output.cpp revision 0:a61af66fc99e
1/* 2 * Copyright 1998-2007 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25#include "incls/_precompiled.incl" 26#include "incls/_output.cpp.incl" 27 28extern uint size_java_to_interp(); 29extern uint reloc_java_to_interp(); 30extern uint size_exception_handler(); 31extern uint size_deopt_handler(); 32 33#ifndef PRODUCT 34#define DEBUG_ARG(x) , x 35#else 36#define DEBUG_ARG(x) 37#endif 38 39extern int emit_exception_handler(CodeBuffer &cbuf); 40extern int emit_deopt_handler(CodeBuffer &cbuf); 41 42//------------------------------Output----------------------------------------- 43// Convert Nodes to instruction bits and pass off to the VM 44void Compile::Output() { 45 // RootNode goes 46 assert( _cfg->_broot->_nodes.size() == 0, "" ); 47 48 // Initialize the space for the BufferBlob used to find and verify 49 // instruction size in MachNode::emit_size() 50 init_scratch_buffer_blob(); 51 52 // Make sure I can find the Start Node 53 Block_Array& bbs = _cfg->_bbs; 54 Block *entry = _cfg->_blocks[1]; 55 Block *broot = _cfg->_broot; 56 57 const StartNode *start = entry->_nodes[0]->as_Start(); 58 59 // Replace StartNode with prolog 60 MachPrologNode *prolog = new (this) MachPrologNode(); 61 entry->_nodes.map( 0, prolog ); 62 bbs.map( prolog->_idx, entry ); 63 bbs.map( start->_idx, NULL ); // start is no longer in any block 64 65 // Virtual methods need an unverified entry point 66 67 if( is_osr_compilation() ) { 68 if( PoisonOSREntry ) { 69 // TODO: Should use a ShouldNotReachHereNode... 70 _cfg->insert( broot, 0, new (this) MachBreakpointNode() ); 71 } 72 } else { 73 if( _method && !_method->flags().is_static() ) { 74 // Insert unvalidated entry point 75 _cfg->insert( broot, 0, new (this) MachUEPNode() ); 76 } 77 78 } 79 80 81 // Break before main entry point 82 if( (_method && _method->break_at_execute()) 83#ifndef PRODUCT 84 ||(OptoBreakpoint && is_method_compilation()) 85 ||(OptoBreakpointOSR && is_osr_compilation()) 86 ||(OptoBreakpointC2R && !_method) 87#endif 88 ) { 89 // checking for _method means that OptoBreakpoint does not apply to 90 // runtime stubs or frame converters 91 _cfg->insert( entry, 1, new (this) MachBreakpointNode() ); 92 } 93 94 // Insert epilogs before every return 95 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 96 Block *b = _cfg->_blocks[i]; 97 if( !b->is_connector() && b->non_connector_successor(0) == _cfg->_broot ) { // Found a program exit point? 98 Node *m = b->end(); 99 if( m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt ) { 100 MachEpilogNode *epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); 101 b->add_inst( epilog ); 102 bbs.map(epilog->_idx, b); 103 //_regalloc->set_bad(epilog->_idx); // Already initialized this way. 104 } 105 } 106 } 107 108# ifdef ENABLE_ZAP_DEAD_LOCALS 109 if ( ZapDeadCompiledLocals ) Insert_zap_nodes(); 110# endif 111 112 ScheduleAndBundle(); 113 114#ifndef PRODUCT 115 if (trace_opto_output()) { 116 tty->print("\n---- After ScheduleAndBundle ----\n"); 117 for (uint i = 0; i < _cfg->_num_blocks; i++) { 118 tty->print("\nBB#%03d:\n", i); 119 Block *bb = _cfg->_blocks[i]; 120 for (uint j = 0; j < bb->_nodes.size(); j++) { 121 Node *n = bb->_nodes[j]; 122 OptoReg::Name reg = _regalloc->get_reg_first(n); 123 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : ""); 124 n->dump(); 125 } 126 } 127 } 128#endif 129 130 if (failing()) return; 131 132 BuildOopMaps(); 133 134 if (failing()) return; 135 136 Fill_buffer(); 137} 138 139bool Compile::need_stack_bang(int frame_size_in_bytes) const { 140 // Determine if we need to generate a stack overflow check. 141 // Do it if the method is not a stub function and 142 // has java calls or has frame size > vm_page_size/8. 143 return (stub_function() == NULL && 144 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3)); 145} 146 147bool Compile::need_register_stack_bang() const { 148 // Determine if we need to generate a register stack overflow check. 149 // This is only used on architectures which have split register 150 // and memory stacks (ie. IA64). 151 // Bang if the method is not a stub function and has java calls 152 return (stub_function() == NULL && has_java_calls()); 153} 154 155# ifdef ENABLE_ZAP_DEAD_LOCALS 156 157 158// In order to catch compiler oop-map bugs, we have implemented 159// a debugging mode called ZapDeadCompilerLocals. 160// This mode causes the compiler to insert a call to a runtime routine, 161// "zap_dead_locals", right before each place in compiled code 162// that could potentially be a gc-point (i.e., a safepoint or oop map point). 163// The runtime routine checks that locations mapped as oops are really 164// oops, that locations mapped as values do not look like oops, 165// and that locations mapped as dead are not used later 166// (by zapping them to an invalid address). 167 168int Compile::_CompiledZap_count = 0; 169 170void Compile::Insert_zap_nodes() { 171 bool skip = false; 172 173 174 // Dink with static counts because code code without the extra 175 // runtime calls is MUCH faster for debugging purposes 176 177 if ( CompileZapFirst == 0 ) ; // nothing special 178 else if ( CompileZapFirst > CompiledZap_count() ) skip = true; 179 else if ( CompileZapFirst == CompiledZap_count() ) 180 warning("starting zap compilation after skipping"); 181 182 if ( CompileZapLast == -1 ) ; // nothing special 183 else if ( CompileZapLast < CompiledZap_count() ) skip = true; 184 else if ( CompileZapLast == CompiledZap_count() ) 185 warning("about to compile last zap"); 186 187 ++_CompiledZap_count; // counts skipped zaps, too 188 189 if ( skip ) return; 190 191 192 if ( _method == NULL ) 193 return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care 194 195 // Insert call to zap runtime stub before every node with an oop map 196 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 197 Block *b = _cfg->_blocks[i]; 198 for ( uint j = 0; j < b->_nodes.size(); ++j ) { 199 Node *n = b->_nodes[j]; 200 201 // Determining if we should insert a zap-a-lot node in output. 202 // We do that for all nodes that has oopmap info, except for calls 203 // to allocation. Calls to allocation passes in the old top-of-eden pointer 204 // and expect the C code to reset it. Hence, there can be no safepoints between 205 // the inlined-allocation and the call to new_Java, etc. 206 // We also cannot zap monitor calls, as they must hold the microlock 207 // during the call to Zap, which also wants to grab the microlock. 208 bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL); 209 if ( insert ) { // it is MachSafePoint 210 if ( !n->is_MachCall() ) { 211 insert = false; 212 } else if ( n->is_MachCall() ) { 213 MachCallNode* call = n->as_MachCall(); 214 if (call->entry_point() == OptoRuntime::new_instance_Java() || 215 call->entry_point() == OptoRuntime::new_array_Java() || 216 call->entry_point() == OptoRuntime::multianewarray2_Java() || 217 call->entry_point() == OptoRuntime::multianewarray3_Java() || 218 call->entry_point() == OptoRuntime::multianewarray4_Java() || 219 call->entry_point() == OptoRuntime::multianewarray5_Java() || 220 call->entry_point() == OptoRuntime::slow_arraycopy_Java() || 221 call->entry_point() == OptoRuntime::complete_monitor_locking_Java() 222 ) { 223 insert = false; 224 } 225 } 226 if (insert) { 227 Node *zap = call_zap_node(n->as_MachSafePoint(), i); 228 b->_nodes.insert( j, zap ); 229 _cfg->_bbs.map( zap->_idx, b ); 230 ++j; 231 } 232 } 233 } 234 } 235} 236 237 238Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) { 239 const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type(); 240 CallStaticJavaNode* ideal_node = 241 new (this, tf->domain()->cnt()) CallStaticJavaNode( tf, 242 OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()), 243 "call zap dead locals stub", 0, TypePtr::BOTTOM); 244 // We need to copy the OopMap from the site we're zapping at. 245 // We have to make a copy, because the zap site might not be 246 // a call site, and zap_dead is a call site. 247 OopMap* clone = node_to_check->oop_map()->deep_copy(); 248 249 // Add the cloned OopMap to the zap node 250 ideal_node->set_oop_map(clone); 251 return _matcher->match_sfpt(ideal_node); 252} 253 254//------------------------------is_node_getting_a_safepoint-------------------- 255bool Compile::is_node_getting_a_safepoint( Node* n) { 256 // This code duplicates the logic prior to the call of add_safepoint 257 // below in this file. 258 if( n->is_MachSafePoint() ) return true; 259 return false; 260} 261 262# endif // ENABLE_ZAP_DEAD_LOCALS 263 264//------------------------------compute_loop_first_inst_sizes------------------ 265// Compute the size of first NumberOfLoopInstrToAlign instructions at head 266// of a loop. When aligning a loop we need to provide enough instructions 267// in cpu's fetch buffer to feed decoders. The loop alignment could be 268// avoided if we have enough instructions in fetch buffer at the head of a loop. 269// By default, the size is set to 999999 by Block's constructor so that 270// a loop will be aligned if the size is not reset here. 271// 272// Note: Mach instructions could contain several HW instructions 273// so the size is estimated only. 274// 275void Compile::compute_loop_first_inst_sizes() { 276 // The next condition is used to gate the loop alignment optimization. 277 // Don't aligned a loop if there are enough instructions at the head of a loop 278 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad 279 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is 280 // equal to 11 bytes which is the largest address NOP instruction. 281 if( MaxLoopPad < OptoLoopAlignment-1 ) { 282 uint last_block = _cfg->_num_blocks-1; 283 for( uint i=1; i <= last_block; i++ ) { 284 Block *b = _cfg->_blocks[i]; 285 // Check the first loop's block which requires an alignment. 286 if( b->head()->is_Loop() && 287 b->code_alignment() > (uint)relocInfo::addr_unit() ) { 288 uint sum_size = 0; 289 uint inst_cnt = NumberOfLoopInstrToAlign; 290 inst_cnt = b->compute_first_inst_size(sum_size, inst_cnt, 291 _regalloc); 292 // Check the next fallthrough block if first loop's block does not have 293 // enough instructions. 294 if( inst_cnt > 0 && i < last_block ) { 295 // First, check if the first loop's block contains whole loop. 296 // LoopNode::LoopBackControl == 2. 297 Block *bx = _cfg->_bbs[b->pred(2)->_idx]; 298 // Skip connector blocks (with limit in case of irreducible loops). 299 int search_limit = 16; 300 while( bx->is_connector() && search_limit-- > 0) { 301 bx = _cfg->_bbs[bx->pred(1)->_idx]; 302 } 303 if( bx != b ) { // loop body is in several blocks. 304 Block *nb = NULL; 305 while( inst_cnt > 0 && i < last_block && nb != bx && 306 !_cfg->_blocks[i+1]->head()->is_Loop() ) { 307 i++; 308 nb = _cfg->_blocks[i]; 309 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, 310 _regalloc); 311 } // while( inst_cnt > 0 && i < last_block ) 312 } // if( bx != b ) 313 } // if( inst_cnt > 0 && i < last_block ) 314 b->set_first_inst_size(sum_size); 315 } // f( b->head()->is_Loop() ) 316 } // for( i <= last_block ) 317 } // if( MaxLoopPad < OptoLoopAlignment-1 ) 318} 319 320//----------------------Shorten_branches--------------------------------------- 321// The architecture description provides short branch variants for some long 322// branch instructions. Replace eligible long branches with short branches. 323void Compile::Shorten_branches(Label *labels, int& code_size, int& reloc_size, int& stub_size, int& const_size) { 324 325 // fill in the nop array for bundling computations 326 MachNode *_nop_list[Bundle::_nop_count]; 327 Bundle::initialize_nops(_nop_list, this); 328 329 // ------------------ 330 // Compute size of each block, method size, and relocation information size 331 uint *jmp_end = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks); 332 uint *blk_starts = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks+1); 333 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks); ) 334 blk_starts[0] = 0; 335 336 // Initialize the sizes to 0 337 code_size = 0; // Size in bytes of generated code 338 stub_size = 0; // Size in bytes of all stub entries 339 // Size in bytes of all relocation entries, including those in local stubs. 340 // Start with 2-bytes of reloc info for the unvalidated entry point 341 reloc_size = 1; // Number of relocation entries 342 const_size = 0; // size of fp constants in words 343 344 // Make three passes. The first computes pessimistic blk_starts, 345 // relative jmp_end, reloc_size and const_size information. 346 // The second performs short branch substitution using the pessimistic 347 // sizing. The third inserts nops where needed. 348 349 Node *nj; // tmp 350 351 // Step one, perform a pessimistic sizing pass. 352 uint i; 353 uint min_offset_from_last_call = 1; // init to a positive value 354 uint nop_size = (new (this) MachNopNode())->size(_regalloc); 355 for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks 356 Block *b = _cfg->_blocks[i]; 357 358 // Sum all instruction sizes to compute block size 359 uint last_inst = b->_nodes.size(); 360 uint blk_size = 0; 361 for( uint j = 0; j<last_inst; j++ ) { 362 nj = b->_nodes[j]; 363 uint inst_size = nj->size(_regalloc); 364 blk_size += inst_size; 365 // Handle machine instruction nodes 366 if( nj->is_Mach() ) { 367 MachNode *mach = nj->as_Mach(); 368 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 369 reloc_size += mach->reloc(); 370 const_size += mach->const_size(); 371 if( mach->is_MachCall() ) { 372 MachCallNode *mcall = mach->as_MachCall(); 373 // This destination address is NOT PC-relative 374 375 mcall->method_set((intptr_t)mcall->entry_point()); 376 377 if( mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method ) { 378 stub_size += size_java_to_interp(); 379 reloc_size += reloc_java_to_interp(); 380 } 381 } else if (mach->is_MachSafePoint()) { 382 // If call/safepoint are adjacent, account for possible 383 // nop to disambiguate the two safepoints. 384 if (min_offset_from_last_call == 0) { 385 blk_size += nop_size; 386 } 387 } 388 } 389 min_offset_from_last_call += inst_size; 390 // Remember end of call offset 391 if (nj->is_MachCall() && nj->as_MachCall()->is_safepoint_node()) { 392 min_offset_from_last_call = 0; 393 } 394 } 395 396 // During short branch replacement, we store the relative (to blk_starts) 397 // end of jump in jmp_end, rather than the absolute end of jump. This 398 // is so that we do not need to recompute sizes of all nodes when we compute 399 // correct blk_starts in our next sizing pass. 400 jmp_end[i] = blk_size; 401 DEBUG_ONLY( jmp_target[i] = 0; ) 402 403 // When the next block starts a loop, we may insert pad NOP 404 // instructions. Since we cannot know our future alignment, 405 // assume the worst. 406 if( i<_cfg->_num_blocks-1 ) { 407 Block *nb = _cfg->_blocks[i+1]; 408 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 409 if( max_loop_pad > 0 ) { 410 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 411 blk_size += max_loop_pad; 412 } 413 } 414 415 // Save block size; update total method size 416 blk_starts[i+1] = blk_starts[i]+blk_size; 417 } 418 419 // Step two, replace eligible long jumps. 420 421 // Note: this will only get the long branches within short branch 422 // range. Another pass might detect more branches that became 423 // candidates because the shortening in the first pass exposed 424 // more opportunities. Unfortunately, this would require 425 // recomputing the starting and ending positions for the blocks 426 for( i=0; i<_cfg->_num_blocks; i++ ) { 427 Block *b = _cfg->_blocks[i]; 428 429 int j; 430 // Find the branch; ignore trailing NOPs. 431 for( j = b->_nodes.size()-1; j>=0; j-- ) { 432 nj = b->_nodes[j]; 433 if( !nj->is_Mach() || nj->as_Mach()->ideal_Opcode() != Op_Con ) 434 break; 435 } 436 437 if (j >= 0) { 438 if( nj->is_Mach() && nj->as_Mach()->may_be_short_branch() ) { 439 MachNode *mach = nj->as_Mach(); 440 // This requires the TRUE branch target be in succs[0] 441 uint bnum = b->non_connector_successor(0)->_pre_order; 442 uintptr_t target = blk_starts[bnum]; 443 if( mach->is_pc_relative() ) { 444 int offset = target-(blk_starts[i] + jmp_end[i]); 445 if (_matcher->is_short_branch_offset(offset)) { 446 // We've got a winner. Replace this branch. 447 MachNode *replacement = mach->short_branch_version(this); 448 b->_nodes.map(j, replacement); 449 450 // Update the jmp_end size to save time in our 451 // next pass. 452 jmp_end[i] -= (mach->size(_regalloc) - replacement->size(_regalloc)); 453 DEBUG_ONLY( jmp_target[i] = bnum; ); 454 } 455 } else { 456#ifndef PRODUCT 457 mach->dump(3); 458#endif 459 Unimplemented(); 460 } 461 } 462 } 463 } 464 465 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 466 // of a loop. It is used to determine the padding for loop alignment. 467 compute_loop_first_inst_sizes(); 468 469 // Step 3, compute the offsets of all the labels 470 uint last_call_adr = max_uint; 471 for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks 472 // copy the offset of the beginning to the corresponding label 473 assert(labels[i].is_unused(), "cannot patch at this point"); 474 labels[i].bind_loc(blk_starts[i], CodeBuffer::SECT_INSTS); 475 476 // insert padding for any instructions that need it 477 Block *b = _cfg->_blocks[i]; 478 uint last_inst = b->_nodes.size(); 479 uint adr = blk_starts[i]; 480 for( uint j = 0; j<last_inst; j++ ) { 481 nj = b->_nodes[j]; 482 if( nj->is_Mach() ) { 483 int padding = nj->as_Mach()->compute_padding(adr); 484 // If call/safepoint are adjacent insert a nop (5010568) 485 if (padding == 0 && nj->is_MachSafePoint() && !nj->is_MachCall() && 486 adr == last_call_adr ) { 487 padding = nop_size; 488 } 489 if(padding > 0) { 490 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 491 int nops_cnt = padding / nop_size; 492 MachNode *nop = new (this) MachNopNode(nops_cnt); 493 b->_nodes.insert(j++, nop); 494 _cfg->_bbs.map( nop->_idx, b ); 495 adr += padding; 496 last_inst++; 497 } 498 } 499 adr += nj->size(_regalloc); 500 501 // Remember end of call offset 502 if (nj->is_MachCall() && nj->as_MachCall()->is_safepoint_node()) { 503 last_call_adr = adr; 504 } 505 } 506 507 if ( i != _cfg->_num_blocks-1) { 508 // Get the size of the block 509 uint blk_size = adr - blk_starts[i]; 510 511 // When the next block starts a loop, we may insert pad NOP 512 // instructions. 513 Block *nb = _cfg->_blocks[i+1]; 514 int current_offset = blk_starts[i] + blk_size; 515 current_offset += nb->alignment_padding(current_offset); 516 // Save block size; update total method size 517 blk_starts[i+1] = current_offset; 518 } 519 } 520 521#ifdef ASSERT 522 for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks 523 if( jmp_target[i] != 0 ) { 524 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_end[i]); 525 if (!_matcher->is_short_branch_offset(offset)) { 526 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]); 527 } 528 assert(_matcher->is_short_branch_offset(offset), "Displacement too large for short jmp"); 529 } 530 } 531#endif 532 533 // ------------------ 534 // Compute size for code buffer 535 code_size = blk_starts[i-1] + jmp_end[i-1]; 536 537 // Relocation records 538 reloc_size += 1; // Relo entry for exception handler 539 540 // Adjust reloc_size to number of record of relocation info 541 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 542 // a relocation index. 543 // The CodeBuffer will expand the locs array if this estimate is too low. 544 reloc_size *= 10 / sizeof(relocInfo); 545 546 // Adjust const_size to number of bytes 547 const_size *= 2*jintSize; // both float and double take two words per entry 548 549} 550 551//------------------------------FillLocArray----------------------------------- 552// Create a bit of debug info and append it to the array. The mapping is from 553// Java local or expression stack to constant, register or stack-slot. For 554// doubles, insert 2 mappings and return 1 (to tell the caller that the next 555// entry has been taken care of and caller should skip it). 556static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 557 // This should never have accepted Bad before 558 assert(OptoReg::is_valid(regnum), "location must be valid"); 559 return (OptoReg::is_reg(regnum)) 560 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 561 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 562} 563 564void Compile::FillLocArray( int idx, Node *local, GrowableArray<ScopeValue*> *array ) { 565 assert( local, "use _top instead of null" ); 566 if (array->length() != idx) { 567 assert(array->length() == idx + 1, "Unexpected array count"); 568 // Old functionality: 569 // return 570 // New functionality: 571 // Assert if the local is not top. In product mode let the new node 572 // override the old entry. 573 assert(local == top(), "LocArray collision"); 574 if (local == top()) { 575 return; 576 } 577 array->pop(); 578 } 579 const Type *t = local->bottom_type(); 580 581 // Grab the register number for the local 582 OptoReg::Name regnum = _regalloc->get_reg_first(local); 583 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 584 // Record the double as two float registers. 585 // The register mask for such a value always specifies two adjacent 586 // float registers, with the lower register number even. 587 // Normally, the allocation of high and low words to these registers 588 // is irrelevant, because nearly all operations on register pairs 589 // (e.g., StoreD) treat them as a single unit. 590 // Here, we assume in addition that the words in these two registers 591 // stored "naturally" (by operations like StoreD and double stores 592 // within the interpreter) such that the lower-numbered register 593 // is written to the lower memory address. This may seem like 594 // a machine dependency, but it is not--it is a requirement on 595 // the author of the <arch>.ad file to ensure that, for every 596 // even/odd double-register pair to which a double may be allocated, 597 // the word in the even single-register is stored to the first 598 // memory word. (Note that register numbers are completely 599 // arbitrary, and are not tied to any machine-level encodings.) 600#ifdef _LP64 601 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 602 array->append(new ConstantIntValue(0)); 603 array->append(new_loc_value( _regalloc, regnum, Location::dbl )); 604 } else if ( t->base() == Type::Long ) { 605 array->append(new ConstantIntValue(0)); 606 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 607 } else if ( t->base() == Type::RawPtr ) { 608 // jsr/ret return address which must be restored into a the full 609 // width 64-bit stack slot. 610 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 611 } 612#else //_LP64 613#ifdef SPARC 614 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) { 615 // For SPARC we have to swap high and low words for 616 // long values stored in a single-register (g0-g7). 617 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 618 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 619 } else 620#endif //SPARC 621 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 622 // Repack the double/long as two jints. 623 // The convention the interpreter uses is that the second local 624 // holds the first raw word of the native double representation. 625 // This is actually reasonable, since locals and stack arrays 626 // grow downwards in all implementations. 627 // (If, on some machine, the interpreter's Java locals or stack 628 // were to grow upwards, the embedded doubles would be word-swapped.) 629 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 630 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 631 } 632#endif //_LP64 633 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 634 OptoReg::is_reg(regnum) ) { 635 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double 636 ? Location::float_in_dbl : Location::normal )); 637 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 638 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long 639 ? Location::int_in_long : Location::normal )); 640 } else { 641 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal )); 642 } 643 return; 644 } 645 646 // No register. It must be constant data. 647 switch (t->base()) { 648 case Type::Half: // Second half of a double 649 ShouldNotReachHere(); // Caller should skip 2nd halves 650 break; 651 case Type::AnyPtr: 652 array->append(new ConstantOopWriteValue(NULL)); 653 break; 654 case Type::AryPtr: 655 case Type::InstPtr: 656 case Type::KlassPtr: // fall through 657 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->encoding())); 658 break; 659 case Type::Int: 660 array->append(new ConstantIntValue(t->is_int()->get_con())); 661 break; 662 case Type::RawPtr: 663 // A return address (T_ADDRESS). 664 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 665#ifdef _LP64 666 // Must be restored to the full-width 64-bit stack slot. 667 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 668#else 669 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 670#endif 671 break; 672 case Type::FloatCon: { 673 float f = t->is_float_constant()->getf(); 674 array->append(new ConstantIntValue(jint_cast(f))); 675 break; 676 } 677 case Type::DoubleCon: { 678 jdouble d = t->is_double_constant()->getd(); 679#ifdef _LP64 680 array->append(new ConstantIntValue(0)); 681 array->append(new ConstantDoubleValue(d)); 682#else 683 // Repack the double as two jints. 684 // The convention the interpreter uses is that the second local 685 // holds the first raw word of the native double representation. 686 // This is actually reasonable, since locals and stack arrays 687 // grow downwards in all implementations. 688 // (If, on some machine, the interpreter's Java locals or stack 689 // were to grow upwards, the embedded doubles would be word-swapped.) 690 jint *dp = (jint*)&d; 691 array->append(new ConstantIntValue(dp[1])); 692 array->append(new ConstantIntValue(dp[0])); 693#endif 694 break; 695 } 696 case Type::Long: { 697 jlong d = t->is_long()->get_con(); 698#ifdef _LP64 699 array->append(new ConstantIntValue(0)); 700 array->append(new ConstantLongValue(d)); 701#else 702 // Repack the long as two jints. 703 // The convention the interpreter uses is that the second local 704 // holds the first raw word of the native double representation. 705 // This is actually reasonable, since locals and stack arrays 706 // grow downwards in all implementations. 707 // (If, on some machine, the interpreter's Java locals or stack 708 // were to grow upwards, the embedded doubles would be word-swapped.) 709 jint *dp = (jint*)&d; 710 array->append(new ConstantIntValue(dp[1])); 711 array->append(new ConstantIntValue(dp[0])); 712#endif 713 break; 714 } 715 case Type::Top: // Add an illegal value here 716 array->append(new LocationValue(Location())); 717 break; 718 default: 719 ShouldNotReachHere(); 720 break; 721 } 722} 723 724// Determine if this node starts a bundle 725bool Compile::starts_bundle(const Node *n) const { 726 return (_node_bundling_limit > n->_idx && 727 _node_bundling_base[n->_idx].starts_bundle()); 728} 729 730//--------------------------Process_OopMap_Node-------------------------------- 731void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) { 732 733 // Handle special safepoint nodes for synchronization 734 MachSafePointNode *sfn = mach->as_MachSafePoint(); 735 MachCallNode *mcall; 736 737#ifdef ENABLE_ZAP_DEAD_LOCALS 738 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative"); 739#endif 740 741 int safepoint_pc_offset = current_offset; 742 743 // Add the safepoint in the DebugInfoRecorder 744 if( !mach->is_MachCall() ) { 745 mcall = NULL; 746 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 747 } else { 748 mcall = mach->as_MachCall(); 749 safepoint_pc_offset += mcall->ret_addr_offset(); 750 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 751 } 752 753 // Loop over the JVMState list to add scope information 754 // Do not skip safepoints with a NULL method, they need monitor info 755 JVMState* youngest_jvms = sfn->jvms(); 756 int max_depth = youngest_jvms->depth(); 757 758 // Visit scopes from oldest to youngest. 759 for (int depth = 1; depth <= max_depth; depth++) { 760 JVMState* jvms = youngest_jvms->of_depth(depth); 761 int idx; 762 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 763 // Safepoints that do not have method() set only provide oop-map and monitor info 764 // to support GC; these do not support deoptimization. 765 int num_locs = (method == NULL) ? 0 : jvms->loc_size(); 766 int num_exps = (method == NULL) ? 0 : jvms->stk_size(); 767 int num_mon = jvms->nof_monitors(); 768 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), 769 "JVMS local count must match that of the method"); 770 771 // Add Local and Expression Stack Information 772 773 // Insert locals into the locarray 774 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 775 for( idx = 0; idx < num_locs; idx++ ) { 776 FillLocArray( idx, sfn->local(jvms, idx), locarray ); 777 } 778 779 // Insert expression stack entries into the exparray 780 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 781 for( idx = 0; idx < num_exps; idx++ ) { 782 FillLocArray( idx, sfn->stack(jvms, idx), exparray ); 783 } 784 785 // Add in mappings of the monitors 786 assert( !method || 787 !method->is_synchronized() || 788 method->is_native() || 789 num_mon > 0 || 790 !GenerateSynchronizationCode, 791 "monitors must always exist for synchronized methods"); 792 793 // Build the growable array of ScopeValues for exp stack 794 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 795 796 // Loop over monitors and insert into array 797 for(idx = 0; idx < num_mon; idx++) { 798 // Grab the node that defines this monitor 799 Node* box_node; 800 Node* obj_node; 801 box_node = sfn->monitor_box(jvms, idx); 802 obj_node = sfn->monitor_obj(jvms, idx); 803 804 // Create ScopeValue for object 805 ScopeValue *scval = NULL; 806 if( !obj_node->is_Con() ) { 807 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); 808 scval = new_loc_value( _regalloc, obj_reg, Location::oop ); 809 } else { 810 scval = new ConstantOopWriteValue(obj_node->bottom_type()->is_instptr()->const_oop()->encoding()); 811 } 812 813 OptoReg::Name box_reg = BoxLockNode::stack_slot(box_node); 814 monarray->append(new MonitorValue(scval, Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)))); 815 } 816 817 // Build first class objects to pass to scope 818 DebugToken *locvals = debug_info()->create_scope_values(locarray); 819 DebugToken *expvals = debug_info()->create_scope_values(exparray); 820 DebugToken *monvals = debug_info()->create_monitor_values(monarray); 821 822 // Make method available for all Safepoints 823 ciMethod* scope_method = method ? method : _method; 824 // Describe the scope here 825 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 826 debug_info()->describe_scope(safepoint_pc_offset,scope_method,jvms->bci(),locvals,expvals,monvals); 827 } // End jvms loop 828 829 // Mark the end of the scope set. 830 debug_info()->end_safepoint(safepoint_pc_offset); 831} 832 833 834 835// A simplified version of Process_OopMap_Node, to handle non-safepoints. 836class NonSafepointEmitter { 837 Compile* C; 838 JVMState* _pending_jvms; 839 int _pending_offset; 840 841 void emit_non_safepoint(); 842 843 public: 844 NonSafepointEmitter(Compile* compile) { 845 this->C = compile; 846 _pending_jvms = NULL; 847 _pending_offset = 0; 848 } 849 850 void observe_instruction(Node* n, int pc_offset) { 851 if (!C->debug_info()->recording_non_safepoints()) return; 852 853 Node_Notes* nn = C->node_notes_at(n->_idx); 854 if (nn == NULL || nn->jvms() == NULL) return; 855 if (_pending_jvms != NULL && 856 _pending_jvms->same_calls_as(nn->jvms())) { 857 // Repeated JVMS? Stretch it up here. 858 _pending_offset = pc_offset; 859 } else { 860 if (_pending_jvms != NULL && 861 _pending_offset < pc_offset) { 862 emit_non_safepoint(); 863 } 864 _pending_jvms = NULL; 865 if (pc_offset > C->debug_info()->last_pc_offset()) { 866 // This is the only way _pending_jvms can become non-NULL: 867 _pending_jvms = nn->jvms(); 868 _pending_offset = pc_offset; 869 } 870 } 871 } 872 873 // Stay out of the way of real safepoints: 874 void observe_safepoint(JVMState* jvms, int pc_offset) { 875 if (_pending_jvms != NULL && 876 !_pending_jvms->same_calls_as(jvms) && 877 _pending_offset < pc_offset) { 878 emit_non_safepoint(); 879 } 880 _pending_jvms = NULL; 881 } 882 883 void flush_at_end() { 884 if (_pending_jvms != NULL) { 885 emit_non_safepoint(); 886 } 887 _pending_jvms = NULL; 888 } 889}; 890 891void NonSafepointEmitter::emit_non_safepoint() { 892 JVMState* youngest_jvms = _pending_jvms; 893 int pc_offset = _pending_offset; 894 895 // Clear it now: 896 _pending_jvms = NULL; 897 898 DebugInformationRecorder* debug_info = C->debug_info(); 899 assert(debug_info->recording_non_safepoints(), "sanity"); 900 901 debug_info->add_non_safepoint(pc_offset); 902 int max_depth = youngest_jvms->depth(); 903 904 // Visit scopes from oldest to youngest. 905 for (int depth = 1; depth <= max_depth; depth++) { 906 JVMState* jvms = youngest_jvms->of_depth(depth); 907 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 908 debug_info->describe_scope(pc_offset, method, jvms->bci()); 909 } 910 911 // Mark the end of the scope set. 912 debug_info->end_non_safepoint(pc_offset); 913} 914 915 916 917// helper for Fill_buffer bailout logic 918static void turn_off_compiler(Compile* C) { 919 if (CodeCache::unallocated_capacity() >= CodeCacheMinimumFreeSpace*10) { 920 // Do not turn off compilation if a single giant method has 921 // blown the code cache size. 922 C->record_failure("excessive request to CodeCache"); 923 } else { 924 UseInterpreter = true; 925 UseCompiler = false; 926 AlwaysCompileLoopMethods = false; 927 C->record_failure("CodeCache is full"); 928 warning("CodeCache is full. Compiling has been disabled"); 929 } 930} 931 932 933//------------------------------Fill_buffer------------------------------------ 934void Compile::Fill_buffer() { 935 936 // Set the initially allocated size 937 int code_req = initial_code_capacity; 938 int locs_req = initial_locs_capacity; 939 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity; 940 int const_req = initial_const_capacity; 941 bool labels_not_set = true; 942 943 int pad_req = NativeCall::instruction_size; 944 // The extra spacing after the code is necessary on some platforms. 945 // Sometimes we need to patch in a jump after the last instruction, 946 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 947 948 uint i; 949 // Compute the byte offset where we can store the deopt pc. 950 if (fixed_slots() != 0) { 951 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 952 } 953 954 // Compute prolog code size 955 _method_size = 0; 956 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; 957#ifdef IA64 958 if (save_argument_registers()) { 959 // 4815101: this is a stub with implicit and unknown precision fp args. 960 // The usual spill mechanism can only generate stfd's in this case, which 961 // doesn't work if the fp reg to spill contains a single-precision denorm. 962 // Instead, we hack around the normal spill mechanism using stfspill's and 963 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate 964 // space here for the fp arg regs (f8-f15) we're going to thusly spill. 965 // 966 // If we ever implement 16-byte 'registers' == stack slots, we can 967 // get rid of this hack and have SpillCopy generate stfspill/ldffill 968 // instead of stfd/stfs/ldfd/ldfs. 969 _frame_slots += 8*(16/BytesPerInt); 970 } 971#endif 972 assert( _frame_slots >= 0 && _frame_slots < 1000000, "sanity check" ); 973 974 // Create an array of unused labels, one for each basic block 975 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, _cfg->_num_blocks+1); 976 977 for( i=0; i <= _cfg->_num_blocks; i++ ) { 978 blk_labels[i].init(); 979 } 980 981 // If this machine supports different size branch offsets, then pre-compute 982 // the length of the blocks 983 if( _matcher->is_short_branch_offset(0) ) { 984 Shorten_branches(blk_labels, code_req, locs_req, stub_req, const_req); 985 labels_not_set = false; 986 } 987 988 // nmethod and CodeBuffer count stubs & constants as part of method's code. 989 int exception_handler_req = size_exception_handler(); 990 int deopt_handler_req = size_deopt_handler(); 991 exception_handler_req += MAX_stubs_size; // add marginal slop for handler 992 deopt_handler_req += MAX_stubs_size; // add marginal slop for handler 993 stub_req += MAX_stubs_size; // ensure per-stub margin 994 code_req += MAX_inst_size; // ensure per-instruction margin 995 if (StressCodeBuffers) 996 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 997 int total_req = code_req + pad_req + stub_req + exception_handler_req + deopt_handler_req + const_req; 998 CodeBuffer* cb = code_buffer(); 999 cb->initialize(total_req, locs_req); 1000 1001 // Have we run out of code space? 1002 if (cb->blob() == NULL) { 1003 turn_off_compiler(this); 1004 return; 1005 } 1006 // Configure the code buffer. 1007 cb->initialize_consts_size(const_req); 1008 cb->initialize_stubs_size(stub_req); 1009 cb->initialize_oop_recorder(env()->oop_recorder()); 1010 1011 // fill in the nop array for bundling computations 1012 MachNode *_nop_list[Bundle::_nop_count]; 1013 Bundle::initialize_nops(_nop_list, this); 1014 1015 // Create oopmap set. 1016 _oop_map_set = new OopMapSet(); 1017 1018 // !!!!! This preserves old handling of oopmaps for now 1019 debug_info()->set_oopmaps(_oop_map_set); 1020 1021 // Count and start of implicit null check instructions 1022 uint inct_cnt = 0; 1023 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, _cfg->_num_blocks+1); 1024 1025 // Count and start of calls 1026 uint *call_returns = NEW_RESOURCE_ARRAY(uint, _cfg->_num_blocks+1); 1027 1028 uint return_offset = 0; 1029 MachNode *nop = new (this) MachNopNode(); 1030 1031 int previous_offset = 0; 1032 int current_offset = 0; 1033 int last_call_offset = -1; 1034 1035 // Create an array of unused labels, one for each basic block, if printing is enabled 1036#ifndef PRODUCT 1037 int *node_offsets = NULL; 1038 uint node_offset_limit = unique(); 1039 1040 1041 if ( print_assembly() ) 1042 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1043#endif 1044 1045 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily 1046 1047 // ------------------ 1048 // Now fill in the code buffer 1049 Node *delay_slot = NULL; 1050 1051 for( i=0; i < _cfg->_num_blocks; i++ ) { 1052 Block *b = _cfg->_blocks[i]; 1053 1054 Node *head = b->head(); 1055 1056 // If this block needs to start aligned (i.e, can be reached other 1057 // than by falling-thru from the previous block), then force the 1058 // start of a new bundle. 1059 if( Pipeline::requires_bundling() && starts_bundle(head) ) 1060 cb->flush_bundle(true); 1061 1062 // Define the label at the beginning of the basic block 1063 if( labels_not_set ) 1064 MacroAssembler(cb).bind( blk_labels[b->_pre_order] ); 1065 1066 else 1067 assert( blk_labels[b->_pre_order].loc_pos() == cb->code_size(), 1068 "label position does not match code offset" ); 1069 1070 uint last_inst = b->_nodes.size(); 1071 1072 // Emit block normally, except for last instruction. 1073 // Emit means "dump code bits into code buffer". 1074 for( uint j = 0; j<last_inst; j++ ) { 1075 1076 // Get the node 1077 Node* n = b->_nodes[j]; 1078 1079 // See if delay slots are supported 1080 if (valid_bundle_info(n) && 1081 node_bundling(n)->used_in_unconditional_delay()) { 1082 assert(delay_slot == NULL, "no use of delay slot node"); 1083 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1084 1085 delay_slot = n; 1086 continue; 1087 } 1088 1089 // If this starts a new instruction group, then flush the current one 1090 // (but allow split bundles) 1091 if( Pipeline::requires_bundling() && starts_bundle(n) ) 1092 cb->flush_bundle(false); 1093 1094 // The following logic is duplicated in the code ifdeffed for 1095 // ENABLE_ZAP_DEAD_LOCALS which apppears above in this file. It 1096 // should be factored out. Or maybe dispersed to the nodes? 1097 1098 // Special handling for SafePoint/Call Nodes 1099 bool is_mcall = false; 1100 if( n->is_Mach() ) { 1101 MachNode *mach = n->as_Mach(); 1102 is_mcall = n->is_MachCall(); 1103 bool is_sfn = n->is_MachSafePoint(); 1104 1105 // If this requires all previous instructions be flushed, then do so 1106 if( is_sfn || is_mcall || mach->alignment_required() != 1) { 1107 cb->flush_bundle(true); 1108 current_offset = cb->code_size(); 1109 } 1110 1111 // align the instruction if necessary 1112 int nop_size = nop->size(_regalloc); 1113 int padding = mach->compute_padding(current_offset); 1114 // Make sure safepoint node for polling is distinct from a call's 1115 // return by adding a nop if needed. 1116 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset ) { 1117 padding = nop_size; 1118 } 1119 assert( labels_not_set || padding == 0, "instruction should already be aligned") 1120 1121 if(padding > 0) { 1122 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1123 int nops_cnt = padding / nop_size; 1124 MachNode *nop = new (this) MachNopNode(nops_cnt); 1125 b->_nodes.insert(j++, nop); 1126 last_inst++; 1127 _cfg->_bbs.map( nop->_idx, b ); 1128 nop->emit(*cb, _regalloc); 1129 cb->flush_bundle(true); 1130 current_offset = cb->code_size(); 1131 } 1132 1133 // Remember the start of the last call in a basic block 1134 if (is_mcall) { 1135 MachCallNode *mcall = mach->as_MachCall(); 1136 1137 // This destination address is NOT PC-relative 1138 mcall->method_set((intptr_t)mcall->entry_point()); 1139 1140 // Save the return address 1141 call_returns[b->_pre_order] = current_offset + mcall->ret_addr_offset(); 1142 1143 if (!mcall->is_safepoint_node()) { 1144 is_mcall = false; 1145 is_sfn = false; 1146 } 1147 } 1148 1149 // sfn will be valid whenever mcall is valid now because of inheritance 1150 if( is_sfn || is_mcall ) { 1151 1152 // Handle special safepoint nodes for synchronization 1153 if( !is_mcall ) { 1154 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1155 // !!!!! Stubs only need an oopmap right now, so bail out 1156 if( sfn->jvms()->method() == NULL) { 1157 // Write the oopmap directly to the code blob??!! 1158# ifdef ENABLE_ZAP_DEAD_LOCALS 1159 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive"); 1160# endif 1161 continue; 1162 } 1163 } // End synchronization 1164 1165 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1166 current_offset); 1167 Process_OopMap_Node(mach, current_offset); 1168 } // End if safepoint 1169 1170 // If this is a null check, then add the start of the previous instruction to the list 1171 else if( mach->is_MachNullCheck() ) { 1172 inct_starts[inct_cnt++] = previous_offset; 1173 } 1174 1175 // If this is a branch, then fill in the label with the target BB's label 1176 else if ( mach->is_Branch() ) { 1177 1178 if ( mach->ideal_Opcode() == Op_Jump ) { 1179 for (uint h = 0; h < b->_num_succs; h++ ) { 1180 Block* succs_block = b->_succs[h]; 1181 for (uint j = 1; j < succs_block->num_preds(); j++) { 1182 Node* jpn = succs_block->pred(j); 1183 if ( jpn->is_JumpProj() && jpn->in(0) == mach ) { 1184 uint block_num = succs_block->non_connector()->_pre_order; 1185 Label *blkLabel = &blk_labels[block_num]; 1186 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1187 } 1188 } 1189 } 1190 } else { 1191 // For Branchs 1192 // This requires the TRUE branch target be in succs[0] 1193 uint block_num = b->non_connector_successor(0)->_pre_order; 1194 mach->label_set( blk_labels[block_num], block_num ); 1195 } 1196 } 1197 1198#ifdef ASSERT 1199 // Check that oop-store preceeds the card-mark 1200 else if( mach->ideal_Opcode() == Op_StoreCM ) { 1201 uint storeCM_idx = j; 1202 Node *oop_store = mach->in(mach->_cnt); // First precedence edge 1203 assert( oop_store != NULL, "storeCM expects a precedence edge"); 1204 uint i4; 1205 for( i4 = 0; i4 < last_inst; ++i4 ) { 1206 if( b->_nodes[i4] == oop_store ) break; 1207 } 1208 // Note: This test can provide a false failure if other precedence 1209 // edges have been added to the storeCMNode. 1210 assert( i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1211 } 1212#endif 1213 1214 else if( !n->is_Proj() ) { 1215 // Remember the begining of the previous instruction, in case 1216 // it's followed by a flag-kill and a null-check. Happens on 1217 // Intel all the time, with add-to-memory kind of opcodes. 1218 previous_offset = current_offset; 1219 } 1220 } 1221 1222 // Verify that there is sufficient space remaining 1223 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1224 if (cb->blob() == NULL) { 1225 turn_off_compiler(this); 1226 return; 1227 } 1228 1229 // Save the offset for the listing 1230#ifndef PRODUCT 1231 if( node_offsets && n->_idx < node_offset_limit ) 1232 node_offsets[n->_idx] = cb->code_size(); 1233#endif 1234 1235 // "Normal" instruction case 1236 n->emit(*cb, _regalloc); 1237 current_offset = cb->code_size(); 1238 non_safepoints.observe_instruction(n, current_offset); 1239 1240 // mcall is last "call" that can be a safepoint 1241 // record it so we can see if a poll will directly follow it 1242 // in which case we'll need a pad to make the PcDesc sites unique 1243 // see 5010568. This can be slightly inaccurate but conservative 1244 // in the case that return address is not actually at current_offset. 1245 // This is a small price to pay. 1246 1247 if (is_mcall) { 1248 last_call_offset = current_offset; 1249 } 1250 1251 // See if this instruction has a delay slot 1252 if ( valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1253 assert(delay_slot != NULL, "expecting delay slot node"); 1254 1255 // Back up 1 instruction 1256 cb->set_code_end( 1257 cb->code_end()-Pipeline::instr_unit_size()); 1258 1259 // Save the offset for the listing 1260#ifndef PRODUCT 1261 if( node_offsets && delay_slot->_idx < node_offset_limit ) 1262 node_offsets[delay_slot->_idx] = cb->code_size(); 1263#endif 1264 1265 // Support a SafePoint in the delay slot 1266 if( delay_slot->is_MachSafePoint() ) { 1267 MachNode *mach = delay_slot->as_Mach(); 1268 // !!!!! Stubs only need an oopmap right now, so bail out 1269 if( !mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL ) { 1270 // Write the oopmap directly to the code blob??!! 1271# ifdef ENABLE_ZAP_DEAD_LOCALS 1272 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive"); 1273# endif 1274 delay_slot = NULL; 1275 continue; 1276 } 1277 1278 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1279 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1280 adjusted_offset); 1281 // Generate an OopMap entry 1282 Process_OopMap_Node(mach, adjusted_offset); 1283 } 1284 1285 // Insert the delay slot instruction 1286 delay_slot->emit(*cb, _regalloc); 1287 1288 // Don't reuse it 1289 delay_slot = NULL; 1290 } 1291 1292 } // End for all instructions in block 1293 1294 // If the next block _starts_ a loop, pad this block out to align 1295 // the loop start a little. Helps prevent pipe stalls at loop starts 1296 int nop_size = (new (this) MachNopNode())->size(_regalloc); 1297 if( i<_cfg->_num_blocks-1 ) { 1298 Block *nb = _cfg->_blocks[i+1]; 1299 uint padding = nb->alignment_padding(current_offset); 1300 if( padding > 0 ) { 1301 MachNode *nop = new (this) MachNopNode(padding / nop_size); 1302 b->_nodes.insert( b->_nodes.size(), nop ); 1303 _cfg->_bbs.map( nop->_idx, b ); 1304 nop->emit(*cb, _regalloc); 1305 current_offset = cb->code_size(); 1306 } 1307 } 1308 1309 } // End of for all blocks 1310 1311 non_safepoints.flush_at_end(); 1312 1313 // Offset too large? 1314 if (failing()) return; 1315 1316 // Define a pseudo-label at the end of the code 1317 MacroAssembler(cb).bind( blk_labels[_cfg->_num_blocks] ); 1318 1319 // Compute the size of the first block 1320 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1321 1322 assert(cb->code_size() < 500000, "method is unreasonably large"); 1323 1324 // ------------------ 1325 1326#ifndef PRODUCT 1327 // Information on the size of the method, without the extraneous code 1328 Scheduling::increment_method_size(cb->code_size()); 1329#endif 1330 1331 // ------------------ 1332 // Fill in exception table entries. 1333 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1334 1335 // Only java methods have exception handlers and deopt handlers 1336 if (_method) { 1337 // Emit the exception handler code. 1338 _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb)); 1339 // Emit the deopt handler code. 1340 _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb)); 1341 } 1342 1343 // One last check for failed CodeBuffer::expand: 1344 if (cb->blob() == NULL) { 1345 turn_off_compiler(this); 1346 return; 1347 } 1348 1349#ifndef PRODUCT 1350 // Dump the assembly code, including basic-block numbers 1351 if (print_assembly()) { 1352 ttyLocker ttyl; // keep the following output all in one block 1353 if (!VMThread::should_terminate()) { // test this under the tty lock 1354 // This output goes directly to the tty, not the compiler log. 1355 // To enable tools to match it up with the compilation activity, 1356 // be sure to tag this tty output with the compile ID. 1357 if (xtty != NULL) { 1358 xtty->head("opto_assembly compile_id='%d'%s", compile_id(), 1359 is_osr_compilation() ? " compile_kind='osr'" : 1360 ""); 1361 } 1362 if (method() != NULL) { 1363 method()->print_oop(); 1364 print_codes(); 1365 } 1366 dump_asm(node_offsets, node_offset_limit); 1367 if (xtty != NULL) { 1368 xtty->tail("opto_assembly"); 1369 } 1370 } 1371 } 1372#endif 1373 1374} 1375 1376void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1377 _inc_table.set_size(cnt); 1378 1379 uint inct_cnt = 0; 1380 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1381 Block *b = _cfg->_blocks[i]; 1382 Node *n = NULL; 1383 int j; 1384 1385 // Find the branch; ignore trailing NOPs. 1386 for( j = b->_nodes.size()-1; j>=0; j-- ) { 1387 n = b->_nodes[j]; 1388 if( !n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con ) 1389 break; 1390 } 1391 1392 // If we didn't find anything, continue 1393 if( j < 0 ) continue; 1394 1395 // Compute ExceptionHandlerTable subtable entry and add it 1396 // (skip empty blocks) 1397 if( n->is_Catch() ) { 1398 1399 // Get the offset of the return from the call 1400 uint call_return = call_returns[b->_pre_order]; 1401#ifdef ASSERT 1402 assert( call_return > 0, "no call seen for this basic block" ); 1403 while( b->_nodes[--j]->Opcode() == Op_MachProj ) ; 1404 assert( b->_nodes[j]->is_Call(), "CatchProj must follow call" ); 1405#endif 1406 // last instruction is a CatchNode, find it's CatchProjNodes 1407 int nof_succs = b->_num_succs; 1408 // allocate space 1409 GrowableArray<intptr_t> handler_bcis(nof_succs); 1410 GrowableArray<intptr_t> handler_pcos(nof_succs); 1411 // iterate through all successors 1412 for (int j = 0; j < nof_succs; j++) { 1413 Block* s = b->_succs[j]; 1414 bool found_p = false; 1415 for( uint k = 1; k < s->num_preds(); k++ ) { 1416 Node *pk = s->pred(k); 1417 if( pk->is_CatchProj() && pk->in(0) == n ) { 1418 const CatchProjNode* p = pk->as_CatchProj(); 1419 found_p = true; 1420 // add the corresponding handler bci & pco information 1421 if( p->_con != CatchProjNode::fall_through_index ) { 1422 // p leads to an exception handler (and is not fall through) 1423 assert(s == _cfg->_blocks[s->_pre_order],"bad numbering"); 1424 // no duplicates, please 1425 if( !handler_bcis.contains(p->handler_bci()) ) { 1426 uint block_num = s->non_connector()->_pre_order; 1427 handler_bcis.append(p->handler_bci()); 1428 handler_pcos.append(blk_labels[block_num].loc_pos()); 1429 } 1430 } 1431 } 1432 } 1433 assert(found_p, "no matching predecessor found"); 1434 // Note: Due to empty block removal, one block may have 1435 // several CatchProj inputs, from the same Catch. 1436 } 1437 1438 // Set the offset of the return from the call 1439 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); 1440 continue; 1441 } 1442 1443 // Handle implicit null exception table updates 1444 if( n->is_MachNullCheck() ) { 1445 uint block_num = b->non_connector_successor(0)->_pre_order; 1446 _inc_table.append( inct_starts[inct_cnt++], blk_labels[block_num].loc_pos() ); 1447 continue; 1448 } 1449 } // End of for all blocks fill in exception table entries 1450} 1451 1452// Static Variables 1453#ifndef PRODUCT 1454uint Scheduling::_total_nop_size = 0; 1455uint Scheduling::_total_method_size = 0; 1456uint Scheduling::_total_branches = 0; 1457uint Scheduling::_total_unconditional_delays = 0; 1458uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1459#endif 1460 1461// Initializer for class Scheduling 1462 1463Scheduling::Scheduling(Arena *arena, Compile &compile) 1464 : _arena(arena), 1465 _cfg(compile.cfg()), 1466 _bbs(compile.cfg()->_bbs), 1467 _regalloc(compile.regalloc()), 1468 _reg_node(arena), 1469 _bundle_instr_count(0), 1470 _bundle_cycle_number(0), 1471 _scheduled(arena), 1472 _available(arena), 1473 _next_node(NULL), 1474 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]), 1475 _pinch_free_list(arena) 1476#ifndef PRODUCT 1477 , _branches(0) 1478 , _unconditional_delays(0) 1479#endif 1480{ 1481 // Create a MachNopNode 1482 _nop = new (&compile) MachNopNode(); 1483 1484 // Now that the nops are in the array, save the count 1485 // (but allow entries for the nops) 1486 _node_bundling_limit = compile.unique(); 1487 uint node_max = _regalloc->node_regs_max_index(); 1488 1489 compile.set_node_bundling_limit(_node_bundling_limit); 1490 1491 // This one is persistant within the Compile class 1492 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1493 1494 // Allocate space for fixed-size arrays 1495 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1496 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1497 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1498 1499 // Clear the arrays 1500 memset(_node_bundling_base, 0, node_max * sizeof(Bundle)); 1501 memset(_node_latency, 0, node_max * sizeof(unsigned short)); 1502 memset(_uses, 0, node_max * sizeof(short)); 1503 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 1504 1505 // Clear the bundling information 1506 memcpy(_bundle_use_elements, 1507 Pipeline_Use::elaborated_elements, 1508 sizeof(Pipeline_Use::elaborated_elements)); 1509 1510 // Get the last node 1511 Block *bb = _cfg->_blocks[_cfg->_blocks.size()-1]; 1512 1513 _next_node = bb->_nodes[bb->_nodes.size()-1]; 1514} 1515 1516#ifndef PRODUCT 1517// Scheduling destructor 1518Scheduling::~Scheduling() { 1519 _total_branches += _branches; 1520 _total_unconditional_delays += _unconditional_delays; 1521} 1522#endif 1523 1524// Step ahead "i" cycles 1525void Scheduling::step(uint i) { 1526 1527 Bundle *bundle = node_bundling(_next_node); 1528 bundle->set_starts_bundle(); 1529 1530 // Update the bundle record, but leave the flags information alone 1531 if (_bundle_instr_count > 0) { 1532 bundle->set_instr_count(_bundle_instr_count); 1533 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1534 } 1535 1536 // Update the state information 1537 _bundle_instr_count = 0; 1538 _bundle_cycle_number += i; 1539 _bundle_use.step(i); 1540} 1541 1542void Scheduling::step_and_clear() { 1543 Bundle *bundle = node_bundling(_next_node); 1544 bundle->set_starts_bundle(); 1545 1546 // Update the bundle record 1547 if (_bundle_instr_count > 0) { 1548 bundle->set_instr_count(_bundle_instr_count); 1549 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1550 1551 _bundle_cycle_number += 1; 1552 } 1553 1554 // Clear the bundling information 1555 _bundle_instr_count = 0; 1556 _bundle_use.reset(); 1557 1558 memcpy(_bundle_use_elements, 1559 Pipeline_Use::elaborated_elements, 1560 sizeof(Pipeline_Use::elaborated_elements)); 1561} 1562 1563//------------------------------ScheduleAndBundle------------------------------ 1564// Perform instruction scheduling and bundling over the sequence of 1565// instructions in backwards order. 1566void Compile::ScheduleAndBundle() { 1567 1568 // Don't optimize this if it isn't a method 1569 if (!_method) 1570 return; 1571 1572 // Don't optimize this if scheduling is disabled 1573 if (!do_scheduling()) 1574 return; 1575 1576 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); ) 1577 1578 // Create a data structure for all the scheduling information 1579 Scheduling scheduling(Thread::current()->resource_area(), *this); 1580 1581 // Walk backwards over each basic block, computing the needed alignment 1582 // Walk over all the basic blocks 1583 scheduling.DoScheduling(); 1584} 1585 1586//------------------------------ComputeLocalLatenciesForward------------------- 1587// Compute the latency of all the instructions. This is fairly simple, 1588// because we already have a legal ordering. Walk over the instructions 1589// from first to last, and compute the latency of the instruction based 1590// on the latency of the preceeding instruction(s). 1591void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { 1592#ifndef PRODUCT 1593 if (_cfg->C->trace_opto_output()) 1594 tty->print("# -> ComputeLocalLatenciesForward\n"); 1595#endif 1596 1597 // Walk over all the schedulable instructions 1598 for( uint j=_bb_start; j < _bb_end; j++ ) { 1599 1600 // This is a kludge, forcing all latency calculations to start at 1. 1601 // Used to allow latency 0 to force an instruction to the beginning 1602 // of the bb 1603 uint latency = 1; 1604 Node *use = bb->_nodes[j]; 1605 uint nlen = use->len(); 1606 1607 // Walk over all the inputs 1608 for ( uint k=0; k < nlen; k++ ) { 1609 Node *def = use->in(k); 1610 if (!def) 1611 continue; 1612 1613 uint l = _node_latency[def->_idx] + use->latency(k); 1614 if (latency < l) 1615 latency = l; 1616 } 1617 1618 _node_latency[use->_idx] = latency; 1619 1620#ifndef PRODUCT 1621 if (_cfg->C->trace_opto_output()) { 1622 tty->print("# latency %4d: ", latency); 1623 use->dump(); 1624 } 1625#endif 1626 } 1627 1628#ifndef PRODUCT 1629 if (_cfg->C->trace_opto_output()) 1630 tty->print("# <- ComputeLocalLatenciesForward\n"); 1631#endif 1632 1633} // end ComputeLocalLatenciesForward 1634 1635// See if this node fits into the present instruction bundle 1636bool Scheduling::NodeFitsInBundle(Node *n) { 1637 uint n_idx = n->_idx; 1638 1639 // If this is the unconditional delay instruction, then it fits 1640 if (n == _unconditional_delay_slot) { 1641#ifndef PRODUCT 1642 if (_cfg->C->trace_opto_output()) 1643 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 1644#endif 1645 return (true); 1646 } 1647 1648 // If the node cannot be scheduled this cycle, skip it 1649 if (_current_latency[n_idx] > _bundle_cycle_number) { 1650#ifndef PRODUCT 1651 if (_cfg->C->trace_opto_output()) 1652 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 1653 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 1654#endif 1655 return (false); 1656 } 1657 1658 const Pipeline *node_pipeline = n->pipeline(); 1659 1660 uint instruction_count = node_pipeline->instructionCount(); 1661 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1662 instruction_count = 0; 1663 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1664 instruction_count++; 1665 1666 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 1667#ifndef PRODUCT 1668 if (_cfg->C->trace_opto_output()) 1669 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 1670 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 1671#endif 1672 return (false); 1673 } 1674 1675 // Don't allow non-machine nodes to be handled this way 1676 if (!n->is_Mach() && instruction_count == 0) 1677 return (false); 1678 1679 // See if there is any overlap 1680 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 1681 1682 if (delay > 0) { 1683#ifndef PRODUCT 1684 if (_cfg->C->trace_opto_output()) 1685 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 1686#endif 1687 return false; 1688 } 1689 1690#ifndef PRODUCT 1691 if (_cfg->C->trace_opto_output()) 1692 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 1693#endif 1694 1695 return true; 1696} 1697 1698Node * Scheduling::ChooseNodeToBundle() { 1699 uint siz = _available.size(); 1700 1701 if (siz == 0) { 1702 1703#ifndef PRODUCT 1704 if (_cfg->C->trace_opto_output()) 1705 tty->print("# ChooseNodeToBundle: NULL\n"); 1706#endif 1707 return (NULL); 1708 } 1709 1710 // Fast path, if only 1 instruction in the bundle 1711 if (siz == 1) { 1712#ifndef PRODUCT 1713 if (_cfg->C->trace_opto_output()) { 1714 tty->print("# ChooseNodeToBundle (only 1): "); 1715 _available[0]->dump(); 1716 } 1717#endif 1718 return (_available[0]); 1719 } 1720 1721 // Don't bother, if the bundle is already full 1722 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 1723 for ( uint i = 0; i < siz; i++ ) { 1724 Node *n = _available[i]; 1725 1726 // Skip projections, we'll handle them another way 1727 if (n->is_Proj()) 1728 continue; 1729 1730 // This presupposed that instructions are inserted into the 1731 // available list in a legality order; i.e. instructions that 1732 // must be inserted first are at the head of the list 1733 if (NodeFitsInBundle(n)) { 1734#ifndef PRODUCT 1735 if (_cfg->C->trace_opto_output()) { 1736 tty->print("# ChooseNodeToBundle: "); 1737 n->dump(); 1738 } 1739#endif 1740 return (n); 1741 } 1742 } 1743 } 1744 1745 // Nothing fits in this bundle, choose the highest priority 1746#ifndef PRODUCT 1747 if (_cfg->C->trace_opto_output()) { 1748 tty->print("# ChooseNodeToBundle: "); 1749 _available[0]->dump(); 1750 } 1751#endif 1752 1753 return _available[0]; 1754} 1755 1756//------------------------------AddNodeToAvailableList------------------------- 1757void Scheduling::AddNodeToAvailableList(Node *n) { 1758 assert( !n->is_Proj(), "projections never directly made available" ); 1759#ifndef PRODUCT 1760 if (_cfg->C->trace_opto_output()) { 1761 tty->print("# AddNodeToAvailableList: "); 1762 n->dump(); 1763 } 1764#endif 1765 1766 int latency = _current_latency[n->_idx]; 1767 1768 // Insert in latency order (insertion sort) 1769 uint i; 1770 for ( i=0; i < _available.size(); i++ ) 1771 if (_current_latency[_available[i]->_idx] > latency) 1772 break; 1773 1774 // Special Check for compares following branches 1775 if( n->is_Mach() && _scheduled.size() > 0 ) { 1776 int op = n->as_Mach()->ideal_Opcode(); 1777 Node *last = _scheduled[0]; 1778 if( last->is_MachIf() && last->in(1) == n && 1779 ( op == Op_CmpI || 1780 op == Op_CmpU || 1781 op == Op_CmpP || 1782 op == Op_CmpF || 1783 op == Op_CmpD || 1784 op == Op_CmpL ) ) { 1785 1786 // Recalculate position, moving to front of same latency 1787 for ( i=0 ; i < _available.size(); i++ ) 1788 if (_current_latency[_available[i]->_idx] >= latency) 1789 break; 1790 } 1791 } 1792 1793 // Insert the node in the available list 1794 _available.insert(i, n); 1795 1796#ifndef PRODUCT 1797 if (_cfg->C->trace_opto_output()) 1798 dump_available(); 1799#endif 1800} 1801 1802//------------------------------DecrementUseCounts----------------------------- 1803void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 1804 for ( uint i=0; i < n->len(); i++ ) { 1805 Node *def = n->in(i); 1806 if (!def) continue; 1807 if( def->is_Proj() ) // If this is a machine projection, then 1808 def = def->in(0); // propagate usage thru to the base instruction 1809 1810 if( _bbs[def->_idx] != bb ) // Ignore if not block-local 1811 continue; 1812 1813 // Compute the latency 1814 uint l = _bundle_cycle_number + n->latency(i); 1815 if (_current_latency[def->_idx] < l) 1816 _current_latency[def->_idx] = l; 1817 1818 // If this does not have uses then schedule it 1819 if ((--_uses[def->_idx]) == 0) 1820 AddNodeToAvailableList(def); 1821 } 1822} 1823 1824//------------------------------AddNodeToBundle-------------------------------- 1825void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 1826#ifndef PRODUCT 1827 if (_cfg->C->trace_opto_output()) { 1828 tty->print("# AddNodeToBundle: "); 1829 n->dump(); 1830 } 1831#endif 1832 1833 // Remove this from the available list 1834 uint i; 1835 for (i = 0; i < _available.size(); i++) 1836 if (_available[i] == n) 1837 break; 1838 assert(i < _available.size(), "entry in _available list not found"); 1839 _available.remove(i); 1840 1841 // See if this fits in the current bundle 1842 const Pipeline *node_pipeline = n->pipeline(); 1843 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 1844 1845 // Check for instructions to be placed in the delay slot. We 1846 // do this before we actually schedule the current instruction, 1847 // because the delay slot follows the current instruction. 1848 if (Pipeline::_branch_has_delay_slot && 1849 node_pipeline->hasBranchDelay() && 1850 !_unconditional_delay_slot) { 1851 1852 uint siz = _available.size(); 1853 1854 // Conditional branches can support an instruction that 1855 // is unconditionally executed and not dependant by the 1856 // branch, OR a conditionally executed instruction if 1857 // the branch is taken. In practice, this means that 1858 // the first instruction at the branch target is 1859 // copied to the delay slot, and the branch goes to 1860 // the instruction after that at the branch target 1861 if ( n->is_Mach() && n->is_Branch() ) { 1862 1863 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 1864 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 1865 1866#ifndef PRODUCT 1867 _branches++; 1868#endif 1869 1870 // At least 1 instruction is on the available list 1871 // that is not dependant on the branch 1872 for (uint i = 0; i < siz; i++) { 1873 Node *d = _available[i]; 1874 const Pipeline *avail_pipeline = d->pipeline(); 1875 1876 // Don't allow safepoints in the branch shadow, that will 1877 // cause a number of difficulties 1878 if ( avail_pipeline->instructionCount() == 1 && 1879 !avail_pipeline->hasMultipleBundles() && 1880 !avail_pipeline->hasBranchDelay() && 1881 Pipeline::instr_has_unit_size() && 1882 d->size(_regalloc) == Pipeline::instr_unit_size() && 1883 NodeFitsInBundle(d) && 1884 !node_bundling(d)->used_in_delay()) { 1885 1886 if (d->is_Mach() && !d->is_MachSafePoint()) { 1887 // A node that fits in the delay slot was found, so we need to 1888 // set the appropriate bits in the bundle pipeline information so 1889 // that it correctly indicates resource usage. Later, when we 1890 // attempt to add this instruction to the bundle, we will skip 1891 // setting the resource usage. 1892 _unconditional_delay_slot = d; 1893 node_bundling(n)->set_use_unconditional_delay(); 1894 node_bundling(d)->set_used_in_unconditional_delay(); 1895 _bundle_use.add_usage(avail_pipeline->resourceUse()); 1896 _current_latency[d->_idx] = _bundle_cycle_number; 1897 _next_node = d; 1898 ++_bundle_instr_count; 1899#ifndef PRODUCT 1900 _unconditional_delays++; 1901#endif 1902 break; 1903 } 1904 } 1905 } 1906 } 1907 1908 // No delay slot, add a nop to the usage 1909 if (!_unconditional_delay_slot) { 1910 // See if adding an instruction in the delay slot will overflow 1911 // the bundle. 1912 if (!NodeFitsInBundle(_nop)) { 1913#ifndef PRODUCT 1914 if (_cfg->C->trace_opto_output()) 1915 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 1916#endif 1917 step(1); 1918 } 1919 1920 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 1921 _next_node = _nop; 1922 ++_bundle_instr_count; 1923 } 1924 1925 // See if the instruction in the delay slot requires a 1926 // step of the bundles 1927 if (!NodeFitsInBundle(n)) { 1928#ifndef PRODUCT 1929 if (_cfg->C->trace_opto_output()) 1930 tty->print("# *** STEP(branch won't fit) ***\n"); 1931#endif 1932 // Update the state information 1933 _bundle_instr_count = 0; 1934 _bundle_cycle_number += 1; 1935 _bundle_use.step(1); 1936 } 1937 } 1938 1939 // Get the number of instructions 1940 uint instruction_count = node_pipeline->instructionCount(); 1941 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1942 instruction_count = 0; 1943 1944 // Compute the latency information 1945 uint delay = 0; 1946 1947 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 1948 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 1949 if (relative_latency < 0) 1950 relative_latency = 0; 1951 1952 delay = _bundle_use.full_latency(relative_latency, node_usage); 1953 1954 // Does not fit in this bundle, start a new one 1955 if (delay > 0) { 1956 step(delay); 1957 1958#ifndef PRODUCT 1959 if (_cfg->C->trace_opto_output()) 1960 tty->print("# *** STEP(%d) ***\n", delay); 1961#endif 1962 } 1963 } 1964 1965 // If this was placed in the delay slot, ignore it 1966 if (n != _unconditional_delay_slot) { 1967 1968 if (delay == 0) { 1969 if (node_pipeline->hasMultipleBundles()) { 1970#ifndef PRODUCT 1971 if (_cfg->C->trace_opto_output()) 1972 tty->print("# *** STEP(multiple instructions) ***\n"); 1973#endif 1974 step(1); 1975 } 1976 1977 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 1978#ifndef PRODUCT 1979 if (_cfg->C->trace_opto_output()) 1980 tty->print("# *** STEP(%d >= %d instructions) ***\n", 1981 instruction_count + _bundle_instr_count, 1982 Pipeline::_max_instrs_per_cycle); 1983#endif 1984 step(1); 1985 } 1986 } 1987 1988 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1989 _bundle_instr_count++; 1990 1991 // Set the node's latency 1992 _current_latency[n->_idx] = _bundle_cycle_number; 1993 1994 // Now merge the functional unit information 1995 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 1996 _bundle_use.add_usage(node_usage); 1997 1998 // Increment the number of instructions in this bundle 1999 _bundle_instr_count += instruction_count; 2000 2001 // Remember this node for later 2002 if (n->is_Mach()) 2003 _next_node = n; 2004 } 2005 2006 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2007 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2008 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2009 // into the block. All other scheduled nodes get put in the schedule here. 2010 int op = n->Opcode(); 2011 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2012 (op != Op_Node && // Not an unused antidepedence node and 2013 // not an unallocated boxlock 2014 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2015 2016 // Push any trailing projections 2017 if( bb->_nodes[bb->_nodes.size()-1] != n ) { 2018 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2019 Node *foi = n->fast_out(i); 2020 if( foi->is_Proj() ) 2021 _scheduled.push(foi); 2022 } 2023 } 2024 2025 // Put the instruction in the schedule list 2026 _scheduled.push(n); 2027 } 2028 2029#ifndef PRODUCT 2030 if (_cfg->C->trace_opto_output()) 2031 dump_available(); 2032#endif 2033 2034 // Walk all the definitions, decrementing use counts, and 2035 // if a definition has a 0 use count, place it in the available list. 2036 DecrementUseCounts(n,bb); 2037} 2038 2039//------------------------------ComputeUseCount-------------------------------- 2040// This method sets the use count within a basic block. We will ignore all 2041// uses outside the current basic block. As we are doing a backwards walk, 2042// any node we reach that has a use count of 0 may be scheduled. This also 2043// avoids the problem of cyclic references from phi nodes, as long as phi 2044// nodes are at the front of the basic block. This method also initializes 2045// the available list to the set of instructions that have no uses within this 2046// basic block. 2047void Scheduling::ComputeUseCount(const Block *bb) { 2048#ifndef PRODUCT 2049 if (_cfg->C->trace_opto_output()) 2050 tty->print("# -> ComputeUseCount\n"); 2051#endif 2052 2053 // Clear the list of available and scheduled instructions, just in case 2054 _available.clear(); 2055 _scheduled.clear(); 2056 2057 // No delay slot specified 2058 _unconditional_delay_slot = NULL; 2059 2060#ifdef ASSERT 2061 for( uint i=0; i < bb->_nodes.size(); i++ ) 2062 assert( _uses[bb->_nodes[i]->_idx] == 0, "_use array not clean" ); 2063#endif 2064 2065 // Force the _uses count to never go to zero for unscheduable pieces 2066 // of the block 2067 for( uint k = 0; k < _bb_start; k++ ) 2068 _uses[bb->_nodes[k]->_idx] = 1; 2069 for( uint l = _bb_end; l < bb->_nodes.size(); l++ ) 2070 _uses[bb->_nodes[l]->_idx] = 1; 2071 2072 // Iterate backwards over the instructions in the block. Don't count the 2073 // branch projections at end or the block header instructions. 2074 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2075 Node *n = bb->_nodes[j]; 2076 if( n->is_Proj() ) continue; // Projections handled another way 2077 2078 // Account for all uses 2079 for ( uint k = 0; k < n->len(); k++ ) { 2080 Node *inp = n->in(k); 2081 if (!inp) continue; 2082 assert(inp != n, "no cycles allowed" ); 2083 if( _bbs[inp->_idx] == bb ) { // Block-local use? 2084 if( inp->is_Proj() ) // Skip through Proj's 2085 inp = inp->in(0); 2086 ++_uses[inp->_idx]; // Count 1 block-local use 2087 } 2088 } 2089 2090 // If this instruction has a 0 use count, then it is available 2091 if (!_uses[n->_idx]) { 2092 _current_latency[n->_idx] = _bundle_cycle_number; 2093 AddNodeToAvailableList(n); 2094 } 2095 2096#ifndef PRODUCT 2097 if (_cfg->C->trace_opto_output()) { 2098 tty->print("# uses: %3d: ", _uses[n->_idx]); 2099 n->dump(); 2100 } 2101#endif 2102 } 2103 2104#ifndef PRODUCT 2105 if (_cfg->C->trace_opto_output()) 2106 tty->print("# <- ComputeUseCount\n"); 2107#endif 2108} 2109 2110// This routine performs scheduling on each basic block in reverse order, 2111// using instruction latencies and taking into account function unit 2112// availability. 2113void Scheduling::DoScheduling() { 2114#ifndef PRODUCT 2115 if (_cfg->C->trace_opto_output()) 2116 tty->print("# -> DoScheduling\n"); 2117#endif 2118 2119 Block *succ_bb = NULL; 2120 Block *bb; 2121 2122 // Walk over all the basic blocks in reverse order 2123 for( int i=_cfg->_num_blocks-1; i >= 0; succ_bb = bb, i-- ) { 2124 bb = _cfg->_blocks[i]; 2125 2126#ifndef PRODUCT 2127 if (_cfg->C->trace_opto_output()) { 2128 tty->print("# Schedule BB#%03d (initial)\n", i); 2129 for (uint j = 0; j < bb->_nodes.size(); j++) 2130 bb->_nodes[j]->dump(); 2131 } 2132#endif 2133 2134 // On the head node, skip processing 2135 if( bb == _cfg->_broot ) 2136 continue; 2137 2138 // Skip empty, connector blocks 2139 if (bb->is_connector()) 2140 continue; 2141 2142 // If the following block is not the sole successor of 2143 // this one, then reset the pipeline information 2144 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2145#ifndef PRODUCT 2146 if (_cfg->C->trace_opto_output()) { 2147 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2148 _next_node->_idx, _bundle_instr_count); 2149 } 2150#endif 2151 step_and_clear(); 2152 } 2153 2154 // Leave untouched the starting instruction, any Phis, a CreateEx node 2155 // or Top. bb->_nodes[_bb_start] is the first schedulable instruction. 2156 _bb_end = bb->_nodes.size()-1; 2157 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2158 Node *n = bb->_nodes[_bb_start]; 2159 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2160 // Also, MachIdealNodes do not get scheduled 2161 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2162 MachNode *mach = n->as_Mach(); 2163 int iop = mach->ideal_Opcode(); 2164 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2165 if( iop == Op_Con ) continue; // Do not schedule Top 2166 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2167 mach->pipeline() == MachNode::pipeline_class() && 2168 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc 2169 continue; 2170 break; // Funny loop structure to be sure... 2171 } 2172 // Compute last "interesting" instruction in block - last instruction we 2173 // might schedule. _bb_end points just after last schedulable inst. We 2174 // normally schedule conditional branches (despite them being forced last 2175 // in the block), because they have delay slots we can fill. Calls all 2176 // have their delay slots filled in the template expansions, so we don't 2177 // bother scheduling them. 2178 Node *last = bb->_nodes[_bb_end]; 2179 if( last->is_Catch() || 2180 (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2181 // There must be a prior call. Skip it. 2182 while( !bb->_nodes[--_bb_end]->is_Call() ) { 2183 assert( bb->_nodes[_bb_end]->is_Proj(), "skipping projections after expected call" ); 2184 } 2185 } else if( last->is_MachNullCheck() ) { 2186 // Backup so the last null-checked memory instruction is 2187 // outside the schedulable range. Skip over the nullcheck, 2188 // projection, and the memory nodes. 2189 Node *mem = last->in(1); 2190 do { 2191 _bb_end--; 2192 } while (mem != bb->_nodes[_bb_end]); 2193 } else { 2194 // Set _bb_end to point after last schedulable inst. 2195 _bb_end++; 2196 } 2197 2198 assert( _bb_start <= _bb_end, "inverted block ends" ); 2199 2200 // Compute the register antidependencies for the basic block 2201 ComputeRegisterAntidependencies(bb); 2202 if (_cfg->C->failing()) return; // too many D-U pinch points 2203 2204 // Compute intra-bb latencies for the nodes 2205 ComputeLocalLatenciesForward(bb); 2206 2207 // Compute the usage within the block, and set the list of all nodes 2208 // in the block that have no uses within the block. 2209 ComputeUseCount(bb); 2210 2211 // Schedule the remaining instructions in the block 2212 while ( _available.size() > 0 ) { 2213 Node *n = ChooseNodeToBundle(); 2214 AddNodeToBundle(n,bb); 2215 } 2216 2217 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2218#ifdef ASSERT 2219 for( uint l = _bb_start; l < _bb_end; l++ ) { 2220 Node *n = bb->_nodes[l]; 2221 uint m; 2222 for( m = 0; m < _bb_end-_bb_start; m++ ) 2223 if( _scheduled[m] == n ) 2224 break; 2225 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2226 } 2227#endif 2228 2229 // Now copy the instructions (in reverse order) back to the block 2230 for ( uint k = _bb_start; k < _bb_end; k++ ) 2231 bb->_nodes.map(k, _scheduled[_bb_end-k-1]); 2232 2233#ifndef PRODUCT 2234 if (_cfg->C->trace_opto_output()) { 2235 tty->print("# Schedule BB#%03d (final)\n", i); 2236 uint current = 0; 2237 for (uint j = 0; j < bb->_nodes.size(); j++) { 2238 Node *n = bb->_nodes[j]; 2239 if( valid_bundle_info(n) ) { 2240 Bundle *bundle = node_bundling(n); 2241 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2242 tty->print("*** Bundle: "); 2243 bundle->dump(); 2244 } 2245 n->dump(); 2246 } 2247 } 2248 } 2249#endif 2250#ifdef ASSERT 2251 verify_good_schedule(bb,"after block local scheduling"); 2252#endif 2253 } 2254 2255#ifndef PRODUCT 2256 if (_cfg->C->trace_opto_output()) 2257 tty->print("# <- DoScheduling\n"); 2258#endif 2259 2260 // Record final node-bundling array location 2261 _regalloc->C->set_node_bundling_base(_node_bundling_base); 2262 2263} // end DoScheduling 2264 2265//------------------------------verify_good_schedule--------------------------- 2266// Verify that no live-range used in the block is killed in the block by a 2267// wrong DEF. This doesn't verify live-ranges that span blocks. 2268 2269// Check for edge existence. Used to avoid adding redundant precedence edges. 2270static bool edge_from_to( Node *from, Node *to ) { 2271 for( uint i=0; i<from->len(); i++ ) 2272 if( from->in(i) == to ) 2273 return true; 2274 return false; 2275} 2276 2277#ifdef ASSERT 2278//------------------------------verify_do_def---------------------------------- 2279void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2280 // Check for bad kills 2281 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2282 Node *prior_use = _reg_node[def]; 2283 if( prior_use && !edge_from_to(prior_use,n) ) { 2284 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2285 n->dump(); 2286 tty->print_cr("..."); 2287 prior_use->dump(); 2288 assert_msg(edge_from_to(prior_use,n),msg); 2289 } 2290 _reg_node.map(def,NULL); // Kill live USEs 2291 } 2292} 2293 2294//------------------------------verify_good_schedule--------------------------- 2295void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2296 2297 // Zap to something reasonable for the verify code 2298 _reg_node.clear(); 2299 2300 // Walk over the block backwards. Check to make sure each DEF doesn't 2301 // kill a live value (other than the one it's supposed to). Add each 2302 // USE to the live set. 2303 for( uint i = b->_nodes.size()-1; i >= _bb_start; i-- ) { 2304 Node *n = b->_nodes[i]; 2305 int n_op = n->Opcode(); 2306 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2307 // Fat-proj kills a slew of registers 2308 RegMask rm = n->out_RegMask();// Make local copy 2309 while( rm.is_NotEmpty() ) { 2310 OptoReg::Name kill = rm.find_first_elem(); 2311 rm.Remove(kill); 2312 verify_do_def( n, kill, msg ); 2313 } 2314 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2315 // Get DEF'd registers the normal way 2316 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2317 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2318 } 2319 2320 // Now make all USEs live 2321 for( uint i=1; i<n->req(); i++ ) { 2322 Node *def = n->in(i); 2323 assert(def != 0, "input edge required"); 2324 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2325 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2326 if( OptoReg::is_valid(reg_lo) ) { 2327 assert_msg(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg ); 2328 _reg_node.map(reg_lo,n); 2329 } 2330 if( OptoReg::is_valid(reg_hi) ) { 2331 assert_msg(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg ); 2332 _reg_node.map(reg_hi,n); 2333 } 2334 } 2335 2336 } 2337 2338 // Zap to something reasonable for the Antidependence code 2339 _reg_node.clear(); 2340} 2341#endif 2342 2343// Conditionally add precedence edges. Avoid putting edges on Projs. 2344static void add_prec_edge_from_to( Node *from, Node *to ) { 2345 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2346 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2347 from = from->in(0); 2348 } 2349 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2350 !edge_from_to( from, to ) ) // Avoid duplicate edge 2351 from->add_prec(to); 2352} 2353 2354//------------------------------anti_do_def------------------------------------ 2355void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2356 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2357 return; 2358 2359 Node *pinch = _reg_node[def_reg]; // Get pinch point 2360 if( !pinch || _bbs[pinch->_idx] != b || // No pinch-point yet? 2361 is_def ) { // Check for a true def (not a kill) 2362 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2363 return; 2364 } 2365 2366 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2367 debug_only( def = (Node*)0xdeadbeef; ) 2368 2369 // After some number of kills there _may_ be a later def 2370 Node *later_def = NULL; 2371 2372 // Finding a kill requires a real pinch-point. 2373 // Check for not already having a pinch-point. 2374 // Pinch points are Op_Node's. 2375 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2376 later_def = pinch; // Must be def/kill as optimistic pinch-point 2377 if ( _pinch_free_list.size() > 0) { 2378 pinch = _pinch_free_list.pop(); 2379 } else { 2380 pinch = new (_cfg->C, 1) Node(1); // Pinch point to-be 2381 } 2382 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2383 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2384 return; 2385 } 2386 _bbs.map(pinch->_idx,b); // Pretend it's valid in this block (lazy init) 2387 _reg_node.map(def_reg,pinch); // Record pinch-point 2388 //_regalloc->set_bad(pinch->_idx); // Already initialized this way. 2389 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2390 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call 2391 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2392 later_def = NULL; // and no later def 2393 } 2394 pinch->set_req(0,later_def); // Hook later def so we can find it 2395 } else { // Else have valid pinch point 2396 if( pinch->in(0) ) // If there is a later-def 2397 later_def = pinch->in(0); // Get it 2398 } 2399 2400 // Add output-dependence edge from later def to kill 2401 if( later_def ) // If there is some original def 2402 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2403 2404 // See if current kill is also a use, and so is forced to be the pinch-point. 2405 if( pinch->Opcode() == Op_Node ) { 2406 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2407 for( uint i=1; i<uses->req(); i++ ) { 2408 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2409 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2410 // Yes, found a use/kill pinch-point 2411 pinch->set_req(0,NULL); // 2412 pinch->replace_by(kill); // Move anti-dep edges up 2413 pinch = kill; 2414 _reg_node.map(def_reg,pinch); 2415 return; 2416 } 2417 } 2418 } 2419 2420 // Add edge from kill to pinch-point 2421 add_prec_edge_from_to(kill,pinch); 2422} 2423 2424//------------------------------anti_do_use------------------------------------ 2425void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2426 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2427 return; 2428 Node *pinch = _reg_node[use_reg]; // Get pinch point 2429 // Check for no later def_reg/kill in block 2430 if( pinch && _bbs[pinch->_idx] == b && 2431 // Use has to be block-local as well 2432 _bbs[use->_idx] == b ) { 2433 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2434 pinch->req() == 1 ) { // pinch not yet in block? 2435 pinch->del_req(0); // yank pointer to later-def, also set flag 2436 // Insert the pinch-point in the block just after the last use 2437 b->_nodes.insert(b->find_node(use)+1,pinch); 2438 _bb_end++; // Increase size scheduled region in block 2439 } 2440 2441 add_prec_edge_from_to(pinch,use); 2442 } 2443} 2444 2445//------------------------------ComputeRegisterAntidependences----------------- 2446// We insert antidependences between the reads and following write of 2447// allocated registers to prevent illegal code motion. Hopefully, the 2448// number of added references should be fairly small, especially as we 2449// are only adding references within the current basic block. 2450void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2451 2452#ifdef ASSERT 2453 verify_good_schedule(b,"before block local scheduling"); 2454#endif 2455 2456 // A valid schedule, for each register independently, is an endless cycle 2457 // of: a def, then some uses (connected to the def by true dependencies), 2458 // then some kills (defs with no uses), finally the cycle repeats with a new 2459 // def. The uses are allowed to float relative to each other, as are the 2460 // kills. No use is allowed to slide past a kill (or def). This requires 2461 // antidependencies between all uses of a single def and all kills that 2462 // follow, up to the next def. More edges are redundant, because later defs 2463 // & kills are already serialized with true or antidependencies. To keep 2464 // the edge count down, we add a 'pinch point' node if there's more than 2465 // one use or more than one kill/def. 2466 2467 // We add dependencies in one bottom-up pass. 2468 2469 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2470 2471 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2472 // register. If not, we record the DEF/KILL in _reg_node, the 2473 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2474 // "pinch point", a new Node that's in the graph but not in the block. 2475 // We put edges from the prior and current DEF/KILLs to the pinch point. 2476 // We put the pinch point in _reg_node. If there's already a pinch point 2477 // we merely add an edge from the current DEF/KILL to the pinch point. 2478 2479 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2480 // put an edge from the pinch point to the USE. 2481 2482 // To be expedient, the _reg_node array is pre-allocated for the whole 2483 // compilation. _reg_node is lazily initialized; it either contains a NULL, 2484 // or a valid def/kill/pinch-point, or a leftover node from some prior 2485 // block. Leftover node from some prior block is treated like a NULL (no 2486 // prior def, so no anti-dependence needed). Valid def is distinguished by 2487 // it being in the current block. 2488 bool fat_proj_seen = false; 2489 uint last_safept = _bb_end-1; 2490 Node* end_node = (_bb_end-1 >= _bb_start) ? b->_nodes[last_safept] : NULL; 2491 Node* last_safept_node = end_node; 2492 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2493 Node *n = b->_nodes[i]; 2494 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2495 if( n->Opcode() == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2496 // Fat-proj kills a slew of registers 2497 // This can add edges to 'n' and obscure whether or not it was a def, 2498 // hence the is_def flag. 2499 fat_proj_seen = true; 2500 RegMask rm = n->out_RegMask();// Make local copy 2501 while( rm.is_NotEmpty() ) { 2502 OptoReg::Name kill = rm.find_first_elem(); 2503 rm.Remove(kill); 2504 anti_do_def( b, n, kill, is_def ); 2505 } 2506 } else { 2507 // Get DEF'd registers the normal way 2508 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2509 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2510 } 2511 2512 // Check each register used by this instruction for a following DEF/KILL 2513 // that must occur afterward and requires an anti-dependence edge. 2514 for( uint j=0; j<n->req(); j++ ) { 2515 Node *def = n->in(j); 2516 if( def ) { 2517 assert( def->Opcode() != Op_MachProj || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2518 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2519 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2520 } 2521 } 2522 // Do not allow defs of new derived values to float above GC 2523 // points unless the base is definitely available at the GC point. 2524 2525 Node *m = b->_nodes[i]; 2526 2527 // Add precedence edge from following safepoint to use of derived pointer 2528 if( last_safept_node != end_node && 2529 m != last_safept_node) { 2530 for (uint k = 1; k < m->req(); k++) { 2531 const Type *t = m->in(k)->bottom_type(); 2532 if( t->isa_oop_ptr() && 2533 t->is_ptr()->offset() != 0 ) { 2534 last_safept_node->add_prec( m ); 2535 break; 2536 } 2537 } 2538 } 2539 2540 if( n->jvms() ) { // Precedence edge from derived to safept 2541 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2542 if( b->_nodes[last_safept] != last_safept_node ) { 2543 last_safept = b->find_node(last_safept_node); 2544 } 2545 for( uint j=last_safept; j > i; j-- ) { 2546 Node *mach = b->_nodes[j]; 2547 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2548 mach->add_prec( n ); 2549 } 2550 last_safept = i; 2551 last_safept_node = m; 2552 } 2553 } 2554 2555 if (fat_proj_seen) { 2556 // Garbage collect pinch nodes that were not consumed. 2557 // They are usually created by a fat kill MachProj for a call. 2558 garbage_collect_pinch_nodes(); 2559 } 2560} 2561 2562//------------------------------garbage_collect_pinch_nodes------------------------------- 2563 2564// Garbage collect pinch nodes for reuse by other blocks. 2565// 2566// The block scheduler's insertion of anti-dependence 2567// edges creates many pinch nodes when the block contains 2568// 2 or more Calls. A pinch node is used to prevent a 2569// combinatorial explosion of edges. If a set of kills for a 2570// register is anti-dependent on a set of uses (or defs), rather 2571// than adding an edge in the graph between each pair of kill 2572// and use (or def), a pinch is inserted between them: 2573// 2574// use1 use2 use3 2575// \ | / 2576// \ | / 2577// pinch 2578// / | \ 2579// / | \ 2580// kill1 kill2 kill3 2581// 2582// One pinch node is created per register killed when 2583// the second call is encountered during a backwards pass 2584// over the block. Most of these pinch nodes are never 2585// wired into the graph because the register is never 2586// used or def'ed in the block. 2587// 2588void Scheduling::garbage_collect_pinch_nodes() { 2589#ifndef PRODUCT 2590 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 2591#endif 2592 int trace_cnt = 0; 2593 for (uint k = 0; k < _reg_node.Size(); k++) { 2594 Node* pinch = _reg_node[k]; 2595 if (pinch != NULL && pinch->Opcode() == Op_Node && 2596 // no predecence input edges 2597 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { 2598 cleanup_pinch(pinch); 2599 _pinch_free_list.push(pinch); 2600 _reg_node.map(k, NULL); 2601#ifndef PRODUCT 2602 if (_cfg->C->trace_opto_output()) { 2603 trace_cnt++; 2604 if (trace_cnt > 40) { 2605 tty->print("\n"); 2606 trace_cnt = 0; 2607 } 2608 tty->print(" %d", pinch->_idx); 2609 } 2610#endif 2611 } 2612 } 2613#ifndef PRODUCT 2614 if (_cfg->C->trace_opto_output()) tty->print("\n"); 2615#endif 2616} 2617 2618// Clean up a pinch node for reuse. 2619void Scheduling::cleanup_pinch( Node *pinch ) { 2620 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 2621 2622 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 2623 Node* use = pinch->last_out(i); 2624 uint uses_found = 0; 2625 for (uint j = use->req(); j < use->len(); j++) { 2626 if (use->in(j) == pinch) { 2627 use->rm_prec(j); 2628 uses_found++; 2629 } 2630 } 2631 assert(uses_found > 0, "must be a precedence edge"); 2632 i -= uses_found; // we deleted 1 or more copies of this edge 2633 } 2634 // May have a later_def entry 2635 pinch->set_req(0, NULL); 2636} 2637 2638//------------------------------print_statistics------------------------------- 2639#ifndef PRODUCT 2640 2641void Scheduling::dump_available() const { 2642 tty->print("#Availist "); 2643 for (uint i = 0; i < _available.size(); i++) 2644 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 2645 tty->cr(); 2646} 2647 2648// Print Scheduling Statistics 2649void Scheduling::print_statistics() { 2650 // Print the size added by nops for bundling 2651 tty->print("Nops added %d bytes to total of %d bytes", 2652 _total_nop_size, _total_method_size); 2653 if (_total_method_size > 0) 2654 tty->print(", for %.2f%%", 2655 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 2656 tty->print("\n"); 2657 2658 // Print the number of branch shadows filled 2659 if (Pipeline::_branch_has_delay_slot) { 2660 tty->print("Of %d branches, %d had unconditional delay slots filled", 2661 _total_branches, _total_unconditional_delays); 2662 if (_total_branches > 0) 2663 tty->print(", for %.2f%%", 2664 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 2665 tty->print("\n"); 2666 } 2667 2668 uint total_instructions = 0, total_bundles = 0; 2669 2670 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 2671 uint bundle_count = _total_instructions_per_bundle[i]; 2672 total_instructions += bundle_count * i; 2673 total_bundles += bundle_count; 2674 } 2675 2676 if (total_bundles > 0) 2677 tty->print("Average ILP (excluding nops) is %.2f\n", 2678 ((double)total_instructions) / ((double)total_bundles)); 2679} 2680#endif 2681