUCT 2195 if (_cfg->C->trace_opto_output()) 2196 dump_available(); 2197 #endif 2198 2199 // Walk all the definitions, decrementing use counts, and 2200 // if a definition has a 0 use count, place it in the available list. 2201 DecrementUseCounts(n,bb); 2202 } 2203 2204 //------------------------------ComputeUseCount-------------------------------- 2205 // This method sets the use count within a basic block. We will ignore all 2206 // uses outside the current basic block. As we are doing a backwards walk, 2207 // any node we reach that has a use count of 0 may be scheduled. This also 2208 // avoids the problem of cyclic references from phi nodes, as long as phi 2209 // nodes are at the front of the basic block. This method also initializes 2210 // the available list to the set of instructions that have no uses within this 2211 // basic block. 2212 void Scheduling::ComputeUseCount(const Block *bb) { 2213 #ifndef PRODUCT 2214 if (_cfg->C->trace_opto_output()) 2215 tty->print("# -> ComputeUseCount\n"); 2216 #endif 2217 2218 // Clear the list of available and scheduled instructions, just in case 2219 _available.clear(); 2220 _scheduled.clear(); 2221 2222 // No delay slot specified 2223 _unconditional_delay_slot = NULL; 2224 2225 #ifdef ASSERT 2226 for( uint i=0; i < bb->_nodes.size(); i++ ) 2227 assert( _uses[bb->_nodes[i]->_idx] == 0, "_use array not clean" ); 2228 #endif 2229 2230 // Force the _uses count to never go to zero for unscheduable pieces 2231 // of the block 2232 for( uint k = 0; k < _bb_start; k++ ) 2233 _uses[bb->_nodes[k]->_idx] = 1; 2234 for( uint l = _bb_end; l < bb->_nodes.size(); l++ ) 2235 _uses[bb->_nodes[l]->_idx] = 1; 2236 2237 // Iterate backwards over the instructions in the block. Don't count the 2238 // branch projections at end or the block header instructions. 2239 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2240 Node *n = bb->_nodes[j]; 2241 if( n->is_Proj() ) continue; // Projections handled another way 2242 2243 // Account for all uses 2244 for ( uint k = 0; k < n->len(); k++ ) { 2245 Node *inp = n->in(k); 2246 if (!inp) continue; 2247 assert(inp != n, "no cycles allowed" ); 2248 if( _bbs[inp->_idx] == bb ) { // Block-local use? 2249 if( inp->is_Proj() ) // Skip through Proj's 2250 inp = inp->in(0); 2251 ++_uses[inp->_idx]; // Count 1 block-local use 2252 } 2253 } 2254 2255 // If this instruction has a 0 use count, then it is available 2256 if (!_uses[n->_idx]) { 2257 _current_latency[n->_idx] = _bundle_cycle_number; 2258 AddNodeToAvailableList(n); 2259 } 2260 2261 #ifndef PRODUCT 2262 if (_cfg->C->trace_opto_output()) { 2263 tty->print("# uses: %3d: ", _uses[n->_idx]); 2264 n->dump(); 2265 } 2266 #endif 2267 } 2268 2269 #ifndef PRODUCT 2270 if (_cfg->C->trace_opto_output()) 2271 tty->print("# <- ComputeUseCount\n"); 2272 #endif 2273 } 2274 2275 // This routine performs scheduling on each basic block in reverse order, 2276 // using instruction latencies and taking into account function unit 2277 // availability. 2278 void Scheduling::DoScheduling() { 2279 #ifndef PRODUCT 2280 if (_cfg->C->trace_opto_output()) 2281 tty->print("# -> DoScheduling\n"); 2282 #endif 2283 2284 Block *succ_bb = NULL; 2285 Block *bb; 2286 2287 // Walk over all the basic blocks in reverse order 2288 for( int i=_cfg->_num_blocks-1; i >= 0; succ_bb = bb, i-- ) { 2289 bb = _cfg->_blocks[i]; 2290 2291 #ifndef PRODUCT 2292 if (_cfg->C->trace_opto_output()) { 2293 tty->print("# Schedule BB#%03d (initial)\n", i); 2294 for (uint j = 0; j < bb->_nodes.size(); j++) 2295 bb->_nodes[j]->dump(); 2296 } 2297 #endif 2298 2299 // On the head node, skip processing 2300 if( bb == _cfg->_broot ) 2301 continue; 2302 2303 // Skip empty, connector blocks 2304 if (bb->is_connector()) 2305 continue; 2306 2307 // If the following block is not the sole successor of 2308 // this one, then reset the pipeline information 2309 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2310 #ifndef PRODUCT 2311 if (_cfg->C->trace_opto_output()) { 2312 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2313 _next_node->_idx, _bundle_instr_count); 2314 } 2315 #endif 2316 step_and_clear(); 2317 } 2318 2319 // Leave untouched the starting instruction, any Phis, a CreateEx node 2320 // or Top. bb->_nodes[_bb_start] is the first schedulable instruction. 2321 _bb_end = bb->_nodes.size()-1; 2322 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2323 Node *n = bb->_nodes[_bb_start]; 2324 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2325 // Also, MachIdealNodes do not get scheduled 2326 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2327 MachNode *mach = n->as_Mach(); 2328 int iop = mach->ideal_Opcode(); 2329 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2330 if( iop == Op_Con ) continue; // Do not schedule Top 2331 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2332 mach->pipeline() == MachNode::pipeline_class() && 2333 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc 2334 continue; 2335 break; // Funny loop structure to be sure... 2336 } 2337 // Compute last "interesting" instruction in block - last instruction we 2338 // might schedule. _bb_end points just after last schedulable inst. We 2339 // normally schedule conditional branches (despite them being forced last 2340 // in the block), because they have delay slots we can fill. Calls all 2341 // have their delay slots filled in the template expansions, so we don't 2342 // bother scheduling them. 2343 Node *last = bb->_nodes[_bb_end]; 2344 if( last->is_Catch() || 2345 // Exclude unreachable path case when Halt node is in a separate block. 2346 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2347 // There must be a prior call. Skip it. 2348 while( !bb->_nodes[--_bb_end]->is_Call() ) { 2349 assert( bb->_nodes[_bb_end]->is_Proj(), "skipping projections after expected call" ); 2350 } 2351 } else if( last->is_MachNullCheck() ) { 2352 // Backup so the last null-checked memory instruction is 2353 // outside the schedulable range. Skip over the nullcheck, 2354 // projection, and the memory nodes. 2355 Node *mem = last->in(1); 2356 do { 2357 _bb_end--; 2358 } while (mem != bb->_nodes[_bb_end]); 2359 } else { 2360 // Set _bb_end to point after last schedulable inst. 2361 _bb_end++; 2362 } 2363 2364 assert( _bb_start <= _bb_end, "inverted block ends" ); 2365 2366 // Compute the register antidependencies for the basic block 2367 ComputeRegisterAntidependencies(bb); 2368 if (_cfg->C->failing()) return; // too many D-U pinch points 2369 2370 // Compute intra-bb latencies for the nodes 2371 ComputeLocalLatenciesForward(bb); 2372 2373 // Compute the usage within the block, and set the list of all nodes 2374 // in the block that have no uses within the block. 2375 ComputeUseCount(bb); 2376 2377 // Schedule the remaining instructions in the block 2378 while ( _available.size() > 0 ) { 2379 Node *n = ChooseNodeToBundle(); 2380 AddNodeToBundle(n,bb); 2381 } 2382 2383 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2384 #ifdef ASSERT 2385 for( uint l = _bb_start; l < _bb_end; l++ ) { 2386 Node *n = bb->_nodes[l]; 2387 uint m; 2388 for( m = 0; m < _bb_end-_bb_start; m++ ) 2389 if( _scheduled[m] == n ) 2390 break; 2391 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2392 } 2393 #endif 2394 2395 // Now copy the instructions (in reverse order) back to the block 2396 for ( uint k = _bb_start; k < _bb_end; k++ ) 2397 bb->_nodes.map(k, _scheduled[_bb_end-k-1]); 2398 2399 #ifndef PRODUCT 2400 if (_cfg->C->trace_opto_output()) { 2401 tty->print("# Schedule BB#%03d (final)\n", i); 2402 uint current = 0; 2403 for (uint j = 0; j < bb->_nodes.size(); j++) { 2404 Node *n = bb->_nodes[j]; 2405 if( valid_bundle_info(n) ) { 2406 Bundle *bundle = node_bundling(n); 2407 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2408 tty->print("*** Bundle: "); 2409 bundle->dump(); 2410 } 2411 n->dump(); 2412 } 2413 } 2414 } 2415 #endif 2416 #ifdef ASSERT 2417 verify_good_schedule(bb,"after block local scheduling"); 2418 #endif 2419 } 2420 2421 #ifndef PRODUCT 2422 if (_cfg->C->trace_opto_output()) 2423 tty->print("# <- DoScheduling\n"); 2424 #endif 2425 2426 // Record final node-bundling array location 2427 _regalloc->C->set_node_bundling_base(_node_bundling_base); 2428 2429 } // end DoScheduling 2430 2431 //------------------------------verify_good_schedule--------------------------- 2432 // Verify that no live-range used in the block is killed in the block by a 2433 // wrong DEF. This doesn't verify live-ranges that span blocks. 2434 2435 // Check for edge existence. Used to avoid adding redundant precedence edges. 2436 static bool edge_from_to( Node *from, Node *to ) { 2437 for( uint i=0; i<from->len(); i++ ) 2438 if( from->in(i) == to ) 2439 return true; 2440 return false; 2441 } 2442 2443 #ifdef ASSERT 2444 //------------------------------verify_do_def---------------------------------- 2445 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2446 // Check for bad kills 2447 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2448 Node *prior_use = _reg_node[def]; 2449 if( prior_use && !edge_from_to(prior_use,n) ) { 2450 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2451 n->dump(); 2452 tty->print_cr("..."); 2453 prior_use->dump(); 2454 assert(edge_from_to(prior_use,n),msg); 2455 } 2456 _reg_node.map(def,NULL); // Kill live USEs 2457 } 2458 } 2459 2460 //------------------------------verify_good_schedule--------------------------- 2461 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2462 2463 // Zap to something reasonable for the verify code 2464 _reg_node.clear(); 2465 2466 // Walk over the block backwards. Check to make sure each DEF doesn't 2467 // kill a live value (other than the one it's supposed to). Add each 2468 // USE to the live set. 2469 for( uint i = b->_nodes.size()-1; i >= _bb_start; i-- ) { 2470 Node *n = b->_nodes[i]; 2471 int n_op = n->Opcode(); 2472 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2473 // Fat-proj kills a slew of registers 2474 RegMask rm = n->out_RegMask();// Make local copy 2475 while( rm.is_NotEmpty() ) { 2476 OptoReg::Name kill = rm.find_first_elem(); 2477 rm.Remove(kill); 2478 verify_do_def( n, kill, msg ); 2479 } 2480 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2481 // Get DEF'd registers the normal way 2482 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2483 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2484 } 2485 2486 // Now make all USEs live 2487 for( uint i=1; i<n->req(); i++ ) { 2488 Node *def = n->in(i); 2489 assert(def != 0, "input edge required"); 2490 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2491 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2492 if( OptoReg::is_valid(reg_lo) ) { 2493 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg); 2494 _reg_node.map(reg_lo,n); 2495 } 2496 if( OptoReg::is_valid(reg_hi) ) { 2497 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg); 2498 _reg_node.map(reg_hi,n); 2499 } 2500 } 2501 2502 } 2503 2504 // Zap to something reasonable for the Antidependence code 2505 _reg_node.clear(); 2506 } 2507 #endif 2508 2509 // Conditionally add precedence edges. Avoid putting edges on Projs. 2510 static void add_prec_edge_from_to( Node *from, Node *to ) { 2511 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2512 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2513 from = from->in(0); 2514 } 2515 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2516 !edge_from_to( from, to ) ) // Avoid duplicate edge 2517 from->add_prec(to); 2518 } 2519 2520 //------------------------------anti_do_def------------------------------------ 2521 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2522 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2523 return; 2524 2525 Node *pinch = _reg_node[def_reg]; // Get pinch point 2526 if( !pinch || _bbs[pinch->_idx] != b || // No pinch-point yet? 2527 is_def ) { // Check for a true def (not a kill) 2528 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2529 return; 2530 } 2531 2532 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2533 debug_only( def = (Node*)0xdeadbeef; ) 2534 2535 // After some number of kills there _may_ be a later def 2536 Node *later_def = NULL; 2537 2538 // Finding a kill requires a real pinch-point. 2539 // Check for not already having a pinch-point. 2540 // Pinch points are Op_Node's. 2541 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2542 later_def = pinch; // Must be def/kill as optimistic pinch-point 2543 if ( _pinch_free_list.size() > 0) { 2544 pinch = _pinch_free_list.pop(); 2545 } else { 2546 pinch = new (_cfg->C, 1) Node(1); // Pinch point to-be 2547 } 2548 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2549 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2550 return; 2551 } 2552 _bbs.map(pinch->_idx,b); // Pretend it's valid in this block (lazy init) 2553 _reg_node.map(def_reg,pinch); // Record pinch-point 2554 //_regalloc->set_bad(pinch->_idx); // Already initialized this way. 2555 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2556 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call 2557 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2558 later_def = NULL; // and no later def 2559 } 2560 pinch->set_req(0,later_def); // Hook later def so we can find it 2561 } else { // Else have valid pinch point 2562 if( pinch->in(0) ) // If there is a later-def 2563 later_def = pinch->in(0); // Get it 2564 } 2565 2566 // Add output-dependence edge from later def to kill 2567 if( later_def ) // If there is some original def 2568 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2569 2570 // See if current kill is also a use, and so is forced to be the pinch-point. 2571 if( pinch->Opcode() == Op_Node ) { 2572 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2573 for( uint i=1; i<uses->req(); i++ ) { 2574 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2575 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2576 // Yes, found a use/kill pinch-point 2577 pinch->set_req(0,NULL); // 2578 pinch->replace_by(kill); // Move anti-dep edges up 2579 pinch = kill; 2580 _reg_node.map(def_reg,pinch); 2581 return; 2582 } 2583 } 2584 } 2585 2586 // Add edge from kill to pinch-point 2587 add_prec_edge_from_to(kill,pinch); 2588 } 2589 2590 //------------------------------anti_do_use------------------------------------ 2591 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2592 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2593 return; 2594 Node *pinch = _reg_node[use_reg]; // Get pinch point 2595 // Check for no later def_reg/kill in block 2596 if( pinch && _bbs[pinch->_idx] == b && 2597 // Use has to be block-local as well 2598 _bbs[use->_idx] == b ) { 2599 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2600 pinch->req() == 1 ) { // pinch not yet in block? 2601 pinch->del_req(0); // yank pointer to later-def, also set flag 2602 // Insert the pinch-point in the block just after the last use 2603 b->_nodes.insert(b->find_node(use)+1,pinch); 2604 _bb_end++; // Increase size scheduled region in block 2605 } 2606 2607 add_prec_edge_from_to(pinch,use); 2608 } 2609 } 2610 2611 //------------------------------ComputeRegisterAntidependences----------------- 2612 // We insert antidependences between the reads and following write of 2613 // allocated registers to prevent illegal code motion. Hopefully, the 2614 // number of added references should be fairly small, especially as we 2615 // are only adding references within the current basic block. 2616 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2617 2618 #ifdef ASSERT 2619 verify_good_schedule(b,"before block local scheduling"); 2620 #endif 2621 2622 // A valid schedule, for each register independently, is an endless cycle 2623 // of: a def, then some uses (connected to the def by true dependencies), 2624 // then some kills (defs with no uses), finally the cycle repeats with a new 2625 // def. The uses are allowed to float relative to each other, as are the 2626 // kills. No use is allowed to slide past a kill (or def). This requires 2627 // antidependencies between all uses of a single def and all kills that 2628 // follow, up to the next def. More edges are redundant, because later defs 2629 // & kills are already serialized with true or antidependencies. To keep 2630 // the edge count down, we add a 'pinch point' node if there's more than 2631 // one use or more than one kill/def. 2632 2633 // We add dependencies in one bottom-up pass. 2634 2635 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2636 2637 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2638 // register. If not, we record the DEF/KILL in _reg_node, the 2639 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2640 // "pinch point", a new Node that's in the graph but not in the block. 2641 // We put edges from the prior and current DEF/KILLs to the pinch point. 2642 // We put the pinch point in _reg_node. If there's already a pinch point 2643 // we merely add an edge from the current DEF/KILL to the pinch point. 2644 2645 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2646 // put an edge from the pinch point to the USE. 2647 2648 // To be expedient, the _reg_node array is pre-allocated for the whole 2649 // compilation. _reg_node is lazily initialized; it either contains a NULL, 2650 // or a valid def/kill/pinch-point, or a leftover node from some prior 2651 // block. Leftover node from some prior block is treated like a NULL (no 2652 // prior def, so no anti-dependence needed). Valid def is distinguished by 2653 // it being in the current block. 2654 bool fat_proj_seen = false; 2655 uint last_safept = _bb_end-1; 2656 Node* end_node = (_bb_end-1 >= _bb_start) ? b->_nodes[last_safept] : NULL; 2657 Node* last_safept_node = end_node; 2658 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2659 Node *n = b->_nodes[i]; 2660 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2661 if( n->Opcode() == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2662 // Fat-proj kills a slew of registers 2663 // This can add edges to 'n' and obscure whether or not it was a def, 2664 // hence the is_def flag. 2665 fat_proj_seen = true; 2666 RegMask rm = n->out_RegMask();// Make local copy 2667 while( rm.is_NotEmpty() ) { 2668 OptoReg::Name kill = rm.find_first_elem(); 2669 rm.Remove(kill); 2670 anti_do_def( b, n, kill, is_def ); 2671 } 2672 } else { 2673 // Get DEF'd registers the normal way 2674 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2675 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2676 } 2677 2678 // Check each register used by this instruction for a following DEF/KILL 2679 // that must occur afterward and requires an anti-dependence edge. 2680 for( uint j=0; j<n->req(); j++ ) { 2681 Node *def = n->in(j); 2682 if( def ) { 2683 assert( def->Opcode() != Op_MachProj || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2684 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2685 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2686 } 2687 } 2688 // Do not allow defs of new derived values to float above GC 2689 // points unless the base is definitely available at the GC point. 2690 2691 Node *m = b->_nodes[i]; 2692 2693 // Add precedence edge from following safepoint to use of derived pointer 2694 if( last_safept_node != end_node && 2695 m != last_safept_node) { 2696 for (uint k = 1; k < m->req(); k++) { 2697 const Type *t = m->in(k)->bottom_type(); 2698 if( t->isa_oop_ptr() && 2699 t->is_ptr()->offset() != 0 ) { 2700 last_safept_node->add_prec( m ); 2701 break; 2702 } 2703 } 2704 } 2705 2706 if( n->jvms() ) { // Precedence edge from derived to safept 2707 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2708 if( b->_nodes[last_safept] != last_safept_node ) { 2709 last_safept = b->find_node(last_safept_node); 2710 } 2711 for( uint j=last_safept; j > i; j-- ) { 2712 Node *mach = b->_nodes[j]; 2713 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2714 mach->add_prec( n ); 2715 } 2716 last_safept = i; 2717 last_safept_node = m; 2718 } 2719 } 2720 2721 if (fat_proj_seen) { 2722 // Garbage collect pinch nodes that were not consumed. 2723 // They are usually created by a fat kill MachProj for a call. 2724 garbage_collect_pinch_nodes(); 2725 } 2726 } 2727 2728 //------------------------------garbage_collect_pinch_nodes------------------------------- 2729 2730 // Garbage collect pinch nodes for reuse by other blocks. 2731 // 2732 // The block scheduler's insertion of anti-dependence 2733 // edges creates many pinch nodes when the block contains 2734 // 2 or more Calls. A pinch node is used to prevent a 2735 // combinatorial explosion of edges. If a set of kills for a 2736 // register is anti-dependent on a set of uses (or defs), rather 2737 // than adding an edge in the graph between each pair of kill 2738 // and use (or def), a pinch is inserted between them: 2739 // 2740 // use1 use2 use3 2741 // \ | / 2742 // \ | / 2743 // pinch 2744 // / | \ 2745 // / | \ 2746 // kill1 kill2 kill3 2747 // 2748 // One pinch node is created per register killed when 2749 // the second call is encountered during a backwards pass 2750 // over the block. Most of these pinch nodes are never 2751 // wired into the graph because the register is never 2752 // used or def'ed in the block. 2753 // 2754 void Scheduling::garbage_collect_pinch_nodes() { 2755 #ifndef PRODUCT 2756 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 2757 #endif 2758 int trace_cnt = 0; 2759 for (uint k = 0; k < _reg_node.Size(); k++) { 2760 Node* pinch = _reg_node[k]; 2761 if (pinch != NULL && pinch->Opcode() == Op_Node && 2762 // no predecence input edges 2763 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { 2764 cleanup_pinch(pinch); 2765 _pinch_free_list.push(pinch); 2766 _reg_node.map(k, NULL); 2767 #ifndef PRODUCT 2768 if (_cfg->C->trace_opto_output()) { 2769 trace_cnt++; 2770 if (trace_cnt > 40) { 2771 tty->print("\n"); 2772 trace_cnt = 0; 2773 } 2774 tty->print(" %d", pinch->_idx); 2775 } 2776 #endif 2777 } 2778 } 2779 #ifndef PRODUCT 2780 if (_cfg->C->trace_opto_output()) tty->print("\n"); 2781 #endif 2782 } 2783 2784 // Clean up a pinch node for reuse. 2785 void Scheduling::cleanup_pinch( Node *pinch ) { 2786 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 2787 2788 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 2789 Node* use = pinch->last_out(i); 2790 uint uses_found = 0; 2791 for (uint j = use->req(); j < use->len(); j++) { 2792 if (use->in(j) == pinch) { 2793 use->rm_prec(j); 2794 uses_found++; 2795 } 2796 } 2797 assert(uses_found > 0, "must be a precedence edge"); 2798 i -= uses_found; // we deleted 1 or more copies of this edge 2799 } 2800 // May have a later_def entry 2801 pinch->set_req(0, NULL); 2802 } 2803 2804 //------------------------------print_statistics------------------------------- 2805 #ifndef PRODUCT 2806 2807 void Scheduling::dump_available() const { 2808 tty->print("#Availist "); 2809 for (uint i = 0; i < _available.size(); i++) 2810 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 2811 tty->cr(); 2812 } 2813 2814 // Print Scheduling Statistics 2815 void Scheduling::print_statistics() { 2816 // Print the size added by nops for bundling 2817 tty->print("Nops added %d bytes to total of %d bytes", 2818 _total_nop_size, _total_method_size); 2819 if (_total_method_size > 0) 2820 tty->print(", for %.2f%%", 2821 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 2822 tty->print("\n"); 2823 2824 // Print the number of branch shadows filled 2825 if (Pipeline::_branch_has_delay_slot) { 2826 tty->print("Of %d branches, %d had unconditional delay slots filled", 2827 _total_branches, _total_unconditional_delays); 2828 if (_total_branches > 0) 2829 tty->print(", for %.2f%%", 2830 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 2831 tty->print("\n"); 2832 } 2833 2834 uint total_instructions = 0, total_bundles = 0; 2835 2836 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 2837 uint bundle_count = _total_instructions_per_bundle[i]; 2838 total_instructions += bundle_count * i; 2839 total_bundles += bundle_count; 2840 } 2841 2842 if (total_bundles > 0) 2843 tty->print("Average ILP (excluding nops) is %.2f\n", 2844 ((double)total_instructions) / ((double)total_bundles)); 2845 } 2846 #endif