1 /* 2 * Copyright (c) 1998, 2012, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/assembler.inline.hpp" 27 #include "code/debugInfo.hpp" 28 #include "code/debugInfoRec.hpp" 29 #include "compiler/compileBroker.hpp" 30 #include "compiler/oopMap.hpp" 31 #include "memory/allocation.inline.hpp" 32 #include "opto/callnode.hpp" 33 #include "opto/cfgnode.hpp" 34 #include "opto/locknode.hpp" 35 #include "opto/machnode.hpp" 36 #include "opto/output.hpp" 37 #include "opto/regalloc.hpp" 38 #include "opto/runtime.hpp" 39 #include "opto/subnode.hpp" 40 #include "opto/type.hpp" 41 #include "runtime/handles.inline.hpp" 42 #include "utilities/xmlstream.hpp" 43 44 extern uint size_java_to_interp(); 45 extern uint reloc_java_to_interp(); 46 extern uint size_exception_handler(); 47 extern uint size_deopt_handler(); 48 49 #ifndef PRODUCT 50 #define DEBUG_ARG(x) , x 51 #else 52 #define DEBUG_ARG(x) 53 #endif 54 55 extern int emit_exception_handler(CodeBuffer &cbuf); 56 extern int emit_deopt_handler(CodeBuffer &cbuf); 57 58 //------------------------------Output----------------------------------------- 59 // Convert Nodes to instruction bits and pass off to the VM 60 void Compile::Output() { 61 // RootNode goes 62 assert( _cfg->_broot->_nodes.size() == 0, "" ); 63 64 // The number of new nodes (mostly MachNop) is proportional to 65 // the number of java calls and inner loops which are aligned. 66 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 + 67 C->inner_loops()*(OptoLoopAlignment-1)), 68 "out of nodes before code generation" ) ) { 69 return; 70 } 71 // Make sure I can find the Start Node 72 Block_Array& bbs = _cfg->_bbs; 73 Block *entry = _cfg->_blocks[1]; 74 Block *broot = _cfg->_broot; 75 76 const StartNode *start = entry->_nodes[0]->as_Start(); 77 78 // Replace StartNode with prolog 79 MachPrologNode *prolog = new (this) MachPrologNode(); 80 entry->_nodes.map( 0, prolog ); 81 bbs.map( prolog->_idx, entry ); 82 bbs.map( start->_idx, NULL ); // start is no longer in any block 83 84 // Virtual methods need an unverified entry point 85 86 if( is_osr_compilation() ) { 87 if( PoisonOSREntry ) { 88 // TODO: Should use a ShouldNotReachHereNode... 89 _cfg->insert( broot, 0, new (this) MachBreakpointNode() ); 90 } 91 } else { 92 if( _method && !_method->flags().is_static() ) { 93 // Insert unvalidated entry point 94 _cfg->insert( broot, 0, new (this) MachUEPNode() ); 95 } 96 97 } 98 99 100 // Break before main entry point 101 if( (_method && _method->break_at_execute()) 102 #ifndef PRODUCT 103 ||(OptoBreakpoint && is_method_compilation()) 104 ||(OptoBreakpointOSR && is_osr_compilation()) 105 ||(OptoBreakpointC2R && !_method) 106 #endif 107 ) { 108 // checking for _method means that OptoBreakpoint does not apply to 109 // runtime stubs or frame converters 110 _cfg->insert( entry, 1, new (this) MachBreakpointNode() ); 111 } 112 113 // Insert epilogs before every return 114 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 115 Block *b = _cfg->_blocks[i]; 116 if( !b->is_connector() && b->non_connector_successor(0) == _cfg->_broot ) { // Found a program exit point? 117 Node *m = b->end(); 118 if( m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt ) { 119 MachEpilogNode *epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); 120 b->add_inst( epilog ); 121 bbs.map(epilog->_idx, b); 122 //_regalloc->set_bad(epilog->_idx); // Already initialized this way. 123 } 124 } 125 } 126 127 # ifdef ENABLE_ZAP_DEAD_LOCALS 128 if ( ZapDeadCompiledLocals ) Insert_zap_nodes(); 129 # endif 130 131 uint* blk_starts = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks+1); 132 blk_starts[0] = 0; 133 134 // Initialize code buffer and process short branches. 135 CodeBuffer* cb = init_buffer(blk_starts); 136 137 if (cb == NULL || failing()) return; 138 139 ScheduleAndBundle(); 140 141 #ifndef PRODUCT 142 if (trace_opto_output()) { 143 tty->print("\n---- After ScheduleAndBundle ----\n"); 144 for (uint i = 0; i < _cfg->_num_blocks; i++) { 145 tty->print("\nBB#%03d:\n", i); 146 Block *bb = _cfg->_blocks[i]; 147 for (uint j = 0; j < bb->_nodes.size(); j++) { 148 Node *n = bb->_nodes[j]; 149 OptoReg::Name reg = _regalloc->get_reg_first(n); 150 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : ""); 151 n->dump(); 152 } 153 } 154 } 155 #endif 156 157 if (failing()) return; 158 159 BuildOopMaps(); 160 161 if (failing()) return; 162 163 fill_buffer(cb, blk_starts); 164 } 165 166 bool Compile::need_stack_bang(int frame_size_in_bytes) const { 167 // Determine if we need to generate a stack overflow check. 168 // Do it if the method is not a stub function and 169 // has java calls or has frame size > vm_page_size/8. 170 return (UseStackBanging && stub_function() == NULL && 171 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3)); 172 } 173 174 bool Compile::need_register_stack_bang() const { 175 // Determine if we need to generate a register stack overflow check. 176 // This is only used on architectures which have split register 177 // and memory stacks (ie. IA64). 178 // Bang if the method is not a stub function and has java calls 179 return (stub_function() == NULL && has_java_calls()); 180 } 181 182 # ifdef ENABLE_ZAP_DEAD_LOCALS 183 184 185 // In order to catch compiler oop-map bugs, we have implemented 186 // a debugging mode called ZapDeadCompilerLocals. 187 // This mode causes the compiler to insert a call to a runtime routine, 188 // "zap_dead_locals", right before each place in compiled code 189 // that could potentially be a gc-point (i.e., a safepoint or oop map point). 190 // The runtime routine checks that locations mapped as oops are really 191 // oops, that locations mapped as values do not look like oops, 192 // and that locations mapped as dead are not used later 193 // (by zapping them to an invalid address). 194 195 int Compile::_CompiledZap_count = 0; 196 197 void Compile::Insert_zap_nodes() { 198 bool skip = false; 199 200 201 // Dink with static counts because code code without the extra 202 // runtime calls is MUCH faster for debugging purposes 203 204 if ( CompileZapFirst == 0 ) ; // nothing special 205 else if ( CompileZapFirst > CompiledZap_count() ) skip = true; 206 else if ( CompileZapFirst == CompiledZap_count() ) 207 warning("starting zap compilation after skipping"); 208 209 if ( CompileZapLast == -1 ) ; // nothing special 210 else if ( CompileZapLast < CompiledZap_count() ) skip = true; 211 else if ( CompileZapLast == CompiledZap_count() ) 212 warning("about to compile last zap"); 213 214 ++_CompiledZap_count; // counts skipped zaps, too 215 216 if ( skip ) return; 217 218 219 if ( _method == NULL ) 220 return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care 221 222 // Insert call to zap runtime stub before every node with an oop map 223 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 224 Block *b = _cfg->_blocks[i]; 225 for ( uint j = 0; j < b->_nodes.size(); ++j ) { 226 Node *n = b->_nodes[j]; 227 228 // Determining if we should insert a zap-a-lot node in output. 229 // We do that for all nodes that has oopmap info, except for calls 230 // to allocation. Calls to allocation passes in the old top-of-eden pointer 231 // and expect the C code to reset it. Hence, there can be no safepoints between 232 // the inlined-allocation and the call to new_Java, etc. 233 // We also cannot zap monitor calls, as they must hold the microlock 234 // during the call to Zap, which also wants to grab the microlock. 235 bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL); 236 if ( insert ) { // it is MachSafePoint 237 if ( !n->is_MachCall() ) { 238 insert = false; 239 } else if ( n->is_MachCall() ) { 240 MachCallNode* call = n->as_MachCall(); 241 if (call->entry_point() == OptoRuntime::new_instance_Java() || 242 call->entry_point() == OptoRuntime::new_array_Java() || 243 call->entry_point() == OptoRuntime::multianewarray2_Java() || 244 call->entry_point() == OptoRuntime::multianewarray3_Java() || 245 call->entry_point() == OptoRuntime::multianewarray4_Java() || 246 call->entry_point() == OptoRuntime::multianewarray5_Java() || 247 call->entry_point() == OptoRuntime::slow_arraycopy_Java() || 248 call->entry_point() == OptoRuntime::complete_monitor_locking_Java() 249 ) { 250 insert = false; 251 } 252 } 253 if (insert) { 254 Node *zap = call_zap_node(n->as_MachSafePoint(), i); 255 b->_nodes.insert( j, zap ); 256 _cfg->_bbs.map( zap->_idx, b ); 257 ++j; 258 } 259 } 260 } 261 } 262 } 263 264 265 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) { 266 const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type(); 267 CallStaticJavaNode* ideal_node = 268 new (this) CallStaticJavaNode( tf, 269 OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()), 270 "call zap dead locals stub", 0, TypePtr::BOTTOM); 271 // We need to copy the OopMap from the site we're zapping at. 272 // We have to make a copy, because the zap site might not be 273 // a call site, and zap_dead is a call site. 274 OopMap* clone = node_to_check->oop_map()->deep_copy(); 275 276 // Add the cloned OopMap to the zap node 277 ideal_node->set_oop_map(clone); 278 return _matcher->match_sfpt(ideal_node); 279 } 280 281 //------------------------------is_node_getting_a_safepoint-------------------- 282 bool Compile::is_node_getting_a_safepoint( Node* n) { 283 // This code duplicates the logic prior to the call of add_safepoint 284 // below in this file. 285 if( n->is_MachSafePoint() ) return true; 286 return false; 287 } 288 289 # endif // ENABLE_ZAP_DEAD_LOCALS 290 291 //------------------------------compute_loop_first_inst_sizes------------------ 292 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top 293 // of a loop. When aligning a loop we need to provide enough instructions 294 // in cpu's fetch buffer to feed decoders. The loop alignment could be 295 // avoided if we have enough instructions in fetch buffer at the head of a loop. 296 // By default, the size is set to 999999 by Block's constructor so that 297 // a loop will be aligned if the size is not reset here. 298 // 299 // Note: Mach instructions could contain several HW instructions 300 // so the size is estimated only. 301 // 302 void Compile::compute_loop_first_inst_sizes() { 303 // The next condition is used to gate the loop alignment optimization. 304 // Don't aligned a loop if there are enough instructions at the head of a loop 305 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad 306 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is 307 // equal to 11 bytes which is the largest address NOP instruction. 308 if( MaxLoopPad < OptoLoopAlignment-1 ) { 309 uint last_block = _cfg->_num_blocks-1; 310 for( uint i=1; i <= last_block; i++ ) { 311 Block *b = _cfg->_blocks[i]; 312 // Check the first loop's block which requires an alignment. 313 if( b->loop_alignment() > (uint)relocInfo::addr_unit() ) { 314 uint sum_size = 0; 315 uint inst_cnt = NumberOfLoopInstrToAlign; 316 inst_cnt = b->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 317 318 // Check subsequent fallthrough blocks if the loop's first 319 // block(s) does not have enough instructions. 320 Block *nb = b; 321 while( inst_cnt > 0 && 322 i < last_block && 323 !_cfg->_blocks[i+1]->has_loop_alignment() && 324 !nb->has_successor(b) ) { 325 i++; 326 nb = _cfg->_blocks[i]; 327 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 328 } // while( inst_cnt > 0 && i < last_block ) 329 330 b->set_first_inst_size(sum_size); 331 } // f( b->head()->is_Loop() ) 332 } // for( i <= last_block ) 333 } // if( MaxLoopPad < OptoLoopAlignment-1 ) 334 } 335 336 //----------------------shorten_branches--------------------------------------- 337 // The architecture description provides short branch variants for some long 338 // branch instructions. Replace eligible long branches with short branches. 339 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) { 340 341 // ------------------ 342 // Compute size of each block, method size, and relocation information size 343 uint nblocks = _cfg->_num_blocks; 344 345 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 346 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 347 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks); 348 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); ) 349 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); ) 350 351 bool has_short_branch_candidate = false; 352 353 // Initialize the sizes to 0 354 code_size = 0; // Size in bytes of generated code 355 stub_size = 0; // Size in bytes of all stub entries 356 // Size in bytes of all relocation entries, including those in local stubs. 357 // Start with 2-bytes of reloc info for the unvalidated entry point 358 reloc_size = 1; // Number of relocation entries 359 360 // Make three passes. The first computes pessimistic blk_starts, 361 // relative jmp_offset and reloc_size information. The second performs 362 // short branch substitution using the pessimistic sizing. The 363 // third inserts nops where needed. 364 365 // Step one, perform a pessimistic sizing pass. 366 uint last_call_adr = max_uint; 367 uint last_avoid_back_to_back_adr = max_uint; 368 uint nop_size = (new (this) MachNopNode())->size(_regalloc); 369 for (uint i = 0; i < nblocks; i++) { // For all blocks 370 Block *b = _cfg->_blocks[i]; 371 372 // During short branch replacement, we store the relative (to blk_starts) 373 // offset of jump in jmp_offset, rather than the absolute offset of jump. 374 // This is so that we do not need to recompute sizes of all nodes when 375 // we compute correct blk_starts in our next sizing pass. 376 jmp_offset[i] = 0; 377 jmp_size[i] = 0; 378 jmp_nidx[i] = -1; 379 DEBUG_ONLY( jmp_target[i] = 0; ) 380 DEBUG_ONLY( jmp_rule[i] = 0; ) 381 382 // Sum all instruction sizes to compute block size 383 uint last_inst = b->_nodes.size(); 384 uint blk_size = 0; 385 for (uint j = 0; j < last_inst; j++) { 386 Node* nj = b->_nodes[j]; 387 // Handle machine instruction nodes 388 if (nj->is_Mach()) { 389 MachNode *mach = nj->as_Mach(); 390 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 391 reloc_size += mach->reloc(); 392 if( mach->is_MachCall() ) { 393 MachCallNode *mcall = mach->as_MachCall(); 394 // This destination address is NOT PC-relative 395 396 mcall->method_set((intptr_t)mcall->entry_point()); 397 398 if( mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method ) { 399 stub_size += size_java_to_interp(); 400 reloc_size += reloc_java_to_interp(); 401 } 402 } else if (mach->is_MachSafePoint()) { 403 // If call/safepoint are adjacent, account for possible 404 // nop to disambiguate the two safepoints. 405 // ScheduleAndBundle() can rearrange nodes in a block, 406 // check for all offsets inside this block. 407 if (last_call_adr >= blk_starts[i]) { 408 blk_size += nop_size; 409 } 410 } 411 if (mach->avoid_back_to_back()) { 412 // Nop is inserted between "avoid back to back" instructions. 413 // ScheduleAndBundle() can rearrange nodes in a block, 414 // check for all offsets inside this block. 415 if (last_avoid_back_to_back_adr >= blk_starts[i]) { 416 blk_size += nop_size; 417 } 418 } 419 if (mach->may_be_short_branch()) { 420 if (!nj->is_MachBranch()) { 421 #ifndef PRODUCT 422 nj->dump(3); 423 #endif 424 Unimplemented(); 425 } 426 assert(jmp_nidx[i] == -1, "block should have only one branch"); 427 jmp_offset[i] = blk_size; 428 jmp_size[i] = nj->size(_regalloc); 429 jmp_nidx[i] = j; 430 has_short_branch_candidate = true; 431 } 432 } 433 blk_size += nj->size(_regalloc); 434 // Remember end of call offset 435 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) { 436 last_call_adr = blk_starts[i]+blk_size; 437 } 438 // Remember end of avoid_back_to_back offset 439 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back()) { 440 last_avoid_back_to_back_adr = blk_starts[i]+blk_size; 441 } 442 } 443 444 // When the next block starts a loop, we may insert pad NOP 445 // instructions. Since we cannot know our future alignment, 446 // assume the worst. 447 if (i< nblocks-1) { 448 Block *nb = _cfg->_blocks[i+1]; 449 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 450 if (max_loop_pad > 0) { 451 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 452 blk_size += max_loop_pad; 453 } 454 } 455 456 // Save block size; update total method size 457 blk_starts[i+1] = blk_starts[i]+blk_size; 458 } 459 460 // Step two, replace eligible long jumps. 461 bool progress = true; 462 uint last_may_be_short_branch_adr = max_uint; 463 while (has_short_branch_candidate && progress) { 464 progress = false; 465 has_short_branch_candidate = false; 466 int adjust_block_start = 0; 467 for (uint i = 0; i < nblocks; i++) { 468 Block *b = _cfg->_blocks[i]; 469 int idx = jmp_nidx[i]; 470 MachNode* mach = (idx == -1) ? NULL: b->_nodes[idx]->as_Mach(); 471 if (mach != NULL && mach->may_be_short_branch()) { 472 #ifdef ASSERT 473 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity"); 474 int j; 475 // Find the branch; ignore trailing NOPs. 476 for (j = b->_nodes.size()-1; j>=0; j--) { 477 Node* n = b->_nodes[j]; 478 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) 479 break; 480 } 481 assert(j >= 0 && j == idx && b->_nodes[j] == (Node*)mach, "sanity"); 482 #endif 483 int br_size = jmp_size[i]; 484 int br_offs = blk_starts[i] + jmp_offset[i]; 485 486 // This requires the TRUE branch target be in succs[0] 487 uint bnum = b->non_connector_successor(0)->_pre_order; 488 int offset = blk_starts[bnum] - br_offs; 489 if (bnum > i) { // adjust following block's offset 490 offset -= adjust_block_start; 491 } 492 // In the following code a nop could be inserted before 493 // the branch which will increase the backward distance. 494 bool needs_padding = ((uint)br_offs == last_may_be_short_branch_adr); 495 if (needs_padding && offset <= 0) 496 offset -= nop_size; 497 498 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 499 // We've got a winner. Replace this branch. 500 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this); 501 502 // Update the jmp_size. 503 int new_size = replacement->size(_regalloc); 504 int diff = br_size - new_size; 505 assert(diff >= (int)nop_size, "short_branch size should be smaller"); 506 // Conservatively take into accound padding between 507 // avoid_back_to_back branches. Previous branch could be 508 // converted into avoid_back_to_back branch during next 509 // rounds. 510 if (needs_padding && replacement->avoid_back_to_back()) { 511 jmp_offset[i] += nop_size; 512 diff -= nop_size; 513 } 514 adjust_block_start += diff; 515 b->_nodes.map(idx, replacement); 516 mach->subsume_by(replacement, C); 517 mach = replacement; 518 progress = true; 519 520 jmp_size[i] = new_size; 521 DEBUG_ONLY( jmp_target[i] = bnum; ); 522 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 523 } else { 524 // The jump distance is not short, try again during next iteration. 525 has_short_branch_candidate = true; 526 } 527 } // (mach->may_be_short_branch()) 528 if (mach != NULL && (mach->may_be_short_branch() || 529 mach->avoid_back_to_back())) { 530 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i]; 531 } 532 blk_starts[i+1] -= adjust_block_start; 533 } 534 } 535 536 #ifdef ASSERT 537 for (uint i = 0; i < nblocks; i++) { // For all blocks 538 if (jmp_target[i] != 0) { 539 int br_size = jmp_size[i]; 540 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 541 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 542 tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]); 543 } 544 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp"); 545 } 546 } 547 #endif 548 549 // Step 3, compute the offsets of all blocks, will be done in fill_buffer() 550 // after ScheduleAndBundle(). 551 552 // ------------------ 553 // Compute size for code buffer 554 code_size = blk_starts[nblocks]; 555 556 // Relocation records 557 reloc_size += 1; // Relo entry for exception handler 558 559 // Adjust reloc_size to number of record of relocation info 560 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 561 // a relocation index. 562 // The CodeBuffer will expand the locs array if this estimate is too low. 563 reloc_size *= 10 / sizeof(relocInfo); 564 } 565 566 //------------------------------FillLocArray----------------------------------- 567 // Create a bit of debug info and append it to the array. The mapping is from 568 // Java local or expression stack to constant, register or stack-slot. For 569 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 570 // entry has been taken care of and caller should skip it). 571 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 572 // This should never have accepted Bad before 573 assert(OptoReg::is_valid(regnum), "location must be valid"); 574 return (OptoReg::is_reg(regnum)) 575 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 576 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 577 } 578 579 580 ObjectValue* 581 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 582 for (int i = 0; i < objs->length(); i++) { 583 assert(objs->at(i)->is_object(), "corrupt object cache"); 584 ObjectValue* sv = (ObjectValue*) objs->at(i); 585 if (sv->id() == id) { 586 return sv; 587 } 588 } 589 // Otherwise.. 590 return NULL; 591 } 592 593 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 594 ObjectValue* sv ) { 595 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition"); 596 objs->append(sv); 597 } 598 599 600 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 601 GrowableArray<ScopeValue*> *array, 602 GrowableArray<ScopeValue*> *objs ) { 603 assert( local, "use _top instead of null" ); 604 if (array->length() != idx) { 605 assert(array->length() == idx + 1, "Unexpected array count"); 606 // Old functionality: 607 // return 608 // New functionality: 609 // Assert if the local is not top. In product mode let the new node 610 // override the old entry. 611 assert(local == top(), "LocArray collision"); 612 if (local == top()) { 613 return; 614 } 615 array->pop(); 616 } 617 const Type *t = local->bottom_type(); 618 619 // Is it a safepoint scalar object node? 620 if (local->is_SafePointScalarObject()) { 621 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 622 623 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx); 624 if (sv == NULL) { 625 ciKlass* cik = t->is_oopptr()->klass(); 626 assert(cik->is_instance_klass() || 627 cik->is_array_klass(), "Not supported allocation."); 628 sv = new ObjectValue(spobj->_idx, 629 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 630 Compile::set_sv_for_object_node(objs, sv); 631 632 uint first_ind = spobj->first_index(); 633 for (uint i = 0; i < spobj->n_fields(); i++) { 634 Node* fld_node = sfpt->in(first_ind+i); 635 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 636 } 637 } 638 array->append(sv); 639 return; 640 } 641 642 // Grab the register number for the local 643 OptoReg::Name regnum = _regalloc->get_reg_first(local); 644 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 645 // Record the double as two float registers. 646 // The register mask for such a value always specifies two adjacent 647 // float registers, with the lower register number even. 648 // Normally, the allocation of high and low words to these registers 649 // is irrelevant, because nearly all operations on register pairs 650 // (e.g., StoreD) treat them as a single unit. 651 // Here, we assume in addition that the words in these two registers 652 // stored "naturally" (by operations like StoreD and double stores 653 // within the interpreter) such that the lower-numbered register 654 // is written to the lower memory address. This may seem like 655 // a machine dependency, but it is not--it is a requirement on 656 // the author of the <arch>.ad file to ensure that, for every 657 // even/odd double-register pair to which a double may be allocated, 658 // the word in the even single-register is stored to the first 659 // memory word. (Note that register numbers are completely 660 // arbitrary, and are not tied to any machine-level encodings.) 661 #ifdef _LP64 662 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 663 array->append(new ConstantIntValue(0)); 664 array->append(new_loc_value( _regalloc, regnum, Location::dbl )); 665 } else if ( t->base() == Type::Long ) { 666 array->append(new ConstantIntValue(0)); 667 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 668 } else if ( t->base() == Type::RawPtr ) { 669 // jsr/ret return address which must be restored into a the full 670 // width 64-bit stack slot. 671 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 672 } 673 #else //_LP64 674 #ifdef SPARC 675 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) { 676 // For SPARC we have to swap high and low words for 677 // long values stored in a single-register (g0-g7). 678 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 679 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 680 } else 681 #endif //SPARC 682 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 683 // Repack the double/long 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 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 691 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 692 } 693 #endif //_LP64 694 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 695 OptoReg::is_reg(regnum) ) { 696 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double() 697 ? Location::float_in_dbl : Location::normal )); 698 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 699 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long 700 ? Location::int_in_long : Location::normal )); 701 } else if( t->base() == Type::NarrowOop ) { 702 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop )); 703 } else { 704 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal )); 705 } 706 return; 707 } 708 709 // No register. It must be constant data. 710 switch (t->base()) { 711 case Type::Half: // Second half of a double 712 ShouldNotReachHere(); // Caller should skip 2nd halves 713 break; 714 case Type::AnyPtr: 715 array->append(new ConstantOopWriteValue(NULL)); 716 break; 717 case Type::AryPtr: 718 case Type::InstPtr: // fall through 719 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 720 break; 721 case Type::NarrowOop: 722 if (t == TypeNarrowOop::NULL_PTR) { 723 array->append(new ConstantOopWriteValue(NULL)); 724 } else { 725 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 726 } 727 break; 728 case Type::Int: 729 array->append(new ConstantIntValue(t->is_int()->get_con())); 730 break; 731 case Type::RawPtr: 732 // A return address (T_ADDRESS). 733 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 734 #ifdef _LP64 735 // Must be restored to the full-width 64-bit stack slot. 736 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 737 #else 738 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 739 #endif 740 break; 741 case Type::FloatCon: { 742 float f = t->is_float_constant()->getf(); 743 array->append(new ConstantIntValue(jint_cast(f))); 744 break; 745 } 746 case Type::DoubleCon: { 747 jdouble d = t->is_double_constant()->getd(); 748 #ifdef _LP64 749 array->append(new ConstantIntValue(0)); 750 array->append(new ConstantDoubleValue(d)); 751 #else 752 // Repack the double as two jints. 753 // The convention the interpreter uses is that the second local 754 // holds the first raw word of the native double representation. 755 // This is actually reasonable, since locals and stack arrays 756 // grow downwards in all implementations. 757 // (If, on some machine, the interpreter's Java locals or stack 758 // were to grow upwards, the embedded doubles would be word-swapped.) 759 jint *dp = (jint*)&d; 760 array->append(new ConstantIntValue(dp[1])); 761 array->append(new ConstantIntValue(dp[0])); 762 #endif 763 break; 764 } 765 case Type::Long: { 766 jlong d = t->is_long()->get_con(); 767 #ifdef _LP64 768 array->append(new ConstantIntValue(0)); 769 array->append(new ConstantLongValue(d)); 770 #else 771 // Repack the long as two jints. 772 // The convention the interpreter uses is that the second local 773 // holds the first raw word of the native double representation. 774 // This is actually reasonable, since locals and stack arrays 775 // grow downwards in all implementations. 776 // (If, on some machine, the interpreter's Java locals or stack 777 // were to grow upwards, the embedded doubles would be word-swapped.) 778 jint *dp = (jint*)&d; 779 array->append(new ConstantIntValue(dp[1])); 780 array->append(new ConstantIntValue(dp[0])); 781 #endif 782 break; 783 } 784 case Type::Top: // Add an illegal value here 785 array->append(new LocationValue(Location())); 786 break; 787 default: 788 ShouldNotReachHere(); 789 break; 790 } 791 } 792 793 // Determine if this node starts a bundle 794 bool Compile::starts_bundle(const Node *n) const { 795 return (_node_bundling_limit > n->_idx && 796 _node_bundling_base[n->_idx].starts_bundle()); 797 } 798 799 //--------------------------Process_OopMap_Node-------------------------------- 800 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) { 801 802 // Handle special safepoint nodes for synchronization 803 MachSafePointNode *sfn = mach->as_MachSafePoint(); 804 MachCallNode *mcall; 805 806 #ifdef ENABLE_ZAP_DEAD_LOCALS 807 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative"); 808 #endif 809 810 int safepoint_pc_offset = current_offset; 811 bool is_method_handle_invoke = false; 812 bool return_oop = false; 813 814 // Add the safepoint in the DebugInfoRecorder 815 if( !mach->is_MachCall() ) { 816 mcall = NULL; 817 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 818 } else { 819 mcall = mach->as_MachCall(); 820 821 // Is the call a MethodHandle call? 822 if (mcall->is_MachCallJava()) { 823 if (mcall->as_MachCallJava()->_method_handle_invoke) { 824 assert(has_method_handle_invokes(), "must have been set during call generation"); 825 is_method_handle_invoke = true; 826 } 827 } 828 829 // Check if a call returns an object. 830 if (mcall->return_value_is_used() && 831 mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) { 832 return_oop = true; 833 } 834 safepoint_pc_offset += mcall->ret_addr_offset(); 835 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 836 } 837 838 // Loop over the JVMState list to add scope information 839 // Do not skip safepoints with a NULL method, they need monitor info 840 JVMState* youngest_jvms = sfn->jvms(); 841 int max_depth = youngest_jvms->depth(); 842 843 // Allocate the object pool for scalar-replaced objects -- the map from 844 // small-integer keys (which can be recorded in the local and ostack 845 // arrays) to descriptions of the object state. 846 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 847 848 // Visit scopes from oldest to youngest. 849 for (int depth = 1; depth <= max_depth; depth++) { 850 JVMState* jvms = youngest_jvms->of_depth(depth); 851 int idx; 852 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 853 // Safepoints that do not have method() set only provide oop-map and monitor info 854 // to support GC; these do not support deoptimization. 855 int num_locs = (method == NULL) ? 0 : jvms->loc_size(); 856 int num_exps = (method == NULL) ? 0 : jvms->stk_size(); 857 int num_mon = jvms->nof_monitors(); 858 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), 859 "JVMS local count must match that of the method"); 860 861 // Add Local and Expression Stack Information 862 863 // Insert locals into the locarray 864 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 865 for( idx = 0; idx < num_locs; idx++ ) { 866 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 867 } 868 869 // Insert expression stack entries into the exparray 870 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 871 for( idx = 0; idx < num_exps; idx++ ) { 872 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 873 } 874 875 // Add in mappings of the monitors 876 assert( !method || 877 !method->is_synchronized() || 878 method->is_native() || 879 num_mon > 0 || 880 !GenerateSynchronizationCode, 881 "monitors must always exist for synchronized methods"); 882 883 // Build the growable array of ScopeValues for exp stack 884 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 885 886 // Loop over monitors and insert into array 887 for(idx = 0; idx < num_mon; idx++) { 888 // Grab the node that defines this monitor 889 Node* box_node = sfn->monitor_box(jvms, idx); 890 Node* obj_node = sfn->monitor_obj(jvms, idx); 891 892 // Create ScopeValue for object 893 ScopeValue *scval = NULL; 894 895 if( obj_node->is_SafePointScalarObject() ) { 896 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 897 scval = Compile::sv_for_node_id(objs, spobj->_idx); 898 if (scval == NULL) { 899 const Type *t = obj_node->bottom_type(); 900 ciKlass* cik = t->is_oopptr()->klass(); 901 assert(cik->is_instance_klass() || 902 cik->is_array_klass(), "Not supported allocation."); 903 ObjectValue* sv = new ObjectValue(spobj->_idx, 904 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 905 Compile::set_sv_for_object_node(objs, sv); 906 907 uint first_ind = spobj->first_index(); 908 for (uint i = 0; i < spobj->n_fields(); i++) { 909 Node* fld_node = sfn->in(first_ind+i); 910 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 911 } 912 scval = sv; 913 } 914 } else if( !obj_node->is_Con() ) { 915 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); 916 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 917 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop ); 918 } else { 919 scval = new_loc_value( _regalloc, obj_reg, Location::oop ); 920 } 921 } else { 922 const TypePtr *tp = obj_node->bottom_type()->make_ptr(); 923 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding()); 924 } 925 926 OptoReg::Name box_reg = BoxLockNode::reg(box_node); 927 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)); 928 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated()); 929 monarray->append(new MonitorValue(scval, basic_lock, eliminated)); 930 } 931 932 // We dump the object pool first, since deoptimization reads it in first. 933 debug_info()->dump_object_pool(objs); 934 935 // Build first class objects to pass to scope 936 DebugToken *locvals = debug_info()->create_scope_values(locarray); 937 DebugToken *expvals = debug_info()->create_scope_values(exparray); 938 DebugToken *monvals = debug_info()->create_monitor_values(monarray); 939 940 // Make method available for all Safepoints 941 ciMethod* scope_method = method ? method : _method; 942 // Describe the scope here 943 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 944 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 945 // Now we can describe the scope. 946 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals); 947 } // End jvms loop 948 949 // Mark the end of the scope set. 950 debug_info()->end_safepoint(safepoint_pc_offset); 951 } 952 953 954 955 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 956 class NonSafepointEmitter { 957 Compile* C; 958 JVMState* _pending_jvms; 959 int _pending_offset; 960 961 void emit_non_safepoint(); 962 963 public: 964 NonSafepointEmitter(Compile* compile) { 965 this->C = compile; 966 _pending_jvms = NULL; 967 _pending_offset = 0; 968 } 969 970 void observe_instruction(Node* n, int pc_offset) { 971 if (!C->debug_info()->recording_non_safepoints()) return; 972 973 Node_Notes* nn = C->node_notes_at(n->_idx); 974 if (nn == NULL || nn->jvms() == NULL) return; 975 if (_pending_jvms != NULL && 976 _pending_jvms->same_calls_as(nn->jvms())) { 977 // Repeated JVMS? Stretch it up here. 978 _pending_offset = pc_offset; 979 } else { 980 if (_pending_jvms != NULL && 981 _pending_offset < pc_offset) { 982 emit_non_safepoint(); 983 } 984 _pending_jvms = NULL; 985 if (pc_offset > C->debug_info()->last_pc_offset()) { 986 // This is the only way _pending_jvms can become non-NULL: 987 _pending_jvms = nn->jvms(); 988 _pending_offset = pc_offset; 989 } 990 } 991 } 992 993 // Stay out of the way of real safepoints: 994 void observe_safepoint(JVMState* jvms, int pc_offset) { 995 if (_pending_jvms != NULL && 996 !_pending_jvms->same_calls_as(jvms) && 997 _pending_offset < pc_offset) { 998 emit_non_safepoint(); 999 } 1000 _pending_jvms = NULL; 1001 } 1002 1003 void flush_at_end() { 1004 if (_pending_jvms != NULL) { 1005 emit_non_safepoint(); 1006 } 1007 _pending_jvms = NULL; 1008 } 1009 }; 1010 1011 void NonSafepointEmitter::emit_non_safepoint() { 1012 JVMState* youngest_jvms = _pending_jvms; 1013 int pc_offset = _pending_offset; 1014 1015 // Clear it now: 1016 _pending_jvms = NULL; 1017 1018 DebugInformationRecorder* debug_info = C->debug_info(); 1019 assert(debug_info->recording_non_safepoints(), "sanity"); 1020 1021 debug_info->add_non_safepoint(pc_offset); 1022 int max_depth = youngest_jvms->depth(); 1023 1024 // Visit scopes from oldest to youngest. 1025 for (int depth = 1; depth <= max_depth; depth++) { 1026 JVMState* jvms = youngest_jvms->of_depth(depth); 1027 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 1028 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 1029 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute()); 1030 } 1031 1032 // Mark the end of the scope set. 1033 debug_info->end_non_safepoint(pc_offset); 1034 } 1035 1036 1037 1038 // helper for fill_buffer bailout logic 1039 static void turn_off_compiler(Compile* C) { 1040 if (CodeCache::largest_free_block() >= CodeCacheMinimumFreeSpace*10) { 1041 // Do not turn off compilation if a single giant method has 1042 // blown the code cache size. 1043 C->record_failure("excessive request to CodeCache"); 1044 } else { 1045 // Let CompilerBroker disable further compilations. 1046 C->record_failure("CodeCache is full"); 1047 } 1048 } 1049 1050 1051 //------------------------------init_buffer------------------------------------ 1052 CodeBuffer* Compile::init_buffer(uint* blk_starts) { 1053 1054 // Set the initially allocated size 1055 int code_req = initial_code_capacity; 1056 int locs_req = initial_locs_capacity; 1057 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity; 1058 int const_req = initial_const_capacity; 1059 1060 int pad_req = NativeCall::instruction_size; 1061 // The extra spacing after the code is necessary on some platforms. 1062 // Sometimes we need to patch in a jump after the last instruction, 1063 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 1064 1065 // Compute the byte offset where we can store the deopt pc. 1066 if (fixed_slots() != 0) { 1067 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 1068 } 1069 1070 // Compute prolog code size 1071 _method_size = 0; 1072 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; 1073 #ifdef IA64 1074 if (save_argument_registers()) { 1075 // 4815101: this is a stub with implicit and unknown precision fp args. 1076 // The usual spill mechanism can only generate stfd's in this case, which 1077 // doesn't work if the fp reg to spill contains a single-precision denorm. 1078 // Instead, we hack around the normal spill mechanism using stfspill's and 1079 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate 1080 // space here for the fp arg regs (f8-f15) we're going to thusly spill. 1081 // 1082 // If we ever implement 16-byte 'registers' == stack slots, we can 1083 // get rid of this hack and have SpillCopy generate stfspill/ldffill 1084 // instead of stfd/stfs/ldfd/ldfs. 1085 _frame_slots += 8*(16/BytesPerInt); 1086 } 1087 #endif 1088 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check"); 1089 1090 if (has_mach_constant_base_node()) { 1091 // Fill the constant table. 1092 // Note: This must happen before shorten_branches. 1093 for (uint i = 0; i < _cfg->_num_blocks; i++) { 1094 Block* b = _cfg->_blocks[i]; 1095 1096 for (uint j = 0; j < b->_nodes.size(); j++) { 1097 Node* n = b->_nodes[j]; 1098 1099 // If the node is a MachConstantNode evaluate the constant 1100 // value section. 1101 if (n->is_MachConstant()) { 1102 MachConstantNode* machcon = n->as_MachConstant(); 1103 machcon->eval_constant(C); 1104 } 1105 } 1106 } 1107 1108 // Calculate the offsets of the constants and the size of the 1109 // constant table (including the padding to the next section). 1110 constant_table().calculate_offsets_and_size(); 1111 const_req = constant_table().size(); 1112 } 1113 1114 // Initialize the space for the BufferBlob used to find and verify 1115 // instruction size in MachNode::emit_size() 1116 init_scratch_buffer_blob(const_req); 1117 if (failing()) return NULL; // Out of memory 1118 1119 // Pre-compute the length of blocks and replace 1120 // long branches with short if machine supports it. 1121 shorten_branches(blk_starts, code_req, locs_req, stub_req); 1122 1123 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1124 int exception_handler_req = size_exception_handler(); 1125 int deopt_handler_req = size_deopt_handler(); 1126 exception_handler_req += MAX_stubs_size; // add marginal slop for handler 1127 deopt_handler_req += MAX_stubs_size; // add marginal slop for handler 1128 stub_req += MAX_stubs_size; // ensure per-stub margin 1129 code_req += MAX_inst_size; // ensure per-instruction margin 1130 1131 if (StressCodeBuffers) 1132 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1133 1134 int total_req = 1135 const_req + 1136 code_req + 1137 pad_req + 1138 stub_req + 1139 exception_handler_req + 1140 deopt_handler_req; // deopt handler 1141 1142 if (has_method_handle_invokes()) 1143 total_req += deopt_handler_req; // deopt MH handler 1144 1145 CodeBuffer* cb = code_buffer(); 1146 cb->initialize(total_req, locs_req); 1147 1148 // Have we run out of code space? 1149 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1150 turn_off_compiler(this); 1151 return NULL; 1152 } 1153 // Configure the code buffer. 1154 cb->initialize_consts_size(const_req); 1155 cb->initialize_stubs_size(stub_req); 1156 cb->initialize_oop_recorder(env()->oop_recorder()); 1157 1158 // fill in the nop array for bundling computations 1159 MachNode *_nop_list[Bundle::_nop_count]; 1160 Bundle::initialize_nops(_nop_list, this); 1161 1162 return cb; 1163 } 1164 1165 //------------------------------fill_buffer------------------------------------ 1166 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) { 1167 // blk_starts[] contains offsets calculated during short branches processing, 1168 // offsets should not be increased during following steps. 1169 1170 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 1171 // of a loop. It is used to determine the padding for loop alignment. 1172 compute_loop_first_inst_sizes(); 1173 1174 // Create oopmap set. 1175 _oop_map_set = new OopMapSet(); 1176 1177 // !!!!! This preserves old handling of oopmaps for now 1178 debug_info()->set_oopmaps(_oop_map_set); 1179 1180 uint nblocks = _cfg->_num_blocks; 1181 // Count and start of implicit null check instructions 1182 uint inct_cnt = 0; 1183 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1184 1185 // Count and start of calls 1186 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1187 1188 uint return_offset = 0; 1189 int nop_size = (new (this) MachNopNode())->size(_regalloc); 1190 1191 int previous_offset = 0; 1192 int current_offset = 0; 1193 int last_call_offset = -1; 1194 int last_avoid_back_to_back_offset = -1; 1195 #ifdef ASSERT 1196 int block_alignment_padding = 0; 1197 1198 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); 1199 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 1200 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 1201 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); 1202 #endif 1203 1204 // Create an array of unused labels, one for each basic block, if printing is enabled 1205 #ifndef PRODUCT 1206 int *node_offsets = NULL; 1207 uint node_offset_limit = unique(); 1208 1209 if (print_assembly()) 1210 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1211 #endif 1212 1213 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily 1214 1215 // Emit the constant table. 1216 if (has_mach_constant_base_node()) { 1217 constant_table().emit(*cb); 1218 } 1219 1220 // Create an array of labels, one for each basic block 1221 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1); 1222 for (uint i=0; i <= nblocks; i++) { 1223 blk_labels[i].init(); 1224 } 1225 1226 // ------------------ 1227 // Now fill in the code buffer 1228 Node *delay_slot = NULL; 1229 1230 for (uint i=0; i < nblocks; i++) { 1231 guarantee(blk_starts[i] >= (uint)cb->insts_size(),"should not increase size"); 1232 1233 Block *b = _cfg->_blocks[i]; 1234 1235 Node *head = b->head(); 1236 1237 // If this block needs to start aligned (i.e, can be reached other 1238 // than by falling-thru from the previous block), then force the 1239 // start of a new bundle. 1240 if (Pipeline::requires_bundling() && starts_bundle(head)) 1241 cb->flush_bundle(true); 1242 1243 #ifdef ASSERT 1244 if (!b->is_connector()) { 1245 stringStream st; 1246 b->dump_head(&_cfg->_bbs, &st); 1247 MacroAssembler(cb).block_comment(st.as_string()); 1248 } 1249 jmp_target[i] = 0; 1250 jmp_offset[i] = 0; 1251 jmp_size[i] = 0; 1252 jmp_rule[i] = 0; 1253 1254 // Maximum alignment padding for loop block was used 1255 // during first round of branches shortening, as result 1256 // padding for nodes (sfpt after call) was not added. 1257 // Take this into account for block's size change check 1258 // and allow increase block's size by the difference 1259 // of maximum and actual alignment paddings. 1260 int orig_blk_size = blk_starts[i+1] - blk_starts[i] + block_alignment_padding; 1261 #endif 1262 int blk_offset = current_offset; 1263 1264 // Define the label at the beginning of the basic block 1265 MacroAssembler(cb).bind(blk_labels[b->_pre_order]); 1266 1267 uint last_inst = b->_nodes.size(); 1268 1269 // Emit block normally, except for last instruction. 1270 // Emit means "dump code bits into code buffer". 1271 for (uint j = 0; j<last_inst; j++) { 1272 1273 // Get the node 1274 Node* n = b->_nodes[j]; 1275 1276 // See if delay slots are supported 1277 if (valid_bundle_info(n) && 1278 node_bundling(n)->used_in_unconditional_delay()) { 1279 assert(delay_slot == NULL, "no use of delay slot node"); 1280 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1281 1282 delay_slot = n; 1283 continue; 1284 } 1285 1286 // If this starts a new instruction group, then flush the current one 1287 // (but allow split bundles) 1288 if (Pipeline::requires_bundling() && starts_bundle(n)) 1289 cb->flush_bundle(false); 1290 1291 // The following logic is duplicated in the code ifdeffed for 1292 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It 1293 // should be factored out. Or maybe dispersed to the nodes? 1294 1295 // Special handling for SafePoint/Call Nodes 1296 bool is_mcall = false; 1297 if (n->is_Mach()) { 1298 MachNode *mach = n->as_Mach(); 1299 is_mcall = n->is_MachCall(); 1300 bool is_sfn = n->is_MachSafePoint(); 1301 1302 // If this requires all previous instructions be flushed, then do so 1303 if (is_sfn || is_mcall || mach->alignment_required() != 1) { 1304 cb->flush_bundle(true); 1305 current_offset = cb->insts_size(); 1306 } 1307 1308 // A padding may be needed again since a previous instruction 1309 // could be moved to delay slot. 1310 1311 // align the instruction if necessary 1312 int padding = mach->compute_padding(current_offset); 1313 // Make sure safepoint node for polling is distinct from a call's 1314 // return by adding a nop if needed. 1315 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) { 1316 padding = nop_size; 1317 } 1318 if (padding == 0 && mach->avoid_back_to_back() && 1319 current_offset == last_avoid_back_to_back_offset) { 1320 // Avoid back to back some instructions. 1321 padding = nop_size; 1322 } 1323 1324 if(padding > 0) { 1325 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1326 int nops_cnt = padding / nop_size; 1327 MachNode *nop = new (this) MachNopNode(nops_cnt); 1328 b->_nodes.insert(j++, nop); 1329 last_inst++; 1330 _cfg->_bbs.map( nop->_idx, b ); 1331 nop->emit(*cb, _regalloc); 1332 cb->flush_bundle(true); 1333 current_offset = cb->insts_size(); 1334 } 1335 1336 // Remember the start of the last call in a basic block 1337 if (is_mcall) { 1338 MachCallNode *mcall = mach->as_MachCall(); 1339 1340 // This destination address is NOT PC-relative 1341 mcall->method_set((intptr_t)mcall->entry_point()); 1342 1343 // Save the return address 1344 call_returns[b->_pre_order] = current_offset + mcall->ret_addr_offset(); 1345 1346 if (mcall->is_MachCallLeaf()) { 1347 is_mcall = false; 1348 is_sfn = false; 1349 } 1350 } 1351 1352 // sfn will be valid whenever mcall is valid now because of inheritance 1353 if (is_sfn || is_mcall) { 1354 1355 // Handle special safepoint nodes for synchronization 1356 if (!is_mcall) { 1357 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1358 // !!!!! Stubs only need an oopmap right now, so bail out 1359 if (sfn->jvms()->method() == NULL) { 1360 // Write the oopmap directly to the code blob??!! 1361 # ifdef ENABLE_ZAP_DEAD_LOCALS 1362 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive"); 1363 # endif 1364 continue; 1365 } 1366 } // End synchronization 1367 1368 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1369 current_offset); 1370 Process_OopMap_Node(mach, current_offset); 1371 } // End if safepoint 1372 1373 // If this is a null check, then add the start of the previous instruction to the list 1374 else if( mach->is_MachNullCheck() ) { 1375 inct_starts[inct_cnt++] = previous_offset; 1376 } 1377 1378 // If this is a branch, then fill in the label with the target BB's label 1379 else if (mach->is_MachBranch()) { 1380 // This requires the TRUE branch target be in succs[0] 1381 uint block_num = b->non_connector_successor(0)->_pre_order; 1382 1383 // Try to replace long branch if delay slot is not used, 1384 // it is mostly for back branches since forward branch's 1385 // distance is not updated yet. 1386 bool delay_slot_is_used = valid_bundle_info(n) && 1387 node_bundling(n)->use_unconditional_delay(); 1388 if (!delay_slot_is_used && mach->may_be_short_branch()) { 1389 assert(delay_slot == NULL, "not expecting delay slot node"); 1390 int br_size = n->size(_regalloc); 1391 int offset = blk_starts[block_num] - current_offset; 1392 if (block_num >= i) { 1393 // Current and following block's offset are not 1394 // finilized yet, adjust distance by the difference 1395 // between calculated and final offsets of current block. 1396 offset -= (blk_starts[i] - blk_offset); 1397 } 1398 // In the following code a nop could be inserted before 1399 // the branch which will increase the backward distance. 1400 bool needs_padding = (current_offset == last_avoid_back_to_back_offset); 1401 if (needs_padding && offset <= 0) 1402 offset -= nop_size; 1403 1404 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 1405 // We've got a winner. Replace this branch. 1406 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this); 1407 1408 // Update the jmp_size. 1409 int new_size = replacement->size(_regalloc); 1410 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller"); 1411 // Insert padding between avoid_back_to_back branches. 1412 if (needs_padding && replacement->avoid_back_to_back()) { 1413 MachNode *nop = new (this) MachNopNode(); 1414 b->_nodes.insert(j++, nop); 1415 _cfg->_bbs.map(nop->_idx, b); 1416 last_inst++; 1417 nop->emit(*cb, _regalloc); 1418 cb->flush_bundle(true); 1419 current_offset = cb->insts_size(); 1420 } 1421 #ifdef ASSERT 1422 jmp_target[i] = block_num; 1423 jmp_offset[i] = current_offset - blk_offset; 1424 jmp_size[i] = new_size; 1425 jmp_rule[i] = mach->rule(); 1426 #endif 1427 b->_nodes.map(j, replacement); 1428 mach->subsume_by(replacement, C); 1429 n = replacement; 1430 mach = replacement; 1431 } 1432 } 1433 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num ); 1434 } else if (mach->ideal_Opcode() == Op_Jump) { 1435 for (uint h = 0; h < b->_num_succs; h++) { 1436 Block* succs_block = b->_succs[h]; 1437 for (uint j = 1; j < succs_block->num_preds(); j++) { 1438 Node* jpn = succs_block->pred(j); 1439 if (jpn->is_JumpProj() && jpn->in(0) == mach) { 1440 uint block_num = succs_block->non_connector()->_pre_order; 1441 Label *blkLabel = &blk_labels[block_num]; 1442 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1443 } 1444 } 1445 } 1446 } 1447 1448 #ifdef ASSERT 1449 // Check that oop-store precedes the card-mark 1450 else if (mach->ideal_Opcode() == Op_StoreCM) { 1451 uint storeCM_idx = j; 1452 int count = 0; 1453 for (uint prec = mach->req(); prec < mach->len(); prec++) { 1454 Node *oop_store = mach->in(prec); // Precedence edge 1455 if (oop_store == NULL) continue; 1456 count++; 1457 uint i4; 1458 for( i4 = 0; i4 < last_inst; ++i4 ) { 1459 if( b->_nodes[i4] == oop_store ) break; 1460 } 1461 // Note: This test can provide a false failure if other precedence 1462 // edges have been added to the storeCMNode. 1463 assert( i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1464 } 1465 assert(count > 0, "storeCM expects at least one precedence edge"); 1466 } 1467 #endif 1468 1469 else if (!n->is_Proj()) { 1470 // Remember the beginning of the previous instruction, in case 1471 // it's followed by a flag-kill and a null-check. Happens on 1472 // Intel all the time, with add-to-memory kind of opcodes. 1473 previous_offset = current_offset; 1474 } 1475 } 1476 1477 // Verify that there is sufficient space remaining 1478 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1479 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1480 turn_off_compiler(this); 1481 return; 1482 } 1483 1484 // Save the offset for the listing 1485 #ifndef PRODUCT 1486 if (node_offsets && n->_idx < node_offset_limit) 1487 node_offsets[n->_idx] = cb->insts_size(); 1488 #endif 1489 1490 // "Normal" instruction case 1491 DEBUG_ONLY( uint instr_offset = cb->insts_size(); ) 1492 n->emit(*cb, _regalloc); 1493 current_offset = cb->insts_size(); 1494 1495 #ifdef ASSERT 1496 if (n->size(_regalloc) < (current_offset-instr_offset)) { 1497 n->dump(); 1498 assert(false, "wrong size of mach node"); 1499 } 1500 #endif 1501 non_safepoints.observe_instruction(n, current_offset); 1502 1503 // mcall is last "call" that can be a safepoint 1504 // record it so we can see if a poll will directly follow it 1505 // in which case we'll need a pad to make the PcDesc sites unique 1506 // see 5010568. This can be slightly inaccurate but conservative 1507 // in the case that return address is not actually at current_offset. 1508 // This is a small price to pay. 1509 1510 if (is_mcall) { 1511 last_call_offset = current_offset; 1512 } 1513 1514 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back()) { 1515 // Avoid back to back some instructions. 1516 last_avoid_back_to_back_offset = current_offset; 1517 } 1518 1519 // See if this instruction has a delay slot 1520 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1521 assert(delay_slot != NULL, "expecting delay slot node"); 1522 1523 // Back up 1 instruction 1524 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size()); 1525 1526 // Save the offset for the listing 1527 #ifndef PRODUCT 1528 if (node_offsets && delay_slot->_idx < node_offset_limit) 1529 node_offsets[delay_slot->_idx] = cb->insts_size(); 1530 #endif 1531 1532 // Support a SafePoint in the delay slot 1533 if (delay_slot->is_MachSafePoint()) { 1534 MachNode *mach = delay_slot->as_Mach(); 1535 // !!!!! Stubs only need an oopmap right now, so bail out 1536 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) { 1537 // Write the oopmap directly to the code blob??!! 1538 # ifdef ENABLE_ZAP_DEAD_LOCALS 1539 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive"); 1540 # endif 1541 delay_slot = NULL; 1542 continue; 1543 } 1544 1545 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1546 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1547 adjusted_offset); 1548 // Generate an OopMap entry 1549 Process_OopMap_Node(mach, adjusted_offset); 1550 } 1551 1552 // Insert the delay slot instruction 1553 delay_slot->emit(*cb, _regalloc); 1554 1555 // Don't reuse it 1556 delay_slot = NULL; 1557 } 1558 1559 } // End for all instructions in block 1560 assert((uint)blk_offset <= blk_starts[i], "shouldn't increase distance"); 1561 blk_starts[i] = blk_offset; 1562 1563 // If the next block is the top of a loop, pad this block out to align 1564 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1565 if (i < nblocks-1) { 1566 Block *nb = _cfg->_blocks[i+1]; 1567 int padding = nb->alignment_padding(current_offset); 1568 if( padding > 0 ) { 1569 MachNode *nop = new (this) MachNopNode(padding / nop_size); 1570 b->_nodes.insert( b->_nodes.size(), nop ); 1571 _cfg->_bbs.map( nop->_idx, b ); 1572 nop->emit(*cb, _regalloc); 1573 current_offset = cb->insts_size(); 1574 } 1575 #ifdef ASSERT 1576 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 1577 block_alignment_padding = (max_loop_pad - padding); 1578 assert(block_alignment_padding >= 0, "sanity"); 1579 #endif 1580 } 1581 // Verify that the distance for generated before forward 1582 // short branches is still valid. 1583 assert(orig_blk_size >= (current_offset - blk_offset), "shouldn't increase block size"); 1584 1585 } // End of for all blocks 1586 blk_starts[nblocks] = current_offset; 1587 1588 non_safepoints.flush_at_end(); 1589 1590 // Offset too large? 1591 if (failing()) return; 1592 1593 // Define a pseudo-label at the end of the code 1594 MacroAssembler(cb).bind( blk_labels[nblocks] ); 1595 1596 // Compute the size of the first block 1597 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1598 1599 assert(cb->insts_size() < 500000, "method is unreasonably large"); 1600 1601 #ifdef ASSERT 1602 for (uint i = 0; i < nblocks; i++) { // For all blocks 1603 if (jmp_target[i] != 0) { 1604 int br_size = jmp_size[i]; 1605 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 1606 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 1607 tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]); 1608 assert(false, "Displacement too large for short jmp"); 1609 } 1610 } 1611 } 1612 #endif 1613 1614 // ------------------ 1615 1616 #ifndef PRODUCT 1617 // Information on the size of the method, without the extraneous code 1618 Scheduling::increment_method_size(cb->insts_size()); 1619 #endif 1620 1621 // ------------------ 1622 // Fill in exception table entries. 1623 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1624 1625 // Only java methods have exception handlers and deopt handlers 1626 if (_method) { 1627 // Emit the exception handler code. 1628 _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb)); 1629 // Emit the deopt handler code. 1630 _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb)); 1631 1632 // Emit the MethodHandle deopt handler code (if required). 1633 if (has_method_handle_invokes()) { 1634 // We can use the same code as for the normal deopt handler, we 1635 // just need a different entry point address. 1636 _code_offsets.set_value(CodeOffsets::DeoptMH, emit_deopt_handler(*cb)); 1637 } 1638 } 1639 1640 // One last check for failed CodeBuffer::expand: 1641 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1642 turn_off_compiler(this); 1643 return; 1644 } 1645 1646 #ifndef PRODUCT 1647 // Dump the assembly code, including basic-block numbers 1648 if (print_assembly()) { 1649 ttyLocker ttyl; // keep the following output all in one block 1650 if (!VMThread::should_terminate()) { // test this under the tty lock 1651 // This output goes directly to the tty, not the compiler log. 1652 // To enable tools to match it up with the compilation activity, 1653 // be sure to tag this tty output with the compile ID. 1654 if (xtty != NULL) { 1655 xtty->head("opto_assembly compile_id='%d'%s", compile_id(), 1656 is_osr_compilation() ? " compile_kind='osr'" : 1657 ""); 1658 } 1659 if (method() != NULL) { 1660 method()->print_metadata(); 1661 } 1662 dump_asm(node_offsets, node_offset_limit); 1663 if (xtty != NULL) { 1664 xtty->tail("opto_assembly"); 1665 } 1666 } 1667 } 1668 #endif 1669 1670 } 1671 1672 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1673 _inc_table.set_size(cnt); 1674 1675 uint inct_cnt = 0; 1676 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1677 Block *b = _cfg->_blocks[i]; 1678 Node *n = NULL; 1679 int j; 1680 1681 // Find the branch; ignore trailing NOPs. 1682 for( j = b->_nodes.size()-1; j>=0; j-- ) { 1683 n = b->_nodes[j]; 1684 if( !n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con ) 1685 break; 1686 } 1687 1688 // If we didn't find anything, continue 1689 if( j < 0 ) continue; 1690 1691 // Compute ExceptionHandlerTable subtable entry and add it 1692 // (skip empty blocks) 1693 if( n->is_Catch() ) { 1694 1695 // Get the offset of the return from the call 1696 uint call_return = call_returns[b->_pre_order]; 1697 #ifdef ASSERT 1698 assert( call_return > 0, "no call seen for this basic block" ); 1699 while( b->_nodes[--j]->is_MachProj() ) ; 1700 assert( b->_nodes[j]->is_MachCall(), "CatchProj must follow call" ); 1701 #endif 1702 // last instruction is a CatchNode, find it's CatchProjNodes 1703 int nof_succs = b->_num_succs; 1704 // allocate space 1705 GrowableArray<intptr_t> handler_bcis(nof_succs); 1706 GrowableArray<intptr_t> handler_pcos(nof_succs); 1707 // iterate through all successors 1708 for (int j = 0; j < nof_succs; j++) { 1709 Block* s = b->_succs[j]; 1710 bool found_p = false; 1711 for( uint k = 1; k < s->num_preds(); k++ ) { 1712 Node *pk = s->pred(k); 1713 if( pk->is_CatchProj() && pk->in(0) == n ) { 1714 const CatchProjNode* p = pk->as_CatchProj(); 1715 found_p = true; 1716 // add the corresponding handler bci & pco information 1717 if( p->_con != CatchProjNode::fall_through_index ) { 1718 // p leads to an exception handler (and is not fall through) 1719 assert(s == _cfg->_blocks[s->_pre_order],"bad numbering"); 1720 // no duplicates, please 1721 if( !handler_bcis.contains(p->handler_bci()) ) { 1722 uint block_num = s->non_connector()->_pre_order; 1723 handler_bcis.append(p->handler_bci()); 1724 handler_pcos.append(blk_labels[block_num].loc_pos()); 1725 } 1726 } 1727 } 1728 } 1729 assert(found_p, "no matching predecessor found"); 1730 // Note: Due to empty block removal, one block may have 1731 // several CatchProj inputs, from the same Catch. 1732 } 1733 1734 // Set the offset of the return from the call 1735 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); 1736 continue; 1737 } 1738 1739 // Handle implicit null exception table updates 1740 if( n->is_MachNullCheck() ) { 1741 uint block_num = b->non_connector_successor(0)->_pre_order; 1742 _inc_table.append( inct_starts[inct_cnt++], blk_labels[block_num].loc_pos() ); 1743 continue; 1744 } 1745 } // End of for all blocks fill in exception table entries 1746 } 1747 1748 // Static Variables 1749 #ifndef PRODUCT 1750 uint Scheduling::_total_nop_size = 0; 1751 uint Scheduling::_total_method_size = 0; 1752 uint Scheduling::_total_branches = 0; 1753 uint Scheduling::_total_unconditional_delays = 0; 1754 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1755 #endif 1756 1757 // Initializer for class Scheduling 1758 1759 Scheduling::Scheduling(Arena *arena, Compile &compile) 1760 : _arena(arena), 1761 _cfg(compile.cfg()), 1762 _bbs(compile.cfg()->_bbs), 1763 _regalloc(compile.regalloc()), 1764 _reg_node(arena), 1765 _bundle_instr_count(0), 1766 _bundle_cycle_number(0), 1767 _scheduled(arena), 1768 _available(arena), 1769 _next_node(NULL), 1770 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]), 1771 _pinch_free_list(arena) 1772 #ifndef PRODUCT 1773 , _branches(0) 1774 , _unconditional_delays(0) 1775 #endif 1776 { 1777 // Create a MachNopNode 1778 _nop = new (&compile) MachNopNode(); 1779 1780 // Now that the nops are in the array, save the count 1781 // (but allow entries for the nops) 1782 _node_bundling_limit = compile.unique(); 1783 uint node_max = _regalloc->node_regs_max_index(); 1784 1785 compile.set_node_bundling_limit(_node_bundling_limit); 1786 1787 // This one is persistent within the Compile class 1788 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1789 1790 // Allocate space for fixed-size arrays 1791 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1792 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1793 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1794 1795 // Clear the arrays 1796 memset(_node_bundling_base, 0, node_max * sizeof(Bundle)); 1797 memset(_node_latency, 0, node_max * sizeof(unsigned short)); 1798 memset(_uses, 0, node_max * sizeof(short)); 1799 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 1800 1801 // Clear the bundling information 1802 memcpy(_bundle_use_elements, 1803 Pipeline_Use::elaborated_elements, 1804 sizeof(Pipeline_Use::elaborated_elements)); 1805 1806 // Get the last node 1807 Block *bb = _cfg->_blocks[_cfg->_blocks.size()-1]; 1808 1809 _next_node = bb->_nodes[bb->_nodes.size()-1]; 1810 } 1811 1812 #ifndef PRODUCT 1813 // Scheduling destructor 1814 Scheduling::~Scheduling() { 1815 _total_branches += _branches; 1816 _total_unconditional_delays += _unconditional_delays; 1817 } 1818 #endif 1819 1820 // Step ahead "i" cycles 1821 void Scheduling::step(uint i) { 1822 1823 Bundle *bundle = node_bundling(_next_node); 1824 bundle->set_starts_bundle(); 1825 1826 // Update the bundle record, but leave the flags information alone 1827 if (_bundle_instr_count > 0) { 1828 bundle->set_instr_count(_bundle_instr_count); 1829 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1830 } 1831 1832 // Update the state information 1833 _bundle_instr_count = 0; 1834 _bundle_cycle_number += i; 1835 _bundle_use.step(i); 1836 } 1837 1838 void Scheduling::step_and_clear() { 1839 Bundle *bundle = node_bundling(_next_node); 1840 bundle->set_starts_bundle(); 1841 1842 // Update the bundle record 1843 if (_bundle_instr_count > 0) { 1844 bundle->set_instr_count(_bundle_instr_count); 1845 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1846 1847 _bundle_cycle_number += 1; 1848 } 1849 1850 // Clear the bundling information 1851 _bundle_instr_count = 0; 1852 _bundle_use.reset(); 1853 1854 memcpy(_bundle_use_elements, 1855 Pipeline_Use::elaborated_elements, 1856 sizeof(Pipeline_Use::elaborated_elements)); 1857 } 1858 1859 //------------------------------ScheduleAndBundle------------------------------ 1860 // Perform instruction scheduling and bundling over the sequence of 1861 // instructions in backwards order. 1862 void Compile::ScheduleAndBundle() { 1863 1864 // Don't optimize this if it isn't a method 1865 if (!_method) 1866 return; 1867 1868 // Don't optimize this if scheduling is disabled 1869 if (!do_scheduling()) 1870 return; 1871 1872 // Scheduling code works only with pairs (8 bytes) maximum. 1873 if (max_vector_size() > 8) 1874 return; 1875 1876 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); ) 1877 1878 // Create a data structure for all the scheduling information 1879 Scheduling scheduling(Thread::current()->resource_area(), *this); 1880 1881 // Walk backwards over each basic block, computing the needed alignment 1882 // Walk over all the basic blocks 1883 scheduling.DoScheduling(); 1884 } 1885 1886 //------------------------------ComputeLocalLatenciesForward------------------- 1887 // Compute the latency of all the instructions. This is fairly simple, 1888 // because we already have a legal ordering. Walk over the instructions 1889 // from first to last, and compute the latency of the instruction based 1890 // on the latency of the preceding instruction(s). 1891 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { 1892 #ifndef PRODUCT 1893 if (_cfg->C->trace_opto_output()) 1894 tty->print("# -> ComputeLocalLatenciesForward\n"); 1895 #endif 1896 1897 // Walk over all the schedulable instructions 1898 for( uint j=_bb_start; j < _bb_end; j++ ) { 1899 1900 // This is a kludge, forcing all latency calculations to start at 1. 1901 // Used to allow latency 0 to force an instruction to the beginning 1902 // of the bb 1903 uint latency = 1; 1904 Node *use = bb->_nodes[j]; 1905 uint nlen = use->len(); 1906 1907 // Walk over all the inputs 1908 for ( uint k=0; k < nlen; k++ ) { 1909 Node *def = use->in(k); 1910 if (!def) 1911 continue; 1912 1913 uint l = _node_latency[def->_idx] + use->latency(k); 1914 if (latency < l) 1915 latency = l; 1916 } 1917 1918 _node_latency[use->_idx] = latency; 1919 1920 #ifndef PRODUCT 1921 if (_cfg->C->trace_opto_output()) { 1922 tty->print("# latency %4d: ", latency); 1923 use->dump(); 1924 } 1925 #endif 1926 } 1927 1928 #ifndef PRODUCT 1929 if (_cfg->C->trace_opto_output()) 1930 tty->print("# <- ComputeLocalLatenciesForward\n"); 1931 #endif 1932 1933 } // end ComputeLocalLatenciesForward 1934 1935 // See if this node fits into the present instruction bundle 1936 bool Scheduling::NodeFitsInBundle(Node *n) { 1937 uint n_idx = n->_idx; 1938 1939 // If this is the unconditional delay instruction, then it fits 1940 if (n == _unconditional_delay_slot) { 1941 #ifndef PRODUCT 1942 if (_cfg->C->trace_opto_output()) 1943 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 1944 #endif 1945 return (true); 1946 } 1947 1948 // If the node cannot be scheduled this cycle, skip it 1949 if (_current_latency[n_idx] > _bundle_cycle_number) { 1950 #ifndef PRODUCT 1951 if (_cfg->C->trace_opto_output()) 1952 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 1953 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 1954 #endif 1955 return (false); 1956 } 1957 1958 const Pipeline *node_pipeline = n->pipeline(); 1959 1960 uint instruction_count = node_pipeline->instructionCount(); 1961 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1962 instruction_count = 0; 1963 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1964 instruction_count++; 1965 1966 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 1967 #ifndef PRODUCT 1968 if (_cfg->C->trace_opto_output()) 1969 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 1970 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 1971 #endif 1972 return (false); 1973 } 1974 1975 // Don't allow non-machine nodes to be handled this way 1976 if (!n->is_Mach() && instruction_count == 0) 1977 return (false); 1978 1979 // See if there is any overlap 1980 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 1981 1982 if (delay > 0) { 1983 #ifndef PRODUCT 1984 if (_cfg->C->trace_opto_output()) 1985 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 1986 #endif 1987 return false; 1988 } 1989 1990 #ifndef PRODUCT 1991 if (_cfg->C->trace_opto_output()) 1992 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 1993 #endif 1994 1995 return true; 1996 } 1997 1998 Node * Scheduling::ChooseNodeToBundle() { 1999 uint siz = _available.size(); 2000 2001 if (siz == 0) { 2002 2003 #ifndef PRODUCT 2004 if (_cfg->C->trace_opto_output()) 2005 tty->print("# ChooseNodeToBundle: NULL\n"); 2006 #endif 2007 return (NULL); 2008 } 2009 2010 // Fast path, if only 1 instruction in the bundle 2011 if (siz == 1) { 2012 #ifndef PRODUCT 2013 if (_cfg->C->trace_opto_output()) { 2014 tty->print("# ChooseNodeToBundle (only 1): "); 2015 _available[0]->dump(); 2016 } 2017 #endif 2018 return (_available[0]); 2019 } 2020 2021 // Don't bother, if the bundle is already full 2022 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 2023 for ( uint i = 0; i < siz; i++ ) { 2024 Node *n = _available[i]; 2025 2026 // Skip projections, we'll handle them another way 2027 if (n->is_Proj()) 2028 continue; 2029 2030 // This presupposed that instructions are inserted into the 2031 // available list in a legality order; i.e. instructions that 2032 // must be inserted first are at the head of the list 2033 if (NodeFitsInBundle(n)) { 2034 #ifndef PRODUCT 2035 if (_cfg->C->trace_opto_output()) { 2036 tty->print("# ChooseNodeToBundle: "); 2037 n->dump(); 2038 } 2039 #endif 2040 return (n); 2041 } 2042 } 2043 } 2044 2045 // Nothing fits in this bundle, choose the highest priority 2046 #ifndef PRODUCT 2047 if (_cfg->C->trace_opto_output()) { 2048 tty->print("# ChooseNodeToBundle: "); 2049 _available[0]->dump(); 2050 } 2051 #endif 2052 2053 return _available[0]; 2054 } 2055 2056 //------------------------------AddNodeToAvailableList------------------------- 2057 void Scheduling::AddNodeToAvailableList(Node *n) { 2058 assert( !n->is_Proj(), "projections never directly made available" ); 2059 #ifndef PRODUCT 2060 if (_cfg->C->trace_opto_output()) { 2061 tty->print("# AddNodeToAvailableList: "); 2062 n->dump(); 2063 } 2064 #endif 2065 2066 int latency = _current_latency[n->_idx]; 2067 2068 // Insert in latency order (insertion sort) 2069 uint i; 2070 for ( i=0; i < _available.size(); i++ ) 2071 if (_current_latency[_available[i]->_idx] > latency) 2072 break; 2073 2074 // Special Check for compares following branches 2075 if( n->is_Mach() && _scheduled.size() > 0 ) { 2076 int op = n->as_Mach()->ideal_Opcode(); 2077 Node *last = _scheduled[0]; 2078 if( last->is_MachIf() && last->in(1) == n && 2079 ( op == Op_CmpI || 2080 op == Op_CmpU || 2081 op == Op_CmpP || 2082 op == Op_CmpF || 2083 op == Op_CmpD || 2084 op == Op_CmpL ) ) { 2085 2086 // Recalculate position, moving to front of same latency 2087 for ( i=0 ; i < _available.size(); i++ ) 2088 if (_current_latency[_available[i]->_idx] >= latency) 2089 break; 2090 } 2091 } 2092 2093 // Insert the node in the available list 2094 _available.insert(i, n); 2095 2096 #ifndef PRODUCT 2097 if (_cfg->C->trace_opto_output()) 2098 dump_available(); 2099 #endif 2100 } 2101 2102 //------------------------------DecrementUseCounts----------------------------- 2103 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 2104 for ( uint i=0; i < n->len(); i++ ) { 2105 Node *def = n->in(i); 2106 if (!def) continue; 2107 if( def->is_Proj() ) // If this is a machine projection, then 2108 def = def->in(0); // propagate usage thru to the base instruction 2109 2110 if( _bbs[def->_idx] != bb ) // Ignore if not block-local 2111 continue; 2112 2113 // Compute the latency 2114 uint l = _bundle_cycle_number + n->latency(i); 2115 if (_current_latency[def->_idx] < l) 2116 _current_latency[def->_idx] = l; 2117 2118 // If this does not have uses then schedule it 2119 if ((--_uses[def->_idx]) == 0) 2120 AddNodeToAvailableList(def); 2121 } 2122 } 2123 2124 //------------------------------AddNodeToBundle-------------------------------- 2125 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 2126 #ifndef PRODUCT 2127 if (_cfg->C->trace_opto_output()) { 2128 tty->print("# AddNodeToBundle: "); 2129 n->dump(); 2130 } 2131 #endif 2132 2133 // Remove this from the available list 2134 uint i; 2135 for (i = 0; i < _available.size(); i++) 2136 if (_available[i] == n) 2137 break; 2138 assert(i < _available.size(), "entry in _available list not found"); 2139 _available.remove(i); 2140 2141 // See if this fits in the current bundle 2142 const Pipeline *node_pipeline = n->pipeline(); 2143 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 2144 2145 // Check for instructions to be placed in the delay slot. We 2146 // do this before we actually schedule the current instruction, 2147 // because the delay slot follows the current instruction. 2148 if (Pipeline::_branch_has_delay_slot && 2149 node_pipeline->hasBranchDelay() && 2150 !_unconditional_delay_slot) { 2151 2152 uint siz = _available.size(); 2153 2154 // Conditional branches can support an instruction that 2155 // is unconditionally executed and not dependent by the 2156 // branch, OR a conditionally executed instruction if 2157 // the branch is taken. In practice, this means that 2158 // the first instruction at the branch target is 2159 // copied to the delay slot, and the branch goes to 2160 // the instruction after that at the branch target 2161 if ( n->is_MachBranch() ) { 2162 2163 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 2164 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 2165 2166 #ifndef PRODUCT 2167 _branches++; 2168 #endif 2169 2170 // At least 1 instruction is on the available list 2171 // that is not dependent on the branch 2172 for (uint i = 0; i < siz; i++) { 2173 Node *d = _available[i]; 2174 const Pipeline *avail_pipeline = d->pipeline(); 2175 2176 // Don't allow safepoints in the branch shadow, that will 2177 // cause a number of difficulties 2178 if ( avail_pipeline->instructionCount() == 1 && 2179 !avail_pipeline->hasMultipleBundles() && 2180 !avail_pipeline->hasBranchDelay() && 2181 Pipeline::instr_has_unit_size() && 2182 d->size(_regalloc) == Pipeline::instr_unit_size() && 2183 NodeFitsInBundle(d) && 2184 !node_bundling(d)->used_in_delay()) { 2185 2186 if (d->is_Mach() && !d->is_MachSafePoint()) { 2187 // A node that fits in the delay slot was found, so we need to 2188 // set the appropriate bits in the bundle pipeline information so 2189 // that it correctly indicates resource usage. Later, when we 2190 // attempt to add this instruction to the bundle, we will skip 2191 // setting the resource usage. 2192 _unconditional_delay_slot = d; 2193 node_bundling(n)->set_use_unconditional_delay(); 2194 node_bundling(d)->set_used_in_unconditional_delay(); 2195 _bundle_use.add_usage(avail_pipeline->resourceUse()); 2196 _current_latency[d->_idx] = _bundle_cycle_number; 2197 _next_node = d; 2198 ++_bundle_instr_count; 2199 #ifndef PRODUCT 2200 _unconditional_delays++; 2201 #endif 2202 break; 2203 } 2204 } 2205 } 2206 } 2207 2208 // No delay slot, add a nop to the usage 2209 if (!_unconditional_delay_slot) { 2210 // See if adding an instruction in the delay slot will overflow 2211 // the bundle. 2212 if (!NodeFitsInBundle(_nop)) { 2213 #ifndef PRODUCT 2214 if (_cfg->C->trace_opto_output()) 2215 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2216 #endif 2217 step(1); 2218 } 2219 2220 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2221 _next_node = _nop; 2222 ++_bundle_instr_count; 2223 } 2224 2225 // See if the instruction in the delay slot requires a 2226 // step of the bundles 2227 if (!NodeFitsInBundle(n)) { 2228 #ifndef PRODUCT 2229 if (_cfg->C->trace_opto_output()) 2230 tty->print("# *** STEP(branch won't fit) ***\n"); 2231 #endif 2232 // Update the state information 2233 _bundle_instr_count = 0; 2234 _bundle_cycle_number += 1; 2235 _bundle_use.step(1); 2236 } 2237 } 2238 2239 // Get the number of instructions 2240 uint instruction_count = node_pipeline->instructionCount(); 2241 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2242 instruction_count = 0; 2243 2244 // Compute the latency information 2245 uint delay = 0; 2246 2247 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2248 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2249 if (relative_latency < 0) 2250 relative_latency = 0; 2251 2252 delay = _bundle_use.full_latency(relative_latency, node_usage); 2253 2254 // Does not fit in this bundle, start a new one 2255 if (delay > 0) { 2256 step(delay); 2257 2258 #ifndef PRODUCT 2259 if (_cfg->C->trace_opto_output()) 2260 tty->print("# *** STEP(%d) ***\n", delay); 2261 #endif 2262 } 2263 } 2264 2265 // If this was placed in the delay slot, ignore it 2266 if (n != _unconditional_delay_slot) { 2267 2268 if (delay == 0) { 2269 if (node_pipeline->hasMultipleBundles()) { 2270 #ifndef PRODUCT 2271 if (_cfg->C->trace_opto_output()) 2272 tty->print("# *** STEP(multiple instructions) ***\n"); 2273 #endif 2274 step(1); 2275 } 2276 2277 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2278 #ifndef PRODUCT 2279 if (_cfg->C->trace_opto_output()) 2280 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2281 instruction_count + _bundle_instr_count, 2282 Pipeline::_max_instrs_per_cycle); 2283 #endif 2284 step(1); 2285 } 2286 } 2287 2288 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2289 _bundle_instr_count++; 2290 2291 // Set the node's latency 2292 _current_latency[n->_idx] = _bundle_cycle_number; 2293 2294 // Now merge the functional unit information 2295 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2296 _bundle_use.add_usage(node_usage); 2297 2298 // Increment the number of instructions in this bundle 2299 _bundle_instr_count += instruction_count; 2300 2301 // Remember this node for later 2302 if (n->is_Mach()) 2303 _next_node = n; 2304 } 2305 2306 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2307 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2308 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2309 // into the block. All other scheduled nodes get put in the schedule here. 2310 int op = n->Opcode(); 2311 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2312 (op != Op_Node && // Not an unused antidepedence node and 2313 // not an unallocated boxlock 2314 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2315 2316 // Push any trailing projections 2317 if( bb->_nodes[bb->_nodes.size()-1] != n ) { 2318 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2319 Node *foi = n->fast_out(i); 2320 if( foi->is_Proj() ) 2321 _scheduled.push(foi); 2322 } 2323 } 2324 2325 // Put the instruction in the schedule list 2326 _scheduled.push(n); 2327 } 2328 2329 #ifndef PRODUCT 2330 if (_cfg->C->trace_opto_output()) 2331 dump_available(); 2332 #endif 2333 2334 // Walk all the definitions, decrementing use counts, and 2335 // if a definition has a 0 use count, place it in the available list. 2336 DecrementUseCounts(n,bb); 2337 } 2338 2339 //------------------------------ComputeUseCount-------------------------------- 2340 // This method sets the use count within a basic block. We will ignore all 2341 // uses outside the current basic block. As we are doing a backwards walk, 2342 // any node we reach that has a use count of 0 may be scheduled. This also 2343 // avoids the problem of cyclic references from phi nodes, as long as phi 2344 // nodes are at the front of the basic block. This method also initializes 2345 // the available list to the set of instructions that have no uses within this 2346 // basic block. 2347 void Scheduling::ComputeUseCount(const Block *bb) { 2348 #ifndef PRODUCT 2349 if (_cfg->C->trace_opto_output()) 2350 tty->print("# -> ComputeUseCount\n"); 2351 #endif 2352 2353 // Clear the list of available and scheduled instructions, just in case 2354 _available.clear(); 2355 _scheduled.clear(); 2356 2357 // No delay slot specified 2358 _unconditional_delay_slot = NULL; 2359 2360 #ifdef ASSERT 2361 for( uint i=0; i < bb->_nodes.size(); i++ ) 2362 assert( _uses[bb->_nodes[i]->_idx] == 0, "_use array not clean" ); 2363 #endif 2364 2365 // Force the _uses count to never go to zero for unscheduable pieces 2366 // of the block 2367 for( uint k = 0; k < _bb_start; k++ ) 2368 _uses[bb->_nodes[k]->_idx] = 1; 2369 for( uint l = _bb_end; l < bb->_nodes.size(); l++ ) 2370 _uses[bb->_nodes[l]->_idx] = 1; 2371 2372 // Iterate backwards over the instructions in the block. Don't count the 2373 // branch projections at end or the block header instructions. 2374 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2375 Node *n = bb->_nodes[j]; 2376 if( n->is_Proj() ) continue; // Projections handled another way 2377 2378 // Account for all uses 2379 for ( uint k = 0; k < n->len(); k++ ) { 2380 Node *inp = n->in(k); 2381 if (!inp) continue; 2382 assert(inp != n, "no cycles allowed" ); 2383 if( _bbs[inp->_idx] == bb ) { // Block-local use? 2384 if( inp->is_Proj() ) // Skip through Proj's 2385 inp = inp->in(0); 2386 ++_uses[inp->_idx]; // Count 1 block-local use 2387 } 2388 } 2389 2390 // If this instruction has a 0 use count, then it is available 2391 if (!_uses[n->_idx]) { 2392 _current_latency[n->_idx] = _bundle_cycle_number; 2393 AddNodeToAvailableList(n); 2394 } 2395 2396 #ifndef PRODUCT 2397 if (_cfg->C->trace_opto_output()) { 2398 tty->print("# uses: %3d: ", _uses[n->_idx]); 2399 n->dump(); 2400 } 2401 #endif 2402 } 2403 2404 #ifndef PRODUCT 2405 if (_cfg->C->trace_opto_output()) 2406 tty->print("# <- ComputeUseCount\n"); 2407 #endif 2408 } 2409 2410 // This routine performs scheduling on each basic block in reverse order, 2411 // using instruction latencies and taking into account function unit 2412 // availability. 2413 void Scheduling::DoScheduling() { 2414 #ifndef PRODUCT 2415 if (_cfg->C->trace_opto_output()) 2416 tty->print("# -> DoScheduling\n"); 2417 #endif 2418 2419 Block *succ_bb = NULL; 2420 Block *bb; 2421 2422 // Walk over all the basic blocks in reverse order 2423 for( int i=_cfg->_num_blocks-1; i >= 0; succ_bb = bb, i-- ) { 2424 bb = _cfg->_blocks[i]; 2425 2426 #ifndef PRODUCT 2427 if (_cfg->C->trace_opto_output()) { 2428 tty->print("# Schedule BB#%03d (initial)\n", i); 2429 for (uint j = 0; j < bb->_nodes.size(); j++) 2430 bb->_nodes[j]->dump(); 2431 } 2432 #endif 2433 2434 // On the head node, skip processing 2435 if( bb == _cfg->_broot ) 2436 continue; 2437 2438 // Skip empty, connector blocks 2439 if (bb->is_connector()) 2440 continue; 2441 2442 // If the following block is not the sole successor of 2443 // this one, then reset the pipeline information 2444 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2445 #ifndef PRODUCT 2446 if (_cfg->C->trace_opto_output()) { 2447 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2448 _next_node->_idx, _bundle_instr_count); 2449 } 2450 #endif 2451 step_and_clear(); 2452 } 2453 2454 // Leave untouched the starting instruction, any Phis, a CreateEx node 2455 // or Top. bb->_nodes[_bb_start] is the first schedulable instruction. 2456 _bb_end = bb->_nodes.size()-1; 2457 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2458 Node *n = bb->_nodes[_bb_start]; 2459 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2460 // Also, MachIdealNodes do not get scheduled 2461 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2462 MachNode *mach = n->as_Mach(); 2463 int iop = mach->ideal_Opcode(); 2464 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2465 if( iop == Op_Con ) continue; // Do not schedule Top 2466 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2467 mach->pipeline() == MachNode::pipeline_class() && 2468 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc 2469 continue; 2470 break; // Funny loop structure to be sure... 2471 } 2472 // Compute last "interesting" instruction in block - last instruction we 2473 // might schedule. _bb_end points just after last schedulable inst. We 2474 // normally schedule conditional branches (despite them being forced last 2475 // in the block), because they have delay slots we can fill. Calls all 2476 // have their delay slots filled in the template expansions, so we don't 2477 // bother scheduling them. 2478 Node *last = bb->_nodes[_bb_end]; 2479 // Ignore trailing NOPs. 2480 while (_bb_end > 0 && last->is_Mach() && 2481 last->as_Mach()->ideal_Opcode() == Op_Con) { 2482 last = bb->_nodes[--_bb_end]; 2483 } 2484 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, ""); 2485 if( last->is_Catch() || 2486 // Exclude unreachable path case when Halt node is in a separate block. 2487 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2488 // There must be a prior call. Skip it. 2489 while( !bb->_nodes[--_bb_end]->is_MachCall() ) { 2490 assert( bb->_nodes[_bb_end]->is_MachProj(), "skipping projections after expected call" ); 2491 } 2492 } else if( last->is_MachNullCheck() ) { 2493 // Backup so the last null-checked memory instruction is 2494 // outside the schedulable range. Skip over the nullcheck, 2495 // projection, and the memory nodes. 2496 Node *mem = last->in(1); 2497 do { 2498 _bb_end--; 2499 } while (mem != bb->_nodes[_bb_end]); 2500 } else { 2501 // Set _bb_end to point after last schedulable inst. 2502 _bb_end++; 2503 } 2504 2505 assert( _bb_start <= _bb_end, "inverted block ends" ); 2506 2507 // Compute the register antidependencies for the basic block 2508 ComputeRegisterAntidependencies(bb); 2509 if (_cfg->C->failing()) return; // too many D-U pinch points 2510 2511 // Compute intra-bb latencies for the nodes 2512 ComputeLocalLatenciesForward(bb); 2513 2514 // Compute the usage within the block, and set the list of all nodes 2515 // in the block that have no uses within the block. 2516 ComputeUseCount(bb); 2517 2518 // Schedule the remaining instructions in the block 2519 while ( _available.size() > 0 ) { 2520 Node *n = ChooseNodeToBundle(); 2521 AddNodeToBundle(n,bb); 2522 } 2523 2524 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2525 #ifdef ASSERT 2526 for( uint l = _bb_start; l < _bb_end; l++ ) { 2527 Node *n = bb->_nodes[l]; 2528 uint m; 2529 for( m = 0; m < _bb_end-_bb_start; m++ ) 2530 if( _scheduled[m] == n ) 2531 break; 2532 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2533 } 2534 #endif 2535 2536 // Now copy the instructions (in reverse order) back to the block 2537 for ( uint k = _bb_start; k < _bb_end; k++ ) 2538 bb->_nodes.map(k, _scheduled[_bb_end-k-1]); 2539 2540 #ifndef PRODUCT 2541 if (_cfg->C->trace_opto_output()) { 2542 tty->print("# Schedule BB#%03d (final)\n", i); 2543 uint current = 0; 2544 for (uint j = 0; j < bb->_nodes.size(); j++) { 2545 Node *n = bb->_nodes[j]; 2546 if( valid_bundle_info(n) ) { 2547 Bundle *bundle = node_bundling(n); 2548 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2549 tty->print("*** Bundle: "); 2550 bundle->dump(); 2551 } 2552 n->dump(); 2553 } 2554 } 2555 } 2556 #endif 2557 #ifdef ASSERT 2558 verify_good_schedule(bb,"after block local scheduling"); 2559 #endif 2560 } 2561 2562 #ifndef PRODUCT 2563 if (_cfg->C->trace_opto_output()) 2564 tty->print("# <- DoScheduling\n"); 2565 #endif 2566 2567 // Record final node-bundling array location 2568 _regalloc->C->set_node_bundling_base(_node_bundling_base); 2569 2570 } // end DoScheduling 2571 2572 //------------------------------verify_good_schedule--------------------------- 2573 // Verify that no live-range used in the block is killed in the block by a 2574 // wrong DEF. This doesn't verify live-ranges that span blocks. 2575 2576 // Check for edge existence. Used to avoid adding redundant precedence edges. 2577 static bool edge_from_to( Node *from, Node *to ) { 2578 for( uint i=0; i<from->len(); i++ ) 2579 if( from->in(i) == to ) 2580 return true; 2581 return false; 2582 } 2583 2584 #ifdef ASSERT 2585 //------------------------------verify_do_def---------------------------------- 2586 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2587 // Check for bad kills 2588 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2589 Node *prior_use = _reg_node[def]; 2590 if( prior_use && !edge_from_to(prior_use,n) ) { 2591 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2592 n->dump(); 2593 tty->print_cr("..."); 2594 prior_use->dump(); 2595 assert(edge_from_to(prior_use,n),msg); 2596 } 2597 _reg_node.map(def,NULL); // Kill live USEs 2598 } 2599 } 2600 2601 //------------------------------verify_good_schedule--------------------------- 2602 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2603 2604 // Zap to something reasonable for the verify code 2605 _reg_node.clear(); 2606 2607 // Walk over the block backwards. Check to make sure each DEF doesn't 2608 // kill a live value (other than the one it's supposed to). Add each 2609 // USE to the live set. 2610 for( uint i = b->_nodes.size()-1; i >= _bb_start; i-- ) { 2611 Node *n = b->_nodes[i]; 2612 int n_op = n->Opcode(); 2613 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2614 // Fat-proj kills a slew of registers 2615 RegMask rm = n->out_RegMask();// Make local copy 2616 while( rm.is_NotEmpty() ) { 2617 OptoReg::Name kill = rm.find_first_elem(); 2618 rm.Remove(kill); 2619 verify_do_def( n, kill, msg ); 2620 } 2621 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2622 // Get DEF'd registers the normal way 2623 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2624 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2625 } 2626 2627 // Now make all USEs live 2628 for( uint i=1; i<n->req(); i++ ) { 2629 Node *def = n->in(i); 2630 assert(def != 0, "input edge required"); 2631 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2632 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2633 if( OptoReg::is_valid(reg_lo) ) { 2634 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg); 2635 _reg_node.map(reg_lo,n); 2636 } 2637 if( OptoReg::is_valid(reg_hi) ) { 2638 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg); 2639 _reg_node.map(reg_hi,n); 2640 } 2641 } 2642 2643 } 2644 2645 // Zap to something reasonable for the Antidependence code 2646 _reg_node.clear(); 2647 } 2648 #endif 2649 2650 // Conditionally add precedence edges. Avoid putting edges on Projs. 2651 static void add_prec_edge_from_to( Node *from, Node *to ) { 2652 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2653 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2654 from = from->in(0); 2655 } 2656 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2657 !edge_from_to( from, to ) ) // Avoid duplicate edge 2658 from->add_prec(to); 2659 } 2660 2661 //------------------------------anti_do_def------------------------------------ 2662 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2663 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2664 return; 2665 2666 Node *pinch = _reg_node[def_reg]; // Get pinch point 2667 if( !pinch || _bbs[pinch->_idx] != b || // No pinch-point yet? 2668 is_def ) { // Check for a true def (not a kill) 2669 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2670 return; 2671 } 2672 2673 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2674 debug_only( def = (Node*)0xdeadbeef; ) 2675 2676 // After some number of kills there _may_ be a later def 2677 Node *later_def = NULL; 2678 2679 // Finding a kill requires a real pinch-point. 2680 // Check for not already having a pinch-point. 2681 // Pinch points are Op_Node's. 2682 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2683 later_def = pinch; // Must be def/kill as optimistic pinch-point 2684 if ( _pinch_free_list.size() > 0) { 2685 pinch = _pinch_free_list.pop(); 2686 } else { 2687 pinch = new (_cfg->C) Node(1); // Pinch point to-be 2688 } 2689 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2690 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2691 return; 2692 } 2693 _bbs.map(pinch->_idx,b); // Pretend it's valid in this block (lazy init) 2694 _reg_node.map(def_reg,pinch); // Record pinch-point 2695 //_regalloc->set_bad(pinch->_idx); // Already initialized this way. 2696 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2697 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call 2698 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2699 later_def = NULL; // and no later def 2700 } 2701 pinch->set_req(0,later_def); // Hook later def so we can find it 2702 } else { // Else have valid pinch point 2703 if( pinch->in(0) ) // If there is a later-def 2704 later_def = pinch->in(0); // Get it 2705 } 2706 2707 // Add output-dependence edge from later def to kill 2708 if( later_def ) // If there is some original def 2709 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2710 2711 // See if current kill is also a use, and so is forced to be the pinch-point. 2712 if( pinch->Opcode() == Op_Node ) { 2713 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2714 for( uint i=1; i<uses->req(); i++ ) { 2715 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2716 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2717 // Yes, found a use/kill pinch-point 2718 pinch->set_req(0,NULL); // 2719 pinch->replace_by(kill); // Move anti-dep edges up 2720 pinch = kill; 2721 _reg_node.map(def_reg,pinch); 2722 return; 2723 } 2724 } 2725 } 2726 2727 // Add edge from kill to pinch-point 2728 add_prec_edge_from_to(kill,pinch); 2729 } 2730 2731 //------------------------------anti_do_use------------------------------------ 2732 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2733 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2734 return; 2735 Node *pinch = _reg_node[use_reg]; // Get pinch point 2736 // Check for no later def_reg/kill in block 2737 if( pinch && _bbs[pinch->_idx] == b && 2738 // Use has to be block-local as well 2739 _bbs[use->_idx] == b ) { 2740 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2741 pinch->req() == 1 ) { // pinch not yet in block? 2742 pinch->del_req(0); // yank pointer to later-def, also set flag 2743 // Insert the pinch-point in the block just after the last use 2744 b->_nodes.insert(b->find_node(use)+1,pinch); 2745 _bb_end++; // Increase size scheduled region in block 2746 } 2747 2748 add_prec_edge_from_to(pinch,use); 2749 } 2750 } 2751 2752 //------------------------------ComputeRegisterAntidependences----------------- 2753 // We insert antidependences between the reads and following write of 2754 // allocated registers to prevent illegal code motion. Hopefully, the 2755 // number of added references should be fairly small, especially as we 2756 // are only adding references within the current basic block. 2757 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2758 2759 #ifdef ASSERT 2760 verify_good_schedule(b,"before block local scheduling"); 2761 #endif 2762 2763 // A valid schedule, for each register independently, is an endless cycle 2764 // of: a def, then some uses (connected to the def by true dependencies), 2765 // then some kills (defs with no uses), finally the cycle repeats with a new 2766 // def. The uses are allowed to float relative to each other, as are the 2767 // kills. No use is allowed to slide past a kill (or def). This requires 2768 // antidependencies between all uses of a single def and all kills that 2769 // follow, up to the next def. More edges are redundant, because later defs 2770 // & kills are already serialized with true or antidependencies. To keep 2771 // the edge count down, we add a 'pinch point' node if there's more than 2772 // one use or more than one kill/def. 2773 2774 // We add dependencies in one bottom-up pass. 2775 2776 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2777 2778 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2779 // register. If not, we record the DEF/KILL in _reg_node, the 2780 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2781 // "pinch point", a new Node that's in the graph but not in the block. 2782 // We put edges from the prior and current DEF/KILLs to the pinch point. 2783 // We put the pinch point in _reg_node. If there's already a pinch point 2784 // we merely add an edge from the current DEF/KILL to the pinch point. 2785 2786 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2787 // put an edge from the pinch point to the USE. 2788 2789 // To be expedient, the _reg_node array is pre-allocated for the whole 2790 // compilation. _reg_node is lazily initialized; it either contains a NULL, 2791 // or a valid def/kill/pinch-point, or a leftover node from some prior 2792 // block. Leftover node from some prior block is treated like a NULL (no 2793 // prior def, so no anti-dependence needed). Valid def is distinguished by 2794 // it being in the current block. 2795 bool fat_proj_seen = false; 2796 uint last_safept = _bb_end-1; 2797 Node* end_node = (_bb_end-1 >= _bb_start) ? b->_nodes[last_safept] : NULL; 2798 Node* last_safept_node = end_node; 2799 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2800 Node *n = b->_nodes[i]; 2801 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2802 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) { 2803 // Fat-proj kills a slew of registers 2804 // This can add edges to 'n' and obscure whether or not it was a def, 2805 // hence the is_def flag. 2806 fat_proj_seen = true; 2807 RegMask rm = n->out_RegMask();// Make local copy 2808 while( rm.is_NotEmpty() ) { 2809 OptoReg::Name kill = rm.find_first_elem(); 2810 rm.Remove(kill); 2811 anti_do_def( b, n, kill, is_def ); 2812 } 2813 } else { 2814 // Get DEF'd registers the normal way 2815 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2816 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2817 } 2818 2819 // Kill projections on a branch should appear to occur on the 2820 // branch, not afterwards, so grab the masks from the projections 2821 // and process them. 2822 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) { 2823 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2824 Node* use = n->fast_out(i); 2825 if (use->is_Proj()) { 2826 RegMask rm = use->out_RegMask();// Make local copy 2827 while( rm.is_NotEmpty() ) { 2828 OptoReg::Name kill = rm.find_first_elem(); 2829 rm.Remove(kill); 2830 anti_do_def( b, n, kill, false ); 2831 } 2832 } 2833 } 2834 } 2835 2836 // Check each register used by this instruction for a following DEF/KILL 2837 // that must occur afterward and requires an anti-dependence edge. 2838 for( uint j=0; j<n->req(); j++ ) { 2839 Node *def = n->in(j); 2840 if( def ) { 2841 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2842 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2843 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2844 } 2845 } 2846 // Do not allow defs of new derived values to float above GC 2847 // points unless the base is definitely available at the GC point. 2848 2849 Node *m = b->_nodes[i]; 2850 2851 // Add precedence edge from following safepoint to use of derived pointer 2852 if( last_safept_node != end_node && 2853 m != last_safept_node) { 2854 for (uint k = 1; k < m->req(); k++) { 2855 const Type *t = m->in(k)->bottom_type(); 2856 if( t->isa_oop_ptr() && 2857 t->is_ptr()->offset() != 0 ) { 2858 last_safept_node->add_prec( m ); 2859 break; 2860 } 2861 } 2862 } 2863 2864 if( n->jvms() ) { // Precedence edge from derived to safept 2865 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2866 if( b->_nodes[last_safept] != last_safept_node ) { 2867 last_safept = b->find_node(last_safept_node); 2868 } 2869 for( uint j=last_safept; j > i; j-- ) { 2870 Node *mach = b->_nodes[j]; 2871 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2872 mach->add_prec( n ); 2873 } 2874 last_safept = i; 2875 last_safept_node = m; 2876 } 2877 } 2878 2879 if (fat_proj_seen) { 2880 // Garbage collect pinch nodes that were not consumed. 2881 // They are usually created by a fat kill MachProj for a call. 2882 garbage_collect_pinch_nodes(); 2883 } 2884 } 2885 2886 //------------------------------garbage_collect_pinch_nodes------------------------------- 2887 2888 // Garbage collect pinch nodes for reuse by other blocks. 2889 // 2890 // The block scheduler's insertion of anti-dependence 2891 // edges creates many pinch nodes when the block contains 2892 // 2 or more Calls. A pinch node is used to prevent a 2893 // combinatorial explosion of edges. If a set of kills for a 2894 // register is anti-dependent on a set of uses (or defs), rather 2895 // than adding an edge in the graph between each pair of kill 2896 // and use (or def), a pinch is inserted between them: 2897 // 2898 // use1 use2 use3 2899 // \ | / 2900 // \ | / 2901 // pinch 2902 // / | \ 2903 // / | \ 2904 // kill1 kill2 kill3 2905 // 2906 // One pinch node is created per register killed when 2907 // the second call is encountered during a backwards pass 2908 // over the block. Most of these pinch nodes are never 2909 // wired into the graph because the register is never 2910 // used or def'ed in the block. 2911 // 2912 void Scheduling::garbage_collect_pinch_nodes() { 2913 #ifndef PRODUCT 2914 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 2915 #endif 2916 int trace_cnt = 0; 2917 for (uint k = 0; k < _reg_node.Size(); k++) { 2918 Node* pinch = _reg_node[k]; 2919 if (pinch != NULL && pinch->Opcode() == Op_Node && 2920 // no predecence input edges 2921 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { 2922 cleanup_pinch(pinch); 2923 _pinch_free_list.push(pinch); 2924 _reg_node.map(k, NULL); 2925 #ifndef PRODUCT 2926 if (_cfg->C->trace_opto_output()) { 2927 trace_cnt++; 2928 if (trace_cnt > 40) { 2929 tty->print("\n"); 2930 trace_cnt = 0; 2931 } 2932 tty->print(" %d", pinch->_idx); 2933 } 2934 #endif 2935 } 2936 } 2937 #ifndef PRODUCT 2938 if (_cfg->C->trace_opto_output()) tty->print("\n"); 2939 #endif 2940 } 2941 2942 // Clean up a pinch node for reuse. 2943 void Scheduling::cleanup_pinch( Node *pinch ) { 2944 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 2945 2946 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 2947 Node* use = pinch->last_out(i); 2948 uint uses_found = 0; 2949 for (uint j = use->req(); j < use->len(); j++) { 2950 if (use->in(j) == pinch) { 2951 use->rm_prec(j); 2952 uses_found++; 2953 } 2954 } 2955 assert(uses_found > 0, "must be a precedence edge"); 2956 i -= uses_found; // we deleted 1 or more copies of this edge 2957 } 2958 // May have a later_def entry 2959 pinch->set_req(0, NULL); 2960 } 2961 2962 //------------------------------print_statistics------------------------------- 2963 #ifndef PRODUCT 2964 2965 void Scheduling::dump_available() const { 2966 tty->print("#Availist "); 2967 for (uint i = 0; i < _available.size(); i++) 2968 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 2969 tty->cr(); 2970 } 2971 2972 // Print Scheduling Statistics 2973 void Scheduling::print_statistics() { 2974 // Print the size added by nops for bundling 2975 tty->print("Nops added %d bytes to total of %d bytes", 2976 _total_nop_size, _total_method_size); 2977 if (_total_method_size > 0) 2978 tty->print(", for %.2f%%", 2979 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 2980 tty->print("\n"); 2981 2982 // Print the number of branch shadows filled 2983 if (Pipeline::_branch_has_delay_slot) { 2984 tty->print("Of %d branches, %d had unconditional delay slots filled", 2985 _total_branches, _total_unconditional_delays); 2986 if (_total_branches > 0) 2987 tty->print(", for %.2f%%", 2988 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 2989 tty->print("\n"); 2990 } 2991 2992 uint total_instructions = 0, total_bundles = 0; 2993 2994 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 2995 uint bundle_count = _total_instructions_per_bundle[i]; 2996 total_instructions += bundle_count * i; 2997 total_bundles += bundle_count; 2998 } 2999 3000 if (total_bundles > 0) 3001 tty->print("Average ILP (excluding nops) is %.2f\n", 3002 ((double)total_instructions) / ((double)total_bundles)); 3003 } 3004 #endif