1 /* 2 * Copyright (c) 1998, 2011, 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); 348 349 // Collect worst case block paddings 350 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks); 351 memset(block_worst_case_pad, 0, nblocks * sizeof(int)); 352 353 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); ) 354 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); ) 355 356 bool has_short_branch_candidate = false; 357 358 // Initialize the sizes to 0 359 code_size = 0; // Size in bytes of generated code 360 stub_size = 0; // Size in bytes of all stub entries 361 // Size in bytes of all relocation entries, including those in local stubs. 362 // Start with 2-bytes of reloc info for the unvalidated entry point 363 reloc_size = 1; // Number of relocation entries 364 365 // Make three passes. The first computes pessimistic blk_starts, 366 // relative jmp_offset and reloc_size information. The second performs 367 // short branch substitution using the pessimistic sizing. The 368 // third inserts nops where needed. 369 370 // Step one, perform a pessimistic sizing pass. 371 uint last_call_adr = max_uint; 372 uint last_avoid_back_to_back_adr = max_uint; 373 uint nop_size = (new (this) MachNopNode())->size(_regalloc); 374 for (uint i = 0; i < nblocks; i++) { // For all blocks 375 Block *b = _cfg->_blocks[i]; 376 377 // During short branch replacement, we store the relative (to blk_starts) 378 // offset of jump in jmp_offset, rather than the absolute offset of jump. 379 // This is so that we do not need to recompute sizes of all nodes when 380 // we compute correct blk_starts in our next sizing pass. 381 jmp_offset[i] = 0; 382 jmp_size[i] = 0; 383 jmp_nidx[i] = -1; 384 DEBUG_ONLY( jmp_target[i] = 0; ) 385 DEBUG_ONLY( jmp_rule[i] = 0; ) 386 387 // Sum all instruction sizes to compute block size 388 uint last_inst = b->_nodes.size(); 389 uint blk_size = 0; 390 for (uint j = 0; j < last_inst; j++) { 391 Node* nj = b->_nodes[j]; 392 // Handle machine instruction nodes 393 if (nj->is_Mach()) { 394 MachNode *mach = nj->as_Mach(); 395 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 396 reloc_size += mach->reloc(); 397 if( mach->is_MachCall() ) { 398 MachCallNode *mcall = mach->as_MachCall(); 399 // This destination address is NOT PC-relative 400 401 mcall->method_set((intptr_t)mcall->entry_point()); 402 403 if( mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method ) { 404 stub_size += size_java_to_interp(); 405 reloc_size += reloc_java_to_interp(); 406 } 407 } else if (mach->is_MachSafePoint()) { 408 // If call/safepoint are adjacent, account for possible 409 // nop to disambiguate the two safepoints. 410 // ScheduleAndBundle() can rearrange nodes in a block, 411 // check for all offsets inside this block. 412 if (last_call_adr >= blk_starts[i]) { 413 blk_size += nop_size; 414 } 415 } 416 if (mach->avoid_back_to_back()) { 417 // Nop is inserted between "avoid back to back" instructions. 418 // ScheduleAndBundle() can rearrange nodes in a block, 419 // check for all offsets inside this block. 420 if (last_avoid_back_to_back_adr >= blk_starts[i]) { 421 blk_size += nop_size; 422 } 423 } 424 if (mach->may_be_short_branch()) { 425 if (!nj->is_MachBranch()) { 426 #ifndef PRODUCT 427 nj->dump(3); 428 #endif 429 Unimplemented(); 430 } 431 assert(jmp_nidx[i] == -1, "block should have only one branch"); 432 jmp_offset[i] = blk_size; 433 jmp_size[i] = nj->size(_regalloc); 434 jmp_nidx[i] = j; 435 has_short_branch_candidate = true; 436 } 437 } 438 blk_size += nj->size(_regalloc); 439 // Remember end of call offset 440 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) { 441 last_call_adr = blk_starts[i]+blk_size; 442 } 443 // Remember end of avoid_back_to_back offset 444 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back()) { 445 last_avoid_back_to_back_adr = blk_starts[i]+blk_size; 446 } 447 } 448 449 // When the next block starts a loop, we may insert pad NOP 450 // instructions. Since we cannot know our future alignment, 451 // assume the worst. 452 if (i< nblocks-1) { 453 Block *nb = _cfg->_blocks[i+1]; 454 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 455 if (max_loop_pad > 0) { 456 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 457 // Adjust last_call_adr and/or last_avoid_back_to_back_adr. 458 // If either is the last instruction in this block, bump by 459 // max_loop_pad in lock-step with blk_size, so sizing 460 // calculations in subsequent blocks still can conservatively 461 // detect that it may the last instruction in this block. 462 if (last_call_adr == blk_starts[i]+blk_size) { 463 last_call_adr += max_loop_pad; 464 } 465 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) { 466 last_avoid_back_to_back_adr += max_loop_pad; 467 } 468 blk_size += max_loop_pad; 469 block_worst_case_pad[i + 1] = max_loop_pad; 470 } 471 } 472 473 // Save block size; update total method size 474 blk_starts[i+1] = blk_starts[i]+blk_size; 475 } 476 477 // Step two, replace eligible long jumps. 478 bool progress = true; 479 uint last_may_be_short_branch_adr = max_uint; 480 while (has_short_branch_candidate && progress) { 481 progress = false; 482 has_short_branch_candidate = false; 483 int adjust_block_start = 0; 484 for (uint i = 0; i < nblocks; i++) { 485 Block *b = _cfg->_blocks[i]; 486 int idx = jmp_nidx[i]; 487 MachNode* mach = (idx == -1) ? NULL: b->_nodes[idx]->as_Mach(); 488 if (mach != NULL && mach->may_be_short_branch()) { 489 #ifdef ASSERT 490 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity"); 491 int j; 492 // Find the branch; ignore trailing NOPs. 493 for (j = b->_nodes.size()-1; j>=0; j--) { 494 Node* n = b->_nodes[j]; 495 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) 496 break; 497 } 498 assert(j >= 0 && j == idx && b->_nodes[j] == (Node*)mach, "sanity"); 499 #endif 500 int br_size = jmp_size[i]; 501 int br_offs = blk_starts[i] + jmp_offset[i]; 502 503 // This requires the TRUE branch target be in succs[0] 504 uint bnum = b->non_connector_successor(0)->_pre_order; 505 int offset = blk_starts[bnum] - br_offs; 506 if (bnum > i) { // adjust following block's offset 507 offset -= adjust_block_start; 508 } 509 510 // This block can be a loop header, account for the padding 511 // in the previous block. 512 int block_padding = block_worst_case_pad[i]; 513 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top"); 514 // In the following code a nop could be inserted before 515 // the branch which will increase the backward distance. 516 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr); 517 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block"); 518 519 if (needs_padding && offset <= 0) 520 offset -= nop_size; 521 522 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 523 // We've got a winner. Replace this branch. 524 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this); 525 526 // Update the jmp_size. 527 int new_size = replacement->size(_regalloc); 528 int diff = br_size - new_size; 529 assert(diff >= (int)nop_size, "short_branch size should be smaller"); 530 // Conservatively take into accound padding between 531 // avoid_back_to_back branches. Previous branch could be 532 // converted into avoid_back_to_back branch during next 533 // rounds. 534 if (needs_padding && replacement->avoid_back_to_back()) { 535 jmp_offset[i] += nop_size; 536 diff -= nop_size; 537 } 538 adjust_block_start += diff; 539 b->_nodes.map(idx, replacement); 540 mach->subsume_by(replacement, C); 541 mach = replacement; 542 progress = true; 543 544 jmp_size[i] = new_size; 545 DEBUG_ONLY( jmp_target[i] = bnum; ); 546 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 547 } else { 548 // The jump distance is not short, try again during next iteration. 549 has_short_branch_candidate = true; 550 } 551 } // (mach->may_be_short_branch()) 552 if (mach != NULL && (mach->may_be_short_branch() || 553 mach->avoid_back_to_back())) { 554 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i]; 555 } 556 blk_starts[i+1] -= adjust_block_start; 557 } 558 } 559 560 #ifdef ASSERT 561 for (uint i = 0; i < nblocks; i++) { // For all blocks 562 if (jmp_target[i] != 0) { 563 int br_size = jmp_size[i]; 564 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 565 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 566 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]); 567 } 568 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp"); 569 } 570 } 571 #endif 572 573 // Step 3, compute the offsets of all blocks, will be done in fill_buffer() 574 // after ScheduleAndBundle(). 575 576 // ------------------ 577 // Compute size for code buffer 578 code_size = blk_starts[nblocks]; 579 580 // Relocation records 581 reloc_size += 1; // Relo entry for exception handler 582 583 // Adjust reloc_size to number of record of relocation info 584 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 585 // a relocation index. 586 // The CodeBuffer will expand the locs array if this estimate is too low. 587 reloc_size *= 10 / sizeof(relocInfo); 588 } 589 590 //------------------------------FillLocArray----------------------------------- 591 // Create a bit of debug info and append it to the array. The mapping is from 592 // Java local or expression stack to constant, register or stack-slot. For 593 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 594 // entry has been taken care of and caller should skip it). 595 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 596 // This should never have accepted Bad before 597 assert(OptoReg::is_valid(regnum), "location must be valid"); 598 return (OptoReg::is_reg(regnum)) 599 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 600 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 601 } 602 603 604 ObjectValue* 605 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 606 for (int i = 0; i < objs->length(); i++) { 607 assert(objs->at(i)->is_object(), "corrupt object cache"); 608 ObjectValue* sv = (ObjectValue*) objs->at(i); 609 if (sv->id() == id) { 610 return sv; 611 } 612 } 613 // Otherwise.. 614 return NULL; 615 } 616 617 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 618 ObjectValue* sv ) { 619 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition"); 620 objs->append(sv); 621 } 622 623 624 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 625 GrowableArray<ScopeValue*> *array, 626 GrowableArray<ScopeValue*> *objs ) { 627 assert( local, "use _top instead of null" ); 628 if (array->length() != idx) { 629 assert(array->length() == idx + 1, "Unexpected array count"); 630 // Old functionality: 631 // return 632 // New functionality: 633 // Assert if the local is not top. In product mode let the new node 634 // override the old entry. 635 assert(local == top(), "LocArray collision"); 636 if (local == top()) { 637 return; 638 } 639 array->pop(); 640 } 641 const Type *t = local->bottom_type(); 642 643 // Is it a safepoint scalar object node? 644 if (local->is_SafePointScalarObject()) { 645 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 646 647 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx); 648 if (sv == NULL) { 649 ciKlass* cik = t->is_oopptr()->klass(); 650 assert(cik->is_instance_klass() || 651 cik->is_array_klass(), "Not supported allocation."); 652 sv = new ObjectValue(spobj->_idx, 653 new ConstantOopWriteValue(cik->constant_encoding())); 654 Compile::set_sv_for_object_node(objs, sv); 655 656 uint first_ind = spobj->first_index(); 657 for (uint i = 0; i < spobj->n_fields(); i++) { 658 Node* fld_node = sfpt->in(first_ind+i); 659 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 660 } 661 } 662 array->append(sv); 663 return; 664 } 665 666 // Grab the register number for the local 667 OptoReg::Name regnum = _regalloc->get_reg_first(local); 668 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 669 // Record the double as two float registers. 670 // The register mask for such a value always specifies two adjacent 671 // float registers, with the lower register number even. 672 // Normally, the allocation of high and low words to these registers 673 // is irrelevant, because nearly all operations on register pairs 674 // (e.g., StoreD) treat them as a single unit. 675 // Here, we assume in addition that the words in these two registers 676 // stored "naturally" (by operations like StoreD and double stores 677 // within the interpreter) such that the lower-numbered register 678 // is written to the lower memory address. This may seem like 679 // a machine dependency, but it is not--it is a requirement on 680 // the author of the <arch>.ad file to ensure that, for every 681 // even/odd double-register pair to which a double may be allocated, 682 // the word in the even single-register is stored to the first 683 // memory word. (Note that register numbers are completely 684 // arbitrary, and are not tied to any machine-level encodings.) 685 #ifdef _LP64 686 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 687 array->append(new ConstantIntValue(0)); 688 array->append(new_loc_value( _regalloc, regnum, Location::dbl )); 689 } else if ( t->base() == Type::Long ) { 690 array->append(new ConstantIntValue(0)); 691 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 692 } else if ( t->base() == Type::RawPtr ) { 693 // jsr/ret return address which must be restored into a the full 694 // width 64-bit stack slot. 695 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 696 } 697 #else //_LP64 698 #ifdef SPARC 699 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) { 700 // For SPARC we have to swap high and low words for 701 // long values stored in a single-register (g0-g7). 702 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 703 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 704 } else 705 #endif //SPARC 706 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 707 // Repack the double/long as two jints. 708 // The convention the interpreter uses is that the second local 709 // holds the first raw word of the native double representation. 710 // This is actually reasonable, since locals and stack arrays 711 // grow downwards in all implementations. 712 // (If, on some machine, the interpreter's Java locals or stack 713 // were to grow upwards, the embedded doubles would be word-swapped.) 714 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 715 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 716 } 717 #endif //_LP64 718 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 719 OptoReg::is_reg(regnum) ) { 720 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double() 721 ? Location::float_in_dbl : Location::normal )); 722 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 723 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long 724 ? Location::int_in_long : Location::normal )); 725 } else if( t->base() == Type::NarrowOop ) { 726 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop )); 727 } else { 728 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal )); 729 } 730 return; 731 } 732 733 // No register. It must be constant data. 734 switch (t->base()) { 735 case Type::Half: // Second half of a double 736 ShouldNotReachHere(); // Caller should skip 2nd halves 737 break; 738 case Type::AnyPtr: 739 array->append(new ConstantOopWriteValue(NULL)); 740 break; 741 case Type::AryPtr: 742 case Type::InstPtr: 743 case Type::KlassPtr: // fall through 744 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 745 break; 746 case Type::NarrowOop: 747 if (t == TypeNarrowOop::NULL_PTR) { 748 array->append(new ConstantOopWriteValue(NULL)); 749 } else { 750 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 751 } 752 break; 753 case Type::Int: 754 array->append(new ConstantIntValue(t->is_int()->get_con())); 755 break; 756 case Type::RawPtr: 757 // A return address (T_ADDRESS). 758 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 759 #ifdef _LP64 760 // Must be restored to the full-width 64-bit stack slot. 761 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 762 #else 763 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 764 #endif 765 break; 766 case Type::FloatCon: { 767 float f = t->is_float_constant()->getf(); 768 array->append(new ConstantIntValue(jint_cast(f))); 769 break; 770 } 771 case Type::DoubleCon: { 772 jdouble d = t->is_double_constant()->getd(); 773 #ifdef _LP64 774 array->append(new ConstantIntValue(0)); 775 array->append(new ConstantDoubleValue(d)); 776 #else 777 // Repack the double as two jints. 778 // The convention the interpreter uses is that the second local 779 // holds the first raw word of the native double representation. 780 // This is actually reasonable, since locals and stack arrays 781 // grow downwards in all implementations. 782 // (If, on some machine, the interpreter's Java locals or stack 783 // were to grow upwards, the embedded doubles would be word-swapped.) 784 jint *dp = (jint*)&d; 785 array->append(new ConstantIntValue(dp[1])); 786 array->append(new ConstantIntValue(dp[0])); 787 #endif 788 break; 789 } 790 case Type::Long: { 791 jlong d = t->is_long()->get_con(); 792 #ifdef _LP64 793 array->append(new ConstantIntValue(0)); 794 array->append(new ConstantLongValue(d)); 795 #else 796 // Repack the long as two jints. 797 // The convention the interpreter uses is that the second local 798 // holds the first raw word of the native double representation. 799 // This is actually reasonable, since locals and stack arrays 800 // grow downwards in all implementations. 801 // (If, on some machine, the interpreter's Java locals or stack 802 // were to grow upwards, the embedded doubles would be word-swapped.) 803 jint *dp = (jint*)&d; 804 array->append(new ConstantIntValue(dp[1])); 805 array->append(new ConstantIntValue(dp[0])); 806 #endif 807 break; 808 } 809 case Type::Top: // Add an illegal value here 810 array->append(new LocationValue(Location())); 811 break; 812 default: 813 ShouldNotReachHere(); 814 break; 815 } 816 } 817 818 // Determine if this node starts a bundle 819 bool Compile::starts_bundle(const Node *n) const { 820 return (_node_bundling_limit > n->_idx && 821 _node_bundling_base[n->_idx].starts_bundle()); 822 } 823 824 //--------------------------Process_OopMap_Node-------------------------------- 825 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) { 826 827 // Handle special safepoint nodes for synchronization 828 MachSafePointNode *sfn = mach->as_MachSafePoint(); 829 MachCallNode *mcall; 830 831 #ifdef ENABLE_ZAP_DEAD_LOCALS 832 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative"); 833 #endif 834 835 int safepoint_pc_offset = current_offset; 836 bool is_method_handle_invoke = false; 837 bool return_oop = false; 838 839 // Add the safepoint in the DebugInfoRecorder 840 if( !mach->is_MachCall() ) { 841 mcall = NULL; 842 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 843 } else { 844 mcall = mach->as_MachCall(); 845 846 // Is the call a MethodHandle call? 847 if (mcall->is_MachCallJava()) { 848 if (mcall->as_MachCallJava()->_method_handle_invoke) { 849 assert(has_method_handle_invokes(), "must have been set during call generation"); 850 is_method_handle_invoke = true; 851 } 852 } 853 854 // Check if a call returns an object. 855 if (mcall->return_value_is_used() && 856 mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) { 857 return_oop = true; 858 } 859 safepoint_pc_offset += mcall->ret_addr_offset(); 860 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 861 } 862 863 // Loop over the JVMState list to add scope information 864 // Do not skip safepoints with a NULL method, they need monitor info 865 JVMState* youngest_jvms = sfn->jvms(); 866 int max_depth = youngest_jvms->depth(); 867 868 // Allocate the object pool for scalar-replaced objects -- the map from 869 // small-integer keys (which can be recorded in the local and ostack 870 // arrays) to descriptions of the object state. 871 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 872 873 // Visit scopes from oldest to youngest. 874 for (int depth = 1; depth <= max_depth; depth++) { 875 JVMState* jvms = youngest_jvms->of_depth(depth); 876 int idx; 877 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 878 // Safepoints that do not have method() set only provide oop-map and monitor info 879 // to support GC; these do not support deoptimization. 880 int num_locs = (method == NULL) ? 0 : jvms->loc_size(); 881 int num_exps = (method == NULL) ? 0 : jvms->stk_size(); 882 int num_mon = jvms->nof_monitors(); 883 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), 884 "JVMS local count must match that of the method"); 885 886 // Add Local and Expression Stack Information 887 888 // Insert locals into the locarray 889 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 890 for( idx = 0; idx < num_locs; idx++ ) { 891 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 892 } 893 894 // Insert expression stack entries into the exparray 895 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 896 for( idx = 0; idx < num_exps; idx++ ) { 897 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 898 } 899 900 // Add in mappings of the monitors 901 assert( !method || 902 !method->is_synchronized() || 903 method->is_native() || 904 num_mon > 0 || 905 !GenerateSynchronizationCode, 906 "monitors must always exist for synchronized methods"); 907 908 // Build the growable array of ScopeValues for exp stack 909 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 910 911 // Loop over monitors and insert into array 912 for(idx = 0; idx < num_mon; idx++) { 913 // Grab the node that defines this monitor 914 Node* box_node = sfn->monitor_box(jvms, idx); 915 Node* obj_node = sfn->monitor_obj(jvms, idx); 916 917 // Create ScopeValue for object 918 ScopeValue *scval = NULL; 919 920 if( obj_node->is_SafePointScalarObject() ) { 921 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 922 scval = Compile::sv_for_node_id(objs, spobj->_idx); 923 if (scval == NULL) { 924 const Type *t = obj_node->bottom_type(); 925 ciKlass* cik = t->is_oopptr()->klass(); 926 assert(cik->is_instance_klass() || 927 cik->is_array_klass(), "Not supported allocation."); 928 ObjectValue* sv = new ObjectValue(spobj->_idx, 929 new ConstantOopWriteValue(cik->constant_encoding())); 930 Compile::set_sv_for_object_node(objs, sv); 931 932 uint first_ind = spobj->first_index(); 933 for (uint i = 0; i < spobj->n_fields(); i++) { 934 Node* fld_node = sfn->in(first_ind+i); 935 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 936 } 937 scval = sv; 938 } 939 } else if( !obj_node->is_Con() ) { 940 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); 941 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 942 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop ); 943 } else { 944 scval = new_loc_value( _regalloc, obj_reg, Location::oop ); 945 } 946 } else { 947 const TypePtr *tp = obj_node->get_ptr_type(); 948 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding()); 949 } 950 951 OptoReg::Name box_reg = BoxLockNode::reg(box_node); 952 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)); 953 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated()); 954 monarray->append(new MonitorValue(scval, basic_lock, eliminated)); 955 } 956 957 // We dump the object pool first, since deoptimization reads it in first. 958 debug_info()->dump_object_pool(objs); 959 960 // Build first class objects to pass to scope 961 DebugToken *locvals = debug_info()->create_scope_values(locarray); 962 DebugToken *expvals = debug_info()->create_scope_values(exparray); 963 DebugToken *monvals = debug_info()->create_monitor_values(monarray); 964 965 // Make method available for all Safepoints 966 ciMethod* scope_method = method ? method : _method; 967 // Describe the scope here 968 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 969 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 970 // Now we can describe the scope. 971 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals); 972 } // End jvms loop 973 974 // Mark the end of the scope set. 975 debug_info()->end_safepoint(safepoint_pc_offset); 976 } 977 978 979 980 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 981 class NonSafepointEmitter { 982 Compile* C; 983 JVMState* _pending_jvms; 984 int _pending_offset; 985 986 void emit_non_safepoint(); 987 988 public: 989 NonSafepointEmitter(Compile* compile) { 990 this->C = compile; 991 _pending_jvms = NULL; 992 _pending_offset = 0; 993 } 994 995 void observe_instruction(Node* n, int pc_offset) { 996 if (!C->debug_info()->recording_non_safepoints()) return; 997 998 Node_Notes* nn = C->node_notes_at(n->_idx); 999 if (nn == NULL || nn->jvms() == NULL) return; 1000 if (_pending_jvms != NULL && 1001 _pending_jvms->same_calls_as(nn->jvms())) { 1002 // Repeated JVMS? Stretch it up here. 1003 _pending_offset = pc_offset; 1004 } else { 1005 if (_pending_jvms != NULL && 1006 _pending_offset < pc_offset) { 1007 emit_non_safepoint(); 1008 } 1009 _pending_jvms = NULL; 1010 if (pc_offset > C->debug_info()->last_pc_offset()) { 1011 // This is the only way _pending_jvms can become non-NULL: 1012 _pending_jvms = nn->jvms(); 1013 _pending_offset = pc_offset; 1014 } 1015 } 1016 } 1017 1018 // Stay out of the way of real safepoints: 1019 void observe_safepoint(JVMState* jvms, int pc_offset) { 1020 if (_pending_jvms != NULL && 1021 !_pending_jvms->same_calls_as(jvms) && 1022 _pending_offset < pc_offset) { 1023 emit_non_safepoint(); 1024 } 1025 _pending_jvms = NULL; 1026 } 1027 1028 void flush_at_end() { 1029 if (_pending_jvms != NULL) { 1030 emit_non_safepoint(); 1031 } 1032 _pending_jvms = NULL; 1033 } 1034 }; 1035 1036 void NonSafepointEmitter::emit_non_safepoint() { 1037 JVMState* youngest_jvms = _pending_jvms; 1038 int pc_offset = _pending_offset; 1039 1040 // Clear it now: 1041 _pending_jvms = NULL; 1042 1043 DebugInformationRecorder* debug_info = C->debug_info(); 1044 assert(debug_info->recording_non_safepoints(), "sanity"); 1045 1046 debug_info->add_non_safepoint(pc_offset); 1047 int max_depth = youngest_jvms->depth(); 1048 1049 // Visit scopes from oldest to youngest. 1050 for (int depth = 1; depth <= max_depth; depth++) { 1051 JVMState* jvms = youngest_jvms->of_depth(depth); 1052 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 1053 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 1054 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute()); 1055 } 1056 1057 // Mark the end of the scope set. 1058 debug_info->end_non_safepoint(pc_offset); 1059 } 1060 1061 1062 1063 // helper for fill_buffer bailout logic 1064 static void turn_off_compiler(Compile* C) { 1065 if (CodeCache::largest_free_block() >= CodeCacheMinimumFreeSpace*10) { 1066 // Do not turn off compilation if a single giant method has 1067 // blown the code cache size. 1068 C->record_failure("excessive request to CodeCache"); 1069 } else { 1070 // Let CompilerBroker disable further compilations. 1071 C->record_failure("CodeCache is full"); 1072 } 1073 } 1074 1075 1076 //------------------------------init_buffer------------------------------------ 1077 CodeBuffer* Compile::init_buffer(uint* blk_starts) { 1078 1079 // Set the initially allocated size 1080 int code_req = initial_code_capacity; 1081 int locs_req = initial_locs_capacity; 1082 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity; 1083 int const_req = initial_const_capacity; 1084 1085 int pad_req = NativeCall::instruction_size; 1086 // The extra spacing after the code is necessary on some platforms. 1087 // Sometimes we need to patch in a jump after the last instruction, 1088 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 1089 1090 // Compute the byte offset where we can store the deopt pc. 1091 if (fixed_slots() != 0) { 1092 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 1093 } 1094 1095 // Compute prolog code size 1096 _method_size = 0; 1097 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; 1098 #if defined(IA64) && !defined(AIX) 1099 if (save_argument_registers()) { 1100 // 4815101: this is a stub with implicit and unknown precision fp args. 1101 // The usual spill mechanism can only generate stfd's in this case, which 1102 // doesn't work if the fp reg to spill contains a single-precision denorm. 1103 // Instead, we hack around the normal spill mechanism using stfspill's and 1104 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate 1105 // space here for the fp arg regs (f8-f15) we're going to thusly spill. 1106 // 1107 // If we ever implement 16-byte 'registers' == stack slots, we can 1108 // get rid of this hack and have SpillCopy generate stfspill/ldffill 1109 // instead of stfd/stfs/ldfd/ldfs. 1110 _frame_slots += 8*(16/BytesPerInt); 1111 } 1112 #endif 1113 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check"); 1114 1115 if (has_mach_constant_base_node()) { 1116 uint add_size = 0; 1117 // Fill the constant table. 1118 // Note: This must happen before shorten_branches. 1119 for (uint i = 0; i < _cfg->_num_blocks; i++) { 1120 Block* b = _cfg->_blocks[i]; 1121 1122 for (uint j = 0; j < b->_nodes.size(); j++) { 1123 Node* n = b->_nodes[j]; 1124 1125 // If the node is a MachConstantNode evaluate the constant 1126 // value section. 1127 if (n->is_MachConstant()) { 1128 MachConstantNode* machcon = n->as_MachConstant(); 1129 machcon->eval_constant(C); 1130 } else if (n->is_Mach()) { 1131 // On Power there are more nodes that issue constants. 1132 add_size += (n->as_Mach()->ins_num_consts() * 8); 1133 } 1134 } 1135 } 1136 1137 // Calculate the offsets of the constants and the size of the 1138 // constant table (including the padding to the next section). 1139 constant_table().calculate_offsets_and_size(); 1140 const_req = constant_table().size() + add_size; 1141 } 1142 1143 #if defined(ASSERT) && defined(PPC64) 1144 // Increase const_req by 40 for better behaviour of constant-pool overflow assertions below. 1145 const_req += 40; 1146 #endif 1147 1148 // Initialize the space for the BufferBlob used to find and verify 1149 // instruction size in MachNode::emit_size() 1150 init_scratch_buffer_blob(const_req); 1151 if (failing()) return NULL; // Out of memory 1152 1153 // Pre-compute the length of blocks and replace 1154 // long branches with short if machine supports it. 1155 shorten_branches(blk_starts, code_req, locs_req, stub_req); 1156 1157 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1158 int exception_handler_req = size_exception_handler(); 1159 int deopt_handler_req = size_deopt_handler(); 1160 exception_handler_req += MAX_stubs_size; // add marginal slop for handler 1161 deopt_handler_req += MAX_stubs_size; // add marginal slop for handler 1162 stub_req += MAX_stubs_size; // ensure per-stub margin 1163 code_req += MAX_inst_size; // ensure per-instruction margin 1164 1165 if (StressCodeBuffers) 1166 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1167 1168 int total_req = 1169 const_req + 1170 code_req + 1171 pad_req + 1172 stub_req + 1173 exception_handler_req + 1174 deopt_handler_req; // deopt handler 1175 1176 if (has_method_handle_invokes()) 1177 total_req += deopt_handler_req; // deopt MH handler 1178 1179 CodeBuffer* cb = code_buffer(); 1180 cb->initialize(total_req, locs_req); 1181 1182 // Have we run out of code space? 1183 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1184 turn_off_compiler(this); 1185 return NULL; 1186 } 1187 // Configure the code buffer. 1188 cb->initialize_consts_size(const_req); 1189 cb->initialize_stubs_size(stub_req); 1190 cb->initialize_oop_recorder(env()->oop_recorder()); 1191 1192 // fill in the nop array for bundling computations 1193 MachNode *_nop_list[Bundle::_nop_count]; 1194 Bundle::initialize_nops(_nop_list, this); 1195 1196 return cb; 1197 } 1198 1199 //------------------------------fill_buffer------------------------------------ 1200 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) { 1201 // blk_starts[] contains offsets calculated during short branches processing, 1202 // offsets should not be increased during following steps. 1203 1204 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 1205 // of a loop. It is used to determine the padding for loop alignment. 1206 compute_loop_first_inst_sizes(); 1207 1208 // Create oopmap set. 1209 _oop_map_set = new OopMapSet(); 1210 1211 // !!!!! This preserves old handling of oopmaps for now 1212 debug_info()->set_oopmaps(_oop_map_set); 1213 1214 uint nblocks = _cfg->_num_blocks; 1215 // Count and start of implicit null check instructions 1216 uint inct_cnt = 0; 1217 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1218 1219 // Count and start of calls 1220 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1221 1222 uint return_offset = 0; 1223 int nop_size = (new (this) MachNopNode())->size(_regalloc); 1224 1225 int previous_offset = 0; 1226 int current_offset = 0; 1227 int last_call_offset = -1; 1228 int last_avoid_back_to_back_offset = -1; 1229 #ifdef ASSERT 1230 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); 1231 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 1232 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 1233 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); 1234 #endif 1235 1236 // Create an array of unused labels, one for each basic block, if printing is enabled 1237 #ifndef PRODUCT 1238 int *node_offsets = NULL; 1239 uint node_offset_limit = unique(); 1240 1241 if (print_assembly()) 1242 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1243 #endif 1244 1245 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily 1246 1247 // Emit the constant table. 1248 if (has_mach_constant_base_node()) { 1249 constant_table().emit(*cb); 1250 } 1251 1252 // Create an array of labels, one for each basic block 1253 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1); 1254 for (uint i=0; i <= nblocks; i++) { 1255 blk_labels[i].init(); 1256 } 1257 1258 // ------------------ 1259 // Now fill in the code buffer 1260 Node *delay_slot = NULL; 1261 1262 for (uint i=0; i < nblocks; i++) { 1263 Block *b = _cfg->_blocks[i]; 1264 1265 Node *head = b->head(); 1266 1267 // If this block needs to start aligned (i.e, can be reached other 1268 // than by falling-thru from the previous block), then force the 1269 // start of a new bundle. 1270 if (Pipeline::requires_bundling() && starts_bundle(head)) 1271 cb->flush_bundle(true); 1272 1273 #ifdef ASSERT 1274 if (!b->is_connector()) { 1275 stringStream st; 1276 b->dump_head(&_cfg->_bbs, &st); 1277 MacroAssembler(cb).block_comment(st.as_string()); 1278 } 1279 jmp_target[i] = 0; 1280 jmp_offset[i] = 0; 1281 jmp_size[i] = 0; 1282 jmp_rule[i] = 0; 1283 #endif 1284 int blk_offset = current_offset; 1285 1286 // Define the label at the beginning of the basic block 1287 MacroAssembler(cb).bind(blk_labels[b->_pre_order]); 1288 1289 uint last_inst = b->_nodes.size(); 1290 1291 // Emit block normally, except for last instruction. 1292 // Emit means "dump code bits into code buffer". 1293 for (uint j = 0; j<last_inst; j++) { 1294 1295 // Get the node 1296 Node* n = b->_nodes[j]; 1297 1298 // See if delay slots are supported 1299 if (valid_bundle_info(n) && 1300 node_bundling(n)->used_in_unconditional_delay()) { 1301 assert(delay_slot == NULL, "no use of delay slot node"); 1302 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1303 1304 delay_slot = n; 1305 continue; 1306 } 1307 1308 // If this starts a new instruction group, then flush the current one 1309 // (but allow split bundles) 1310 if (Pipeline::requires_bundling() && starts_bundle(n)) 1311 cb->flush_bundle(false); 1312 1313 // The following logic is duplicated in the code ifdeffed for 1314 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It 1315 // should be factored out. Or maybe dispersed to the nodes? 1316 1317 // Special handling for SafePoint/Call Nodes 1318 bool is_mcall = false; 1319 if (n->is_Mach()) { 1320 MachNode *mach = n->as_Mach(); 1321 is_mcall = n->is_MachCall(); 1322 bool is_sfn = n->is_MachSafePoint(); 1323 1324 // If this requires all previous instructions be flushed, then do so 1325 if (is_sfn || is_mcall || mach->alignment_required() != 1) { 1326 cb->flush_bundle(true); 1327 current_offset = cb->insts_size(); 1328 } 1329 1330 // A padding may be needed again since a previous instruction 1331 // could be moved to delay slot. 1332 1333 // align the instruction if necessary 1334 int padding = mach->compute_padding(current_offset); 1335 // Make sure safepoint node for polling is distinct from a call's 1336 // return by adding a nop if needed. 1337 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) { 1338 padding = nop_size; 1339 } 1340 if (padding == 0 && mach->avoid_back_to_back() && 1341 current_offset == last_avoid_back_to_back_offset) { 1342 // Avoid back to back some instructions. 1343 padding = nop_size; 1344 } 1345 1346 if(padding > 0) { 1347 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1348 int nops_cnt = padding / nop_size; 1349 MachNode *nop = new (this) MachNopNode(nops_cnt); 1350 b->_nodes.insert(j++, nop); 1351 last_inst++; 1352 _cfg->_bbs.map( nop->_idx, b ); 1353 nop->emit(*cb, _regalloc); 1354 cb->flush_bundle(true); 1355 current_offset = cb->insts_size(); 1356 } 1357 1358 // Remember the start of the last call in a basic block 1359 if (is_mcall) { 1360 MachCallNode *mcall = mach->as_MachCall(); 1361 1362 // This destination address is NOT PC-relative 1363 mcall->method_set((intptr_t)mcall->entry_point()); 1364 1365 // Save the return address 1366 call_returns[b->_pre_order] = current_offset + mcall->ret_addr_offset(); 1367 1368 if (mcall->is_MachCallLeaf()) { 1369 is_mcall = false; 1370 is_sfn = false; 1371 } 1372 } 1373 1374 // sfn will be valid whenever mcall is valid now because of inheritance 1375 if (is_sfn || is_mcall) { 1376 1377 // Handle special safepoint nodes for synchronization 1378 if (!is_mcall) { 1379 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1380 // !!!!! Stubs only need an oopmap right now, so bail out 1381 if (sfn->jvms()->method() == NULL) { 1382 // Write the oopmap directly to the code blob??!! 1383 # ifdef ENABLE_ZAP_DEAD_LOCALS 1384 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive"); 1385 # endif 1386 continue; 1387 } 1388 } // End synchronization 1389 1390 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1391 current_offset); 1392 Process_OopMap_Node(mach, current_offset); 1393 } // End if safepoint 1394 1395 // If this is a null check, then add the start of the previous instruction to the list 1396 else if( mach->is_MachNullCheck() ) { 1397 inct_starts[inct_cnt++] = previous_offset; 1398 } 1399 1400 // If this is a branch, then fill in the label with the target BB's label 1401 else if (mach->is_MachBranch()) { 1402 // This requires the TRUE branch target be in succs[0] 1403 uint block_num = b->non_connector_successor(0)->_pre_order; 1404 1405 // Try to replace long branch if delay slot is not used, 1406 // it is mostly for back branches since forward branch's 1407 // distance is not updated yet. 1408 bool delay_slot_is_used = valid_bundle_info(n) && 1409 node_bundling(n)->use_unconditional_delay(); 1410 if (!delay_slot_is_used && mach->may_be_short_branch()) { 1411 assert(delay_slot == NULL, "not expecting delay slot node"); 1412 int br_size = n->size(_regalloc); 1413 int offset = blk_starts[block_num] - current_offset; 1414 if (block_num >= i) { 1415 // Current and following block's offset are not 1416 // finalized yet, adjust distance by the difference 1417 // between calculated and final offsets of current block. 1418 offset -= (blk_starts[i] - blk_offset); 1419 } 1420 // In the following code a nop could be inserted before 1421 // the branch which will increase the backward distance. 1422 bool needs_padding = (current_offset == last_avoid_back_to_back_offset); 1423 if (needs_padding && offset <= 0) 1424 offset -= nop_size; 1425 1426 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 1427 // We've got a winner. Replace this branch. 1428 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this); 1429 1430 // Update the jmp_size. 1431 int new_size = replacement->size(_regalloc); 1432 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller"); 1433 // Insert padding between avoid_back_to_back branches. 1434 if (needs_padding && replacement->avoid_back_to_back()) { 1435 MachNode *nop = new (this) MachNopNode(); 1436 b->_nodes.insert(j++, nop); 1437 _cfg->_bbs.map(nop->_idx, b); 1438 last_inst++; 1439 nop->emit(*cb, _regalloc); 1440 cb->flush_bundle(true); 1441 current_offset = cb->insts_size(); 1442 } 1443 #ifdef ASSERT 1444 jmp_target[i] = block_num; 1445 jmp_offset[i] = current_offset - blk_offset; 1446 jmp_size[i] = new_size; 1447 jmp_rule[i] = mach->rule(); 1448 #endif 1449 b->_nodes.map(j, replacement); 1450 mach->subsume_by(replacement, C); 1451 n = replacement; 1452 mach = replacement; 1453 } 1454 } 1455 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num ); 1456 } else if (mach->ideal_Opcode() == Op_Jump) { 1457 for (uint h = 0; h < b->_num_succs; h++) { 1458 Block* succs_block = b->_succs[h]; 1459 for (uint j = 1; j < succs_block->num_preds(); j++) { 1460 Node* jpn = succs_block->pred(j); 1461 if (jpn->is_JumpProj() && jpn->in(0) == mach) { 1462 uint block_num = succs_block->non_connector()->_pre_order; 1463 Label *blkLabel = &blk_labels[block_num]; 1464 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1465 } 1466 } 1467 } 1468 } 1469 1470 #ifdef ASSERT 1471 // Check that oop-store precedes the card-mark 1472 else if (mach->ideal_Opcode() == Op_StoreCM) { 1473 uint storeCM_idx = j; 1474 int count = 0; 1475 for (uint prec = mach->req(); prec < mach->len(); prec++) { 1476 Node *oop_store = mach->in(prec); // Precedence edge 1477 if (oop_store == NULL) continue; 1478 count++; 1479 uint i4; 1480 for( i4 = 0; i4 < last_inst; ++i4 ) { 1481 if( b->_nodes[i4] == oop_store ) break; 1482 } 1483 // Note: This test can provide a false failure if other precedence 1484 // edges have been added to the storeCMNode. 1485 assert( i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1486 } 1487 assert(count > 0, "storeCM expects at least one precedence edge"); 1488 } 1489 #endif 1490 1491 else if (!n->is_Proj()) { 1492 // Remember the beginning of the previous instruction, in case 1493 // it's followed by a flag-kill and a null-check. Happens on 1494 // Intel all the time, with add-to-memory kind of opcodes. 1495 previous_offset = current_offset; 1496 } 1497 1498 // Not an else-if! 1499 // If this is a trap based cmp then add its offset to the list. 1500 if (mach->is_TrapBasedCheckNode()) { 1501 inct_starts[inct_cnt++] = current_offset; 1502 } 1503 } 1504 1505 // Verify that there is sufficient space remaining 1506 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1507 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1508 turn_off_compiler(this); 1509 return; 1510 } 1511 1512 // Save the offset for the listing 1513 #ifndef PRODUCT 1514 if (node_offsets && n->_idx < node_offset_limit) 1515 node_offsets[n->_idx] = cb->insts_size(); 1516 #endif 1517 1518 // "Normal" instruction case 1519 DEBUG_ONLY( uint instr_offset = cb->insts_size(); ) 1520 n->emit(*cb, _regalloc); 1521 current_offset = cb->insts_size(); 1522 1523 #ifdef ASSERT 1524 if (n->size(_regalloc) < (current_offset-instr_offset)) { 1525 n->dump(); 1526 assert(false, "wrong size of mach node"); 1527 } 1528 #endif 1529 non_safepoints.observe_instruction(n, current_offset); 1530 1531 // mcall is last "call" that can be a safepoint 1532 // record it so we can see if a poll will directly follow it 1533 // in which case we'll need a pad to make the PcDesc sites unique 1534 // see 5010568. This can be slightly inaccurate but conservative 1535 // in the case that return address is not actually at current_offset. 1536 // This is a small price to pay. 1537 1538 if (is_mcall) { 1539 last_call_offset = current_offset; 1540 } 1541 1542 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back()) { 1543 // Avoid back to back some instructions. 1544 last_avoid_back_to_back_offset = current_offset; 1545 } 1546 1547 // See if this instruction has a delay slot 1548 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1549 assert(delay_slot != NULL, "expecting delay slot node"); 1550 1551 // Back up 1 instruction 1552 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size()); 1553 1554 // Save the offset for the listing 1555 #ifndef PRODUCT 1556 if (node_offsets && delay_slot->_idx < node_offset_limit) 1557 node_offsets[delay_slot->_idx] = cb->insts_size(); 1558 #endif 1559 1560 // Support a SafePoint in the delay slot 1561 if (delay_slot->is_MachSafePoint()) { 1562 MachNode *mach = delay_slot->as_Mach(); 1563 // !!!!! Stubs only need an oopmap right now, so bail out 1564 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) { 1565 // Write the oopmap directly to the code blob??!! 1566 # ifdef ENABLE_ZAP_DEAD_LOCALS 1567 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive"); 1568 # endif 1569 delay_slot = NULL; 1570 continue; 1571 } 1572 1573 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1574 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1575 adjusted_offset); 1576 // Generate an OopMap entry 1577 Process_OopMap_Node(mach, adjusted_offset); 1578 } 1579 1580 // Insert the delay slot instruction 1581 delay_slot->emit(*cb, _regalloc); 1582 1583 // Don't reuse it 1584 delay_slot = NULL; 1585 } 1586 1587 } // End for all instructions in block 1588 1589 // If the next block is the top of a loop, pad this block out to align 1590 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1591 if (i < nblocks-1) { 1592 Block *nb = _cfg->_blocks[i+1]; 1593 int padding = nb->alignment_padding(current_offset); 1594 if( padding > 0 ) { 1595 MachNode *nop = new (this) MachNopNode(padding / nop_size); 1596 b->_nodes.insert( b->_nodes.size(), nop ); 1597 _cfg->_bbs.map( nop->_idx, b ); 1598 nop->emit(*cb, _regalloc); 1599 current_offset = cb->insts_size(); 1600 } 1601 } 1602 // Verify that the distance for generated before forward 1603 // short branches is still valid. 1604 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size"); 1605 1606 // Save new block start offset 1607 blk_starts[i] = blk_offset; 1608 } // End of for all blocks 1609 blk_starts[nblocks] = current_offset; 1610 1611 non_safepoints.flush_at_end(); 1612 1613 // Offset too large? 1614 if (failing()) return; 1615 1616 // Define a pseudo-label at the end of the code 1617 MacroAssembler(cb).bind( blk_labels[nblocks] ); 1618 1619 // Compute the size of the first block 1620 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1621 1622 assert(cb->insts_size() < 500000, "method is unreasonably large"); 1623 1624 #ifdef ASSERT 1625 for (uint i = 0; i < nblocks; i++) { // For all blocks 1626 if (jmp_target[i] != 0) { 1627 int br_size = jmp_size[i]; 1628 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 1629 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 1630 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]); 1631 assert(false, "Displacement too large for short jmp"); 1632 } 1633 } 1634 } 1635 #endif 1636 1637 // ------------------ 1638 1639 #ifndef PRODUCT 1640 // Information on the size of the method, without the extraneous code 1641 Scheduling::increment_method_size(cb->insts_size()); 1642 #endif 1643 1644 // ------------------ 1645 // Fill in exception table entries. 1646 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1647 1648 // Only java methods have exception handlers and deopt handlers 1649 if (_method) { 1650 // Emit the exception handler code. 1651 _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb)); 1652 // Emit the deopt handler code. 1653 _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb)); 1654 1655 // Emit the MethodHandle deopt handler code (if required). 1656 if (has_method_handle_invokes()) { 1657 // We can use the same code as for the normal deopt handler, we 1658 // just need a different entry point address. 1659 _code_offsets.set_value(CodeOffsets::DeoptMH, emit_deopt_handler(*cb)); 1660 } 1661 } 1662 1663 // One last check for failed CodeBuffer::expand: 1664 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1665 turn_off_compiler(this); 1666 return; 1667 } 1668 1669 #ifndef PRODUCT 1670 // Dump the assembly code, including basic-block numbers 1671 if (print_assembly()) { 1672 ttyLocker ttyl; // keep the following output all in one block 1673 if (!VMThread::should_terminate()) { // test this under the tty lock 1674 // This output goes directly to the tty, not the compiler log. 1675 // To enable tools to match it up with the compilation activity, 1676 // be sure to tag this tty output with the compile ID. 1677 if (xtty != NULL) { 1678 xtty->head("opto_assembly compile_id='%d'%s", compile_id(), 1679 is_osr_compilation() ? " compile_kind='osr'" : 1680 ""); 1681 } 1682 if (method() != NULL) { 1683 method()->print_oop(); 1684 print_codes(); 1685 } 1686 dump_asm(node_offsets, node_offset_limit); 1687 if (xtty != NULL) { 1688 xtty->tail("opto_assembly"); 1689 } 1690 } 1691 } 1692 #endif 1693 1694 } 1695 1696 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1697 _inc_table.set_size(cnt); 1698 1699 uint inct_cnt = 0; 1700 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1701 Block *b = _cfg->_blocks[i]; 1702 Node *n = NULL; 1703 int j; 1704 1705 // Find the branch; ignore trailing NOPs. 1706 for( j = b->_nodes.size()-1; j>=0; j-- ) { 1707 n = b->_nodes[j]; 1708 if( !n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con ) 1709 break; 1710 } 1711 1712 // If we didn't find anything, continue 1713 if( j < 0 ) continue; 1714 1715 // Compute ExceptionHandlerTable subtable entry and add it 1716 // (skip empty blocks) 1717 if( n->is_Catch() ) { 1718 1719 // Get the offset of the return from the call 1720 uint call_return = call_returns[b->_pre_order]; 1721 #ifdef ASSERT 1722 assert( call_return > 0, "no call seen for this basic block" ); 1723 while( b->_nodes[--j]->is_MachProj() ) ; 1724 assert( b->_nodes[j]->is_MachCall(), "CatchProj must follow call" ); 1725 #endif 1726 // last instruction is a CatchNode, find it's CatchProjNodes 1727 int nof_succs = b->_num_succs; 1728 // allocate space 1729 GrowableArray<intptr_t> handler_bcis(nof_succs); 1730 GrowableArray<intptr_t> handler_pcos(nof_succs); 1731 // iterate through all successors 1732 for (int j = 0; j < nof_succs; j++) { 1733 Block* s = b->_succs[j]; 1734 bool found_p = false; 1735 for( uint k = 1; k < s->num_preds(); k++ ) { 1736 Node *pk = s->pred(k); 1737 if( pk->is_CatchProj() && pk->in(0) == n ) { 1738 const CatchProjNode* p = pk->as_CatchProj(); 1739 found_p = true; 1740 // add the corresponding handler bci & pco information 1741 if( p->_con != CatchProjNode::fall_through_index ) { 1742 // p leads to an exception handler (and is not fall through) 1743 assert(s == _cfg->_blocks[s->_pre_order],"bad numbering"); 1744 // no duplicates, please 1745 if( !handler_bcis.contains(p->handler_bci()) ) { 1746 uint block_num = s->non_connector()->_pre_order; 1747 handler_bcis.append(p->handler_bci()); 1748 handler_pcos.append(blk_labels[block_num].loc_pos()); 1749 } 1750 } 1751 } 1752 } 1753 assert(found_p, "no matching predecessor found"); 1754 // Note: Due to empty block removal, one block may have 1755 // several CatchProj inputs, from the same Catch. 1756 } 1757 1758 // Set the offset of the return from the call 1759 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); 1760 continue; 1761 } 1762 1763 // Handle implicit null exception table updates 1764 if( n->is_MachNullCheck() ) { 1765 uint block_num = b->non_connector_successor(0)->_pre_order; 1766 _inc_table.append( inct_starts[inct_cnt++], blk_labels[block_num].loc_pos() ); 1767 continue; 1768 } 1769 // Handle implicit exception table updates: trap instructions. 1770 if (n->is_TrapBasedCheckNode()) { 1771 uint block_num = b->non_connector_successor(0)->_pre_order; 1772 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1773 continue; 1774 } 1775 } // End of for all blocks fill in exception table entries 1776 } 1777 1778 // Static Variables 1779 #ifndef PRODUCT 1780 uint Scheduling::_total_nop_size = 0; 1781 uint Scheduling::_total_method_size = 0; 1782 uint Scheduling::_total_branches = 0; 1783 uint Scheduling::_total_unconditional_delays = 0; 1784 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1785 #endif 1786 1787 // Initializer for class Scheduling 1788 1789 Scheduling::Scheduling(Arena *arena, Compile &compile) 1790 : _arena(arena), 1791 _cfg(compile.cfg()), 1792 _bbs(compile.cfg()->_bbs), 1793 _regalloc(compile.regalloc()), 1794 _reg_node(arena), 1795 _bundle_instr_count(0), 1796 _bundle_cycle_number(0), 1797 _scheduled(arena), 1798 _available(arena), 1799 _next_node(NULL), 1800 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]), 1801 _pinch_free_list(arena) 1802 #ifndef PRODUCT 1803 , _branches(0) 1804 , _unconditional_delays(0) 1805 #endif 1806 { 1807 // Create a MachNopNode 1808 _nop = new (&compile) MachNopNode(); 1809 1810 // Now that the nops are in the array, save the count 1811 // (but allow entries for the nops) 1812 _node_bundling_limit = compile.unique(); 1813 uint node_max = _regalloc->node_regs_max_index(); 1814 1815 compile.set_node_bundling_limit(_node_bundling_limit); 1816 1817 // This one is persistent within the Compile class 1818 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1819 1820 // Allocate space for fixed-size arrays 1821 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1822 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1823 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1824 1825 // Clear the arrays 1826 memset(_node_bundling_base, 0, node_max * sizeof(Bundle)); 1827 memset(_node_latency, 0, node_max * sizeof(unsigned short)); 1828 memset(_uses, 0, node_max * sizeof(short)); 1829 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 1830 1831 // Clear the bundling information 1832 memcpy(_bundle_use_elements, 1833 Pipeline_Use::elaborated_elements, 1834 sizeof(Pipeline_Use::elaborated_elements)); 1835 1836 // Get the last node 1837 Block *bb = _cfg->_blocks[_cfg->_blocks.size()-1]; 1838 1839 _next_node = bb->_nodes[bb->_nodes.size()-1]; 1840 } 1841 1842 #ifndef PRODUCT 1843 // Scheduling destructor 1844 Scheduling::~Scheduling() { 1845 _total_branches += _branches; 1846 _total_unconditional_delays += _unconditional_delays; 1847 } 1848 #endif 1849 1850 // Step ahead "i" cycles 1851 void Scheduling::step(uint i) { 1852 1853 Bundle *bundle = node_bundling(_next_node); 1854 bundle->set_starts_bundle(); 1855 1856 // Update the bundle record, but leave the flags information alone 1857 if (_bundle_instr_count > 0) { 1858 bundle->set_instr_count(_bundle_instr_count); 1859 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1860 } 1861 1862 // Update the state information 1863 _bundle_instr_count = 0; 1864 _bundle_cycle_number += i; 1865 _bundle_use.step(i); 1866 } 1867 1868 void Scheduling::step_and_clear() { 1869 Bundle *bundle = node_bundling(_next_node); 1870 bundle->set_starts_bundle(); 1871 1872 // Update the bundle record 1873 if (_bundle_instr_count > 0) { 1874 bundle->set_instr_count(_bundle_instr_count); 1875 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1876 1877 _bundle_cycle_number += 1; 1878 } 1879 1880 // Clear the bundling information 1881 _bundle_instr_count = 0; 1882 _bundle_use.reset(); 1883 1884 memcpy(_bundle_use_elements, 1885 Pipeline_Use::elaborated_elements, 1886 sizeof(Pipeline_Use::elaborated_elements)); 1887 } 1888 1889 //------------------------------ScheduleAndBundle------------------------------ 1890 // Perform instruction scheduling and bundling over the sequence of 1891 // instructions in backwards order. 1892 void Compile::ScheduleAndBundle() { 1893 1894 // Don't optimize this if it isn't a method 1895 if (!_method) 1896 return; 1897 1898 // Don't optimize this if scheduling is disabled 1899 if (!do_scheduling()) 1900 return; 1901 1902 // Scheduling code works only with pairs (8 bytes) maximum. 1903 if (max_vector_size() > 8) 1904 return; 1905 1906 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); ) 1907 1908 // Create a data structure for all the scheduling information 1909 Scheduling scheduling(Thread::current()->resource_area(), *this); 1910 1911 // Walk backwards over each basic block, computing the needed alignment 1912 // Walk over all the basic blocks 1913 scheduling.DoScheduling(); 1914 } 1915 1916 //------------------------------ComputeLocalLatenciesForward------------------- 1917 // Compute the latency of all the instructions. This is fairly simple, 1918 // because we already have a legal ordering. Walk over the instructions 1919 // from first to last, and compute the latency of the instruction based 1920 // on the latency of the preceding instruction(s). 1921 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { 1922 #ifndef PRODUCT 1923 if (_cfg->C->trace_opto_output()) 1924 tty->print("# -> ComputeLocalLatenciesForward\n"); 1925 #endif 1926 1927 // Walk over all the schedulable instructions 1928 for( uint j=_bb_start; j < _bb_end; j++ ) { 1929 1930 // This is a kludge, forcing all latency calculations to start at 1. 1931 // Used to allow latency 0 to force an instruction to the beginning 1932 // of the bb 1933 uint latency = 1; 1934 Node *use = bb->_nodes[j]; 1935 uint nlen = use->len(); 1936 1937 // Walk over all the inputs 1938 for ( uint k=0; k < nlen; k++ ) { 1939 Node *def = use->in(k); 1940 if (!def) 1941 continue; 1942 1943 uint l = _node_latency[def->_idx] + use->latency(k); 1944 if (latency < l) 1945 latency = l; 1946 } 1947 1948 _node_latency[use->_idx] = latency; 1949 1950 #ifndef PRODUCT 1951 if (_cfg->C->trace_opto_output()) { 1952 tty->print("# latency %4d: ", latency); 1953 use->dump(); 1954 } 1955 #endif 1956 } 1957 1958 #ifndef PRODUCT 1959 if (_cfg->C->trace_opto_output()) 1960 tty->print("# <- ComputeLocalLatenciesForward\n"); 1961 #endif 1962 1963 } // end ComputeLocalLatenciesForward 1964 1965 // See if this node fits into the present instruction bundle 1966 bool Scheduling::NodeFitsInBundle(Node *n) { 1967 uint n_idx = n->_idx; 1968 1969 // If this is the unconditional delay instruction, then it fits 1970 if (n == _unconditional_delay_slot) { 1971 #ifndef PRODUCT 1972 if (_cfg->C->trace_opto_output()) 1973 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 1974 #endif 1975 return (true); 1976 } 1977 1978 // If the node cannot be scheduled this cycle, skip it 1979 if (_current_latency[n_idx] > _bundle_cycle_number) { 1980 #ifndef PRODUCT 1981 if (_cfg->C->trace_opto_output()) 1982 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 1983 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 1984 #endif 1985 return (false); 1986 } 1987 1988 const Pipeline *node_pipeline = n->pipeline(); 1989 1990 uint instruction_count = node_pipeline->instructionCount(); 1991 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1992 instruction_count = 0; 1993 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1994 instruction_count++; 1995 1996 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 1997 #ifndef PRODUCT 1998 if (_cfg->C->trace_opto_output()) 1999 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 2000 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 2001 #endif 2002 return (false); 2003 } 2004 2005 // Don't allow non-machine nodes to be handled this way 2006 if (!n->is_Mach() && instruction_count == 0) 2007 return (false); 2008 2009 // See if there is any overlap 2010 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 2011 2012 if (delay > 0) { 2013 #ifndef PRODUCT 2014 if (_cfg->C->trace_opto_output()) 2015 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 2016 #endif 2017 return false; 2018 } 2019 2020 #ifndef PRODUCT 2021 if (_cfg->C->trace_opto_output()) 2022 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 2023 #endif 2024 2025 return true; 2026 } 2027 2028 Node * Scheduling::ChooseNodeToBundle() { 2029 uint siz = _available.size(); 2030 2031 if (siz == 0) { 2032 2033 #ifndef PRODUCT 2034 if (_cfg->C->trace_opto_output()) 2035 tty->print("# ChooseNodeToBundle: NULL\n"); 2036 #endif 2037 return (NULL); 2038 } 2039 2040 // Fast path, if only 1 instruction in the bundle 2041 if (siz == 1) { 2042 #ifndef PRODUCT 2043 if (_cfg->C->trace_opto_output()) { 2044 tty->print("# ChooseNodeToBundle (only 1): "); 2045 _available[0]->dump(); 2046 } 2047 #endif 2048 return (_available[0]); 2049 } 2050 2051 // Don't bother, if the bundle is already full 2052 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 2053 for ( uint i = 0; i < siz; i++ ) { 2054 Node *n = _available[i]; 2055 2056 // Skip projections, we'll handle them another way 2057 if (n->is_Proj()) 2058 continue; 2059 2060 // This presupposed that instructions are inserted into the 2061 // available list in a legality order; i.e. instructions that 2062 // must be inserted first are at the head of the list 2063 if (NodeFitsInBundle(n)) { 2064 #ifndef PRODUCT 2065 if (_cfg->C->trace_opto_output()) { 2066 tty->print("# ChooseNodeToBundle: "); 2067 n->dump(); 2068 } 2069 #endif 2070 return (n); 2071 } 2072 } 2073 } 2074 2075 // Nothing fits in this bundle, choose the highest priority 2076 #ifndef PRODUCT 2077 if (_cfg->C->trace_opto_output()) { 2078 tty->print("# ChooseNodeToBundle: "); 2079 _available[0]->dump(); 2080 } 2081 #endif 2082 2083 return _available[0]; 2084 } 2085 2086 //------------------------------AddNodeToAvailableList------------------------- 2087 void Scheduling::AddNodeToAvailableList(Node *n) { 2088 assert( !n->is_Proj(), "projections never directly made available" ); 2089 #ifndef PRODUCT 2090 if (_cfg->C->trace_opto_output()) { 2091 tty->print("# AddNodeToAvailableList: "); 2092 n->dump(); 2093 } 2094 #endif 2095 2096 int latency = _current_latency[n->_idx]; 2097 2098 // Insert in latency order (insertion sort) 2099 uint i; 2100 for ( i=0; i < _available.size(); i++ ) 2101 if (_current_latency[_available[i]->_idx] > latency) 2102 break; 2103 2104 // Special Check for compares following branches 2105 if( n->is_Mach() && _scheduled.size() > 0 ) { 2106 int op = n->as_Mach()->ideal_Opcode(); 2107 Node *last = _scheduled[0]; 2108 if( last->is_MachIf() && last->in(1) == n && 2109 ( op == Op_CmpI || 2110 op == Op_CmpU || 2111 op == Op_CmpP || 2112 op == Op_CmpF || 2113 op == Op_CmpD || 2114 op == Op_CmpL ) ) { 2115 2116 // Recalculate position, moving to front of same latency 2117 for ( i=0 ; i < _available.size(); i++ ) 2118 if (_current_latency[_available[i]->_idx] >= latency) 2119 break; 2120 } 2121 } 2122 2123 // Insert the node in the available list 2124 _available.insert(i, n); 2125 2126 #ifndef PRODUCT 2127 if (_cfg->C->trace_opto_output()) 2128 dump_available(); 2129 #endif 2130 } 2131 2132 //------------------------------DecrementUseCounts----------------------------- 2133 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 2134 for ( uint i=0; i < n->len(); i++ ) { 2135 Node *def = n->in(i); 2136 if (!def) continue; 2137 if( def->is_Proj() ) // If this is a machine projection, then 2138 def = def->in(0); // propagate usage thru to the base instruction 2139 2140 if( _bbs[def->_idx] != bb ) // Ignore if not block-local 2141 continue; 2142 2143 // Compute the latency 2144 uint l = _bundle_cycle_number + n->latency(i); 2145 if (_current_latency[def->_idx] < l) 2146 _current_latency[def->_idx] = l; 2147 2148 // If this does not have uses then schedule it 2149 if ((--_uses[def->_idx]) == 0) 2150 AddNodeToAvailableList(def); 2151 } 2152 } 2153 2154 //------------------------------AddNodeToBundle-------------------------------- 2155 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 2156 #ifndef PRODUCT 2157 if (_cfg->C->trace_opto_output()) { 2158 tty->print("# AddNodeToBundle: "); 2159 n->dump(); 2160 } 2161 #endif 2162 2163 // Remove this from the available list 2164 uint i; 2165 for (i = 0; i < _available.size(); i++) 2166 if (_available[i] == n) 2167 break; 2168 assert(i < _available.size(), "entry in _available list not found"); 2169 _available.remove(i); 2170 2171 // See if this fits in the current bundle 2172 const Pipeline *node_pipeline = n->pipeline(); 2173 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 2174 2175 // Check for instructions to be placed in the delay slot. We 2176 // do this before we actually schedule the current instruction, 2177 // because the delay slot follows the current instruction. 2178 if (Pipeline::_branch_has_delay_slot && 2179 node_pipeline->hasBranchDelay() && 2180 !_unconditional_delay_slot) { 2181 2182 uint siz = _available.size(); 2183 2184 // Conditional branches can support an instruction that 2185 // is unconditionally executed and not dependent by the 2186 // branch, OR a conditionally executed instruction if 2187 // the branch is taken. In practice, this means that 2188 // the first instruction at the branch target is 2189 // copied to the delay slot, and the branch goes to 2190 // the instruction after that at the branch target 2191 if ( n->is_MachBranch() ) { 2192 2193 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 2194 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 2195 2196 #ifndef PRODUCT 2197 _branches++; 2198 #endif 2199 2200 // At least 1 instruction is on the available list 2201 // that is not dependent on the branch 2202 for (uint i = 0; i < siz; i++) { 2203 Node *d = _available[i]; 2204 const Pipeline *avail_pipeline = d->pipeline(); 2205 2206 // Don't allow safepoints in the branch shadow, that will 2207 // cause a number of difficulties 2208 if ( avail_pipeline->instructionCount() == 1 && 2209 !avail_pipeline->hasMultipleBundles() && 2210 !avail_pipeline->hasBranchDelay() && 2211 Pipeline::instr_has_unit_size() && 2212 d->size(_regalloc) == Pipeline::instr_unit_size() && 2213 NodeFitsInBundle(d) && 2214 !node_bundling(d)->used_in_delay()) { 2215 2216 if (d->is_Mach() && !d->is_MachSafePoint()) { 2217 // A node that fits in the delay slot was found, so we need to 2218 // set the appropriate bits in the bundle pipeline information so 2219 // that it correctly indicates resource usage. Later, when we 2220 // attempt to add this instruction to the bundle, we will skip 2221 // setting the resource usage. 2222 _unconditional_delay_slot = d; 2223 node_bundling(n)->set_use_unconditional_delay(); 2224 node_bundling(d)->set_used_in_unconditional_delay(); 2225 _bundle_use.add_usage(avail_pipeline->resourceUse()); 2226 _current_latency[d->_idx] = _bundle_cycle_number; 2227 _next_node = d; 2228 ++_bundle_instr_count; 2229 #ifndef PRODUCT 2230 _unconditional_delays++; 2231 #endif 2232 break; 2233 } 2234 } 2235 } 2236 } 2237 2238 // No delay slot, add a nop to the usage 2239 if (!_unconditional_delay_slot) { 2240 // See if adding an instruction in the delay slot will overflow 2241 // the bundle. 2242 if (!NodeFitsInBundle(_nop)) { 2243 #ifndef PRODUCT 2244 if (_cfg->C->trace_opto_output()) 2245 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2246 #endif 2247 step(1); 2248 } 2249 2250 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2251 _next_node = _nop; 2252 ++_bundle_instr_count; 2253 } 2254 2255 // See if the instruction in the delay slot requires a 2256 // step of the bundles 2257 if (!NodeFitsInBundle(n)) { 2258 #ifndef PRODUCT 2259 if (_cfg->C->trace_opto_output()) 2260 tty->print("# *** STEP(branch won't fit) ***\n"); 2261 #endif 2262 // Update the state information 2263 _bundle_instr_count = 0; 2264 _bundle_cycle_number += 1; 2265 _bundle_use.step(1); 2266 } 2267 } 2268 2269 // Get the number of instructions 2270 uint instruction_count = node_pipeline->instructionCount(); 2271 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2272 instruction_count = 0; 2273 2274 // Compute the latency information 2275 uint delay = 0; 2276 2277 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2278 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2279 if (relative_latency < 0) 2280 relative_latency = 0; 2281 2282 delay = _bundle_use.full_latency(relative_latency, node_usage); 2283 2284 // Does not fit in this bundle, start a new one 2285 if (delay > 0) { 2286 step(delay); 2287 2288 #ifndef PRODUCT 2289 if (_cfg->C->trace_opto_output()) 2290 tty->print("# *** STEP(%d) ***\n", delay); 2291 #endif 2292 } 2293 } 2294 2295 // If this was placed in the delay slot, ignore it 2296 if (n != _unconditional_delay_slot) { 2297 2298 if (delay == 0) { 2299 if (node_pipeline->hasMultipleBundles()) { 2300 #ifndef PRODUCT 2301 if (_cfg->C->trace_opto_output()) 2302 tty->print("# *** STEP(multiple instructions) ***\n"); 2303 #endif 2304 step(1); 2305 } 2306 2307 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2308 #ifndef PRODUCT 2309 if (_cfg->C->trace_opto_output()) 2310 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2311 instruction_count + _bundle_instr_count, 2312 Pipeline::_max_instrs_per_cycle); 2313 #endif 2314 step(1); 2315 } 2316 } 2317 2318 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2319 _bundle_instr_count++; 2320 2321 // Set the node's latency 2322 _current_latency[n->_idx] = _bundle_cycle_number; 2323 2324 // Now merge the functional unit information 2325 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2326 _bundle_use.add_usage(node_usage); 2327 2328 // Increment the number of instructions in this bundle 2329 _bundle_instr_count += instruction_count; 2330 2331 // Remember this node for later 2332 if (n->is_Mach()) 2333 _next_node = n; 2334 } 2335 2336 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2337 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2338 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2339 // into the block. All other scheduled nodes get put in the schedule here. 2340 int op = n->Opcode(); 2341 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2342 (op != Op_Node && // Not an unused antidepedence node and 2343 // not an unallocated boxlock 2344 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2345 2346 // Push any trailing projections 2347 if( bb->_nodes[bb->_nodes.size()-1] != n ) { 2348 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2349 Node *foi = n->fast_out(i); 2350 if( foi->is_Proj() ) 2351 _scheduled.push(foi); 2352 } 2353 } 2354 2355 // Put the instruction in the schedule list 2356 _scheduled.push(n); 2357 } 2358 2359 #ifndef PRODUCT 2360 if (_cfg->C->trace_opto_output()) 2361 dump_available(); 2362 #endif 2363 2364 // Walk all the definitions, decrementing use counts, and 2365 // if a definition has a 0 use count, place it in the available list. 2366 DecrementUseCounts(n,bb); 2367 } 2368 2369 //------------------------------ComputeUseCount-------------------------------- 2370 // This method sets the use count within a basic block. We will ignore all 2371 // uses outside the current basic block. As we are doing a backwards walk, 2372 // any node we reach that has a use count of 0 may be scheduled. This also 2373 // avoids the problem of cyclic references from phi nodes, as long as phi 2374 // nodes are at the front of the basic block. This method also initializes 2375 // the available list to the set of instructions that have no uses within this 2376 // basic block. 2377 void Scheduling::ComputeUseCount(const Block *bb) { 2378 #ifndef PRODUCT 2379 if (_cfg->C->trace_opto_output()) 2380 tty->print("# -> ComputeUseCount\n"); 2381 #endif 2382 2383 // Clear the list of available and scheduled instructions, just in case 2384 _available.clear(); 2385 _scheduled.clear(); 2386 2387 // No delay slot specified 2388 _unconditional_delay_slot = NULL; 2389 2390 #ifdef ASSERT 2391 for( uint i=0; i < bb->_nodes.size(); i++ ) 2392 assert( _uses[bb->_nodes[i]->_idx] == 0, "_use array not clean" ); 2393 #endif 2394 2395 // Force the _uses count to never go to zero for unscheduable pieces 2396 // of the block 2397 for( uint k = 0; k < _bb_start; k++ ) 2398 _uses[bb->_nodes[k]->_idx] = 1; 2399 for( uint l = _bb_end; l < bb->_nodes.size(); l++ ) 2400 _uses[bb->_nodes[l]->_idx] = 1; 2401 2402 // Iterate backwards over the instructions in the block. Don't count the 2403 // branch projections at end or the block header instructions. 2404 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2405 Node *n = bb->_nodes[j]; 2406 if( n->is_Proj() ) continue; // Projections handled another way 2407 2408 // Account for all uses 2409 for ( uint k = 0; k < n->len(); k++ ) { 2410 Node *inp = n->in(k); 2411 if (!inp) continue; 2412 assert(inp != n, "no cycles allowed" ); 2413 if( _bbs[inp->_idx] == bb ) { // Block-local use? 2414 if( inp->is_Proj() ) // Skip through Proj's 2415 inp = inp->in(0); 2416 ++_uses[inp->_idx]; // Count 1 block-local use 2417 } 2418 } 2419 2420 // If this instruction has a 0 use count, then it is available 2421 if (!_uses[n->_idx]) { 2422 _current_latency[n->_idx] = _bundle_cycle_number; 2423 AddNodeToAvailableList(n); 2424 } 2425 2426 #ifndef PRODUCT 2427 if (_cfg->C->trace_opto_output()) { 2428 tty->print("# uses: %3d: ", _uses[n->_idx]); 2429 n->dump(); 2430 } 2431 #endif 2432 } 2433 2434 #ifndef PRODUCT 2435 if (_cfg->C->trace_opto_output()) 2436 tty->print("# <- ComputeUseCount\n"); 2437 #endif 2438 } 2439 2440 // This routine performs scheduling on each basic block in reverse order, 2441 // using instruction latencies and taking into account function unit 2442 // availability. 2443 void Scheduling::DoScheduling() { 2444 #ifndef PRODUCT 2445 if (_cfg->C->trace_opto_output()) 2446 tty->print("# -> DoScheduling\n"); 2447 #endif 2448 2449 Block *succ_bb = NULL; 2450 Block *bb; 2451 2452 // Walk over all the basic blocks in reverse order 2453 for( int i=_cfg->_num_blocks-1; i >= 0; succ_bb = bb, i-- ) { 2454 bb = _cfg->_blocks[i]; 2455 2456 #ifndef PRODUCT 2457 if (_cfg->C->trace_opto_output()) { 2458 tty->print("# Schedule BB#%03d (initial)\n", i); 2459 for (uint j = 0; j < bb->_nodes.size(); j++) 2460 bb->_nodes[j]->dump(); 2461 } 2462 #endif 2463 2464 // On the head node, skip processing 2465 if( bb == _cfg->_broot ) 2466 continue; 2467 2468 // Skip empty, connector blocks 2469 if (bb->is_connector()) 2470 continue; 2471 2472 // If the following block is not the sole successor of 2473 // this one, then reset the pipeline information 2474 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2475 #ifndef PRODUCT 2476 if (_cfg->C->trace_opto_output()) { 2477 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2478 _next_node->_idx, _bundle_instr_count); 2479 } 2480 #endif 2481 step_and_clear(); 2482 } 2483 2484 // Leave untouched the starting instruction, any Phis, a CreateEx node 2485 // or Top. bb->_nodes[_bb_start] is the first schedulable instruction. 2486 _bb_end = bb->_nodes.size()-1; 2487 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2488 Node *n = bb->_nodes[_bb_start]; 2489 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2490 // Also, MachIdealNodes do not get scheduled 2491 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2492 MachNode *mach = n->as_Mach(); 2493 int iop = mach->ideal_Opcode(); 2494 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2495 if( iop == Op_Con ) continue; // Do not schedule Top 2496 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2497 mach->pipeline() == MachNode::pipeline_class() && 2498 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc 2499 continue; 2500 break; // Funny loop structure to be sure... 2501 } 2502 // Compute last "interesting" instruction in block - last instruction we 2503 // might schedule. _bb_end points just after last schedulable inst. We 2504 // normally schedule conditional branches (despite them being forced last 2505 // in the block), because they have delay slots we can fill. Calls all 2506 // have their delay slots filled in the template expansions, so we don't 2507 // bother scheduling them. 2508 Node *last = bb->_nodes[_bb_end]; 2509 // Ignore trailing NOPs. 2510 while (_bb_end > 0 && last->is_Mach() && 2511 last->as_Mach()->ideal_Opcode() == Op_Con) { 2512 last = bb->_nodes[--_bb_end]; 2513 } 2514 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, ""); 2515 if( last->is_Catch() || 2516 // Exclude unreachable path case when Halt node is in a separate block. 2517 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2518 // There must be a prior call. Skip it. 2519 while( !bb->_nodes[--_bb_end]->is_MachCall() ) { 2520 assert( bb->_nodes[_bb_end]->is_MachProj(), "skipping projections after expected call" ); 2521 } 2522 } else if( last->is_MachNullCheck() ) { 2523 // Backup so the last null-checked memory instruction is 2524 // outside the schedulable range. Skip over the nullcheck, 2525 // projection, and the memory nodes. 2526 Node *mem = last->in(1); 2527 do { 2528 _bb_end--; 2529 } while (mem != bb->_nodes[_bb_end]); 2530 } else { 2531 // Set _bb_end to point after last schedulable inst. 2532 _bb_end++; 2533 } 2534 2535 assert( _bb_start <= _bb_end, "inverted block ends" ); 2536 2537 // Compute the register antidependencies for the basic block 2538 ComputeRegisterAntidependencies(bb); 2539 if (_cfg->C->failing()) return; // too many D-U pinch points 2540 2541 // Compute intra-bb latencies for the nodes 2542 ComputeLocalLatenciesForward(bb); 2543 2544 // Compute the usage within the block, and set the list of all nodes 2545 // in the block that have no uses within the block. 2546 ComputeUseCount(bb); 2547 2548 // Schedule the remaining instructions in the block 2549 while ( _available.size() > 0 ) { 2550 Node *n = ChooseNodeToBundle(); 2551 AddNodeToBundle(n,bb); 2552 } 2553 2554 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2555 #ifdef ASSERT 2556 for( uint l = _bb_start; l < _bb_end; l++ ) { 2557 Node *n = bb->_nodes[l]; 2558 uint m; 2559 for( m = 0; m < _bb_end-_bb_start; m++ ) 2560 if( _scheduled[m] == n ) 2561 break; 2562 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2563 } 2564 #endif 2565 2566 // Now copy the instructions (in reverse order) back to the block 2567 for ( uint k = _bb_start; k < _bb_end; k++ ) 2568 bb->_nodes.map(k, _scheduled[_bb_end-k-1]); 2569 2570 #ifndef PRODUCT 2571 if (_cfg->C->trace_opto_output()) { 2572 tty->print("# Schedule BB#%03d (final)\n", i); 2573 uint current = 0; 2574 for (uint j = 0; j < bb->_nodes.size(); j++) { 2575 Node *n = bb->_nodes[j]; 2576 if( valid_bundle_info(n) ) { 2577 Bundle *bundle = node_bundling(n); 2578 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2579 tty->print("*** Bundle: "); 2580 bundle->dump(); 2581 } 2582 n->dump(); 2583 } 2584 } 2585 } 2586 #endif 2587 #ifdef ASSERT 2588 verify_good_schedule(bb,"after block local scheduling"); 2589 #endif 2590 } 2591 2592 #ifndef PRODUCT 2593 if (_cfg->C->trace_opto_output()) 2594 tty->print("# <- DoScheduling\n"); 2595 #endif 2596 2597 // Record final node-bundling array location 2598 _regalloc->C->set_node_bundling_base(_node_bundling_base); 2599 2600 } // end DoScheduling 2601 2602 //------------------------------verify_good_schedule--------------------------- 2603 // Verify that no live-range used in the block is killed in the block by a 2604 // wrong DEF. This doesn't verify live-ranges that span blocks. 2605 2606 // Check for edge existence. Used to avoid adding redundant precedence edges. 2607 static bool edge_from_to( Node *from, Node *to ) { 2608 for( uint i=0; i<from->len(); i++ ) 2609 if( from->in(i) == to ) 2610 return true; 2611 return false; 2612 } 2613 2614 #ifdef ASSERT 2615 //------------------------------verify_do_def---------------------------------- 2616 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2617 // Check for bad kills 2618 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2619 Node *prior_use = _reg_node[def]; 2620 if( prior_use && !edge_from_to(prior_use,n) ) { 2621 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2622 n->dump(); 2623 tty->print_cr("..."); 2624 prior_use->dump(); 2625 assert(edge_from_to(prior_use,n),msg); 2626 } 2627 _reg_node.map(def,NULL); // Kill live USEs 2628 } 2629 } 2630 2631 //------------------------------verify_good_schedule--------------------------- 2632 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2633 2634 // Zap to something reasonable for the verify code 2635 _reg_node.clear(); 2636 2637 // Walk over the block backwards. Check to make sure each DEF doesn't 2638 // kill a live value (other than the one it's supposed to). Add each 2639 // USE to the live set. 2640 for( uint i = b->_nodes.size()-1; i >= _bb_start; i-- ) { 2641 Node *n = b->_nodes[i]; 2642 int n_op = n->Opcode(); 2643 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2644 // Fat-proj kills a slew of registers 2645 RegMask rm = n->out_RegMask();// Make local copy 2646 while( rm.is_NotEmpty() ) { 2647 OptoReg::Name kill = rm.find_first_elem(); 2648 rm.Remove(kill); 2649 verify_do_def( n, kill, msg ); 2650 } 2651 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2652 // Get DEF'd registers the normal way 2653 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2654 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2655 } 2656 2657 // Now make all USEs live 2658 for( uint i=1; i<n->req(); i++ ) { 2659 Node *def = n->in(i); 2660 assert(def != 0, "input edge required"); 2661 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2662 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2663 if( OptoReg::is_valid(reg_lo) ) { 2664 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg); 2665 _reg_node.map(reg_lo,n); 2666 } 2667 if( OptoReg::is_valid(reg_hi) ) { 2668 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg); 2669 _reg_node.map(reg_hi,n); 2670 } 2671 } 2672 2673 } 2674 2675 // Zap to something reasonable for the Antidependence code 2676 _reg_node.clear(); 2677 } 2678 #endif 2679 2680 // Conditionally add precedence edges. Avoid putting edges on Projs. 2681 static void add_prec_edge_from_to( Node *from, Node *to ) { 2682 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2683 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2684 from = from->in(0); 2685 } 2686 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2687 !edge_from_to( from, to ) ) // Avoid duplicate edge 2688 from->add_prec(to); 2689 } 2690 2691 //------------------------------anti_do_def------------------------------------ 2692 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2693 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2694 return; 2695 2696 Node *pinch = _reg_node[def_reg]; // Get pinch point 2697 if( !pinch || _bbs[pinch->_idx] != b || // No pinch-point yet? 2698 is_def ) { // Check for a true def (not a kill) 2699 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2700 return; 2701 } 2702 2703 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2704 debug_only( def = (Node*)0xdeadbeef; ) 2705 2706 // After some number of kills there _may_ be a later def 2707 Node *later_def = NULL; 2708 2709 // Finding a kill requires a real pinch-point. 2710 // Check for not already having a pinch-point. 2711 // Pinch points are Op_Node's. 2712 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2713 later_def = pinch; // Must be def/kill as optimistic pinch-point 2714 if ( _pinch_free_list.size() > 0) { 2715 pinch = _pinch_free_list.pop(); 2716 } else { 2717 pinch = new (_cfg->C) Node(1); // Pinch point to-be 2718 } 2719 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2720 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2721 return; 2722 } 2723 _bbs.map(pinch->_idx,b); // Pretend it's valid in this block (lazy init) 2724 _reg_node.map(def_reg,pinch); // Record pinch-point 2725 //_regalloc->set_bad(pinch->_idx); // Already initialized this way. 2726 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2727 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call 2728 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2729 later_def = NULL; // and no later def 2730 } 2731 pinch->set_req(0,later_def); // Hook later def so we can find it 2732 } else { // Else have valid pinch point 2733 if( pinch->in(0) ) // If there is a later-def 2734 later_def = pinch->in(0); // Get it 2735 } 2736 2737 // Add output-dependence edge from later def to kill 2738 if( later_def ) // If there is some original def 2739 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2740 2741 // See if current kill is also a use, and so is forced to be the pinch-point. 2742 if( pinch->Opcode() == Op_Node ) { 2743 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2744 for( uint i=1; i<uses->req(); i++ ) { 2745 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2746 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2747 // Yes, found a use/kill pinch-point 2748 pinch->set_req(0,NULL); // 2749 pinch->replace_by(kill); // Move anti-dep edges up 2750 pinch = kill; 2751 _reg_node.map(def_reg,pinch); 2752 return; 2753 } 2754 } 2755 } 2756 2757 // Add edge from kill to pinch-point 2758 add_prec_edge_from_to(kill,pinch); 2759 } 2760 2761 //------------------------------anti_do_use------------------------------------ 2762 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2763 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2764 return; 2765 Node *pinch = _reg_node[use_reg]; // Get pinch point 2766 // Check for no later def_reg/kill in block 2767 if( pinch && _bbs[pinch->_idx] == b && 2768 // Use has to be block-local as well 2769 _bbs[use->_idx] == b ) { 2770 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2771 pinch->req() == 1 ) { // pinch not yet in block? 2772 pinch->del_req(0); // yank pointer to later-def, also set flag 2773 // Insert the pinch-point in the block just after the last use 2774 b->_nodes.insert(b->find_node(use)+1,pinch); 2775 _bb_end++; // Increase size scheduled region in block 2776 } 2777 2778 add_prec_edge_from_to(pinch,use); 2779 } 2780 } 2781 2782 //------------------------------ComputeRegisterAntidependences----------------- 2783 // We insert antidependences between the reads and following write of 2784 // allocated registers to prevent illegal code motion. Hopefully, the 2785 // number of added references should be fairly small, especially as we 2786 // are only adding references within the current basic block. 2787 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2788 2789 #ifdef ASSERT 2790 verify_good_schedule(b,"before block local scheduling"); 2791 #endif 2792 2793 // A valid schedule, for each register independently, is an endless cycle 2794 // of: a def, then some uses (connected to the def by true dependencies), 2795 // then some kills (defs with no uses), finally the cycle repeats with a new 2796 // def. The uses are allowed to float relative to each other, as are the 2797 // kills. No use is allowed to slide past a kill (or def). This requires 2798 // antidependencies between all uses of a single def and all kills that 2799 // follow, up to the next def. More edges are redundant, because later defs 2800 // & kills are already serialized with true or antidependencies. To keep 2801 // the edge count down, we add a 'pinch point' node if there's more than 2802 // one use or more than one kill/def. 2803 2804 // We add dependencies in one bottom-up pass. 2805 2806 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2807 2808 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2809 // register. If not, we record the DEF/KILL in _reg_node, the 2810 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2811 // "pinch point", a new Node that's in the graph but not in the block. 2812 // We put edges from the prior and current DEF/KILLs to the pinch point. 2813 // We put the pinch point in _reg_node. If there's already a pinch point 2814 // we merely add an edge from the current DEF/KILL to the pinch point. 2815 2816 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2817 // put an edge from the pinch point to the USE. 2818 2819 // To be expedient, the _reg_node array is pre-allocated for the whole 2820 // compilation. _reg_node is lazily initialized; it either contains a NULL, 2821 // or a valid def/kill/pinch-point, or a leftover node from some prior 2822 // block. Leftover node from some prior block is treated like a NULL (no 2823 // prior def, so no anti-dependence needed). Valid def is distinguished by 2824 // it being in the current block. 2825 bool fat_proj_seen = false; 2826 uint last_safept = _bb_end-1; 2827 Node* end_node = (_bb_end-1 >= _bb_start) ? b->_nodes[last_safept] : NULL; 2828 Node* last_safept_node = end_node; 2829 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2830 Node *n = b->_nodes[i]; 2831 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2832 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) { 2833 // Fat-proj kills a slew of registers 2834 // This can add edges to 'n' and obscure whether or not it was a def, 2835 // hence the is_def flag. 2836 fat_proj_seen = true; 2837 RegMask rm = n->out_RegMask();// Make local copy 2838 while( rm.is_NotEmpty() ) { 2839 OptoReg::Name kill = rm.find_first_elem(); 2840 rm.Remove(kill); 2841 anti_do_def( b, n, kill, is_def ); 2842 } 2843 } else { 2844 // Get DEF'd registers the normal way 2845 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2846 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2847 } 2848 2849 // Kill projections on a branch should appear to occur on the 2850 // branch, not afterwards, so grab the masks from the projections 2851 // and process them. 2852 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) { 2853 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2854 Node* use = n->fast_out(i); 2855 if (use->is_Proj()) { 2856 RegMask rm = use->out_RegMask();// Make local copy 2857 while( rm.is_NotEmpty() ) { 2858 OptoReg::Name kill = rm.find_first_elem(); 2859 rm.Remove(kill); 2860 anti_do_def( b, n, kill, false ); 2861 } 2862 } 2863 } 2864 } 2865 2866 // Check each register used by this instruction for a following DEF/KILL 2867 // that must occur afterward and requires an anti-dependence edge. 2868 for( uint j=0; j<n->req(); j++ ) { 2869 Node *def = n->in(j); 2870 if( def ) { 2871 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2872 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2873 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2874 } 2875 } 2876 // Do not allow defs of new derived values to float above GC 2877 // points unless the base is definitely available at the GC point. 2878 2879 Node *m = b->_nodes[i]; 2880 2881 // Add precedence edge from following safepoint to use of derived pointer 2882 if( last_safept_node != end_node && 2883 m != last_safept_node) { 2884 for (uint k = 1; k < m->req(); k++) { 2885 const Type *t = m->in(k)->bottom_type(); 2886 if( t->isa_oop_ptr() && 2887 t->is_ptr()->offset() != 0 ) { 2888 last_safept_node->add_prec( m ); 2889 break; 2890 } 2891 } 2892 } 2893 2894 if( n->jvms() ) { // Precedence edge from derived to safept 2895 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2896 if( b->_nodes[last_safept] != last_safept_node ) { 2897 last_safept = b->find_node(last_safept_node); 2898 } 2899 for( uint j=last_safept; j > i; j-- ) { 2900 Node *mach = b->_nodes[j]; 2901 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2902 mach->add_prec( n ); 2903 } 2904 last_safept = i; 2905 last_safept_node = m; 2906 } 2907 } 2908 2909 if (fat_proj_seen) { 2910 // Garbage collect pinch nodes that were not consumed. 2911 // They are usually created by a fat kill MachProj for a call. 2912 garbage_collect_pinch_nodes(); 2913 } 2914 } 2915 2916 //------------------------------garbage_collect_pinch_nodes------------------------------- 2917 2918 // Garbage collect pinch nodes for reuse by other blocks. 2919 // 2920 // The block scheduler's insertion of anti-dependence 2921 // edges creates many pinch nodes when the block contains 2922 // 2 or more Calls. A pinch node is used to prevent a 2923 // combinatorial explosion of edges. If a set of kills for a 2924 // register is anti-dependent on a set of uses (or defs), rather 2925 // than adding an edge in the graph between each pair of kill 2926 // and use (or def), a pinch is inserted between them: 2927 // 2928 // use1 use2 use3 2929 // \ | / 2930 // \ | / 2931 // pinch 2932 // / | \ 2933 // / | \ 2934 // kill1 kill2 kill3 2935 // 2936 // One pinch node is created per register killed when 2937 // the second call is encountered during a backwards pass 2938 // over the block. Most of these pinch nodes are never 2939 // wired into the graph because the register is never 2940 // used or def'ed in the block. 2941 // 2942 void Scheduling::garbage_collect_pinch_nodes() { 2943 #ifndef PRODUCT 2944 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 2945 #endif 2946 int trace_cnt = 0; 2947 for (uint k = 0; k < _reg_node.Size(); k++) { 2948 Node* pinch = _reg_node[k]; 2949 if (pinch != NULL && pinch->Opcode() == Op_Node && 2950 // no predecence input edges 2951 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { 2952 cleanup_pinch(pinch); 2953 _pinch_free_list.push(pinch); 2954 _reg_node.map(k, NULL); 2955 #ifndef PRODUCT 2956 if (_cfg->C->trace_opto_output()) { 2957 trace_cnt++; 2958 if (trace_cnt > 40) { 2959 tty->print("\n"); 2960 trace_cnt = 0; 2961 } 2962 tty->print(" %d", pinch->_idx); 2963 } 2964 #endif 2965 } 2966 } 2967 #ifndef PRODUCT 2968 if (_cfg->C->trace_opto_output()) tty->print("\n"); 2969 #endif 2970 } 2971 2972 // Clean up a pinch node for reuse. 2973 void Scheduling::cleanup_pinch( Node *pinch ) { 2974 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 2975 2976 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 2977 Node* use = pinch->last_out(i); 2978 uint uses_found = 0; 2979 for (uint j = use->req(); j < use->len(); j++) { 2980 if (use->in(j) == pinch) { 2981 use->rm_prec(j); 2982 uses_found++; 2983 } 2984 } 2985 assert(uses_found > 0, "must be a precedence edge"); 2986 i -= uses_found; // we deleted 1 or more copies of this edge 2987 } 2988 // May have a later_def entry 2989 pinch->set_req(0, NULL); 2990 } 2991 2992 //------------------------------print_statistics------------------------------- 2993 #ifndef PRODUCT 2994 2995 void Scheduling::dump_available() const { 2996 tty->print("#Availist "); 2997 for (uint i = 0; i < _available.size(); i++) 2998 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 2999 tty->cr(); 3000 } 3001 3002 // Print Scheduling Statistics 3003 void Scheduling::print_statistics() { 3004 // Print the size added by nops for bundling 3005 tty->print("Nops added %d bytes to total of %d bytes", 3006 _total_nop_size, _total_method_size); 3007 if (_total_method_size > 0) 3008 tty->print(", for %.2f%%", 3009 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 3010 tty->print("\n"); 3011 3012 // Print the number of branch shadows filled 3013 if (Pipeline::_branch_has_delay_slot) { 3014 tty->print("Of %d branches, %d had unconditional delay slots filled", 3015 _total_branches, _total_unconditional_delays); 3016 if (_total_branches > 0) 3017 tty->print(", for %.2f%%", 3018 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 3019 tty->print("\n"); 3020 } 3021 3022 uint total_instructions = 0, total_bundles = 0; 3023 3024 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 3025 uint bundle_count = _total_instructions_per_bundle[i]; 3026 total_instructions += bundle_count * i; 3027 total_bundles += bundle_count; 3028 } 3029 3030 if (total_bundles > 0) 3031 tty->print("Average ILP (excluding nops) is %.2f\n", 3032 ((double)total_instructions) / ((double)total_bundles)); 3033 } 3034 #endif