1 /* 2 * Copyright (c) 1998, 2013, 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/compiledIC.hpp" 28 #include "code/debugInfo.hpp" 29 #include "code/debugInfoRec.hpp" 30 #include "compiler/compileBroker.hpp" 31 #include "compiler/oopMap.hpp" 32 #include "memory/allocation.inline.hpp" 33 #include "opto/callnode.hpp" 34 #include "opto/cfgnode.hpp" 35 #include "opto/locknode.hpp" 36 #include "opto/machnode.hpp" 37 #include "opto/output.hpp" 38 #include "opto/regalloc.hpp" 39 #include "opto/runtime.hpp" 40 #include "opto/subnode.hpp" 41 #include "opto/type.hpp" 42 #include "runtime/handles.inline.hpp" 43 #include "utilities/xmlstream.hpp" 44 45 extern uint size_exception_handler(); 46 extern uint size_deopt_handler(); 47 48 #ifndef PRODUCT 49 #define DEBUG_ARG(x) , x 50 #else 51 #define DEBUG_ARG(x) 52 #endif 53 54 extern int emit_exception_handler(CodeBuffer &cbuf); 55 extern int emit_deopt_handler(CodeBuffer &cbuf); 56 57 // Convert Nodes to instruction bits and pass off to the VM 58 void Compile::Output() { 59 // RootNode goes 60 assert( _cfg->get_root_block()->number_of_nodes() == 0, "" ); 61 62 // The number of new nodes (mostly MachNop) is proportional to 63 // the number of java calls and inner loops which are aligned. 64 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 + 65 C->inner_loops()*(OptoLoopAlignment-1)), 66 "out of nodes before code generation" ) ) { 67 return; 68 } 69 // Make sure I can find the Start Node 70 Block *entry = _cfg->get_block(1); 71 Block *broot = _cfg->get_root_block(); 72 73 const StartNode *start = entry->head()->as_Start(); 74 75 // Replace StartNode with prolog 76 MachPrologNode *prolog = new (this) MachPrologNode(); 77 entry->map_node(prolog, 0); 78 _cfg->map_node_to_block(prolog, entry); 79 _cfg->unmap_node_from_block(start); // start is no longer in any block 80 81 // Virtual methods need an unverified entry point 82 83 if( is_osr_compilation() ) { 84 if( PoisonOSREntry ) { 85 // TODO: Should use a ShouldNotReachHereNode... 86 _cfg->insert( broot, 0, new (this) MachBreakpointNode() ); 87 } 88 } else { 89 if( _method && !_method->flags().is_static() ) { 90 // Insert unvalidated entry point 91 _cfg->insert( broot, 0, new (this) MachUEPNode() ); 92 } 93 94 } 95 96 97 // Break before main entry point 98 if( (_method && _method->break_at_execute()) 99 #ifndef PRODUCT 100 ||(OptoBreakpoint && is_method_compilation()) 101 ||(OptoBreakpointOSR && is_osr_compilation()) 102 ||(OptoBreakpointC2R && !_method) 103 #endif 104 ) { 105 // checking for _method means that OptoBreakpoint does not apply to 106 // runtime stubs or frame converters 107 _cfg->insert( entry, 1, new (this) MachBreakpointNode() ); 108 } 109 110 // Insert epilogs before every return 111 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 112 Block* block = _cfg->get_block(i); 113 if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point? 114 Node* m = block->end(); 115 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) { 116 MachEpilogNode* epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); 117 block->add_inst(epilog); 118 _cfg->map_node_to_block(epilog, block); 119 } 120 } 121 } 122 123 # ifdef ENABLE_ZAP_DEAD_LOCALS 124 if (ZapDeadCompiledLocals) { 125 Insert_zap_nodes(); 126 } 127 # endif 128 129 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1); 130 blk_starts[0] = 0; 131 132 // Initialize code buffer and process short branches. 133 CodeBuffer* cb = init_buffer(blk_starts); 134 135 if (cb == NULL || failing()) { 136 return; 137 } 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->number_of_blocks(); i++) { 145 tty->print("\nBB#%03d:\n", i); 146 Block* block = _cfg->get_block(i); 147 for (uint j = 0; j < block->number_of_nodes(); j++) { 148 Node* n = block->get_node(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()) { 158 return; 159 } 160 161 BuildOopMaps(); 162 163 if (failing()) { 164 return; 165 } 166 167 fill_buffer(cb, blk_starts); 168 } 169 170 bool Compile::need_stack_bang(int frame_size_in_bytes) const { 171 // Determine if we need to generate a stack overflow check. 172 // Do it if the method is not a stub function and 173 // has java calls or has frame size > vm_page_size/8. 174 return (UseStackBanging && stub_function() == NULL && 175 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3)); 176 } 177 178 bool Compile::need_register_stack_bang() const { 179 // Determine if we need to generate a register stack overflow check. 180 // This is only used on architectures which have split register 181 // and memory stacks (ie. IA64). 182 // Bang if the method is not a stub function and has java calls 183 return (stub_function() == NULL && has_java_calls()); 184 } 185 186 # ifdef ENABLE_ZAP_DEAD_LOCALS 187 188 189 // In order to catch compiler oop-map bugs, we have implemented 190 // a debugging mode called ZapDeadCompilerLocals. 191 // This mode causes the compiler to insert a call to a runtime routine, 192 // "zap_dead_locals", right before each place in compiled code 193 // that could potentially be a gc-point (i.e., a safepoint or oop map point). 194 // The runtime routine checks that locations mapped as oops are really 195 // oops, that locations mapped as values do not look like oops, 196 // and that locations mapped as dead are not used later 197 // (by zapping them to an invalid address). 198 199 int Compile::_CompiledZap_count = 0; 200 201 void Compile::Insert_zap_nodes() { 202 bool skip = false; 203 204 205 // Dink with static counts because code code without the extra 206 // runtime calls is MUCH faster for debugging purposes 207 208 if ( CompileZapFirst == 0 ) ; // nothing special 209 else if ( CompileZapFirst > CompiledZap_count() ) skip = true; 210 else if ( CompileZapFirst == CompiledZap_count() ) 211 warning("starting zap compilation after skipping"); 212 213 if ( CompileZapLast == -1 ) ; // nothing special 214 else if ( CompileZapLast < CompiledZap_count() ) skip = true; 215 else if ( CompileZapLast == CompiledZap_count() ) 216 warning("about to compile last zap"); 217 218 ++_CompiledZap_count; // counts skipped zaps, too 219 220 if ( skip ) return; 221 222 223 if ( _method == NULL ) 224 return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care 225 226 // Insert call to zap runtime stub before every node with an oop map 227 for( uint i=0; i<_cfg->number_of_blocks(); i++ ) { 228 Block *b = _cfg->get_block(i); 229 for ( uint j = 0; j < b->number_of_nodes(); ++j ) { 230 Node *n = b->get_node(j); 231 232 // Determining if we should insert a zap-a-lot node in output. 233 // We do that for all nodes that has oopmap info, except for calls 234 // to allocation. Calls to allocation passes in the old top-of-eden pointer 235 // and expect the C code to reset it. Hence, there can be no safepoints between 236 // the inlined-allocation and the call to new_Java, etc. 237 // We also cannot zap monitor calls, as they must hold the microlock 238 // during the call to Zap, which also wants to grab the microlock. 239 bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL); 240 if ( insert ) { // it is MachSafePoint 241 if ( !n->is_MachCall() ) { 242 insert = false; 243 } else if ( n->is_MachCall() ) { 244 MachCallNode* call = n->as_MachCall(); 245 if (call->entry_point() == OptoRuntime::new_instance_Java() || 246 call->entry_point() == OptoRuntime::new_array_Java() || 247 call->entry_point() == OptoRuntime::multianewarray2_Java() || 248 call->entry_point() == OptoRuntime::multianewarray3_Java() || 249 call->entry_point() == OptoRuntime::multianewarray4_Java() || 250 call->entry_point() == OptoRuntime::multianewarray5_Java() || 251 call->entry_point() == OptoRuntime::slow_arraycopy_Java() || 252 call->entry_point() == OptoRuntime::complete_monitor_locking_Java() 253 ) { 254 insert = false; 255 } 256 } 257 if (insert) { 258 Node *zap = call_zap_node(n->as_MachSafePoint(), i); 259 b->insert_node(zap, j); 260 _cfg->map_node_to_block(zap, b); 261 ++j; 262 } 263 } 264 } 265 } 266 } 267 268 269 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) { 270 const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type(); 271 CallStaticJavaNode* ideal_node = 272 new (this) CallStaticJavaNode( tf, 273 OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()), 274 "call zap dead locals stub", 0, TypePtr::BOTTOM); 275 // We need to copy the OopMap from the site we're zapping at. 276 // We have to make a copy, because the zap site might not be 277 // a call site, and zap_dead is a call site. 278 OopMap* clone = node_to_check->oop_map()->deep_copy(); 279 280 // Add the cloned OopMap to the zap node 281 ideal_node->set_oop_map(clone); 282 return _matcher->match_sfpt(ideal_node); 283 } 284 285 bool Compile::is_node_getting_a_safepoint( Node* n) { 286 // This code duplicates the logic prior to the call of add_safepoint 287 // below in this file. 288 if( n->is_MachSafePoint() ) return true; 289 return false; 290 } 291 292 # endif // ENABLE_ZAP_DEAD_LOCALS 293 294 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top 295 // of a loop. When aligning a loop we need to provide enough instructions 296 // in cpu's fetch buffer to feed decoders. The loop alignment could be 297 // avoided if we have enough instructions in fetch buffer at the head of a loop. 298 // By default, the size is set to 999999 by Block's constructor so that 299 // a loop will be aligned if the size is not reset here. 300 // 301 // Note: Mach instructions could contain several HW instructions 302 // so the size is estimated only. 303 // 304 void Compile::compute_loop_first_inst_sizes() { 305 // The next condition is used to gate the loop alignment optimization. 306 // Don't aligned a loop if there are enough instructions at the head of a loop 307 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad 308 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is 309 // equal to 11 bytes which is the largest address NOP instruction. 310 if (MaxLoopPad < OptoLoopAlignment - 1) { 311 uint last_block = _cfg->number_of_blocks() - 1; 312 for (uint i = 1; i <= last_block; i++) { 313 Block* block = _cfg->get_block(i); 314 // Check the first loop's block which requires an alignment. 315 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) { 316 uint sum_size = 0; 317 uint inst_cnt = NumberOfLoopInstrToAlign; 318 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 319 320 // Check subsequent fallthrough blocks if the loop's first 321 // block(s) does not have enough instructions. 322 Block *nb = block; 323 while(inst_cnt > 0 && 324 i < last_block && 325 !_cfg->get_block(i + 1)->has_loop_alignment() && 326 !nb->has_successor(block)) { 327 i++; 328 nb = _cfg->get_block(i); 329 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 330 } // while( inst_cnt > 0 && i < last_block ) 331 332 block->set_first_inst_size(sum_size); 333 } // f( b->head()->is_Loop() ) 334 } // for( i <= last_block ) 335 } // if( MaxLoopPad < OptoLoopAlignment-1 ) 336 } 337 338 // The architecture description provides short branch variants for some long 339 // branch instructions. Replace eligible long branches with short branches. 340 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) { 341 // Compute size of each block, method size, and relocation information size 342 uint nblocks = _cfg->number_of_blocks(); 343 344 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 345 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 346 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks); 347 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); ) 348 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); ) 349 350 bool has_short_branch_candidate = false; 351 352 // Initialize the sizes to 0 353 code_size = 0; // Size in bytes of generated code 354 stub_size = 0; // Size in bytes of all stub entries 355 // Size in bytes of all relocation entries, including those in local stubs. 356 // Start with 2-bytes of reloc info for the unvalidated entry point 357 reloc_size = 1; // Number of relocation entries 358 359 // Make three passes. The first computes pessimistic blk_starts, 360 // relative jmp_offset and reloc_size information. The second performs 361 // short branch substitution using the pessimistic sizing. The 362 // third inserts nops where needed. 363 364 // Step one, perform a pessimistic sizing pass. 365 uint last_call_adr = max_uint; 366 uint last_avoid_back_to_back_adr = max_uint; 367 uint nop_size = (new (this) MachNopNode())->size(_regalloc); 368 for (uint i = 0; i < nblocks; i++) { // For all blocks 369 Block* block = _cfg->get_block(i); 370 371 // During short branch replacement, we store the relative (to blk_starts) 372 // offset of jump in jmp_offset, rather than the absolute offset of jump. 373 // This is so that we do not need to recompute sizes of all nodes when 374 // we compute correct blk_starts in our next sizing pass. 375 jmp_offset[i] = 0; 376 jmp_size[i] = 0; 377 jmp_nidx[i] = -1; 378 DEBUG_ONLY( jmp_target[i] = 0; ) 379 DEBUG_ONLY( jmp_rule[i] = 0; ) 380 381 // Sum all instruction sizes to compute block size 382 uint last_inst = block->number_of_nodes(); 383 uint blk_size = 0; 384 for (uint j = 0; j < last_inst; j++) { 385 Node* nj = block->get_node(j); 386 // Handle machine instruction nodes 387 if (nj->is_Mach()) { 388 MachNode *mach = nj->as_Mach(); 389 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 390 reloc_size += mach->reloc(); 391 if (mach->is_MachCall()) { 392 MachCallNode *mcall = mach->as_MachCall(); 393 // This destination address is NOT PC-relative 394 395 mcall->method_set((intptr_t)mcall->entry_point()); 396 397 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) { 398 stub_size += CompiledStaticCall::to_interp_stub_size(); 399 reloc_size += CompiledStaticCall::reloc_to_interp_stub(); 400 } 401 } else if (mach->is_MachSafePoint()) { 402 // If call/safepoint are adjacent, account for possible 403 // nop to disambiguate the two safepoints. 404 // ScheduleAndBundle() can rearrange nodes in a block, 405 // check for all offsets inside this block. 406 if (last_call_adr >= blk_starts[i]) { 407 blk_size += nop_size; 408 } 409 } 410 if (mach->avoid_back_to_back()) { 411 // Nop is inserted between "avoid back to back" instructions. 412 // ScheduleAndBundle() can rearrange nodes in a block, 413 // check for all offsets inside this block. 414 if (last_avoid_back_to_back_adr >= blk_starts[i]) { 415 blk_size += nop_size; 416 } 417 } 418 if (mach->may_be_short_branch()) { 419 if (!nj->is_MachBranch()) { 420 #ifndef PRODUCT 421 nj->dump(3); 422 #endif 423 Unimplemented(); 424 } 425 assert(jmp_nidx[i] == -1, "block should have only one branch"); 426 jmp_offset[i] = blk_size; 427 jmp_size[i] = nj->size(_regalloc); 428 jmp_nidx[i] = j; 429 has_short_branch_candidate = true; 430 } 431 } 432 blk_size += nj->size(_regalloc); 433 // Remember end of call offset 434 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) { 435 last_call_adr = blk_starts[i]+blk_size; 436 } 437 // Remember end of avoid_back_to_back offset 438 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back()) { 439 last_avoid_back_to_back_adr = blk_starts[i]+blk_size; 440 } 441 } 442 443 // When the next block starts a loop, we may insert pad NOP 444 // instructions. Since we cannot know our future alignment, 445 // assume the worst. 446 if (i < nblocks - 1) { 447 Block* nb = _cfg->get_block(i + 1); 448 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 449 if (max_loop_pad > 0) { 450 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 451 // Adjust last_call_adr and/or last_avoid_back_to_back_adr. 452 // If either is the last instruction in this block, bump by 453 // max_loop_pad in lock-step with blk_size, so sizing 454 // calculations in subsequent blocks still can conservatively 455 // detect that it may the last instruction in this block. 456 if (last_call_adr == blk_starts[i]+blk_size) { 457 last_call_adr += max_loop_pad; 458 } 459 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) { 460 last_avoid_back_to_back_adr += max_loop_pad; 461 } 462 blk_size += max_loop_pad; 463 } 464 } 465 466 // Save block size; update total method size 467 blk_starts[i+1] = blk_starts[i]+blk_size; 468 } 469 470 // Step two, replace eligible long jumps. 471 bool progress = true; 472 uint last_may_be_short_branch_adr = max_uint; 473 while (has_short_branch_candidate && progress) { 474 progress = false; 475 has_short_branch_candidate = false; 476 int adjust_block_start = 0; 477 for (uint i = 0; i < nblocks; i++) { 478 Block* block = _cfg->get_block(i); 479 int idx = jmp_nidx[i]; 480 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach(); 481 if (mach != NULL && mach->may_be_short_branch()) { 482 #ifdef ASSERT 483 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity"); 484 int j; 485 // Find the branch; ignore trailing NOPs. 486 for (j = block->number_of_nodes()-1; j>=0; j--) { 487 Node* n = block->get_node(j); 488 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) 489 break; 490 } 491 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity"); 492 #endif 493 int br_size = jmp_size[i]; 494 int br_offs = blk_starts[i] + jmp_offset[i]; 495 496 // This requires the TRUE branch target be in succs[0] 497 uint bnum = block->non_connector_successor(0)->_pre_order; 498 int offset = blk_starts[bnum] - br_offs; 499 if (bnum > i) { // adjust following block's offset 500 offset -= adjust_block_start; 501 } 502 503 // This block can be a loop header, account for the padding 504 // in the previous block. 505 int prev_block_loop_pad = block->code_alignment() - relocInfo::addr_unit(); 506 if (i > 0 && prev_block_loop_pad > 0) { 507 assert(br_offs >= prev_block_loop_pad, "Should have at least a padding on top"); 508 } else { 509 // First block or not a loop 510 prev_block_loop_pad = 0; 511 } 512 // In the following code a nop could be inserted before 513 // the branch which will increase the backward distance. 514 bool needs_padding = ((uint)(br_offs - prev_block_loop_pad) == last_may_be_short_branch_adr); 515 if (needs_padding && offset <= 0) 516 offset -= nop_size; 517 518 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 519 // We've got a winner. Replace this branch. 520 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this); 521 522 // Update the jmp_size. 523 int new_size = replacement->size(_regalloc); 524 int diff = br_size - new_size; 525 assert(diff >= (int)nop_size, "short_branch size should be smaller"); 526 // Conservatively take into accound padding between 527 // avoid_back_to_back branches. Previous branch could be 528 // converted into avoid_back_to_back branch during next 529 // rounds. 530 if (needs_padding && replacement->avoid_back_to_back()) { 531 jmp_offset[i] += nop_size; 532 diff -= nop_size; 533 } 534 adjust_block_start += diff; 535 block->map_node(replacement, idx); 536 mach->subsume_by(replacement, C); 537 mach = replacement; 538 progress = true; 539 540 jmp_size[i] = new_size; 541 DEBUG_ONLY( jmp_target[i] = bnum; ); 542 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 543 } else { 544 // The jump distance is not short, try again during next iteration. 545 has_short_branch_candidate = true; 546 } 547 } // (mach->may_be_short_branch()) 548 if (mach != NULL && (mach->may_be_short_branch() || 549 mach->avoid_back_to_back())) { 550 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i]; 551 } 552 blk_starts[i+1] -= adjust_block_start; 553 } 554 } 555 556 #ifdef ASSERT 557 for (uint i = 0; i < nblocks; i++) { // For all blocks 558 if (jmp_target[i] != 0) { 559 int br_size = jmp_size[i]; 560 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 561 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 562 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]); 563 } 564 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp"); 565 } 566 } 567 #endif 568 569 // Step 3, compute the offsets of all blocks, will be done in fill_buffer() 570 // after ScheduleAndBundle(). 571 572 // ------------------ 573 // Compute size for code buffer 574 code_size = blk_starts[nblocks]; 575 576 // Relocation records 577 reloc_size += 1; // Relo entry for exception handler 578 579 // Adjust reloc_size to number of record of relocation info 580 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 581 // a relocation index. 582 // The CodeBuffer will expand the locs array if this estimate is too low. 583 reloc_size *= 10 / sizeof(relocInfo); 584 } 585 586 //------------------------------FillLocArray----------------------------------- 587 // Create a bit of debug info and append it to the array. The mapping is from 588 // Java local or expression stack to constant, register or stack-slot. For 589 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 590 // entry has been taken care of and caller should skip it). 591 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 592 // This should never have accepted Bad before 593 assert(OptoReg::is_valid(regnum), "location must be valid"); 594 return (OptoReg::is_reg(regnum)) 595 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 596 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 597 } 598 599 600 ObjectValue* 601 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 602 for (int i = 0; i < objs->length(); i++) { 603 assert(objs->at(i)->is_object(), "corrupt object cache"); 604 ObjectValue* sv = (ObjectValue*) objs->at(i); 605 if (sv->id() == id) { 606 return sv; 607 } 608 } 609 // Otherwise.. 610 return NULL; 611 } 612 613 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 614 ObjectValue* sv ) { 615 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition"); 616 objs->append(sv); 617 } 618 619 620 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 621 GrowableArray<ScopeValue*> *array, 622 GrowableArray<ScopeValue*> *objs ) { 623 assert( local, "use _top instead of null" ); 624 if (array->length() != idx) { 625 assert(array->length() == idx + 1, "Unexpected array count"); 626 // Old functionality: 627 // return 628 // New functionality: 629 // Assert if the local is not top. In product mode let the new node 630 // override the old entry. 631 assert(local == top(), "LocArray collision"); 632 if (local == top()) { 633 return; 634 } 635 array->pop(); 636 } 637 const Type *t = local->bottom_type(); 638 639 // Is it a safepoint scalar object node? 640 if (local->is_SafePointScalarObject()) { 641 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 642 643 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx); 644 if (sv == NULL) { 645 ciKlass* cik = t->is_oopptr()->klass(); 646 assert(cik->is_instance_klass() || 647 cik->is_array_klass(), "Not supported allocation."); 648 sv = new ObjectValue(spobj->_idx, 649 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 650 Compile::set_sv_for_object_node(objs, sv); 651 652 uint first_ind = spobj->first_index(sfpt->jvms()); 653 for (uint i = 0; i < spobj->n_fields(); i++) { 654 Node* fld_node = sfpt->in(first_ind+i); 655 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 656 } 657 } 658 array->append(sv); 659 return; 660 } 661 662 // Grab the register number for the local 663 OptoReg::Name regnum = _regalloc->get_reg_first(local); 664 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 665 // Record the double as two float registers. 666 // The register mask for such a value always specifies two adjacent 667 // float registers, with the lower register number even. 668 // Normally, the allocation of high and low words to these registers 669 // is irrelevant, because nearly all operations on register pairs 670 // (e.g., StoreD) treat them as a single unit. 671 // Here, we assume in addition that the words in these two registers 672 // stored "naturally" (by operations like StoreD and double stores 673 // within the interpreter) such that the lower-numbered register 674 // is written to the lower memory address. This may seem like 675 // a machine dependency, but it is not--it is a requirement on 676 // the author of the <arch>.ad file to ensure that, for every 677 // even/odd double-register pair to which a double may be allocated, 678 // the word in the even single-register is stored to the first 679 // memory word. (Note that register numbers are completely 680 // arbitrary, and are not tied to any machine-level encodings.) 681 #ifdef _LP64 682 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 683 array->append(new ConstantIntValue(0)); 684 array->append(new_loc_value( _regalloc, regnum, Location::dbl )); 685 } else if ( t->base() == Type::Long ) { 686 array->append(new ConstantIntValue(0)); 687 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 688 } else if ( t->base() == Type::RawPtr ) { 689 // jsr/ret return address which must be restored into a the full 690 // width 64-bit stack slot. 691 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 692 } 693 #else //_LP64 694 #ifdef SPARC 695 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) { 696 // For SPARC we have to swap high and low words for 697 // long values stored in a single-register (g0-g7). 698 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 699 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 700 } else 701 #endif //SPARC 702 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 703 // Repack the double/long as two jints. 704 // The convention the interpreter uses is that the second local 705 // holds the first raw word of the native double representation. 706 // This is actually reasonable, since locals and stack arrays 707 // grow downwards in all implementations. 708 // (If, on some machine, the interpreter's Java locals or stack 709 // were to grow upwards, the embedded doubles would be word-swapped.) 710 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 711 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 712 } 713 #endif //_LP64 714 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 715 OptoReg::is_reg(regnum) ) { 716 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double() 717 ? Location::float_in_dbl : Location::normal )); 718 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 719 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long 720 ? Location::int_in_long : Location::normal )); 721 } else if( t->base() == Type::NarrowOop ) { 722 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop )); 723 } else { 724 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal )); 725 } 726 return; 727 } 728 729 // No register. It must be constant data. 730 switch (t->base()) { 731 case Type::Half: // Second half of a double 732 ShouldNotReachHere(); // Caller should skip 2nd halves 733 break; 734 case Type::AnyPtr: 735 array->append(new ConstantOopWriteValue(NULL)); 736 break; 737 case Type::AryPtr: 738 case Type::InstPtr: // fall through 739 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 740 break; 741 case Type::NarrowOop: 742 if (t == TypeNarrowOop::NULL_PTR) { 743 array->append(new ConstantOopWriteValue(NULL)); 744 } else { 745 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 746 } 747 break; 748 case Type::Int: 749 array->append(new ConstantIntValue(t->is_int()->get_con())); 750 break; 751 case Type::RawPtr: 752 // A return address (T_ADDRESS). 753 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 754 #ifdef _LP64 755 // Must be restored to the full-width 64-bit stack slot. 756 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 757 #else 758 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 759 #endif 760 break; 761 case Type::FloatCon: { 762 float f = t->is_float_constant()->getf(); 763 array->append(new ConstantIntValue(jint_cast(f))); 764 break; 765 } 766 case Type::DoubleCon: { 767 jdouble d = t->is_double_constant()->getd(); 768 #ifdef _LP64 769 array->append(new ConstantIntValue(0)); 770 array->append(new ConstantDoubleValue(d)); 771 #else 772 // Repack the double as two jints. 773 // The convention the interpreter uses is that the second local 774 // holds the first raw word of the native double representation. 775 // This is actually reasonable, since locals and stack arrays 776 // grow downwards in all implementations. 777 // (If, on some machine, the interpreter's Java locals or stack 778 // were to grow upwards, the embedded doubles would be word-swapped.) 779 jint *dp = (jint*)&d; 780 array->append(new ConstantIntValue(dp[1])); 781 array->append(new ConstantIntValue(dp[0])); 782 #endif 783 break; 784 } 785 case Type::Long: { 786 jlong d = t->is_long()->get_con(); 787 #ifdef _LP64 788 array->append(new ConstantIntValue(0)); 789 array->append(new ConstantLongValue(d)); 790 #else 791 // Repack the long as two jints. 792 // The convention the interpreter uses is that the second local 793 // holds the first raw word of the native double representation. 794 // This is actually reasonable, since locals and stack arrays 795 // grow downwards in all implementations. 796 // (If, on some machine, the interpreter's Java locals or stack 797 // were to grow upwards, the embedded doubles would be word-swapped.) 798 jint *dp = (jint*)&d; 799 array->append(new ConstantIntValue(dp[1])); 800 array->append(new ConstantIntValue(dp[0])); 801 #endif 802 break; 803 } 804 case Type::Top: // Add an illegal value here 805 array->append(new LocationValue(Location())); 806 break; 807 default: 808 ShouldNotReachHere(); 809 break; 810 } 811 } 812 813 // Determine if this node starts a bundle 814 bool Compile::starts_bundle(const Node *n) const { 815 return (_node_bundling_limit > n->_idx && 816 _node_bundling_base[n->_idx].starts_bundle()); 817 } 818 819 //--------------------------Process_OopMap_Node-------------------------------- 820 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) { 821 822 // Handle special safepoint nodes for synchronization 823 MachSafePointNode *sfn = mach->as_MachSafePoint(); 824 MachCallNode *mcall; 825 826 #ifdef ENABLE_ZAP_DEAD_LOCALS 827 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative"); 828 #endif 829 830 int safepoint_pc_offset = current_offset; 831 bool is_method_handle_invoke = false; 832 bool return_oop = false; 833 834 // Add the safepoint in the DebugInfoRecorder 835 if( !mach->is_MachCall() ) { 836 mcall = NULL; 837 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 838 } else { 839 mcall = mach->as_MachCall(); 840 841 // Is the call a MethodHandle call? 842 if (mcall->is_MachCallJava()) { 843 if (mcall->as_MachCallJava()->_method_handle_invoke) { 844 assert(has_method_handle_invokes(), "must have been set during call generation"); 845 is_method_handle_invoke = true; 846 } 847 } 848 849 // Check if a call returns an object. 850 if (mcall->return_value_is_used() && 851 mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) { 852 return_oop = true; 853 } 854 safepoint_pc_offset += mcall->ret_addr_offset(); 855 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 856 } 857 858 // Loop over the JVMState list to add scope information 859 // Do not skip safepoints with a NULL method, they need monitor info 860 JVMState* youngest_jvms = sfn->jvms(); 861 int max_depth = youngest_jvms->depth(); 862 863 // Allocate the object pool for scalar-replaced objects -- the map from 864 // small-integer keys (which can be recorded in the local and ostack 865 // arrays) to descriptions of the object state. 866 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 867 868 // Visit scopes from oldest to youngest. 869 for (int depth = 1; depth <= max_depth; depth++) { 870 JVMState* jvms = youngest_jvms->of_depth(depth); 871 int idx; 872 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 873 // Safepoints that do not have method() set only provide oop-map and monitor info 874 // to support GC; these do not support deoptimization. 875 int num_locs = (method == NULL) ? 0 : jvms->loc_size(); 876 int num_exps = (method == NULL) ? 0 : jvms->stk_size(); 877 int num_mon = jvms->nof_monitors(); 878 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), 879 "JVMS local count must match that of the method"); 880 881 // Add Local and Expression Stack Information 882 883 // Insert locals into the locarray 884 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 885 for( idx = 0; idx < num_locs; idx++ ) { 886 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 887 } 888 889 // Insert expression stack entries into the exparray 890 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 891 for( idx = 0; idx < num_exps; idx++ ) { 892 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 893 } 894 895 // Add in mappings of the monitors 896 assert( !method || 897 !method->is_synchronized() || 898 method->is_native() || 899 num_mon > 0 || 900 !GenerateSynchronizationCode, 901 "monitors must always exist for synchronized methods"); 902 903 // Build the growable array of ScopeValues for exp stack 904 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 905 906 // Loop over monitors and insert into array 907 for (idx = 0; idx < num_mon; idx++) { 908 // Grab the node that defines this monitor 909 Node* box_node = sfn->monitor_box(jvms, idx); 910 Node* obj_node = sfn->monitor_obj(jvms, idx); 911 912 // Create ScopeValue for object 913 ScopeValue *scval = NULL; 914 915 if (obj_node->is_SafePointScalarObject()) { 916 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 917 scval = Compile::sv_for_node_id(objs, spobj->_idx); 918 if (scval == NULL) { 919 const Type *t = spobj->bottom_type(); 920 ciKlass* cik = t->is_oopptr()->klass(); 921 assert(cik->is_instance_klass() || 922 cik->is_array_klass(), "Not supported allocation."); 923 ObjectValue* sv = new ObjectValue(spobj->_idx, 924 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 925 Compile::set_sv_for_object_node(objs, sv); 926 927 uint first_ind = spobj->first_index(youngest_jvms); 928 for (uint i = 0; i < spobj->n_fields(); i++) { 929 Node* fld_node = sfn->in(first_ind+i); 930 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 931 } 932 scval = sv; 933 } 934 } else if (!obj_node->is_Con()) { 935 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); 936 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 937 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop ); 938 } else { 939 scval = new_loc_value( _regalloc, obj_reg, Location::oop ); 940 } 941 } else { 942 const TypePtr *tp = obj_node->get_ptr_type(); 943 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding()); 944 } 945 946 OptoReg::Name box_reg = BoxLockNode::reg(box_node); 947 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)); 948 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated()); 949 monarray->append(new MonitorValue(scval, basic_lock, eliminated)); 950 } 951 952 // We dump the object pool first, since deoptimization reads it in first. 953 debug_info()->dump_object_pool(objs); 954 955 // Build first class objects to pass to scope 956 DebugToken *locvals = debug_info()->create_scope_values(locarray); 957 DebugToken *expvals = debug_info()->create_scope_values(exparray); 958 DebugToken *monvals = debug_info()->create_monitor_values(monarray); 959 960 // Make method available for all Safepoints 961 ciMethod* scope_method = method ? method : _method; 962 // Describe the scope here 963 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 964 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 965 // Now we can describe the scope. 966 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals); 967 } // End jvms loop 968 969 // Mark the end of the scope set. 970 debug_info()->end_safepoint(safepoint_pc_offset); 971 } 972 973 974 975 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 976 class NonSafepointEmitter { 977 Compile* C; 978 JVMState* _pending_jvms; 979 int _pending_offset; 980 981 void emit_non_safepoint(); 982 983 public: 984 NonSafepointEmitter(Compile* compile) { 985 this->C = compile; 986 _pending_jvms = NULL; 987 _pending_offset = 0; 988 } 989 990 void observe_instruction(Node* n, int pc_offset) { 991 if (!C->debug_info()->recording_non_safepoints()) return; 992 993 Node_Notes* nn = C->node_notes_at(n->_idx); 994 if (nn == NULL || nn->jvms() == NULL) return; 995 if (_pending_jvms != NULL && 996 _pending_jvms->same_calls_as(nn->jvms())) { 997 // Repeated JVMS? Stretch it up here. 998 _pending_offset = pc_offset; 999 } else { 1000 if (_pending_jvms != NULL && 1001 _pending_offset < pc_offset) { 1002 emit_non_safepoint(); 1003 } 1004 _pending_jvms = NULL; 1005 if (pc_offset > C->debug_info()->last_pc_offset()) { 1006 // This is the only way _pending_jvms can become non-NULL: 1007 _pending_jvms = nn->jvms(); 1008 _pending_offset = pc_offset; 1009 } 1010 } 1011 } 1012 1013 // Stay out of the way of real safepoints: 1014 void observe_safepoint(JVMState* jvms, int pc_offset) { 1015 if (_pending_jvms != NULL && 1016 !_pending_jvms->same_calls_as(jvms) && 1017 _pending_offset < pc_offset) { 1018 emit_non_safepoint(); 1019 } 1020 _pending_jvms = NULL; 1021 } 1022 1023 void flush_at_end() { 1024 if (_pending_jvms != NULL) { 1025 emit_non_safepoint(); 1026 } 1027 _pending_jvms = NULL; 1028 } 1029 }; 1030 1031 void NonSafepointEmitter::emit_non_safepoint() { 1032 JVMState* youngest_jvms = _pending_jvms; 1033 int pc_offset = _pending_offset; 1034 1035 // Clear it now: 1036 _pending_jvms = NULL; 1037 1038 DebugInformationRecorder* debug_info = C->debug_info(); 1039 assert(debug_info->recording_non_safepoints(), "sanity"); 1040 1041 debug_info->add_non_safepoint(pc_offset); 1042 int max_depth = youngest_jvms->depth(); 1043 1044 // Visit scopes from oldest to youngest. 1045 for (int depth = 1; depth <= max_depth; depth++) { 1046 JVMState* jvms = youngest_jvms->of_depth(depth); 1047 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 1048 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 1049 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute()); 1050 } 1051 1052 // Mark the end of the scope set. 1053 debug_info->end_non_safepoint(pc_offset); 1054 } 1055 1056 //------------------------------init_buffer------------------------------------ 1057 CodeBuffer* Compile::init_buffer(uint* blk_starts) { 1058 1059 // Set the initially allocated size 1060 int code_req = initial_code_capacity; 1061 int locs_req = initial_locs_capacity; 1062 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity; 1063 int const_req = initial_const_capacity; 1064 1065 int pad_req = NativeCall::instruction_size; 1066 // The extra spacing after the code is necessary on some platforms. 1067 // Sometimes we need to patch in a jump after the last instruction, 1068 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 1069 1070 // Compute the byte offset where we can store the deopt pc. 1071 if (fixed_slots() != 0) { 1072 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 1073 } 1074 1075 // Compute prolog code size 1076 _method_size = 0; 1077 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; 1078 #if defined(IA64) && !defined(AIX) 1079 if (save_argument_registers()) { 1080 // 4815101: this is a stub with implicit and unknown precision fp args. 1081 // The usual spill mechanism can only generate stfd's in this case, which 1082 // doesn't work if the fp reg to spill contains a single-precision denorm. 1083 // Instead, we hack around the normal spill mechanism using stfspill's and 1084 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate 1085 // space here for the fp arg regs (f8-f15) we're going to thusly spill. 1086 // 1087 // If we ever implement 16-byte 'registers' == stack slots, we can 1088 // get rid of this hack and have SpillCopy generate stfspill/ldffill 1089 // instead of stfd/stfs/ldfd/ldfs. 1090 _frame_slots += 8*(16/BytesPerInt); 1091 } 1092 #endif 1093 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check"); 1094 1095 if (has_mach_constant_base_node()) { 1096 uint add_size = 0; 1097 // Fill the constant table. 1098 // Note: This must happen before shorten_branches. 1099 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 1100 Block* b = _cfg->get_block(i); 1101 1102 for (uint j = 0; j < b->number_of_nodes(); j++) { 1103 Node* n = b->get_node(j); 1104 1105 // If the node is a MachConstantNode evaluate the constant 1106 // value section. 1107 if (n->is_MachConstant()) { 1108 MachConstantNode* machcon = n->as_MachConstant(); 1109 machcon->eval_constant(C); 1110 } else if (n->is_Mach()) { 1111 // On Power there are more nodes that issue constants. 1112 add_size += (n->as_Mach()->ins_num_consts() * 8); 1113 } 1114 } 1115 } 1116 1117 // Calculate the offsets of the constants and the size of the 1118 // constant table (including the padding to the next section). 1119 constant_table().calculate_offsets_and_size(); 1120 const_req = constant_table().size() + add_size; 1121 } 1122 1123 // Initialize the space for the BufferBlob used to find and verify 1124 // instruction size in MachNode::emit_size() 1125 init_scratch_buffer_blob(const_req); 1126 if (failing()) return NULL; // Out of memory 1127 1128 // Pre-compute the length of blocks and replace 1129 // long branches with short if machine supports it. 1130 shorten_branches(blk_starts, code_req, locs_req, stub_req); 1131 1132 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1133 int exception_handler_req = size_exception_handler(); 1134 int deopt_handler_req = size_deopt_handler(); 1135 exception_handler_req += MAX_stubs_size; // add marginal slop for handler 1136 deopt_handler_req += MAX_stubs_size; // add marginal slop for handler 1137 stub_req += MAX_stubs_size; // ensure per-stub margin 1138 code_req += MAX_inst_size; // ensure per-instruction margin 1139 1140 if (StressCodeBuffers) 1141 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1142 1143 int total_req = 1144 const_req + 1145 code_req + 1146 pad_req + 1147 stub_req + 1148 exception_handler_req + 1149 deopt_handler_req; // deopt handler 1150 1151 if (has_method_handle_invokes()) 1152 total_req += deopt_handler_req; // deopt MH handler 1153 1154 CodeBuffer* cb = code_buffer(); 1155 cb->initialize(total_req, locs_req); 1156 1157 // Have we run out of code space? 1158 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1159 C->record_failure("CodeCache is full"); 1160 return NULL; 1161 } 1162 // Configure the code buffer. 1163 cb->initialize_consts_size(const_req); 1164 cb->initialize_stubs_size(stub_req); 1165 cb->initialize_oop_recorder(env()->oop_recorder()); 1166 1167 // fill in the nop array for bundling computations 1168 MachNode *_nop_list[Bundle::_nop_count]; 1169 Bundle::initialize_nops(_nop_list, this); 1170 1171 return cb; 1172 } 1173 1174 //------------------------------fill_buffer------------------------------------ 1175 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) { 1176 // blk_starts[] contains offsets calculated during short branches processing, 1177 // offsets should not be increased during following steps. 1178 1179 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 1180 // of a loop. It is used to determine the padding for loop alignment. 1181 compute_loop_first_inst_sizes(); 1182 1183 // Create oopmap set. 1184 _oop_map_set = new OopMapSet(); 1185 1186 // !!!!! This preserves old handling of oopmaps for now 1187 debug_info()->set_oopmaps(_oop_map_set); 1188 1189 uint nblocks = _cfg->number_of_blocks(); 1190 // Count and start of implicit null check instructions 1191 uint inct_cnt = 0; 1192 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1193 1194 // Count and start of calls 1195 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1196 1197 uint return_offset = 0; 1198 int nop_size = (new (this) MachNopNode())->size(_regalloc); 1199 1200 int previous_offset = 0; 1201 int current_offset = 0; 1202 int last_call_offset = -1; 1203 int last_avoid_back_to_back_offset = -1; 1204 #ifdef ASSERT 1205 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); 1206 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 1207 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 1208 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); 1209 #endif 1210 1211 // Create an array of unused labels, one for each basic block, if printing is enabled 1212 #ifndef PRODUCT 1213 int *node_offsets = NULL; 1214 uint node_offset_limit = unique(); 1215 1216 if (print_assembly()) 1217 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1218 #endif 1219 1220 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily 1221 1222 // Emit the constant table. 1223 if (has_mach_constant_base_node()) { 1224 constant_table().emit(*cb); 1225 } 1226 1227 // Create an array of labels, one for each basic block 1228 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1); 1229 for (uint i=0; i <= nblocks; i++) { 1230 blk_labels[i].init(); 1231 } 1232 1233 // ------------------ 1234 // Now fill in the code buffer 1235 Node *delay_slot = NULL; 1236 1237 for (uint i = 0; i < nblocks; i++) { 1238 Block* block = _cfg->get_block(i); 1239 Node* head = block->head(); 1240 1241 // If this block needs to start aligned (i.e, can be reached other 1242 // than by falling-thru from the previous block), then force the 1243 // start of a new bundle. 1244 if (Pipeline::requires_bundling() && starts_bundle(head)) { 1245 cb->flush_bundle(true); 1246 } 1247 1248 #ifdef ASSERT 1249 if (!block->is_connector()) { 1250 stringStream st; 1251 block->dump_head(_cfg, &st); 1252 MacroAssembler(cb).block_comment(st.as_string()); 1253 } 1254 jmp_target[i] = 0; 1255 jmp_offset[i] = 0; 1256 jmp_size[i] = 0; 1257 jmp_rule[i] = 0; 1258 #endif 1259 int blk_offset = current_offset; 1260 1261 // Define the label at the beginning of the basic block 1262 MacroAssembler(cb).bind(blk_labels[block->_pre_order]); 1263 1264 uint last_inst = block->number_of_nodes(); 1265 1266 // Emit block normally, except for last instruction. 1267 // Emit means "dump code bits into code buffer". 1268 for (uint j = 0; j<last_inst; j++) { 1269 1270 // Get the node 1271 Node* n = block->get_node(j); 1272 1273 // See if delay slots are supported 1274 if (valid_bundle_info(n) && 1275 node_bundling(n)->used_in_unconditional_delay()) { 1276 assert(delay_slot == NULL, "no use of delay slot node"); 1277 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1278 1279 delay_slot = n; 1280 continue; 1281 } 1282 1283 // If this starts a new instruction group, then flush the current one 1284 // (but allow split bundles) 1285 if (Pipeline::requires_bundling() && starts_bundle(n)) 1286 cb->flush_bundle(false); 1287 1288 // The following logic is duplicated in the code ifdeffed for 1289 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It 1290 // should be factored out. Or maybe dispersed to the nodes? 1291 1292 // Special handling for SafePoint/Call Nodes 1293 bool is_mcall = false; 1294 if (n->is_Mach()) { 1295 MachNode *mach = n->as_Mach(); 1296 is_mcall = n->is_MachCall(); 1297 bool is_sfn = n->is_MachSafePoint(); 1298 1299 // If this requires all previous instructions be flushed, then do so 1300 if (is_sfn || is_mcall || mach->alignment_required() != 1) { 1301 cb->flush_bundle(true); 1302 current_offset = cb->insts_size(); 1303 } 1304 1305 // A padding may be needed again since a previous instruction 1306 // could be moved to delay slot. 1307 1308 // align the instruction if necessary 1309 int padding = mach->compute_padding(current_offset); 1310 // Make sure safepoint node for polling is distinct from a call's 1311 // return by adding a nop if needed. 1312 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) { 1313 padding = nop_size; 1314 } 1315 if (padding == 0 && mach->avoid_back_to_back() && 1316 current_offset == last_avoid_back_to_back_offset) { 1317 // Avoid back to back some instructions. 1318 padding = nop_size; 1319 } 1320 1321 if(padding > 0) { 1322 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1323 int nops_cnt = padding / nop_size; 1324 MachNode *nop = new (this) MachNopNode(nops_cnt); 1325 block->insert_node(nop, j++); 1326 last_inst++; 1327 _cfg->map_node_to_block(nop, block); 1328 nop->emit(*cb, _regalloc); 1329 cb->flush_bundle(true); 1330 current_offset = cb->insts_size(); 1331 } 1332 1333 // Remember the start of the last call in a basic block 1334 if (is_mcall) { 1335 MachCallNode *mcall = mach->as_MachCall(); 1336 1337 // This destination address is NOT PC-relative 1338 mcall->method_set((intptr_t)mcall->entry_point()); 1339 1340 // Save the return address 1341 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset(); 1342 1343 if (mcall->is_MachCallLeaf()) { 1344 is_mcall = false; 1345 is_sfn = false; 1346 } 1347 } 1348 1349 // sfn will be valid whenever mcall is valid now because of inheritance 1350 if (is_sfn || is_mcall) { 1351 1352 // Handle special safepoint nodes for synchronization 1353 if (!is_mcall) { 1354 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1355 // !!!!! Stubs only need an oopmap right now, so bail out 1356 if (sfn->jvms()->method() == NULL) { 1357 // Write the oopmap directly to the code blob??!! 1358 # ifdef ENABLE_ZAP_DEAD_LOCALS 1359 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive"); 1360 # endif 1361 continue; 1362 } 1363 } // End synchronization 1364 1365 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1366 current_offset); 1367 Process_OopMap_Node(mach, current_offset); 1368 } // End if safepoint 1369 1370 // If this is a null check, then add the start of the previous instruction to the list 1371 else if( mach->is_MachNullCheck() ) { 1372 inct_starts[inct_cnt++] = previous_offset; 1373 } 1374 1375 // If this is a branch, then fill in the label with the target BB's label 1376 else if (mach->is_MachBranch()) { 1377 // This requires the TRUE branch target be in succs[0] 1378 uint block_num = block->non_connector_successor(0)->_pre_order; 1379 1380 // Try to replace long branch if delay slot is not used, 1381 // it is mostly for back branches since forward branch's 1382 // distance is not updated yet. 1383 bool delay_slot_is_used = valid_bundle_info(n) && 1384 node_bundling(n)->use_unconditional_delay(); 1385 if (!delay_slot_is_used && mach->may_be_short_branch()) { 1386 assert(delay_slot == NULL, "not expecting delay slot node"); 1387 int br_size = n->size(_regalloc); 1388 int offset = blk_starts[block_num] - current_offset; 1389 if (block_num >= i) { 1390 // Current and following block's offset are not 1391 // finalized yet, adjust distance by the difference 1392 // between calculated and final offsets of current block. 1393 offset -= (blk_starts[i] - blk_offset); 1394 } 1395 // In the following code a nop could be inserted before 1396 // the branch which will increase the backward distance. 1397 bool needs_padding = (current_offset == last_avoid_back_to_back_offset); 1398 if (needs_padding && offset <= 0) 1399 offset -= nop_size; 1400 1401 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 1402 // We've got a winner. Replace this branch. 1403 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this); 1404 1405 // Update the jmp_size. 1406 int new_size = replacement->size(_regalloc); 1407 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller"); 1408 // Insert padding between avoid_back_to_back branches. 1409 if (needs_padding && replacement->avoid_back_to_back()) { 1410 MachNode *nop = new (this) MachNopNode(); 1411 block->insert_node(nop, j++); 1412 _cfg->map_node_to_block(nop, block); 1413 last_inst++; 1414 nop->emit(*cb, _regalloc); 1415 cb->flush_bundle(true); 1416 current_offset = cb->insts_size(); 1417 } 1418 #ifdef ASSERT 1419 jmp_target[i] = block_num; 1420 jmp_offset[i] = current_offset - blk_offset; 1421 jmp_size[i] = new_size; 1422 jmp_rule[i] = mach->rule(); 1423 #endif 1424 block->map_node(replacement, j); 1425 mach->subsume_by(replacement, C); 1426 n = replacement; 1427 mach = replacement; 1428 } 1429 } 1430 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num ); 1431 } else if (mach->ideal_Opcode() == Op_Jump) { 1432 for (uint h = 0; h < block->_num_succs; h++) { 1433 Block* succs_block = block->_succs[h]; 1434 for (uint j = 1; j < succs_block->num_preds(); j++) { 1435 Node* jpn = succs_block->pred(j); 1436 if (jpn->is_JumpProj() && jpn->in(0) == mach) { 1437 uint block_num = succs_block->non_connector()->_pre_order; 1438 Label *blkLabel = &blk_labels[block_num]; 1439 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1440 } 1441 } 1442 } 1443 } 1444 #ifdef ASSERT 1445 // Check that oop-store precedes the card-mark 1446 else if (mach->ideal_Opcode() == Op_StoreCM) { 1447 uint storeCM_idx = j; 1448 int count = 0; 1449 for (uint prec = mach->req(); prec < mach->len(); prec++) { 1450 Node *oop_store = mach->in(prec); // Precedence edge 1451 if (oop_store == NULL) continue; 1452 count++; 1453 uint i4; 1454 for (i4 = 0; i4 < last_inst; ++i4) { 1455 if (block->get_node(i4) == oop_store) { 1456 break; 1457 } 1458 } 1459 // Note: This test can provide a false failure if other precedence 1460 // edges have been added to the storeCMNode. 1461 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1462 } 1463 assert(count > 0, "storeCM expects at least one precedence edge"); 1464 } 1465 #endif 1466 else if (!n->is_Proj()) { 1467 // Remember the beginning of the previous instruction, in case 1468 // it's followed by a flag-kill and a null-check. Happens on 1469 // Intel all the time, with add-to-memory kind of opcodes. 1470 previous_offset = current_offset; 1471 } 1472 1473 // Not an else-if! 1474 // If this is a trap based cmp then add its offset to the list. 1475 if (mach->is_TrapBasedCheckNode()) { 1476 inct_starts[inct_cnt++] = current_offset; 1477 } 1478 } 1479 1480 // Verify that there is sufficient space remaining 1481 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1482 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1483 C->record_failure("CodeCache is full"); 1484 return; 1485 } 1486 1487 // Save the offset for the listing 1488 #ifndef PRODUCT 1489 if (node_offsets && n->_idx < node_offset_limit) 1490 node_offsets[n->_idx] = cb->insts_size(); 1491 #endif 1492 1493 // "Normal" instruction case 1494 DEBUG_ONLY( uint instr_offset = cb->insts_size(); ) 1495 n->emit(*cb, _regalloc); 1496 current_offset = cb->insts_size(); 1497 1498 #ifdef ASSERT 1499 if (n->size(_regalloc) < (current_offset-instr_offset)) { 1500 n->dump(); 1501 assert(false, "wrong size of mach node"); 1502 } 1503 #endif 1504 non_safepoints.observe_instruction(n, current_offset); 1505 1506 // mcall is last "call" that can be a safepoint 1507 // record it so we can see if a poll will directly follow it 1508 // in which case we'll need a pad to make the PcDesc sites unique 1509 // see 5010568. This can be slightly inaccurate but conservative 1510 // in the case that return address is not actually at current_offset. 1511 // This is a small price to pay. 1512 1513 if (is_mcall) { 1514 last_call_offset = current_offset; 1515 } 1516 1517 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back()) { 1518 // Avoid back to back some instructions. 1519 last_avoid_back_to_back_offset = current_offset; 1520 } 1521 1522 // See if this instruction has a delay slot 1523 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1524 assert(delay_slot != NULL, "expecting delay slot node"); 1525 1526 // Back up 1 instruction 1527 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size()); 1528 1529 // Save the offset for the listing 1530 #ifndef PRODUCT 1531 if (node_offsets && delay_slot->_idx < node_offset_limit) 1532 node_offsets[delay_slot->_idx] = cb->insts_size(); 1533 #endif 1534 1535 // Support a SafePoint in the delay slot 1536 if (delay_slot->is_MachSafePoint()) { 1537 MachNode *mach = delay_slot->as_Mach(); 1538 // !!!!! Stubs only need an oopmap right now, so bail out 1539 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) { 1540 // Write the oopmap directly to the code blob??!! 1541 # ifdef ENABLE_ZAP_DEAD_LOCALS 1542 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive"); 1543 # endif 1544 delay_slot = NULL; 1545 continue; 1546 } 1547 1548 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1549 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1550 adjusted_offset); 1551 // Generate an OopMap entry 1552 Process_OopMap_Node(mach, adjusted_offset); 1553 } 1554 1555 // Insert the delay slot instruction 1556 delay_slot->emit(*cb, _regalloc); 1557 1558 // Don't reuse it 1559 delay_slot = NULL; 1560 } 1561 1562 } // End for all instructions in block 1563 1564 // If the next block is the top of a loop, pad this block out to align 1565 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1566 if (i < nblocks-1) { 1567 Block *nb = _cfg->get_block(i + 1); 1568 int padding = nb->alignment_padding(current_offset); 1569 if( padding > 0 ) { 1570 MachNode *nop = new (this) MachNopNode(padding / nop_size); 1571 block->insert_node(nop, block->number_of_nodes()); 1572 _cfg->map_node_to_block(nop, block); 1573 nop->emit(*cb, _regalloc); 1574 current_offset = cb->insts_size(); 1575 } 1576 } 1577 // Verify that the distance for generated before forward 1578 // short branches is still valid. 1579 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size"); 1580 1581 // Save new block start offset 1582 blk_starts[i] = blk_offset; 1583 } // End of for all blocks 1584 blk_starts[nblocks] = current_offset; 1585 1586 non_safepoints.flush_at_end(); 1587 1588 // Offset too large? 1589 if (failing()) return; 1590 1591 // Define a pseudo-label at the end of the code 1592 MacroAssembler(cb).bind( blk_labels[nblocks] ); 1593 1594 // Compute the size of the first block 1595 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1596 1597 assert(cb->insts_size() < 500000, "method is unreasonably large"); 1598 1599 #ifdef ASSERT 1600 for (uint i = 0; i < nblocks; i++) { // For all blocks 1601 if (jmp_target[i] != 0) { 1602 int br_size = jmp_size[i]; 1603 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 1604 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 1605 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]); 1606 assert(false, "Displacement too large for short jmp"); 1607 } 1608 } 1609 } 1610 #endif 1611 1612 #ifndef PRODUCT 1613 // Information on the size of the method, without the extraneous code 1614 Scheduling::increment_method_size(cb->insts_size()); 1615 #endif 1616 1617 // ------------------ 1618 // Fill in exception table entries. 1619 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1620 1621 // Only java methods have exception handlers and deopt handlers 1622 if (_method) { 1623 // Emit the exception handler code. 1624 _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb)); 1625 // Emit the deopt handler code. 1626 _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb)); 1627 1628 // Emit the MethodHandle deopt handler code (if required). 1629 if (has_method_handle_invokes()) { 1630 // We can use the same code as for the normal deopt handler, we 1631 // just need a different entry point address. 1632 _code_offsets.set_value(CodeOffsets::DeoptMH, emit_deopt_handler(*cb)); 1633 } 1634 } 1635 1636 // One last check for failed CodeBuffer::expand: 1637 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1638 C->record_failure("CodeCache is full"); 1639 return; 1640 } 1641 1642 #ifndef PRODUCT 1643 // Dump the assembly code, including basic-block numbers 1644 if (print_assembly()) { 1645 ttyLocker ttyl; // keep the following output all in one block 1646 if (!VMThread::should_terminate()) { // test this under the tty lock 1647 // This output goes directly to the tty, not the compiler log. 1648 // To enable tools to match it up with the compilation activity, 1649 // be sure to tag this tty output with the compile ID. 1650 if (xtty != NULL) { 1651 xtty->head("opto_assembly compile_id='%d'%s", compile_id(), 1652 is_osr_compilation() ? " compile_kind='osr'" : 1653 ""); 1654 } 1655 if (method() != NULL) { 1656 method()->print_metadata(); 1657 } 1658 dump_asm(node_offsets, node_offset_limit); 1659 if (xtty != NULL) { 1660 xtty->tail("opto_assembly"); 1661 } 1662 } 1663 } 1664 #endif 1665 1666 } 1667 1668 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1669 _inc_table.set_size(cnt); 1670 1671 uint inct_cnt = 0; 1672 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 1673 Block* block = _cfg->get_block(i); 1674 Node *n = NULL; 1675 int j; 1676 1677 // Find the branch; ignore trailing NOPs. 1678 for (j = block->number_of_nodes() - 1; j >= 0; j--) { 1679 n = block->get_node(j); 1680 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) { 1681 break; 1682 } 1683 } 1684 1685 // If we didn't find anything, continue 1686 if (j < 0) { 1687 continue; 1688 } 1689 1690 // Compute ExceptionHandlerTable subtable entry and add it 1691 // (skip empty blocks) 1692 if (n->is_Catch()) { 1693 1694 // Get the offset of the return from the call 1695 uint call_return = call_returns[block->_pre_order]; 1696 #ifdef ASSERT 1697 assert( call_return > 0, "no call seen for this basic block" ); 1698 while (block->get_node(--j)->is_MachProj()) ; 1699 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call"); 1700 #endif 1701 // last instruction is a CatchNode, find it's CatchProjNodes 1702 int nof_succs = block->_num_succs; 1703 // allocate space 1704 GrowableArray<intptr_t> handler_bcis(nof_succs); 1705 GrowableArray<intptr_t> handler_pcos(nof_succs); 1706 // iterate through all successors 1707 for (int j = 0; j < nof_succs; j++) { 1708 Block* s = block->_succs[j]; 1709 bool found_p = false; 1710 for (uint k = 1; k < s->num_preds(); k++) { 1711 Node* pk = s->pred(k); 1712 if (pk->is_CatchProj() && pk->in(0) == n) { 1713 const CatchProjNode* p = pk->as_CatchProj(); 1714 found_p = true; 1715 // add the corresponding handler bci & pco information 1716 if (p->_con != CatchProjNode::fall_through_index) { 1717 // p leads to an exception handler (and is not fall through) 1718 assert(s == _cfg->get_block(s->_pre_order), "bad numbering"); 1719 // no duplicates, please 1720 if (!handler_bcis.contains(p->handler_bci())) { 1721 uint block_num = s->non_connector()->_pre_order; 1722 handler_bcis.append(p->handler_bci()); 1723 handler_pcos.append(blk_labels[block_num].loc_pos()); 1724 } 1725 } 1726 } 1727 } 1728 assert(found_p, "no matching predecessor found"); 1729 // Note: Due to empty block removal, one block may have 1730 // several CatchProj inputs, from the same Catch. 1731 } 1732 1733 // Set the offset of the return from the call 1734 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); 1735 continue; 1736 } 1737 1738 // Handle implicit null exception table updates 1739 if (n->is_MachNullCheck()) { 1740 uint block_num = block->non_connector_successor(0)->_pre_order; 1741 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1742 continue; 1743 } 1744 // Handle implicit exception table updates: trap instructions. 1745 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) { 1746 uint block_num = block->non_connector_successor(0)->_pre_order; 1747 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1748 continue; 1749 } 1750 } // End of for all blocks fill in exception table entries 1751 } 1752 1753 // Static Variables 1754 #ifndef PRODUCT 1755 uint Scheduling::_total_nop_size = 0; 1756 uint Scheduling::_total_method_size = 0; 1757 uint Scheduling::_total_branches = 0; 1758 uint Scheduling::_total_unconditional_delays = 0; 1759 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1760 #endif 1761 1762 // Initializer for class Scheduling 1763 1764 Scheduling::Scheduling(Arena *arena, Compile &compile) 1765 : _arena(arena), 1766 _cfg(compile.cfg()), 1767 _regalloc(compile.regalloc()), 1768 _reg_node(arena), 1769 _bundle_instr_count(0), 1770 _bundle_cycle_number(0), 1771 _scheduled(arena), 1772 _available(arena), 1773 _next_node(NULL), 1774 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]), 1775 _pinch_free_list(arena) 1776 #ifndef PRODUCT 1777 , _branches(0) 1778 , _unconditional_delays(0) 1779 #endif 1780 { 1781 // Create a MachNopNode 1782 _nop = new (&compile) MachNopNode(); 1783 1784 // Now that the nops are in the array, save the count 1785 // (but allow entries for the nops) 1786 _node_bundling_limit = compile.unique(); 1787 uint node_max = _regalloc->node_regs_max_index(); 1788 1789 compile.set_node_bundling_limit(_node_bundling_limit); 1790 1791 // This one is persistent within the Compile class 1792 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1793 1794 // Allocate space for fixed-size arrays 1795 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1796 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1797 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1798 1799 // Clear the arrays 1800 memset(_node_bundling_base, 0, node_max * sizeof(Bundle)); 1801 memset(_node_latency, 0, node_max * sizeof(unsigned short)); 1802 memset(_uses, 0, node_max * sizeof(short)); 1803 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 1804 1805 // Clear the bundling information 1806 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements)); 1807 1808 // Get the last node 1809 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1); 1810 1811 _next_node = block->get_node(block->number_of_nodes() - 1); 1812 } 1813 1814 #ifndef PRODUCT 1815 // Scheduling destructor 1816 Scheduling::~Scheduling() { 1817 _total_branches += _branches; 1818 _total_unconditional_delays += _unconditional_delays; 1819 } 1820 #endif 1821 1822 // Step ahead "i" cycles 1823 void Scheduling::step(uint i) { 1824 1825 Bundle *bundle = node_bundling(_next_node); 1826 bundle->set_starts_bundle(); 1827 1828 // Update the bundle record, but leave the flags information alone 1829 if (_bundle_instr_count > 0) { 1830 bundle->set_instr_count(_bundle_instr_count); 1831 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1832 } 1833 1834 // Update the state information 1835 _bundle_instr_count = 0; 1836 _bundle_cycle_number += i; 1837 _bundle_use.step(i); 1838 } 1839 1840 void Scheduling::step_and_clear() { 1841 Bundle *bundle = node_bundling(_next_node); 1842 bundle->set_starts_bundle(); 1843 1844 // Update the bundle record 1845 if (_bundle_instr_count > 0) { 1846 bundle->set_instr_count(_bundle_instr_count); 1847 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1848 1849 _bundle_cycle_number += 1; 1850 } 1851 1852 // Clear the bundling information 1853 _bundle_instr_count = 0; 1854 _bundle_use.reset(); 1855 1856 memcpy(_bundle_use_elements, 1857 Pipeline_Use::elaborated_elements, 1858 sizeof(Pipeline_Use::elaborated_elements)); 1859 } 1860 1861 // Perform instruction scheduling and bundling over the sequence of 1862 // instructions in backwards order. 1863 void Compile::ScheduleAndBundle() { 1864 1865 // Don't optimize this if it isn't a method 1866 if (!_method) 1867 return; 1868 1869 // Don't optimize this if scheduling is disabled 1870 if (!do_scheduling()) 1871 return; 1872 1873 // Scheduling code works only with pairs (8 bytes) maximum. 1874 if (max_vector_size() > 8) 1875 return; 1876 1877 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); ) 1878 1879 // Create a data structure for all the scheduling information 1880 Scheduling scheduling(Thread::current()->resource_area(), *this); 1881 1882 // Walk backwards over each basic block, computing the needed alignment 1883 // Walk over all the basic blocks 1884 scheduling.DoScheduling(); 1885 } 1886 1887 // Compute the latency of all the instructions. This is fairly simple, 1888 // because we already have a legal ordering. Walk over the instructions 1889 // from first to last, and compute the latency of the instruction based 1890 // on the latency of the preceding instruction(s). 1891 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { 1892 #ifndef PRODUCT 1893 if (_cfg->C->trace_opto_output()) 1894 tty->print("# -> ComputeLocalLatenciesForward\n"); 1895 #endif 1896 1897 // Walk over all the schedulable instructions 1898 for( uint j=_bb_start; j < _bb_end; j++ ) { 1899 1900 // This is a kludge, forcing all latency calculations to start at 1. 1901 // Used to allow latency 0 to force an instruction to the beginning 1902 // of the bb 1903 uint latency = 1; 1904 Node *use = bb->get_node(j); 1905 uint nlen = use->len(); 1906 1907 // Walk over all the inputs 1908 for ( uint k=0; k < nlen; k++ ) { 1909 Node *def = use->in(k); 1910 if (!def) 1911 continue; 1912 1913 uint l = _node_latency[def->_idx] + use->latency(k); 1914 if (latency < l) 1915 latency = l; 1916 } 1917 1918 _node_latency[use->_idx] = latency; 1919 1920 #ifndef PRODUCT 1921 if (_cfg->C->trace_opto_output()) { 1922 tty->print("# latency %4d: ", latency); 1923 use->dump(); 1924 } 1925 #endif 1926 } 1927 1928 #ifndef PRODUCT 1929 if (_cfg->C->trace_opto_output()) 1930 tty->print("# <- ComputeLocalLatenciesForward\n"); 1931 #endif 1932 1933 } // end ComputeLocalLatenciesForward 1934 1935 // See if this node fits into the present instruction bundle 1936 bool Scheduling::NodeFitsInBundle(Node *n) { 1937 uint n_idx = n->_idx; 1938 1939 // If this is the unconditional delay instruction, then it fits 1940 if (n == _unconditional_delay_slot) { 1941 #ifndef PRODUCT 1942 if (_cfg->C->trace_opto_output()) 1943 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 1944 #endif 1945 return (true); 1946 } 1947 1948 // If the node cannot be scheduled this cycle, skip it 1949 if (_current_latency[n_idx] > _bundle_cycle_number) { 1950 #ifndef PRODUCT 1951 if (_cfg->C->trace_opto_output()) 1952 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 1953 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 1954 #endif 1955 return (false); 1956 } 1957 1958 const Pipeline *node_pipeline = n->pipeline(); 1959 1960 uint instruction_count = node_pipeline->instructionCount(); 1961 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1962 instruction_count = 0; 1963 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1964 instruction_count++; 1965 1966 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 1967 #ifndef PRODUCT 1968 if (_cfg->C->trace_opto_output()) 1969 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 1970 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 1971 #endif 1972 return (false); 1973 } 1974 1975 // Don't allow non-machine nodes to be handled this way 1976 if (!n->is_Mach() && instruction_count == 0) 1977 return (false); 1978 1979 // See if there is any overlap 1980 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 1981 1982 if (delay > 0) { 1983 #ifndef PRODUCT 1984 if (_cfg->C->trace_opto_output()) 1985 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 1986 #endif 1987 return false; 1988 } 1989 1990 #ifndef PRODUCT 1991 if (_cfg->C->trace_opto_output()) 1992 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 1993 #endif 1994 1995 return true; 1996 } 1997 1998 Node * Scheduling::ChooseNodeToBundle() { 1999 uint siz = _available.size(); 2000 2001 if (siz == 0) { 2002 2003 #ifndef PRODUCT 2004 if (_cfg->C->trace_opto_output()) 2005 tty->print("# ChooseNodeToBundle: NULL\n"); 2006 #endif 2007 return (NULL); 2008 } 2009 2010 // Fast path, if only 1 instruction in the bundle 2011 if (siz == 1) { 2012 #ifndef PRODUCT 2013 if (_cfg->C->trace_opto_output()) { 2014 tty->print("# ChooseNodeToBundle (only 1): "); 2015 _available[0]->dump(); 2016 } 2017 #endif 2018 return (_available[0]); 2019 } 2020 2021 // Don't bother, if the bundle is already full 2022 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 2023 for ( uint i = 0; i < siz; i++ ) { 2024 Node *n = _available[i]; 2025 2026 // Skip projections, we'll handle them another way 2027 if (n->is_Proj()) 2028 continue; 2029 2030 // This presupposed that instructions are inserted into the 2031 // available list in a legality order; i.e. instructions that 2032 // must be inserted first are at the head of the list 2033 if (NodeFitsInBundle(n)) { 2034 #ifndef PRODUCT 2035 if (_cfg->C->trace_opto_output()) { 2036 tty->print("# ChooseNodeToBundle: "); 2037 n->dump(); 2038 } 2039 #endif 2040 return (n); 2041 } 2042 } 2043 } 2044 2045 // Nothing fits in this bundle, choose the highest priority 2046 #ifndef PRODUCT 2047 if (_cfg->C->trace_opto_output()) { 2048 tty->print("# ChooseNodeToBundle: "); 2049 _available[0]->dump(); 2050 } 2051 #endif 2052 2053 return _available[0]; 2054 } 2055 2056 void Scheduling::AddNodeToAvailableList(Node *n) { 2057 assert( !n->is_Proj(), "projections never directly made available" ); 2058 #ifndef PRODUCT 2059 if (_cfg->C->trace_opto_output()) { 2060 tty->print("# AddNodeToAvailableList: "); 2061 n->dump(); 2062 } 2063 #endif 2064 2065 int latency = _current_latency[n->_idx]; 2066 2067 // Insert in latency order (insertion sort) 2068 uint i; 2069 for ( i=0; i < _available.size(); i++ ) 2070 if (_current_latency[_available[i]->_idx] > latency) 2071 break; 2072 2073 // Special Check for compares following branches 2074 if( n->is_Mach() && _scheduled.size() > 0 ) { 2075 int op = n->as_Mach()->ideal_Opcode(); 2076 Node *last = _scheduled[0]; 2077 if( last->is_MachIf() && last->in(1) == n && 2078 ( op == Op_CmpI || 2079 op == Op_CmpU || 2080 op == Op_CmpP || 2081 op == Op_CmpF || 2082 op == Op_CmpD || 2083 op == Op_CmpL ) ) { 2084 2085 // Recalculate position, moving to front of same latency 2086 for ( i=0 ; i < _available.size(); i++ ) 2087 if (_current_latency[_available[i]->_idx] >= latency) 2088 break; 2089 } 2090 } 2091 2092 // Insert the node in the available list 2093 _available.insert(i, n); 2094 2095 #ifndef PRODUCT 2096 if (_cfg->C->trace_opto_output()) 2097 dump_available(); 2098 #endif 2099 } 2100 2101 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 2102 for ( uint i=0; i < n->len(); i++ ) { 2103 Node *def = n->in(i); 2104 if (!def) continue; 2105 if( def->is_Proj() ) // If this is a machine projection, then 2106 def = def->in(0); // propagate usage thru to the base instruction 2107 2108 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local 2109 continue; 2110 } 2111 2112 // Compute the latency 2113 uint l = _bundle_cycle_number + n->latency(i); 2114 if (_current_latency[def->_idx] < l) 2115 _current_latency[def->_idx] = l; 2116 2117 // If this does not have uses then schedule it 2118 if ((--_uses[def->_idx]) == 0) 2119 AddNodeToAvailableList(def); 2120 } 2121 } 2122 2123 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 2124 #ifndef PRODUCT 2125 if (_cfg->C->trace_opto_output()) { 2126 tty->print("# AddNodeToBundle: "); 2127 n->dump(); 2128 } 2129 #endif 2130 2131 // Remove this from the available list 2132 uint i; 2133 for (i = 0; i < _available.size(); i++) 2134 if (_available[i] == n) 2135 break; 2136 assert(i < _available.size(), "entry in _available list not found"); 2137 _available.remove(i); 2138 2139 // See if this fits in the current bundle 2140 const Pipeline *node_pipeline = n->pipeline(); 2141 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 2142 2143 // Check for instructions to be placed in the delay slot. We 2144 // do this before we actually schedule the current instruction, 2145 // because the delay slot follows the current instruction. 2146 if (Pipeline::_branch_has_delay_slot && 2147 node_pipeline->hasBranchDelay() && 2148 !_unconditional_delay_slot) { 2149 2150 uint siz = _available.size(); 2151 2152 // Conditional branches can support an instruction that 2153 // is unconditionally executed and not dependent by the 2154 // branch, OR a conditionally executed instruction if 2155 // the branch is taken. In practice, this means that 2156 // the first instruction at the branch target is 2157 // copied to the delay slot, and the branch goes to 2158 // the instruction after that at the branch target 2159 if ( n->is_MachBranch() ) { 2160 2161 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 2162 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 2163 2164 #ifndef PRODUCT 2165 _branches++; 2166 #endif 2167 2168 // At least 1 instruction is on the available list 2169 // that is not dependent on the branch 2170 for (uint i = 0; i < siz; i++) { 2171 Node *d = _available[i]; 2172 const Pipeline *avail_pipeline = d->pipeline(); 2173 2174 // Don't allow safepoints in the branch shadow, that will 2175 // cause a number of difficulties 2176 if ( avail_pipeline->instructionCount() == 1 && 2177 !avail_pipeline->hasMultipleBundles() && 2178 !avail_pipeline->hasBranchDelay() && 2179 Pipeline::instr_has_unit_size() && 2180 d->size(_regalloc) == Pipeline::instr_unit_size() && 2181 NodeFitsInBundle(d) && 2182 !node_bundling(d)->used_in_delay()) { 2183 2184 if (d->is_Mach() && !d->is_MachSafePoint()) { 2185 // A node that fits in the delay slot was found, so we need to 2186 // set the appropriate bits in the bundle pipeline information so 2187 // that it correctly indicates resource usage. Later, when we 2188 // attempt to add this instruction to the bundle, we will skip 2189 // setting the resource usage. 2190 _unconditional_delay_slot = d; 2191 node_bundling(n)->set_use_unconditional_delay(); 2192 node_bundling(d)->set_used_in_unconditional_delay(); 2193 _bundle_use.add_usage(avail_pipeline->resourceUse()); 2194 _current_latency[d->_idx] = _bundle_cycle_number; 2195 _next_node = d; 2196 ++_bundle_instr_count; 2197 #ifndef PRODUCT 2198 _unconditional_delays++; 2199 #endif 2200 break; 2201 } 2202 } 2203 } 2204 } 2205 2206 // No delay slot, add a nop to the usage 2207 if (!_unconditional_delay_slot) { 2208 // See if adding an instruction in the delay slot will overflow 2209 // the bundle. 2210 if (!NodeFitsInBundle(_nop)) { 2211 #ifndef PRODUCT 2212 if (_cfg->C->trace_opto_output()) 2213 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2214 #endif 2215 step(1); 2216 } 2217 2218 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2219 _next_node = _nop; 2220 ++_bundle_instr_count; 2221 } 2222 2223 // See if the instruction in the delay slot requires a 2224 // step of the bundles 2225 if (!NodeFitsInBundle(n)) { 2226 #ifndef PRODUCT 2227 if (_cfg->C->trace_opto_output()) 2228 tty->print("# *** STEP(branch won't fit) ***\n"); 2229 #endif 2230 // Update the state information 2231 _bundle_instr_count = 0; 2232 _bundle_cycle_number += 1; 2233 _bundle_use.step(1); 2234 } 2235 } 2236 2237 // Get the number of instructions 2238 uint instruction_count = node_pipeline->instructionCount(); 2239 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2240 instruction_count = 0; 2241 2242 // Compute the latency information 2243 uint delay = 0; 2244 2245 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2246 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2247 if (relative_latency < 0) 2248 relative_latency = 0; 2249 2250 delay = _bundle_use.full_latency(relative_latency, node_usage); 2251 2252 // Does not fit in this bundle, start a new one 2253 if (delay > 0) { 2254 step(delay); 2255 2256 #ifndef PRODUCT 2257 if (_cfg->C->trace_opto_output()) 2258 tty->print("# *** STEP(%d) ***\n", delay); 2259 #endif 2260 } 2261 } 2262 2263 // If this was placed in the delay slot, ignore it 2264 if (n != _unconditional_delay_slot) { 2265 2266 if (delay == 0) { 2267 if (node_pipeline->hasMultipleBundles()) { 2268 #ifndef PRODUCT 2269 if (_cfg->C->trace_opto_output()) 2270 tty->print("# *** STEP(multiple instructions) ***\n"); 2271 #endif 2272 step(1); 2273 } 2274 2275 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2276 #ifndef PRODUCT 2277 if (_cfg->C->trace_opto_output()) 2278 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2279 instruction_count + _bundle_instr_count, 2280 Pipeline::_max_instrs_per_cycle); 2281 #endif 2282 step(1); 2283 } 2284 } 2285 2286 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2287 _bundle_instr_count++; 2288 2289 // Set the node's latency 2290 _current_latency[n->_idx] = _bundle_cycle_number; 2291 2292 // Now merge the functional unit information 2293 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2294 _bundle_use.add_usage(node_usage); 2295 2296 // Increment the number of instructions in this bundle 2297 _bundle_instr_count += instruction_count; 2298 2299 // Remember this node for later 2300 if (n->is_Mach()) 2301 _next_node = n; 2302 } 2303 2304 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2305 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2306 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2307 // into the block. All other scheduled nodes get put in the schedule here. 2308 int op = n->Opcode(); 2309 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2310 (op != Op_Node && // Not an unused antidepedence node and 2311 // not an unallocated boxlock 2312 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2313 2314 // Push any trailing projections 2315 if( bb->get_node(bb->number_of_nodes()-1) != n ) { 2316 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2317 Node *foi = n->fast_out(i); 2318 if( foi->is_Proj() ) 2319 _scheduled.push(foi); 2320 } 2321 } 2322 2323 // Put the instruction in the schedule list 2324 _scheduled.push(n); 2325 } 2326 2327 #ifndef PRODUCT 2328 if (_cfg->C->trace_opto_output()) 2329 dump_available(); 2330 #endif 2331 2332 // Walk all the definitions, decrementing use counts, and 2333 // if a definition has a 0 use count, place it in the available list. 2334 DecrementUseCounts(n,bb); 2335 } 2336 2337 // This method sets the use count within a basic block. We will ignore all 2338 // uses outside the current basic block. As we are doing a backwards walk, 2339 // any node we reach that has a use count of 0 may be scheduled. This also 2340 // avoids the problem of cyclic references from phi nodes, as long as phi 2341 // nodes are at the front of the basic block. This method also initializes 2342 // the available list to the set of instructions that have no uses within this 2343 // basic block. 2344 void Scheduling::ComputeUseCount(const Block *bb) { 2345 #ifndef PRODUCT 2346 if (_cfg->C->trace_opto_output()) 2347 tty->print("# -> ComputeUseCount\n"); 2348 #endif 2349 2350 // Clear the list of available and scheduled instructions, just in case 2351 _available.clear(); 2352 _scheduled.clear(); 2353 2354 // No delay slot specified 2355 _unconditional_delay_slot = NULL; 2356 2357 #ifdef ASSERT 2358 for( uint i=0; i < bb->number_of_nodes(); i++ ) 2359 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" ); 2360 #endif 2361 2362 // Force the _uses count to never go to zero for unscheduable pieces 2363 // of the block 2364 for( uint k = 0; k < _bb_start; k++ ) 2365 _uses[bb->get_node(k)->_idx] = 1; 2366 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ ) 2367 _uses[bb->get_node(l)->_idx] = 1; 2368 2369 // Iterate backwards over the instructions in the block. Don't count the 2370 // branch projections at end or the block header instructions. 2371 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2372 Node *n = bb->get_node(j); 2373 if( n->is_Proj() ) continue; // Projections handled another way 2374 2375 // Account for all uses 2376 for ( uint k = 0; k < n->len(); k++ ) { 2377 Node *inp = n->in(k); 2378 if (!inp) continue; 2379 assert(inp != n, "no cycles allowed" ); 2380 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use? 2381 if (inp->is_Proj()) { // Skip through Proj's 2382 inp = inp->in(0); 2383 } 2384 ++_uses[inp->_idx]; // Count 1 block-local use 2385 } 2386 } 2387 2388 // If this instruction has a 0 use count, then it is available 2389 if (!_uses[n->_idx]) { 2390 _current_latency[n->_idx] = _bundle_cycle_number; 2391 AddNodeToAvailableList(n); 2392 } 2393 2394 #ifndef PRODUCT 2395 if (_cfg->C->trace_opto_output()) { 2396 tty->print("# uses: %3d: ", _uses[n->_idx]); 2397 n->dump(); 2398 } 2399 #endif 2400 } 2401 2402 #ifndef PRODUCT 2403 if (_cfg->C->trace_opto_output()) 2404 tty->print("# <- ComputeUseCount\n"); 2405 #endif 2406 } 2407 2408 // This routine performs scheduling on each basic block in reverse order, 2409 // using instruction latencies and taking into account function unit 2410 // availability. 2411 void Scheduling::DoScheduling() { 2412 #ifndef PRODUCT 2413 if (_cfg->C->trace_opto_output()) 2414 tty->print("# -> DoScheduling\n"); 2415 #endif 2416 2417 Block *succ_bb = NULL; 2418 Block *bb; 2419 2420 // Walk over all the basic blocks in reverse order 2421 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) { 2422 bb = _cfg->get_block(i); 2423 2424 #ifndef PRODUCT 2425 if (_cfg->C->trace_opto_output()) { 2426 tty->print("# Schedule BB#%03d (initial)\n", i); 2427 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2428 bb->get_node(j)->dump(); 2429 } 2430 } 2431 #endif 2432 2433 // On the head node, skip processing 2434 if (bb == _cfg->get_root_block()) { 2435 continue; 2436 } 2437 2438 // Skip empty, connector blocks 2439 if (bb->is_connector()) 2440 continue; 2441 2442 // If the following block is not the sole successor of 2443 // this one, then reset the pipeline information 2444 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2445 #ifndef PRODUCT 2446 if (_cfg->C->trace_opto_output()) { 2447 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2448 _next_node->_idx, _bundle_instr_count); 2449 } 2450 #endif 2451 step_and_clear(); 2452 } 2453 2454 // Leave untouched the starting instruction, any Phis, a CreateEx node 2455 // or Top. bb->get_node(_bb_start) is the first schedulable instruction. 2456 _bb_end = bb->number_of_nodes()-1; 2457 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2458 Node *n = bb->get_node(_bb_start); 2459 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2460 // Also, MachIdealNodes do not get scheduled 2461 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2462 MachNode *mach = n->as_Mach(); 2463 int iop = mach->ideal_Opcode(); 2464 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2465 if( iop == Op_Con ) continue; // Do not schedule Top 2466 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2467 mach->pipeline() == MachNode::pipeline_class() && 2468 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc 2469 continue; 2470 break; // Funny loop structure to be sure... 2471 } 2472 // Compute last "interesting" instruction in block - last instruction we 2473 // might schedule. _bb_end points just after last schedulable inst. We 2474 // normally schedule conditional branches (despite them being forced last 2475 // in the block), because they have delay slots we can fill. Calls all 2476 // have their delay slots filled in the template expansions, so we don't 2477 // bother scheduling them. 2478 Node *last = bb->get_node(_bb_end); 2479 // Ignore trailing NOPs. 2480 while (_bb_end > 0 && last->is_Mach() && 2481 last->as_Mach()->ideal_Opcode() == Op_Con) { 2482 last = bb->get_node(--_bb_end); 2483 } 2484 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, ""); 2485 if( last->is_Catch() || 2486 // Exclude unreachable path case when Halt node is in a separate block. 2487 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2488 // There must be a prior call. Skip it. 2489 while( !bb->get_node(--_bb_end)->is_MachCall() ) { 2490 assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" ); 2491 } 2492 } else if( last->is_MachNullCheck() ) { 2493 // Backup so the last null-checked memory instruction is 2494 // outside the schedulable range. Skip over the nullcheck, 2495 // projection, and the memory nodes. 2496 Node *mem = last->in(1); 2497 do { 2498 _bb_end--; 2499 } while (mem != bb->get_node(_bb_end)); 2500 } else { 2501 // Set _bb_end to point after last schedulable inst. 2502 _bb_end++; 2503 } 2504 2505 assert( _bb_start <= _bb_end, "inverted block ends" ); 2506 2507 // Compute the register antidependencies for the basic block 2508 ComputeRegisterAntidependencies(bb); 2509 if (_cfg->C->failing()) return; // too many D-U pinch points 2510 2511 // Compute intra-bb latencies for the nodes 2512 ComputeLocalLatenciesForward(bb); 2513 2514 // Compute the usage within the block, and set the list of all nodes 2515 // in the block that have no uses within the block. 2516 ComputeUseCount(bb); 2517 2518 // Schedule the remaining instructions in the block 2519 while ( _available.size() > 0 ) { 2520 Node *n = ChooseNodeToBundle(); 2521 guarantee(n != NULL, "no nodes available"); 2522 AddNodeToBundle(n,bb); 2523 } 2524 2525 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2526 #ifdef ASSERT 2527 for( uint l = _bb_start; l < _bb_end; l++ ) { 2528 Node *n = bb->get_node(l); 2529 uint m; 2530 for( m = 0; m < _bb_end-_bb_start; m++ ) 2531 if( _scheduled[m] == n ) 2532 break; 2533 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2534 } 2535 #endif 2536 2537 // Now copy the instructions (in reverse order) back to the block 2538 for ( uint k = _bb_start; k < _bb_end; k++ ) 2539 bb->map_node(_scheduled[_bb_end-k-1], k); 2540 2541 #ifndef PRODUCT 2542 if (_cfg->C->trace_opto_output()) { 2543 tty->print("# Schedule BB#%03d (final)\n", i); 2544 uint current = 0; 2545 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2546 Node *n = bb->get_node(j); 2547 if( valid_bundle_info(n) ) { 2548 Bundle *bundle = node_bundling(n); 2549 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2550 tty->print("*** Bundle: "); 2551 bundle->dump(); 2552 } 2553 n->dump(); 2554 } 2555 } 2556 } 2557 #endif 2558 #ifdef ASSERT 2559 verify_good_schedule(bb,"after block local scheduling"); 2560 #endif 2561 } 2562 2563 #ifndef PRODUCT 2564 if (_cfg->C->trace_opto_output()) 2565 tty->print("# <- DoScheduling\n"); 2566 #endif 2567 2568 // Record final node-bundling array location 2569 _regalloc->C->set_node_bundling_base(_node_bundling_base); 2570 2571 } // end DoScheduling 2572 2573 // Verify that no live-range used in the block is killed in the block by a 2574 // wrong DEF. This doesn't verify live-ranges that span blocks. 2575 2576 // Check for edge existence. Used to avoid adding redundant precedence edges. 2577 static bool edge_from_to( Node *from, Node *to ) { 2578 for( uint i=0; i<from->len(); i++ ) 2579 if( from->in(i) == to ) 2580 return true; 2581 return false; 2582 } 2583 2584 #ifdef ASSERT 2585 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2586 // Check for bad kills 2587 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2588 Node *prior_use = _reg_node[def]; 2589 if( prior_use && !edge_from_to(prior_use,n) ) { 2590 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2591 n->dump(); 2592 tty->print_cr("..."); 2593 prior_use->dump(); 2594 assert(edge_from_to(prior_use,n),msg); 2595 } 2596 _reg_node.map(def,NULL); // Kill live USEs 2597 } 2598 } 2599 2600 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2601 2602 // Zap to something reasonable for the verify code 2603 _reg_node.clear(); 2604 2605 // Walk over the block backwards. Check to make sure each DEF doesn't 2606 // kill a live value (other than the one it's supposed to). Add each 2607 // USE to the live set. 2608 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) { 2609 Node *n = b->get_node(i); 2610 int n_op = n->Opcode(); 2611 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2612 // Fat-proj kills a slew of registers 2613 RegMask rm = n->out_RegMask();// Make local copy 2614 while( rm.is_NotEmpty() ) { 2615 OptoReg::Name kill = rm.find_first_elem(); 2616 rm.Remove(kill); 2617 verify_do_def( n, kill, msg ); 2618 } 2619 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2620 // Get DEF'd registers the normal way 2621 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2622 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2623 } 2624 2625 // Now make all USEs live 2626 for( uint i=1; i<n->req(); i++ ) { 2627 Node *def = n->in(i); 2628 assert(def != 0, "input edge required"); 2629 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2630 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2631 if( OptoReg::is_valid(reg_lo) ) { 2632 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg); 2633 _reg_node.map(reg_lo,n); 2634 } 2635 if( OptoReg::is_valid(reg_hi) ) { 2636 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg); 2637 _reg_node.map(reg_hi,n); 2638 } 2639 } 2640 2641 } 2642 2643 // Zap to something reasonable for the Antidependence code 2644 _reg_node.clear(); 2645 } 2646 #endif 2647 2648 // Conditionally add precedence edges. Avoid putting edges on Projs. 2649 static void add_prec_edge_from_to( Node *from, Node *to ) { 2650 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2651 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2652 from = from->in(0); 2653 } 2654 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2655 !edge_from_to( from, to ) ) // Avoid duplicate edge 2656 from->add_prec(to); 2657 } 2658 2659 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2660 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2661 return; 2662 2663 Node *pinch = _reg_node[def_reg]; // Get pinch point 2664 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet? 2665 is_def ) { // Check for a true def (not a kill) 2666 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2667 return; 2668 } 2669 2670 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2671 debug_only( def = (Node*)0xdeadbeef; ) 2672 2673 // After some number of kills there _may_ be a later def 2674 Node *later_def = NULL; 2675 2676 // Finding a kill requires a real pinch-point. 2677 // Check for not already having a pinch-point. 2678 // Pinch points are Op_Node's. 2679 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2680 later_def = pinch; // Must be def/kill as optimistic pinch-point 2681 if ( _pinch_free_list.size() > 0) { 2682 pinch = _pinch_free_list.pop(); 2683 } else { 2684 pinch = new (_cfg->C) Node(1); // Pinch point to-be 2685 } 2686 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2687 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2688 return; 2689 } 2690 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init) 2691 _reg_node.map(def_reg,pinch); // Record pinch-point 2692 //_regalloc->set_bad(pinch->_idx); // Already initialized this way. 2693 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2694 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call 2695 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2696 later_def = NULL; // and no later def 2697 } 2698 pinch->set_req(0,later_def); // Hook later def so we can find it 2699 } else { // Else have valid pinch point 2700 if( pinch->in(0) ) // If there is a later-def 2701 later_def = pinch->in(0); // Get it 2702 } 2703 2704 // Add output-dependence edge from later def to kill 2705 if( later_def ) // If there is some original def 2706 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2707 2708 // See if current kill is also a use, and so is forced to be the pinch-point. 2709 if( pinch->Opcode() == Op_Node ) { 2710 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2711 for( uint i=1; i<uses->req(); i++ ) { 2712 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2713 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2714 // Yes, found a use/kill pinch-point 2715 pinch->set_req(0,NULL); // 2716 pinch->replace_by(kill); // Move anti-dep edges up 2717 pinch = kill; 2718 _reg_node.map(def_reg,pinch); 2719 return; 2720 } 2721 } 2722 } 2723 2724 // Add edge from kill to pinch-point 2725 add_prec_edge_from_to(kill,pinch); 2726 } 2727 2728 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2729 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2730 return; 2731 Node *pinch = _reg_node[use_reg]; // Get pinch point 2732 // Check for no later def_reg/kill in block 2733 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b && 2734 // Use has to be block-local as well 2735 _cfg->get_block_for_node(use) == b) { 2736 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2737 pinch->req() == 1 ) { // pinch not yet in block? 2738 pinch->del_req(0); // yank pointer to later-def, also set flag 2739 // Insert the pinch-point in the block just after the last use 2740 b->insert_node(pinch, b->find_node(use) + 1); 2741 _bb_end++; // Increase size scheduled region in block 2742 } 2743 2744 add_prec_edge_from_to(pinch,use); 2745 } 2746 } 2747 2748 // We insert antidependences between the reads and following write of 2749 // allocated registers to prevent illegal code motion. Hopefully, the 2750 // number of added references should be fairly small, especially as we 2751 // are only adding references within the current basic block. 2752 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2753 2754 #ifdef ASSERT 2755 verify_good_schedule(b,"before block local scheduling"); 2756 #endif 2757 2758 // A valid schedule, for each register independently, is an endless cycle 2759 // of: a def, then some uses (connected to the def by true dependencies), 2760 // then some kills (defs with no uses), finally the cycle repeats with a new 2761 // def. The uses are allowed to float relative to each other, as are the 2762 // kills. No use is allowed to slide past a kill (or def). This requires 2763 // antidependencies between all uses of a single def and all kills that 2764 // follow, up to the next def. More edges are redundant, because later defs 2765 // & kills are already serialized with true or antidependencies. To keep 2766 // the edge count down, we add a 'pinch point' node if there's more than 2767 // one use or more than one kill/def. 2768 2769 // We add dependencies in one bottom-up pass. 2770 2771 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2772 2773 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2774 // register. If not, we record the DEF/KILL in _reg_node, the 2775 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2776 // "pinch point", a new Node that's in the graph but not in the block. 2777 // We put edges from the prior and current DEF/KILLs to the pinch point. 2778 // We put the pinch point in _reg_node. If there's already a pinch point 2779 // we merely add an edge from the current DEF/KILL to the pinch point. 2780 2781 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2782 // put an edge from the pinch point to the USE. 2783 2784 // To be expedient, the _reg_node array is pre-allocated for the whole 2785 // compilation. _reg_node is lazily initialized; it either contains a NULL, 2786 // or a valid def/kill/pinch-point, or a leftover node from some prior 2787 // block. Leftover node from some prior block is treated like a NULL (no 2788 // prior def, so no anti-dependence needed). Valid def is distinguished by 2789 // it being in the current block. 2790 bool fat_proj_seen = false; 2791 uint last_safept = _bb_end-1; 2792 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL; 2793 Node* last_safept_node = end_node; 2794 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2795 Node *n = b->get_node(i); 2796 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2797 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) { 2798 // Fat-proj kills a slew of registers 2799 // This can add edges to 'n' and obscure whether or not it was a def, 2800 // hence the is_def flag. 2801 fat_proj_seen = true; 2802 RegMask rm = n->out_RegMask();// Make local copy 2803 while( rm.is_NotEmpty() ) { 2804 OptoReg::Name kill = rm.find_first_elem(); 2805 rm.Remove(kill); 2806 anti_do_def( b, n, kill, is_def ); 2807 } 2808 } else { 2809 // Get DEF'd registers the normal way 2810 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2811 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2812 } 2813 2814 // Kill projections on a branch should appear to occur on the 2815 // branch, not afterwards, so grab the masks from the projections 2816 // and process them. 2817 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) { 2818 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2819 Node* use = n->fast_out(i); 2820 if (use->is_Proj()) { 2821 RegMask rm = use->out_RegMask();// Make local copy 2822 while( rm.is_NotEmpty() ) { 2823 OptoReg::Name kill = rm.find_first_elem(); 2824 rm.Remove(kill); 2825 anti_do_def( b, n, kill, false ); 2826 } 2827 } 2828 } 2829 } 2830 2831 // Check each register used by this instruction for a following DEF/KILL 2832 // that must occur afterward and requires an anti-dependence edge. 2833 for( uint j=0; j<n->req(); j++ ) { 2834 Node *def = n->in(j); 2835 if( def ) { 2836 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2837 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2838 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2839 } 2840 } 2841 // Do not allow defs of new derived values to float above GC 2842 // points unless the base is definitely available at the GC point. 2843 2844 Node *m = b->get_node(i); 2845 2846 // Add precedence edge from following safepoint to use of derived pointer 2847 if( last_safept_node != end_node && 2848 m != last_safept_node) { 2849 for (uint k = 1; k < m->req(); k++) { 2850 const Type *t = m->in(k)->bottom_type(); 2851 if( t->isa_oop_ptr() && 2852 t->is_ptr()->offset() != 0 ) { 2853 last_safept_node->add_prec( m ); 2854 break; 2855 } 2856 } 2857 } 2858 2859 if( n->jvms() ) { // Precedence edge from derived to safept 2860 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2861 if( b->get_node(last_safept) != last_safept_node ) { 2862 last_safept = b->find_node(last_safept_node); 2863 } 2864 for( uint j=last_safept; j > i; j-- ) { 2865 Node *mach = b->get_node(j); 2866 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2867 mach->add_prec( n ); 2868 } 2869 last_safept = i; 2870 last_safept_node = m; 2871 } 2872 } 2873 2874 if (fat_proj_seen) { 2875 // Garbage collect pinch nodes that were not consumed. 2876 // They are usually created by a fat kill MachProj for a call. 2877 garbage_collect_pinch_nodes(); 2878 } 2879 } 2880 2881 // Garbage collect pinch nodes for reuse by other blocks. 2882 // 2883 // The block scheduler's insertion of anti-dependence 2884 // edges creates many pinch nodes when the block contains 2885 // 2 or more Calls. A pinch node is used to prevent a 2886 // combinatorial explosion of edges. If a set of kills for a 2887 // register is anti-dependent on a set of uses (or defs), rather 2888 // than adding an edge in the graph between each pair of kill 2889 // and use (or def), a pinch is inserted between them: 2890 // 2891 // use1 use2 use3 2892 // \ | / 2893 // \ | / 2894 // pinch 2895 // / | \ 2896 // / | \ 2897 // kill1 kill2 kill3 2898 // 2899 // One pinch node is created per register killed when 2900 // the second call is encountered during a backwards pass 2901 // over the block. Most of these pinch nodes are never 2902 // wired into the graph because the register is never 2903 // used or def'ed in the block. 2904 // 2905 void Scheduling::garbage_collect_pinch_nodes() { 2906 #ifndef PRODUCT 2907 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 2908 #endif 2909 int trace_cnt = 0; 2910 for (uint k = 0; k < _reg_node.Size(); k++) { 2911 Node* pinch = _reg_node[k]; 2912 if ((pinch != NULL) && pinch->Opcode() == Op_Node && 2913 // no predecence input edges 2914 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { 2915 cleanup_pinch(pinch); 2916 _pinch_free_list.push(pinch); 2917 _reg_node.map(k, NULL); 2918 #ifndef PRODUCT 2919 if (_cfg->C->trace_opto_output()) { 2920 trace_cnt++; 2921 if (trace_cnt > 40) { 2922 tty->print("\n"); 2923 trace_cnt = 0; 2924 } 2925 tty->print(" %d", pinch->_idx); 2926 } 2927 #endif 2928 } 2929 } 2930 #ifndef PRODUCT 2931 if (_cfg->C->trace_opto_output()) tty->print("\n"); 2932 #endif 2933 } 2934 2935 // Clean up a pinch node for reuse. 2936 void Scheduling::cleanup_pinch( Node *pinch ) { 2937 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 2938 2939 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 2940 Node* use = pinch->last_out(i); 2941 uint uses_found = 0; 2942 for (uint j = use->req(); j < use->len(); j++) { 2943 if (use->in(j) == pinch) { 2944 use->rm_prec(j); 2945 uses_found++; 2946 } 2947 } 2948 assert(uses_found > 0, "must be a precedence edge"); 2949 i -= uses_found; // we deleted 1 or more copies of this edge 2950 } 2951 // May have a later_def entry 2952 pinch->set_req(0, NULL); 2953 } 2954 2955 #ifndef PRODUCT 2956 2957 void Scheduling::dump_available() const { 2958 tty->print("#Availist "); 2959 for (uint i = 0; i < _available.size(); i++) 2960 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 2961 tty->cr(); 2962 } 2963 2964 // Print Scheduling Statistics 2965 void Scheduling::print_statistics() { 2966 // Print the size added by nops for bundling 2967 tty->print("Nops added %d bytes to total of %d bytes", 2968 _total_nop_size, _total_method_size); 2969 if (_total_method_size > 0) 2970 tty->print(", for %.2f%%", 2971 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 2972 tty->print("\n"); 2973 2974 // Print the number of branch shadows filled 2975 if (Pipeline::_branch_has_delay_slot) { 2976 tty->print("Of %d branches, %d had unconditional delay slots filled", 2977 _total_branches, _total_unconditional_delays); 2978 if (_total_branches > 0) 2979 tty->print(", for %.2f%%", 2980 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 2981 tty->print("\n"); 2982 } 2983 2984 uint total_instructions = 0, total_bundles = 0; 2985 2986 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 2987 uint bundle_count = _total_instructions_per_bundle[i]; 2988 total_instructions += bundle_count * i; 2989 total_bundles += bundle_count; 2990 } 2991 2992 if (total_bundles > 0) 2993 tty->print("Average ILP (excluding nops) is %.2f\n", 2994 ((double)total_instructions) / ((double)total_bundles)); 2995 } 2996 #endif