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); 347 348 // Collect worst case block paddings 349 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks); 350 memset(block_worst_case_pad, 0, nblocks * sizeof(int)); 351 352 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); ) 353 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); ) 354 355 bool has_short_branch_candidate = false; 356 357 // Initialize the sizes to 0 358 code_size = 0; // Size in bytes of generated code 359 stub_size = 0; // Size in bytes of all stub entries 360 // Size in bytes of all relocation entries, including those in local stubs. 361 // Start with 2-bytes of reloc info for the unvalidated entry point 362 reloc_size = 1; // Number of relocation entries 363 364 // Make three passes. The first computes pessimistic blk_starts, 365 // relative jmp_offset and reloc_size information. The second performs 366 // short branch substitution using the pessimistic sizing. The 367 // third inserts nops where needed. 368 369 // Step one, perform a pessimistic sizing pass. 370 uint last_call_adr = max_uint; 371 uint last_avoid_back_to_back_adr = max_uint; 372 uint nop_size = (new (this) MachNopNode())->size(_regalloc); 373 for (uint i = 0; i < nblocks; i++) { // For all blocks 374 Block* block = _cfg->get_block(i); 375 376 // During short branch replacement, we store the relative (to blk_starts) 377 // offset of jump in jmp_offset, rather than the absolute offset of jump. 378 // This is so that we do not need to recompute sizes of all nodes when 379 // we compute correct blk_starts in our next sizing pass. 380 jmp_offset[i] = 0; 381 jmp_size[i] = 0; 382 jmp_nidx[i] = -1; 383 DEBUG_ONLY( jmp_target[i] = 0; ) 384 DEBUG_ONLY( jmp_rule[i] = 0; ) 385 386 // Sum all instruction sizes to compute block size 387 uint last_inst = block->number_of_nodes(); 388 uint blk_size = 0; 389 for (uint j = 0; j < last_inst; j++) { 390 Node* nj = block->get_node(j); 391 // Handle machine instruction nodes 392 if (nj->is_Mach()) { 393 MachNode *mach = nj->as_Mach(); 394 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 395 reloc_size += mach->reloc(); 396 if (mach->is_MachCall()) { 397 MachCallNode *mcall = mach->as_MachCall(); 398 // This destination address is NOT PC-relative 399 400 mcall->method_set((intptr_t)mcall->entry_point()); 401 402 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) { 403 stub_size += CompiledStaticCall::to_interp_stub_size(); 404 reloc_size += CompiledStaticCall::reloc_to_interp_stub(); 405 } 406 } else if (mach->is_MachSafePoint()) { 407 // If call/safepoint are adjacent, account for possible 408 // nop to disambiguate the two safepoints. 409 // ScheduleAndBundle() can rearrange nodes in a block, 410 // check for all offsets inside this block. 411 if (last_call_adr >= blk_starts[i]) { 412 blk_size += nop_size; 413 } 414 } 415 if (mach->avoid_back_to_back()) { 416 // Nop is inserted between "avoid back to back" instructions. 417 // ScheduleAndBundle() can rearrange nodes in a block, 418 // check for all offsets inside this block. 419 if (last_avoid_back_to_back_adr >= blk_starts[i]) { 420 blk_size += nop_size; 421 } 422 } 423 if (mach->may_be_short_branch()) { 424 if (!nj->is_MachBranch()) { 425 #ifndef PRODUCT 426 nj->dump(3); 427 #endif 428 Unimplemented(); 429 } 430 assert(jmp_nidx[i] == -1, "block should have only one branch"); 431 jmp_offset[i] = blk_size; 432 jmp_size[i] = nj->size(_regalloc); 433 jmp_nidx[i] = j; 434 has_short_branch_candidate = true; 435 } 436 } 437 blk_size += nj->size(_regalloc); 438 // Remember end of call offset 439 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) { 440 last_call_adr = blk_starts[i]+blk_size; 441 } 442 // Remember end of avoid_back_to_back offset 443 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back()) { 444 last_avoid_back_to_back_adr = blk_starts[i]+blk_size; 445 } 446 } 447 448 // When the next block starts a loop, we may insert pad NOP 449 // instructions. Since we cannot know our future alignment, 450 // assume the worst. 451 if (i < nblocks - 1) { 452 Block* nb = _cfg->get_block(i + 1); 453 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 454 if (max_loop_pad > 0) { 455 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 456 // Adjust last_call_adr and/or last_avoid_back_to_back_adr. 457 // If either is the last instruction in this block, bump by 458 // max_loop_pad in lock-step with blk_size, so sizing 459 // calculations in subsequent blocks still can conservatively 460 // detect that it may the last instruction in this block. 461 if (last_call_adr == blk_starts[i]+blk_size) { 462 last_call_adr += max_loop_pad; 463 } 464 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) { 465 last_avoid_back_to_back_adr += max_loop_pad; 466 } 467 blk_size += max_loop_pad; 468 block_worst_case_pad[i + 1] = max_loop_pad; 469 } 470 } 471 472 // Save block size; update total method size 473 blk_starts[i+1] = blk_starts[i]+blk_size; 474 } 475 476 // Step two, replace eligible long jumps. 477 bool progress = true; 478 uint last_may_be_short_branch_adr = max_uint; 479 while (has_short_branch_candidate && progress) { 480 progress = false; 481 has_short_branch_candidate = false; 482 int adjust_block_start = 0; 483 for (uint i = 0; i < nblocks; i++) { 484 Block* block = _cfg->get_block(i); 485 int idx = jmp_nidx[i]; 486 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach(); 487 if (mach != NULL && mach->may_be_short_branch()) { 488 #ifdef ASSERT 489 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity"); 490 int j; 491 // Find the branch; ignore trailing NOPs. 492 for (j = block->number_of_nodes()-1; j>=0; j--) { 493 Node* n = block->get_node(j); 494 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) 495 break; 496 } 497 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity"); 498 #endif 499 int br_size = jmp_size[i]; 500 int br_offs = blk_starts[i] + jmp_offset[i]; 501 502 // This requires the TRUE branch target be in succs[0] 503 uint bnum = block->non_connector_successor(0)->_pre_order; 504 int offset = blk_starts[bnum] - br_offs; 505 if (bnum > i) { // adjust following block's offset 506 offset -= adjust_block_start; 507 } 508 509 // This block can be a loop header, account for the padding 510 // in the previous block. 511 int block_padding = block_worst_case_pad[i]; 512 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top"); 513 // In the following code a nop could be inserted before 514 // the branch which will increase the backward distance. 515 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr); 516 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block"); 517 518 if (needs_padding && offset <= 0) 519 offset -= nop_size; 520 521 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 522 // We've got a winner. Replace this branch. 523 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this); 524 525 // Update the jmp_size. 526 int new_size = replacement->size(_regalloc); 527 int diff = br_size - new_size; 528 assert(diff >= (int)nop_size, "short_branch size should be smaller"); 529 // Conservatively take into accound padding between 530 // avoid_back_to_back branches. Previous branch could be 531 // converted into avoid_back_to_back branch during next 532 // rounds. 533 if (needs_padding && replacement->avoid_back_to_back()) { 534 jmp_offset[i] += nop_size; 535 diff -= nop_size; 536 } 537 adjust_block_start += diff; 538 block->map_node(replacement, idx); 539 mach->subsume_by(replacement, C); 540 mach = replacement; 541 progress = true; 542 543 jmp_size[i] = new_size; 544 DEBUG_ONLY( jmp_target[i] = bnum; ); 545 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 546 } else { 547 // The jump distance is not short, try again during next iteration. 548 has_short_branch_candidate = true; 549 } 550 } // (mach->may_be_short_branch()) 551 if (mach != NULL && (mach->may_be_short_branch() || 552 mach->avoid_back_to_back())) { 553 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i]; 554 } 555 blk_starts[i+1] -= adjust_block_start; 556 } 557 } 558 559 #ifdef ASSERT 560 for (uint i = 0; i < nblocks; i++) { // For all blocks 561 if (jmp_target[i] != 0) { 562 int br_size = jmp_size[i]; 563 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 564 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 565 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]); 566 } 567 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp"); 568 } 569 } 570 #endif 571 572 // Step 3, compute the offsets of all blocks, will be done in fill_buffer() 573 // after ScheduleAndBundle(). 574 575 // ------------------ 576 // Compute size for code buffer 577 code_size = blk_starts[nblocks]; 578 579 // Relocation records 580 reloc_size += 1; // Relo entry for exception handler 581 582 // Adjust reloc_size to number of record of relocation info 583 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 584 // a relocation index. 585 // The CodeBuffer will expand the locs array if this estimate is too low. 586 reloc_size *= 10 / sizeof(relocInfo); 587 } 588 589 //------------------------------FillLocArray----------------------------------- 590 // Create a bit of debug info and append it to the array. The mapping is from 591 // Java local or expression stack to constant, register or stack-slot. For 592 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 593 // entry has been taken care of and caller should skip it). 594 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 595 // This should never have accepted Bad before 596 assert(OptoReg::is_valid(regnum), "location must be valid"); 597 return (OptoReg::is_reg(regnum)) 598 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 599 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 600 } 601 602 603 ObjectValue* 604 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 605 for (int i = 0; i < objs->length(); i++) { 606 assert(objs->at(i)->is_object(), "corrupt object cache"); 607 ObjectValue* sv = (ObjectValue*) objs->at(i); 608 if (sv->id() == id) { 609 return sv; 610 } 611 } 612 // Otherwise.. 613 return NULL; 614 } 615 616 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 617 ObjectValue* sv ) { 618 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition"); 619 objs->append(sv); 620 } 621 622 623 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 624 GrowableArray<ScopeValue*> *array, 625 GrowableArray<ScopeValue*> *objs ) { 626 assert( local, "use _top instead of null" ); 627 if (array->length() != idx) { 628 assert(array->length() == idx + 1, "Unexpected array count"); 629 // Old functionality: 630 // return 631 // New functionality: 632 // Assert if the local is not top. In product mode let the new node 633 // override the old entry. 634 assert(local == top(), "LocArray collision"); 635 if (local == top()) { 636 return; 637 } 638 array->pop(); 639 } 640 const Type *t = local->bottom_type(); 641 642 // Is it a safepoint scalar object node? 643 if (local->is_SafePointScalarObject()) { 644 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 645 646 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx); 647 if (sv == NULL) { 648 ciKlass* cik = t->is_oopptr()->klass(); 649 assert(cik->is_instance_klass() || 650 cik->is_array_klass(), "Not supported allocation."); 651 sv = new ObjectValue(spobj->_idx, 652 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 653 Compile::set_sv_for_object_node(objs, sv); 654 655 uint first_ind = spobj->first_index(sfpt->jvms()); 656 for (uint i = 0; i < spobj->n_fields(); i++) { 657 Node* fld_node = sfpt->in(first_ind+i); 658 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 659 } 660 } 661 array->append(sv); 662 return; 663 } 664 665 // Grab the register number for the local 666 OptoReg::Name regnum = _regalloc->get_reg_first(local); 667 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 668 // Record the double as two float registers. 669 // The register mask for such a value always specifies two adjacent 670 // float registers, with the lower register number even. 671 // Normally, the allocation of high and low words to these registers 672 // is irrelevant, because nearly all operations on register pairs 673 // (e.g., StoreD) treat them as a single unit. 674 // Here, we assume in addition that the words in these two registers 675 // stored "naturally" (by operations like StoreD and double stores 676 // within the interpreter) such that the lower-numbered register 677 // is written to the lower memory address. This may seem like 678 // a machine dependency, but it is not--it is a requirement on 679 // the author of the <arch>.ad file to ensure that, for every 680 // even/odd double-register pair to which a double may be allocated, 681 // the word in the even single-register is stored to the first 682 // memory word. (Note that register numbers are completely 683 // arbitrary, and are not tied to any machine-level encodings.) 684 #ifdef _LP64 685 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 686 array->append(new ConstantIntValue(0)); 687 array->append(new_loc_value( _regalloc, regnum, Location::dbl )); 688 } else if ( t->base() == Type::Long ) { 689 array->append(new ConstantIntValue(0)); 690 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 691 } else if ( t->base() == Type::RawPtr ) { 692 // jsr/ret return address which must be restored into a the full 693 // width 64-bit stack slot. 694 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 695 } 696 #else //_LP64 697 #ifdef SPARC 698 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) { 699 // For SPARC we have to swap high and low words for 700 // long values stored in a single-register (g0-g7). 701 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 702 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 703 } else 704 #endif //SPARC 705 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 706 // Repack the double/long as two jints. 707 // The convention the interpreter uses is that the second local 708 // holds the first raw word of the native double representation. 709 // This is actually reasonable, since locals and stack arrays 710 // grow downwards in all implementations. 711 // (If, on some machine, the interpreter's Java locals or stack 712 // were to grow upwards, the embedded doubles would be word-swapped.) 713 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 714 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 715 } 716 #endif //_LP64 717 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 718 OptoReg::is_reg(regnum) ) { 719 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double() 720 ? Location::float_in_dbl : Location::normal )); 721 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 722 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long 723 ? Location::int_in_long : Location::normal )); 724 } else if( t->base() == Type::NarrowOop ) { 725 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop )); 726 } else { 727 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal )); 728 } 729 return; 730 } 731 732 // No register. It must be constant data. 733 switch (t->base()) { 734 case Type::Half: // Second half of a double 735 ShouldNotReachHere(); // Caller should skip 2nd halves 736 break; 737 case Type::AnyPtr: 738 array->append(new ConstantOopWriteValue(NULL)); 739 break; 740 case Type::AryPtr: 741 case Type::InstPtr: // fall through 742 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 743 break; 744 case Type::NarrowOop: 745 if (t == TypeNarrowOop::NULL_PTR) { 746 array->append(new ConstantOopWriteValue(NULL)); 747 } else { 748 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 749 } 750 break; 751 case Type::Int: 752 array->append(new ConstantIntValue(t->is_int()->get_con())); 753 break; 754 case Type::RawPtr: 755 // A return address (T_ADDRESS). 756 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 757 #ifdef _LP64 758 // Must be restored to the full-width 64-bit stack slot. 759 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 760 #else 761 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 762 #endif 763 break; 764 case Type::FloatCon: { 765 float f = t->is_float_constant()->getf(); 766 array->append(new ConstantIntValue(jint_cast(f))); 767 break; 768 } 769 case Type::DoubleCon: { 770 jdouble d = t->is_double_constant()->getd(); 771 #ifdef _LP64 772 array->append(new ConstantIntValue(0)); 773 array->append(new ConstantDoubleValue(d)); 774 #else 775 // Repack the double as two jints. 776 // The convention the interpreter uses is that the second local 777 // holds the first raw word of the native double representation. 778 // This is actually reasonable, since locals and stack arrays 779 // grow downwards in all implementations. 780 // (If, on some machine, the interpreter's Java locals or stack 781 // were to grow upwards, the embedded doubles would be word-swapped.) 782 jint *dp = (jint*)&d; 783 array->append(new ConstantIntValue(dp[1])); 784 array->append(new ConstantIntValue(dp[0])); 785 #endif 786 break; 787 } 788 case Type::Long: { 789 jlong d = t->is_long()->get_con(); 790 #ifdef _LP64 791 array->append(new ConstantIntValue(0)); 792 array->append(new ConstantLongValue(d)); 793 #else 794 // Repack the long as two jints. 795 // The convention the interpreter uses is that the second local 796 // holds the first raw word of the native double representation. 797 // This is actually reasonable, since locals and stack arrays 798 // grow downwards in all implementations. 799 // (If, on some machine, the interpreter's Java locals or stack 800 // were to grow upwards, the embedded doubles would be word-swapped.) 801 jint *dp = (jint*)&d; 802 array->append(new ConstantIntValue(dp[1])); 803 array->append(new ConstantIntValue(dp[0])); 804 #endif 805 break; 806 } 807 case Type::Top: // Add an illegal value here 808 array->append(new LocationValue(Location())); 809 break; 810 default: 811 ShouldNotReachHere(); 812 break; 813 } 814 } 815 816 // Determine if this node starts a bundle 817 bool Compile::starts_bundle(const Node *n) const { 818 return (_node_bundling_limit > n->_idx && 819 _node_bundling_base[n->_idx].starts_bundle()); 820 } 821 822 //--------------------------Process_OopMap_Node-------------------------------- 823 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) { 824 825 // Handle special safepoint nodes for synchronization 826 MachSafePointNode *sfn = mach->as_MachSafePoint(); 827 MachCallNode *mcall; 828 829 #ifdef ENABLE_ZAP_DEAD_LOCALS 830 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative"); 831 #endif 832 833 int safepoint_pc_offset = current_offset; 834 bool is_method_handle_invoke = false; 835 bool return_oop = false; 836 837 // Add the safepoint in the DebugInfoRecorder 838 if( !mach->is_MachCall() ) { 839 mcall = NULL; 840 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 841 } else { 842 mcall = mach->as_MachCall(); 843 844 // Is the call a MethodHandle call? 845 if (mcall->is_MachCallJava()) { 846 if (mcall->as_MachCallJava()->_method_handle_invoke) { 847 assert(has_method_handle_invokes(), "must have been set during call generation"); 848 is_method_handle_invoke = true; 849 } 850 } 851 852 // Check if a call returns an object. 853 if (mcall->return_value_is_used() && 854 mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) { 855 return_oop = true; 856 } 857 safepoint_pc_offset += mcall->ret_addr_offset(); 858 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 859 } 860 861 // Loop over the JVMState list to add scope information 862 // Do not skip safepoints with a NULL method, they need monitor info 863 JVMState* youngest_jvms = sfn->jvms(); 864 int max_depth = youngest_jvms->depth(); 865 866 // Allocate the object pool for scalar-replaced objects -- the map from 867 // small-integer keys (which can be recorded in the local and ostack 868 // arrays) to descriptions of the object state. 869 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 870 871 // Visit scopes from oldest to youngest. 872 for (int depth = 1; depth <= max_depth; depth++) { 873 JVMState* jvms = youngest_jvms->of_depth(depth); 874 int idx; 875 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 876 // Safepoints that do not have method() set only provide oop-map and monitor info 877 // to support GC; these do not support deoptimization. 878 int num_locs = (method == NULL) ? 0 : jvms->loc_size(); 879 int num_exps = (method == NULL) ? 0 : jvms->stk_size(); 880 int num_mon = jvms->nof_monitors(); 881 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), 882 "JVMS local count must match that of the method"); 883 884 // Add Local and Expression Stack Information 885 886 // Insert locals into the locarray 887 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 888 for( idx = 0; idx < num_locs; idx++ ) { 889 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 890 } 891 892 // Insert expression stack entries into the exparray 893 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 894 for( idx = 0; idx < num_exps; idx++ ) { 895 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 896 } 897 898 // Add in mappings of the monitors 899 assert( !method || 900 !method->is_synchronized() || 901 method->is_native() || 902 num_mon > 0 || 903 !GenerateSynchronizationCode, 904 "monitors must always exist for synchronized methods"); 905 906 // Build the growable array of ScopeValues for exp stack 907 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 908 909 // Loop over monitors and insert into array 910 for (idx = 0; idx < num_mon; idx++) { 911 // Grab the node that defines this monitor 912 Node* box_node = sfn->monitor_box(jvms, idx); 913 Node* obj_node = sfn->monitor_obj(jvms, idx); 914 915 // Create ScopeValue for object 916 ScopeValue *scval = NULL; 917 918 if (obj_node->is_SafePointScalarObject()) { 919 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 920 scval = Compile::sv_for_node_id(objs, spobj->_idx); 921 if (scval == NULL) { 922 const Type *t = spobj->bottom_type(); 923 ciKlass* cik = t->is_oopptr()->klass(); 924 assert(cik->is_instance_klass() || 925 cik->is_array_klass(), "Not supported allocation."); 926 ObjectValue* sv = new ObjectValue(spobj->_idx, 927 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 928 Compile::set_sv_for_object_node(objs, sv); 929 930 uint first_ind = spobj->first_index(youngest_jvms); 931 for (uint i = 0; i < spobj->n_fields(); i++) { 932 Node* fld_node = sfn->in(first_ind+i); 933 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 934 } 935 scval = sv; 936 } 937 } else if (!obj_node->is_Con()) { 938 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); 939 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 940 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop ); 941 } else { 942 scval = new_loc_value( _regalloc, obj_reg, Location::oop ); 943 } 944 } else { 945 const TypePtr *tp = obj_node->get_ptr_type(); 946 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding()); 947 } 948 949 OptoReg::Name box_reg = BoxLockNode::reg(box_node); 950 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)); 951 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated()); 952 monarray->append(new MonitorValue(scval, basic_lock, eliminated)); 953 } 954 955 // We dump the object pool first, since deoptimization reads it in first. 956 debug_info()->dump_object_pool(objs); 957 958 // Build first class objects to pass to scope 959 DebugToken *locvals = debug_info()->create_scope_values(locarray); 960 DebugToken *expvals = debug_info()->create_scope_values(exparray); 961 DebugToken *monvals = debug_info()->create_monitor_values(monarray); 962 963 // Make method available for all Safepoints 964 ciMethod* scope_method = method ? method : _method; 965 // Describe the scope here 966 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 967 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 968 // Now we can describe the scope. 969 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals); 970 } // End jvms loop 971 972 // Mark the end of the scope set. 973 debug_info()->end_safepoint(safepoint_pc_offset); 974 } 975 976 977 978 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 979 class NonSafepointEmitter { 980 Compile* C; 981 JVMState* _pending_jvms; 982 int _pending_offset; 983 984 void emit_non_safepoint(); 985 986 public: 987 NonSafepointEmitter(Compile* compile) { 988 this->C = compile; 989 _pending_jvms = NULL; 990 _pending_offset = 0; 991 } 992 993 void observe_instruction(Node* n, int pc_offset) { 994 if (!C->debug_info()->recording_non_safepoints()) return; 995 996 Node_Notes* nn = C->node_notes_at(n->_idx); 997 if (nn == NULL || nn->jvms() == NULL) return; 998 if (_pending_jvms != NULL && 999 _pending_jvms->same_calls_as(nn->jvms())) { 1000 // Repeated JVMS? Stretch it up here. 1001 _pending_offset = pc_offset; 1002 } else { 1003 if (_pending_jvms != NULL && 1004 _pending_offset < pc_offset) { 1005 emit_non_safepoint(); 1006 } 1007 _pending_jvms = NULL; 1008 if (pc_offset > C->debug_info()->last_pc_offset()) { 1009 // This is the only way _pending_jvms can become non-NULL: 1010 _pending_jvms = nn->jvms(); 1011 _pending_offset = pc_offset; 1012 } 1013 } 1014 } 1015 1016 // Stay out of the way of real safepoints: 1017 void observe_safepoint(JVMState* jvms, int pc_offset) { 1018 if (_pending_jvms != NULL && 1019 !_pending_jvms->same_calls_as(jvms) && 1020 _pending_offset < pc_offset) { 1021 emit_non_safepoint(); 1022 } 1023 _pending_jvms = NULL; 1024 } 1025 1026 void flush_at_end() { 1027 if (_pending_jvms != NULL) { 1028 emit_non_safepoint(); 1029 } 1030 _pending_jvms = NULL; 1031 } 1032 }; 1033 1034 void NonSafepointEmitter::emit_non_safepoint() { 1035 JVMState* youngest_jvms = _pending_jvms; 1036 int pc_offset = _pending_offset; 1037 1038 // Clear it now: 1039 _pending_jvms = NULL; 1040 1041 DebugInformationRecorder* debug_info = C->debug_info(); 1042 assert(debug_info->recording_non_safepoints(), "sanity"); 1043 1044 debug_info->add_non_safepoint(pc_offset); 1045 int max_depth = youngest_jvms->depth(); 1046 1047 // Visit scopes from oldest to youngest. 1048 for (int depth = 1; depth <= max_depth; depth++) { 1049 JVMState* jvms = youngest_jvms->of_depth(depth); 1050 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 1051 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 1052 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute()); 1053 } 1054 1055 // Mark the end of the scope set. 1056 debug_info->end_non_safepoint(pc_offset); 1057 } 1058 1059 //------------------------------init_buffer------------------------------------ 1060 CodeBuffer* Compile::init_buffer(uint* blk_starts) { 1061 1062 // Set the initially allocated size 1063 int code_req = initial_code_capacity; 1064 int locs_req = initial_locs_capacity; 1065 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity; 1066 int const_req = initial_const_capacity; 1067 1068 int pad_req = NativeCall::instruction_size; 1069 // The extra spacing after the code is necessary on some platforms. 1070 // Sometimes we need to patch in a jump after the last instruction, 1071 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 1072 1073 // Compute the byte offset where we can store the deopt pc. 1074 if (fixed_slots() != 0) { 1075 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 1076 } 1077 1078 // Compute prolog code size 1079 _method_size = 0; 1080 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; 1081 #if defined(IA64) && !defined(AIX) 1082 if (save_argument_registers()) { 1083 // 4815101: this is a stub with implicit and unknown precision fp args. 1084 // The usual spill mechanism can only generate stfd's in this case, which 1085 // doesn't work if the fp reg to spill contains a single-precision denorm. 1086 // Instead, we hack around the normal spill mechanism using stfspill's and 1087 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate 1088 // space here for the fp arg regs (f8-f15) we're going to thusly spill. 1089 // 1090 // If we ever implement 16-byte 'registers' == stack slots, we can 1091 // get rid of this hack and have SpillCopy generate stfspill/ldffill 1092 // instead of stfd/stfs/ldfd/ldfs. 1093 _frame_slots += 8*(16/BytesPerInt); 1094 } 1095 #endif 1096 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check"); 1097 1098 if (has_mach_constant_base_node()) { 1099 uint add_size = 0; 1100 // Fill the constant table. 1101 // Note: This must happen before shorten_branches. 1102 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 1103 Block* b = _cfg->get_block(i); 1104 1105 for (uint j = 0; j < b->number_of_nodes(); j++) { 1106 Node* n = b->get_node(j); 1107 1108 // If the node is a MachConstantNode evaluate the constant 1109 // value section. 1110 if (n->is_MachConstant()) { 1111 MachConstantNode* machcon = n->as_MachConstant(); 1112 machcon->eval_constant(C); 1113 } else if (n->is_Mach()) { 1114 // On Power there are more nodes that issue constants. 1115 add_size += (n->as_Mach()->ins_num_consts() * 8); 1116 } 1117 } 1118 } 1119 1120 // Calculate the offsets of the constants and the size of the 1121 // constant table (including the padding to the next section). 1122 constant_table().calculate_offsets_and_size(); 1123 const_req = constant_table().size() + add_size; 1124 } 1125 1126 // Initialize the space for the BufferBlob used to find and verify 1127 // instruction size in MachNode::emit_size() 1128 init_scratch_buffer_blob(const_req); 1129 if (failing()) return NULL; // Out of memory 1130 1131 // Pre-compute the length of blocks and replace 1132 // long branches with short if machine supports it. 1133 shorten_branches(blk_starts, code_req, locs_req, stub_req); 1134 1135 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1136 int exception_handler_req = size_exception_handler(); 1137 int deopt_handler_req = size_deopt_handler(); 1138 exception_handler_req += MAX_stubs_size; // add marginal slop for handler 1139 deopt_handler_req += MAX_stubs_size; // add marginal slop for handler 1140 stub_req += MAX_stubs_size; // ensure per-stub margin 1141 code_req += MAX_inst_size; // ensure per-instruction margin 1142 1143 if (StressCodeBuffers) 1144 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1145 1146 int total_req = 1147 const_req + 1148 code_req + 1149 pad_req + 1150 stub_req + 1151 exception_handler_req + 1152 deopt_handler_req; // deopt handler 1153 1154 if (has_method_handle_invokes()) 1155 total_req += deopt_handler_req; // deopt MH handler 1156 1157 CodeBuffer* cb = code_buffer(); 1158 cb->initialize(total_req, locs_req); 1159 1160 // Have we run out of code space? 1161 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1162 C->record_failure("CodeCache is full"); 1163 return NULL; 1164 } 1165 // Configure the code buffer. 1166 cb->initialize_consts_size(const_req); 1167 cb->initialize_stubs_size(stub_req); 1168 cb->initialize_oop_recorder(env()->oop_recorder()); 1169 1170 // fill in the nop array for bundling computations 1171 MachNode *_nop_list[Bundle::_nop_count]; 1172 Bundle::initialize_nops(_nop_list, this); 1173 1174 return cb; 1175 } 1176 1177 //------------------------------fill_buffer------------------------------------ 1178 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) { 1179 // blk_starts[] contains offsets calculated during short branches processing, 1180 // offsets should not be increased during following steps. 1181 1182 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 1183 // of a loop. It is used to determine the padding for loop alignment. 1184 compute_loop_first_inst_sizes(); 1185 1186 // Create oopmap set. 1187 _oop_map_set = new OopMapSet(); 1188 1189 // !!!!! This preserves old handling of oopmaps for now 1190 debug_info()->set_oopmaps(_oop_map_set); 1191 1192 uint nblocks = _cfg->number_of_blocks(); 1193 // Count and start of implicit null check instructions 1194 uint inct_cnt = 0; 1195 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1196 1197 // Count and start of calls 1198 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1199 1200 uint return_offset = 0; 1201 int nop_size = (new (this) MachNopNode())->size(_regalloc); 1202 1203 int previous_offset = 0; 1204 int current_offset = 0; 1205 int last_call_offset = -1; 1206 int last_avoid_back_to_back_offset = -1; 1207 #ifdef ASSERT 1208 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); 1209 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 1210 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 1211 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); 1212 #endif 1213 1214 // Create an array of unused labels, one for each basic block, if printing is enabled 1215 #ifndef PRODUCT 1216 int *node_offsets = NULL; 1217 uint node_offset_limit = unique(); 1218 1219 if (print_assembly()) 1220 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1221 #endif 1222 1223 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily 1224 1225 // Emit the constant table. 1226 if (has_mach_constant_base_node()) { 1227 constant_table().emit(*cb); 1228 } 1229 1230 // Create an array of labels, one for each basic block 1231 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1); 1232 for (uint i=0; i <= nblocks; i++) { 1233 blk_labels[i].init(); 1234 } 1235 1236 // ------------------ 1237 // Now fill in the code buffer 1238 Node *delay_slot = NULL; 1239 1240 for (uint i = 0; i < nblocks; i++) { 1241 Block* block = _cfg->get_block(i); 1242 Node* head = block->head(); 1243 1244 // If this block needs to start aligned (i.e, can be reached other 1245 // than by falling-thru from the previous block), then force the 1246 // start of a new bundle. 1247 if (Pipeline::requires_bundling() && starts_bundle(head)) { 1248 cb->flush_bundle(true); 1249 } 1250 1251 #ifdef ASSERT 1252 if (!block->is_connector()) { 1253 stringStream st; 1254 block->dump_head(_cfg, &st); 1255 MacroAssembler(cb).block_comment(st.as_string()); 1256 } 1257 jmp_target[i] = 0; 1258 jmp_offset[i] = 0; 1259 jmp_size[i] = 0; 1260 jmp_rule[i] = 0; 1261 #endif 1262 int blk_offset = current_offset; 1263 1264 // Define the label at the beginning of the basic block 1265 MacroAssembler(cb).bind(blk_labels[block->_pre_order]); 1266 1267 uint last_inst = block->number_of_nodes(); 1268 1269 // Emit block normally, except for last instruction. 1270 // Emit means "dump code bits into code buffer". 1271 for (uint j = 0; j<last_inst; j++) { 1272 1273 // Get the node 1274 Node* n = block->get_node(j); 1275 1276 // See if delay slots are supported 1277 if (valid_bundle_info(n) && 1278 node_bundling(n)->used_in_unconditional_delay()) { 1279 assert(delay_slot == NULL, "no use of delay slot node"); 1280 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1281 1282 delay_slot = n; 1283 continue; 1284 } 1285 1286 // If this starts a new instruction group, then flush the current one 1287 // (but allow split bundles) 1288 if (Pipeline::requires_bundling() && starts_bundle(n)) 1289 cb->flush_bundle(false); 1290 1291 // The following logic is duplicated in the code ifdeffed for 1292 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It 1293 // should be factored out. Or maybe dispersed to the nodes? 1294 1295 // Special handling for SafePoint/Call Nodes 1296 bool is_mcall = false; 1297 if (n->is_Mach()) { 1298 MachNode *mach = n->as_Mach(); 1299 is_mcall = n->is_MachCall(); 1300 bool is_sfn = n->is_MachSafePoint(); 1301 1302 // If this requires all previous instructions be flushed, then do so 1303 if (is_sfn || is_mcall || mach->alignment_required() != 1) { 1304 cb->flush_bundle(true); 1305 current_offset = cb->insts_size(); 1306 } 1307 1308 // A padding may be needed again since a previous instruction 1309 // could be moved to delay slot. 1310 1311 // align the instruction if necessary 1312 int padding = mach->compute_padding(current_offset); 1313 // Make sure safepoint node for polling is distinct from a call's 1314 // return by adding a nop if needed. 1315 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) { 1316 padding = nop_size; 1317 } 1318 if (padding == 0 && mach->avoid_back_to_back() && 1319 current_offset == last_avoid_back_to_back_offset) { 1320 // Avoid back to back some instructions. 1321 padding = nop_size; 1322 } 1323 1324 if(padding > 0) { 1325 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1326 int nops_cnt = padding / nop_size; 1327 MachNode *nop = new (this) MachNopNode(nops_cnt); 1328 block->insert_node(nop, j++); 1329 last_inst++; 1330 _cfg->map_node_to_block(nop, block); 1331 nop->emit(*cb, _regalloc); 1332 cb->flush_bundle(true); 1333 current_offset = cb->insts_size(); 1334 } 1335 1336 // Remember the start of the last call in a basic block 1337 if (is_mcall) { 1338 MachCallNode *mcall = mach->as_MachCall(); 1339 1340 // This destination address is NOT PC-relative 1341 mcall->method_set((intptr_t)mcall->entry_point()); 1342 1343 // Save the return address 1344 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset(); 1345 1346 if (mcall->is_MachCallLeaf()) { 1347 is_mcall = false; 1348 is_sfn = false; 1349 } 1350 } 1351 1352 // sfn will be valid whenever mcall is valid now because of inheritance 1353 if (is_sfn || is_mcall) { 1354 1355 // Handle special safepoint nodes for synchronization 1356 if (!is_mcall) { 1357 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1358 // !!!!! Stubs only need an oopmap right now, so bail out 1359 if (sfn->jvms()->method() == NULL) { 1360 // Write the oopmap directly to the code blob??!! 1361 # ifdef ENABLE_ZAP_DEAD_LOCALS 1362 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive"); 1363 # endif 1364 continue; 1365 } 1366 } // End synchronization 1367 1368 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1369 current_offset); 1370 Process_OopMap_Node(mach, current_offset); 1371 } // End if safepoint 1372 1373 // If this is a null check, then add the start of the previous instruction to the list 1374 else if( mach->is_MachNullCheck() ) { 1375 inct_starts[inct_cnt++] = previous_offset; 1376 } 1377 1378 // If this is a branch, then fill in the label with the target BB's label 1379 else if (mach->is_MachBranch()) { 1380 // This requires the TRUE branch target be in succs[0] 1381 uint block_num = block->non_connector_successor(0)->_pre_order; 1382 1383 // Try to replace long branch if delay slot is not used, 1384 // it is mostly for back branches since forward branch's 1385 // distance is not updated yet. 1386 bool delay_slot_is_used = valid_bundle_info(n) && 1387 node_bundling(n)->use_unconditional_delay(); 1388 if (!delay_slot_is_used && mach->may_be_short_branch()) { 1389 assert(delay_slot == NULL, "not expecting delay slot node"); 1390 int br_size = n->size(_regalloc); 1391 int offset = blk_starts[block_num] - current_offset; 1392 if (block_num >= i) { 1393 // Current and following block's offset are not 1394 // finalized yet, adjust distance by the difference 1395 // between calculated and final offsets of current block. 1396 offset -= (blk_starts[i] - blk_offset); 1397 } 1398 // In the following code a nop could be inserted before 1399 // the branch which will increase the backward distance. 1400 bool needs_padding = (current_offset == last_avoid_back_to_back_offset); 1401 if (needs_padding && offset <= 0) 1402 offset -= nop_size; 1403 1404 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 1405 // We've got a winner. Replace this branch. 1406 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this); 1407 1408 // Update the jmp_size. 1409 int new_size = replacement->size(_regalloc); 1410 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller"); 1411 // Insert padding between avoid_back_to_back branches. 1412 if (needs_padding && replacement->avoid_back_to_back()) { 1413 MachNode *nop = new (this) MachNopNode(); 1414 block->insert_node(nop, j++); 1415 _cfg->map_node_to_block(nop, block); 1416 last_inst++; 1417 nop->emit(*cb, _regalloc); 1418 cb->flush_bundle(true); 1419 current_offset = cb->insts_size(); 1420 } 1421 #ifdef ASSERT 1422 jmp_target[i] = block_num; 1423 jmp_offset[i] = current_offset - blk_offset; 1424 jmp_size[i] = new_size; 1425 jmp_rule[i] = mach->rule(); 1426 #endif 1427 block->map_node(replacement, j); 1428 mach->subsume_by(replacement, C); 1429 n = replacement; 1430 mach = replacement; 1431 } 1432 } 1433 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num ); 1434 } else if (mach->ideal_Opcode() == Op_Jump) { 1435 for (uint h = 0; h < block->_num_succs; h++) { 1436 Block* succs_block = block->_succs[h]; 1437 for (uint j = 1; j < succs_block->num_preds(); j++) { 1438 Node* jpn = succs_block->pred(j); 1439 if (jpn->is_JumpProj() && jpn->in(0) == mach) { 1440 uint block_num = succs_block->non_connector()->_pre_order; 1441 Label *blkLabel = &blk_labels[block_num]; 1442 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1443 } 1444 } 1445 } 1446 } 1447 #ifdef ASSERT 1448 // Check that oop-store precedes the card-mark 1449 else if (mach->ideal_Opcode() == Op_StoreCM) { 1450 uint storeCM_idx = j; 1451 int count = 0; 1452 for (uint prec = mach->req(); prec < mach->len(); prec++) { 1453 Node *oop_store = mach->in(prec); // Precedence edge 1454 if (oop_store == NULL) continue; 1455 count++; 1456 uint i4; 1457 for (i4 = 0; i4 < last_inst; ++i4) { 1458 if (block->get_node(i4) == oop_store) { 1459 break; 1460 } 1461 } 1462 // Note: This test can provide a false failure if other precedence 1463 // edges have been added to the storeCMNode. 1464 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1465 } 1466 assert(count > 0, "storeCM expects at least one precedence edge"); 1467 } 1468 #endif 1469 else if (!n->is_Proj()) { 1470 // Remember the beginning of the previous instruction, in case 1471 // it's followed by a flag-kill and a null-check. Happens on 1472 // Intel all the time, with add-to-memory kind of opcodes. 1473 previous_offset = current_offset; 1474 } 1475 1476 // Not an else-if! 1477 // If this is a trap based cmp then add its offset to the list. 1478 if (mach->is_TrapBasedCheckNode()) { 1479 inct_starts[inct_cnt++] = current_offset; 1480 } 1481 } 1482 1483 // Verify that there is sufficient space remaining 1484 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1485 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1486 C->record_failure("CodeCache is full"); 1487 return; 1488 } 1489 1490 // Save the offset for the listing 1491 #ifndef PRODUCT 1492 if (node_offsets && n->_idx < node_offset_limit) 1493 node_offsets[n->_idx] = cb->insts_size(); 1494 #endif 1495 1496 // "Normal" instruction case 1497 DEBUG_ONLY( uint instr_offset = cb->insts_size(); ) 1498 n->emit(*cb, _regalloc); 1499 current_offset = cb->insts_size(); 1500 1501 #ifdef ASSERT 1502 if (n->size(_regalloc) < (current_offset-instr_offset)) { 1503 n->dump(); 1504 assert(false, "wrong size of mach node"); 1505 } 1506 #endif 1507 non_safepoints.observe_instruction(n, current_offset); 1508 1509 // mcall is last "call" that can be a safepoint 1510 // record it so we can see if a poll will directly follow it 1511 // in which case we'll need a pad to make the PcDesc sites unique 1512 // see 5010568. This can be slightly inaccurate but conservative 1513 // in the case that return address is not actually at current_offset. 1514 // This is a small price to pay. 1515 1516 if (is_mcall) { 1517 last_call_offset = current_offset; 1518 } 1519 1520 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back()) { 1521 // Avoid back to back some instructions. 1522 last_avoid_back_to_back_offset = current_offset; 1523 } 1524 1525 // See if this instruction has a delay slot 1526 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1527 assert(delay_slot != NULL, "expecting delay slot node"); 1528 1529 // Back up 1 instruction 1530 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size()); 1531 1532 // Save the offset for the listing 1533 #ifndef PRODUCT 1534 if (node_offsets && delay_slot->_idx < node_offset_limit) 1535 node_offsets[delay_slot->_idx] = cb->insts_size(); 1536 #endif 1537 1538 // Support a SafePoint in the delay slot 1539 if (delay_slot->is_MachSafePoint()) { 1540 MachNode *mach = delay_slot->as_Mach(); 1541 // !!!!! Stubs only need an oopmap right now, so bail out 1542 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) { 1543 // Write the oopmap directly to the code blob??!! 1544 # ifdef ENABLE_ZAP_DEAD_LOCALS 1545 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive"); 1546 # endif 1547 delay_slot = NULL; 1548 continue; 1549 } 1550 1551 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1552 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1553 adjusted_offset); 1554 // Generate an OopMap entry 1555 Process_OopMap_Node(mach, adjusted_offset); 1556 } 1557 1558 // Insert the delay slot instruction 1559 delay_slot->emit(*cb, _regalloc); 1560 1561 // Don't reuse it 1562 delay_slot = NULL; 1563 } 1564 1565 } // End for all instructions in block 1566 1567 // If the next block is the top of a loop, pad this block out to align 1568 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1569 if (i < nblocks-1) { 1570 Block *nb = _cfg->get_block(i + 1); 1571 int padding = nb->alignment_padding(current_offset); 1572 if( padding > 0 ) { 1573 MachNode *nop = new (this) MachNopNode(padding / nop_size); 1574 block->insert_node(nop, block->number_of_nodes()); 1575 _cfg->map_node_to_block(nop, block); 1576 nop->emit(*cb, _regalloc); 1577 current_offset = cb->insts_size(); 1578 } 1579 } 1580 // Verify that the distance for generated before forward 1581 // short branches is still valid. 1582 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size"); 1583 1584 // Save new block start offset 1585 blk_starts[i] = blk_offset; 1586 } // End of for all blocks 1587 blk_starts[nblocks] = current_offset; 1588 1589 non_safepoints.flush_at_end(); 1590 1591 // Offset too large? 1592 if (failing()) return; 1593 1594 // Define a pseudo-label at the end of the code 1595 MacroAssembler(cb).bind( blk_labels[nblocks] ); 1596 1597 // Compute the size of the first block 1598 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1599 1600 assert(cb->insts_size() < 500000, "method is unreasonably large"); 1601 1602 #ifdef ASSERT 1603 for (uint i = 0; i < nblocks; i++) { // For all blocks 1604 if (jmp_target[i] != 0) { 1605 int br_size = jmp_size[i]; 1606 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 1607 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 1608 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]); 1609 assert(false, "Displacement too large for short jmp"); 1610 } 1611 } 1612 } 1613 #endif 1614 1615 #ifndef PRODUCT 1616 // Information on the size of the method, without the extraneous code 1617 Scheduling::increment_method_size(cb->insts_size()); 1618 #endif 1619 1620 // ------------------ 1621 // Fill in exception table entries. 1622 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1623 1624 // Only java methods have exception handlers and deopt handlers 1625 if (_method) { 1626 // Emit the exception handler code. 1627 _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb)); 1628 // Emit the deopt handler code. 1629 _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb)); 1630 1631 // Emit the MethodHandle deopt handler code (if required). 1632 if (has_method_handle_invokes()) { 1633 // We can use the same code as for the normal deopt handler, we 1634 // just need a different entry point address. 1635 _code_offsets.set_value(CodeOffsets::DeoptMH, emit_deopt_handler(*cb)); 1636 } 1637 } 1638 1639 // One last check for failed CodeBuffer::expand: 1640 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1641 C->record_failure("CodeCache is full"); 1642 return; 1643 } 1644 1645 #ifndef PRODUCT 1646 // Dump the assembly code, including basic-block numbers 1647 if (print_assembly()) { 1648 ttyLocker ttyl; // keep the following output all in one block 1649 if (!VMThread::should_terminate()) { // test this under the tty lock 1650 // This output goes directly to the tty, not the compiler log. 1651 // To enable tools to match it up with the compilation activity, 1652 // be sure to tag this tty output with the compile ID. 1653 if (xtty != NULL) { 1654 xtty->head("opto_assembly compile_id='%d'%s", compile_id(), 1655 is_osr_compilation() ? " compile_kind='osr'" : 1656 ""); 1657 } 1658 if (method() != NULL) { 1659 method()->print_metadata(); 1660 } 1661 dump_asm(node_offsets, node_offset_limit); 1662 if (xtty != NULL) { 1663 xtty->tail("opto_assembly"); 1664 } 1665 } 1666 } 1667 #endif 1668 1669 } 1670 1671 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1672 _inc_table.set_size(cnt); 1673 1674 uint inct_cnt = 0; 1675 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 1676 Block* block = _cfg->get_block(i); 1677 Node *n = NULL; 1678 int j; 1679 1680 // Find the branch; ignore trailing NOPs. 1681 for (j = block->number_of_nodes() - 1; j >= 0; j--) { 1682 n = block->get_node(j); 1683 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) { 1684 break; 1685 } 1686 } 1687 1688 // If we didn't find anything, continue 1689 if (j < 0) { 1690 continue; 1691 } 1692 1693 // Compute ExceptionHandlerTable subtable entry and add it 1694 // (skip empty blocks) 1695 if (n->is_Catch()) { 1696 1697 // Get the offset of the return from the call 1698 uint call_return = call_returns[block->_pre_order]; 1699 #ifdef ASSERT 1700 assert( call_return > 0, "no call seen for this basic block" ); 1701 while (block->get_node(--j)->is_MachProj()) ; 1702 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call"); 1703 #endif 1704 // last instruction is a CatchNode, find it's CatchProjNodes 1705 int nof_succs = block->_num_succs; 1706 // allocate space 1707 GrowableArray<intptr_t> handler_bcis(nof_succs); 1708 GrowableArray<intptr_t> handler_pcos(nof_succs); 1709 // iterate through all successors 1710 for (int j = 0; j < nof_succs; j++) { 1711 Block* s = block->_succs[j]; 1712 bool found_p = false; 1713 for (uint k = 1; k < s->num_preds(); k++) { 1714 Node* pk = s->pred(k); 1715 if (pk->is_CatchProj() && pk->in(0) == n) { 1716 const CatchProjNode* p = pk->as_CatchProj(); 1717 found_p = true; 1718 // add the corresponding handler bci & pco information 1719 if (p->_con != CatchProjNode::fall_through_index) { 1720 // p leads to an exception handler (and is not fall through) 1721 assert(s == _cfg->get_block(s->_pre_order), "bad numbering"); 1722 // no duplicates, please 1723 if (!handler_bcis.contains(p->handler_bci())) { 1724 uint block_num = s->non_connector()->_pre_order; 1725 handler_bcis.append(p->handler_bci()); 1726 handler_pcos.append(blk_labels[block_num].loc_pos()); 1727 } 1728 } 1729 } 1730 } 1731 assert(found_p, "no matching predecessor found"); 1732 // Note: Due to empty block removal, one block may have 1733 // several CatchProj inputs, from the same Catch. 1734 } 1735 1736 // Set the offset of the return from the call 1737 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); 1738 continue; 1739 } 1740 1741 // Handle implicit null exception table updates 1742 if (n->is_MachNullCheck()) { 1743 uint block_num = block->non_connector_successor(0)->_pre_order; 1744 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1745 continue; 1746 } 1747 // Handle implicit exception table updates: trap instructions. 1748 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) { 1749 uint block_num = block->non_connector_successor(0)->_pre_order; 1750 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1751 continue; 1752 } 1753 } // End of for all blocks fill in exception table entries 1754 } 1755 1756 // Static Variables 1757 #ifndef PRODUCT 1758 uint Scheduling::_total_nop_size = 0; 1759 uint Scheduling::_total_method_size = 0; 1760 uint Scheduling::_total_branches = 0; 1761 uint Scheduling::_total_unconditional_delays = 0; 1762 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1763 #endif 1764 1765 // Initializer for class Scheduling 1766 1767 Scheduling::Scheduling(Arena *arena, Compile &compile) 1768 : _arena(arena), 1769 _cfg(compile.cfg()), 1770 _regalloc(compile.regalloc()), 1771 _reg_node(arena), 1772 _bundle_instr_count(0), 1773 _bundle_cycle_number(0), 1774 _scheduled(arena), 1775 _available(arena), 1776 _next_node(NULL), 1777 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]), 1778 _pinch_free_list(arena) 1779 #ifndef PRODUCT 1780 , _branches(0) 1781 , _unconditional_delays(0) 1782 #endif 1783 { 1784 // Create a MachNopNode 1785 _nop = new (&compile) MachNopNode(); 1786 1787 // Now that the nops are in the array, save the count 1788 // (but allow entries for the nops) 1789 _node_bundling_limit = compile.unique(); 1790 uint node_max = _regalloc->node_regs_max_index(); 1791 1792 compile.set_node_bundling_limit(_node_bundling_limit); 1793 1794 // This one is persistent within the Compile class 1795 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1796 1797 // Allocate space for fixed-size arrays 1798 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1799 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1800 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1801 1802 // Clear the arrays 1803 memset(_node_bundling_base, 0, node_max * sizeof(Bundle)); 1804 memset(_node_latency, 0, node_max * sizeof(unsigned short)); 1805 memset(_uses, 0, node_max * sizeof(short)); 1806 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 1807 1808 // Clear the bundling information 1809 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements)); 1810 1811 // Get the last node 1812 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1); 1813 1814 _next_node = block->get_node(block->number_of_nodes() - 1); 1815 } 1816 1817 #ifndef PRODUCT 1818 // Scheduling destructor 1819 Scheduling::~Scheduling() { 1820 _total_branches += _branches; 1821 _total_unconditional_delays += _unconditional_delays; 1822 } 1823 #endif 1824 1825 // Step ahead "i" cycles 1826 void Scheduling::step(uint i) { 1827 1828 Bundle *bundle = node_bundling(_next_node); 1829 bundle->set_starts_bundle(); 1830 1831 // Update the bundle record, but leave the flags information alone 1832 if (_bundle_instr_count > 0) { 1833 bundle->set_instr_count(_bundle_instr_count); 1834 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1835 } 1836 1837 // Update the state information 1838 _bundle_instr_count = 0; 1839 _bundle_cycle_number += i; 1840 _bundle_use.step(i); 1841 } 1842 1843 void Scheduling::step_and_clear() { 1844 Bundle *bundle = node_bundling(_next_node); 1845 bundle->set_starts_bundle(); 1846 1847 // Update the bundle record 1848 if (_bundle_instr_count > 0) { 1849 bundle->set_instr_count(_bundle_instr_count); 1850 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1851 1852 _bundle_cycle_number += 1; 1853 } 1854 1855 // Clear the bundling information 1856 _bundle_instr_count = 0; 1857 _bundle_use.reset(); 1858 1859 memcpy(_bundle_use_elements, 1860 Pipeline_Use::elaborated_elements, 1861 sizeof(Pipeline_Use::elaborated_elements)); 1862 } 1863 1864 // Perform instruction scheduling and bundling over the sequence of 1865 // instructions in backwards order. 1866 void Compile::ScheduleAndBundle() { 1867 1868 // Don't optimize this if it isn't a method 1869 if (!_method) 1870 return; 1871 1872 // Don't optimize this if scheduling is disabled 1873 if (!do_scheduling()) 1874 return; 1875 1876 // Scheduling code works only with pairs (8 bytes) maximum. 1877 if (max_vector_size() > 8) 1878 return; 1879 1880 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); ) 1881 1882 // Create a data structure for all the scheduling information 1883 Scheduling scheduling(Thread::current()->resource_area(), *this); 1884 1885 // Walk backwards over each basic block, computing the needed alignment 1886 // Walk over all the basic blocks 1887 scheduling.DoScheduling(); 1888 } 1889 1890 // Compute the latency of all the instructions. This is fairly simple, 1891 // because we already have a legal ordering. Walk over the instructions 1892 // from first to last, and compute the latency of the instruction based 1893 // on the latency of the preceding instruction(s). 1894 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { 1895 #ifndef PRODUCT 1896 if (_cfg->C->trace_opto_output()) 1897 tty->print("# -> ComputeLocalLatenciesForward\n"); 1898 #endif 1899 1900 // Walk over all the schedulable instructions 1901 for( uint j=_bb_start; j < _bb_end; j++ ) { 1902 1903 // This is a kludge, forcing all latency calculations to start at 1. 1904 // Used to allow latency 0 to force an instruction to the beginning 1905 // of the bb 1906 uint latency = 1; 1907 Node *use = bb->get_node(j); 1908 uint nlen = use->len(); 1909 1910 // Walk over all the inputs 1911 for ( uint k=0; k < nlen; k++ ) { 1912 Node *def = use->in(k); 1913 if (!def) 1914 continue; 1915 1916 uint l = _node_latency[def->_idx] + use->latency(k); 1917 if (latency < l) 1918 latency = l; 1919 } 1920 1921 _node_latency[use->_idx] = latency; 1922 1923 #ifndef PRODUCT 1924 if (_cfg->C->trace_opto_output()) { 1925 tty->print("# latency %4d: ", latency); 1926 use->dump(); 1927 } 1928 #endif 1929 } 1930 1931 #ifndef PRODUCT 1932 if (_cfg->C->trace_opto_output()) 1933 tty->print("# <- ComputeLocalLatenciesForward\n"); 1934 #endif 1935 1936 } // end ComputeLocalLatenciesForward 1937 1938 // See if this node fits into the present instruction bundle 1939 bool Scheduling::NodeFitsInBundle(Node *n) { 1940 uint n_idx = n->_idx; 1941 1942 // If this is the unconditional delay instruction, then it fits 1943 if (n == _unconditional_delay_slot) { 1944 #ifndef PRODUCT 1945 if (_cfg->C->trace_opto_output()) 1946 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 1947 #endif 1948 return (true); 1949 } 1950 1951 // If the node cannot be scheduled this cycle, skip it 1952 if (_current_latency[n_idx] > _bundle_cycle_number) { 1953 #ifndef PRODUCT 1954 if (_cfg->C->trace_opto_output()) 1955 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 1956 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 1957 #endif 1958 return (false); 1959 } 1960 1961 const Pipeline *node_pipeline = n->pipeline(); 1962 1963 uint instruction_count = node_pipeline->instructionCount(); 1964 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1965 instruction_count = 0; 1966 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1967 instruction_count++; 1968 1969 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 1970 #ifndef PRODUCT 1971 if (_cfg->C->trace_opto_output()) 1972 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 1973 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 1974 #endif 1975 return (false); 1976 } 1977 1978 // Don't allow non-machine nodes to be handled this way 1979 if (!n->is_Mach() && instruction_count == 0) 1980 return (false); 1981 1982 // See if there is any overlap 1983 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 1984 1985 if (delay > 0) { 1986 #ifndef PRODUCT 1987 if (_cfg->C->trace_opto_output()) 1988 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 1989 #endif 1990 return false; 1991 } 1992 1993 #ifndef PRODUCT 1994 if (_cfg->C->trace_opto_output()) 1995 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 1996 #endif 1997 1998 return true; 1999 } 2000 2001 Node * Scheduling::ChooseNodeToBundle() { 2002 uint siz = _available.size(); 2003 2004 if (siz == 0) { 2005 2006 #ifndef PRODUCT 2007 if (_cfg->C->trace_opto_output()) 2008 tty->print("# ChooseNodeToBundle: NULL\n"); 2009 #endif 2010 return (NULL); 2011 } 2012 2013 // Fast path, if only 1 instruction in the bundle 2014 if (siz == 1) { 2015 #ifndef PRODUCT 2016 if (_cfg->C->trace_opto_output()) { 2017 tty->print("# ChooseNodeToBundle (only 1): "); 2018 _available[0]->dump(); 2019 } 2020 #endif 2021 return (_available[0]); 2022 } 2023 2024 // Don't bother, if the bundle is already full 2025 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 2026 for ( uint i = 0; i < siz; i++ ) { 2027 Node *n = _available[i]; 2028 2029 // Skip projections, we'll handle them another way 2030 if (n->is_Proj()) 2031 continue; 2032 2033 // This presupposed that instructions are inserted into the 2034 // available list in a legality order; i.e. instructions that 2035 // must be inserted first are at the head of the list 2036 if (NodeFitsInBundle(n)) { 2037 #ifndef PRODUCT 2038 if (_cfg->C->trace_opto_output()) { 2039 tty->print("# ChooseNodeToBundle: "); 2040 n->dump(); 2041 } 2042 #endif 2043 return (n); 2044 } 2045 } 2046 } 2047 2048 // Nothing fits in this bundle, choose the highest priority 2049 #ifndef PRODUCT 2050 if (_cfg->C->trace_opto_output()) { 2051 tty->print("# ChooseNodeToBundle: "); 2052 _available[0]->dump(); 2053 } 2054 #endif 2055 2056 return _available[0]; 2057 } 2058 2059 void Scheduling::AddNodeToAvailableList(Node *n) { 2060 assert( !n->is_Proj(), "projections never directly made available" ); 2061 #ifndef PRODUCT 2062 if (_cfg->C->trace_opto_output()) { 2063 tty->print("# AddNodeToAvailableList: "); 2064 n->dump(); 2065 } 2066 #endif 2067 2068 int latency = _current_latency[n->_idx]; 2069 2070 // Insert in latency order (insertion sort) 2071 uint i; 2072 for ( i=0; i < _available.size(); i++ ) 2073 if (_current_latency[_available[i]->_idx] > latency) 2074 break; 2075 2076 // Special Check for compares following branches 2077 if( n->is_Mach() && _scheduled.size() > 0 ) { 2078 int op = n->as_Mach()->ideal_Opcode(); 2079 Node *last = _scheduled[0]; 2080 if( last->is_MachIf() && last->in(1) == n && 2081 ( op == Op_CmpI || 2082 op == Op_CmpU || 2083 op == Op_CmpP || 2084 op == Op_CmpF || 2085 op == Op_CmpD || 2086 op == Op_CmpL ) ) { 2087 2088 // Recalculate position, moving to front of same latency 2089 for ( i=0 ; i < _available.size(); i++ ) 2090 if (_current_latency[_available[i]->_idx] >= latency) 2091 break; 2092 } 2093 } 2094 2095 // Insert the node in the available list 2096 _available.insert(i, n); 2097 2098 #ifndef PRODUCT 2099 if (_cfg->C->trace_opto_output()) 2100 dump_available(); 2101 #endif 2102 } 2103 2104 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 2105 for ( uint i=0; i < n->len(); i++ ) { 2106 Node *def = n->in(i); 2107 if (!def) continue; 2108 if( def->is_Proj() ) // If this is a machine projection, then 2109 def = def->in(0); // propagate usage thru to the base instruction 2110 2111 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local 2112 continue; 2113 } 2114 2115 // Compute the latency 2116 uint l = _bundle_cycle_number + n->latency(i); 2117 if (_current_latency[def->_idx] < l) 2118 _current_latency[def->_idx] = l; 2119 2120 // If this does not have uses then schedule it 2121 if ((--_uses[def->_idx]) == 0) 2122 AddNodeToAvailableList(def); 2123 } 2124 } 2125 2126 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 2127 #ifndef PRODUCT 2128 if (_cfg->C->trace_opto_output()) { 2129 tty->print("# AddNodeToBundle: "); 2130 n->dump(); 2131 } 2132 #endif 2133 2134 // Remove this from the available list 2135 uint i; 2136 for (i = 0; i < _available.size(); i++) 2137 if (_available[i] == n) 2138 break; 2139 assert(i < _available.size(), "entry in _available list not found"); 2140 _available.remove(i); 2141 2142 // See if this fits in the current bundle 2143 const Pipeline *node_pipeline = n->pipeline(); 2144 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 2145 2146 // Check for instructions to be placed in the delay slot. We 2147 // do this before we actually schedule the current instruction, 2148 // because the delay slot follows the current instruction. 2149 if (Pipeline::_branch_has_delay_slot && 2150 node_pipeline->hasBranchDelay() && 2151 !_unconditional_delay_slot) { 2152 2153 uint siz = _available.size(); 2154 2155 // Conditional branches can support an instruction that 2156 // is unconditionally executed and not dependent by the 2157 // branch, OR a conditionally executed instruction if 2158 // the branch is taken. In practice, this means that 2159 // the first instruction at the branch target is 2160 // copied to the delay slot, and the branch goes to 2161 // the instruction after that at the branch target 2162 if ( n->is_MachBranch() ) { 2163 2164 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 2165 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 2166 2167 #ifndef PRODUCT 2168 _branches++; 2169 #endif 2170 2171 // At least 1 instruction is on the available list 2172 // that is not dependent on the branch 2173 for (uint i = 0; i < siz; i++) { 2174 Node *d = _available[i]; 2175 const Pipeline *avail_pipeline = d->pipeline(); 2176 2177 // Don't allow safepoints in the branch shadow, that will 2178 // cause a number of difficulties 2179 if ( avail_pipeline->instructionCount() == 1 && 2180 !avail_pipeline->hasMultipleBundles() && 2181 !avail_pipeline->hasBranchDelay() && 2182 Pipeline::instr_has_unit_size() && 2183 d->size(_regalloc) == Pipeline::instr_unit_size() && 2184 NodeFitsInBundle(d) && 2185 !node_bundling(d)->used_in_delay()) { 2186 2187 if (d->is_Mach() && !d->is_MachSafePoint()) { 2188 // A node that fits in the delay slot was found, so we need to 2189 // set the appropriate bits in the bundle pipeline information so 2190 // that it correctly indicates resource usage. Later, when we 2191 // attempt to add this instruction to the bundle, we will skip 2192 // setting the resource usage. 2193 _unconditional_delay_slot = d; 2194 node_bundling(n)->set_use_unconditional_delay(); 2195 node_bundling(d)->set_used_in_unconditional_delay(); 2196 _bundle_use.add_usage(avail_pipeline->resourceUse()); 2197 _current_latency[d->_idx] = _bundle_cycle_number; 2198 _next_node = d; 2199 ++_bundle_instr_count; 2200 #ifndef PRODUCT 2201 _unconditional_delays++; 2202 #endif 2203 break; 2204 } 2205 } 2206 } 2207 } 2208 2209 // No delay slot, add a nop to the usage 2210 if (!_unconditional_delay_slot) { 2211 // See if adding an instruction in the delay slot will overflow 2212 // the bundle. 2213 if (!NodeFitsInBundle(_nop)) { 2214 #ifndef PRODUCT 2215 if (_cfg->C->trace_opto_output()) 2216 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2217 #endif 2218 step(1); 2219 } 2220 2221 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2222 _next_node = _nop; 2223 ++_bundle_instr_count; 2224 } 2225 2226 // See if the instruction in the delay slot requires a 2227 // step of the bundles 2228 if (!NodeFitsInBundle(n)) { 2229 #ifndef PRODUCT 2230 if (_cfg->C->trace_opto_output()) 2231 tty->print("# *** STEP(branch won't fit) ***\n"); 2232 #endif 2233 // Update the state information 2234 _bundle_instr_count = 0; 2235 _bundle_cycle_number += 1; 2236 _bundle_use.step(1); 2237 } 2238 } 2239 2240 // Get the number of instructions 2241 uint instruction_count = node_pipeline->instructionCount(); 2242 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2243 instruction_count = 0; 2244 2245 // Compute the latency information 2246 uint delay = 0; 2247 2248 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2249 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2250 if (relative_latency < 0) 2251 relative_latency = 0; 2252 2253 delay = _bundle_use.full_latency(relative_latency, node_usage); 2254 2255 // Does not fit in this bundle, start a new one 2256 if (delay > 0) { 2257 step(delay); 2258 2259 #ifndef PRODUCT 2260 if (_cfg->C->trace_opto_output()) 2261 tty->print("# *** STEP(%d) ***\n", delay); 2262 #endif 2263 } 2264 } 2265 2266 // If this was placed in the delay slot, ignore it 2267 if (n != _unconditional_delay_slot) { 2268 2269 if (delay == 0) { 2270 if (node_pipeline->hasMultipleBundles()) { 2271 #ifndef PRODUCT 2272 if (_cfg->C->trace_opto_output()) 2273 tty->print("# *** STEP(multiple instructions) ***\n"); 2274 #endif 2275 step(1); 2276 } 2277 2278 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2279 #ifndef PRODUCT 2280 if (_cfg->C->trace_opto_output()) 2281 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2282 instruction_count + _bundle_instr_count, 2283 Pipeline::_max_instrs_per_cycle); 2284 #endif 2285 step(1); 2286 } 2287 } 2288 2289 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2290 _bundle_instr_count++; 2291 2292 // Set the node's latency 2293 _current_latency[n->_idx] = _bundle_cycle_number; 2294 2295 // Now merge the functional unit information 2296 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2297 _bundle_use.add_usage(node_usage); 2298 2299 // Increment the number of instructions in this bundle 2300 _bundle_instr_count += instruction_count; 2301 2302 // Remember this node for later 2303 if (n->is_Mach()) 2304 _next_node = n; 2305 } 2306 2307 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2308 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2309 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2310 // into the block. All other scheduled nodes get put in the schedule here. 2311 int op = n->Opcode(); 2312 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2313 (op != Op_Node && // Not an unused antidepedence node and 2314 // not an unallocated boxlock 2315 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2316 2317 // Push any trailing projections 2318 if( bb->get_node(bb->number_of_nodes()-1) != n ) { 2319 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2320 Node *foi = n->fast_out(i); 2321 if( foi->is_Proj() ) 2322 _scheduled.push(foi); 2323 } 2324 } 2325 2326 // Put the instruction in the schedule list 2327 _scheduled.push(n); 2328 } 2329 2330 #ifndef PRODUCT 2331 if (_cfg->C->trace_opto_output()) 2332 dump_available(); 2333 #endif 2334 2335 // Walk all the definitions, decrementing use counts, and 2336 // if a definition has a 0 use count, place it in the available list. 2337 DecrementUseCounts(n,bb); 2338 } 2339 2340 // This method sets the use count within a basic block. We will ignore all 2341 // uses outside the current basic block. As we are doing a backwards walk, 2342 // any node we reach that has a use count of 0 may be scheduled. This also 2343 // avoids the problem of cyclic references from phi nodes, as long as phi 2344 // nodes are at the front of the basic block. This method also initializes 2345 // the available list to the set of instructions that have no uses within this 2346 // basic block. 2347 void Scheduling::ComputeUseCount(const Block *bb) { 2348 #ifndef PRODUCT 2349 if (_cfg->C->trace_opto_output()) 2350 tty->print("# -> ComputeUseCount\n"); 2351 #endif 2352 2353 // Clear the list of available and scheduled instructions, just in case 2354 _available.clear(); 2355 _scheduled.clear(); 2356 2357 // No delay slot specified 2358 _unconditional_delay_slot = NULL; 2359 2360 #ifdef ASSERT 2361 for( uint i=0; i < bb->number_of_nodes(); i++ ) 2362 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" ); 2363 #endif 2364 2365 // Force the _uses count to never go to zero for unscheduable pieces 2366 // of the block 2367 for( uint k = 0; k < _bb_start; k++ ) 2368 _uses[bb->get_node(k)->_idx] = 1; 2369 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ ) 2370 _uses[bb->get_node(l)->_idx] = 1; 2371 2372 // Iterate backwards over the instructions in the block. Don't count the 2373 // branch projections at end or the block header instructions. 2374 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2375 Node *n = bb->get_node(j); 2376 if( n->is_Proj() ) continue; // Projections handled another way 2377 2378 // Account for all uses 2379 for ( uint k = 0; k < n->len(); k++ ) { 2380 Node *inp = n->in(k); 2381 if (!inp) continue; 2382 assert(inp != n, "no cycles allowed" ); 2383 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use? 2384 if (inp->is_Proj()) { // Skip through Proj's 2385 inp = inp->in(0); 2386 } 2387 ++_uses[inp->_idx]; // Count 1 block-local use 2388 } 2389 } 2390 2391 // If this instruction has a 0 use count, then it is available 2392 if (!_uses[n->_idx]) { 2393 _current_latency[n->_idx] = _bundle_cycle_number; 2394 AddNodeToAvailableList(n); 2395 } 2396 2397 #ifndef PRODUCT 2398 if (_cfg->C->trace_opto_output()) { 2399 tty->print("# uses: %3d: ", _uses[n->_idx]); 2400 n->dump(); 2401 } 2402 #endif 2403 } 2404 2405 #ifndef PRODUCT 2406 if (_cfg->C->trace_opto_output()) 2407 tty->print("# <- ComputeUseCount\n"); 2408 #endif 2409 } 2410 2411 // This routine performs scheduling on each basic block in reverse order, 2412 // using instruction latencies and taking into account function unit 2413 // availability. 2414 void Scheduling::DoScheduling() { 2415 #ifndef PRODUCT 2416 if (_cfg->C->trace_opto_output()) 2417 tty->print("# -> DoScheduling\n"); 2418 #endif 2419 2420 Block *succ_bb = NULL; 2421 Block *bb; 2422 2423 // Walk over all the basic blocks in reverse order 2424 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) { 2425 bb = _cfg->get_block(i); 2426 2427 #ifndef PRODUCT 2428 if (_cfg->C->trace_opto_output()) { 2429 tty->print("# Schedule BB#%03d (initial)\n", i); 2430 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2431 bb->get_node(j)->dump(); 2432 } 2433 } 2434 #endif 2435 2436 // On the head node, skip processing 2437 if (bb == _cfg->get_root_block()) { 2438 continue; 2439 } 2440 2441 // Skip empty, connector blocks 2442 if (bb->is_connector()) 2443 continue; 2444 2445 // If the following block is not the sole successor of 2446 // this one, then reset the pipeline information 2447 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2448 #ifndef PRODUCT 2449 if (_cfg->C->trace_opto_output()) { 2450 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2451 _next_node->_idx, _bundle_instr_count); 2452 } 2453 #endif 2454 step_and_clear(); 2455 } 2456 2457 // Leave untouched the starting instruction, any Phis, a CreateEx node 2458 // or Top. bb->get_node(_bb_start) is the first schedulable instruction. 2459 _bb_end = bb->number_of_nodes()-1; 2460 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2461 Node *n = bb->get_node(_bb_start); 2462 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2463 // Also, MachIdealNodes do not get scheduled 2464 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2465 MachNode *mach = n->as_Mach(); 2466 int iop = mach->ideal_Opcode(); 2467 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2468 if( iop == Op_Con ) continue; // Do not schedule Top 2469 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2470 mach->pipeline() == MachNode::pipeline_class() && 2471 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc 2472 continue; 2473 break; // Funny loop structure to be sure... 2474 } 2475 // Compute last "interesting" instruction in block - last instruction we 2476 // might schedule. _bb_end points just after last schedulable inst. We 2477 // normally schedule conditional branches (despite them being forced last 2478 // in the block), because they have delay slots we can fill. Calls all 2479 // have their delay slots filled in the template expansions, so we don't 2480 // bother scheduling them. 2481 Node *last = bb->get_node(_bb_end); 2482 // Ignore trailing NOPs. 2483 while (_bb_end > 0 && last->is_Mach() && 2484 last->as_Mach()->ideal_Opcode() == Op_Con) { 2485 last = bb->get_node(--_bb_end); 2486 } 2487 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, ""); 2488 if( last->is_Catch() || 2489 // Exclude unreachable path case when Halt node is in a separate block. 2490 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2491 // There must be a prior call. Skip it. 2492 while( !bb->get_node(--_bb_end)->is_MachCall() ) { 2493 assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" ); 2494 } 2495 } else if( last->is_MachNullCheck() ) { 2496 // Backup so the last null-checked memory instruction is 2497 // outside the schedulable range. Skip over the nullcheck, 2498 // projection, and the memory nodes. 2499 Node *mem = last->in(1); 2500 do { 2501 _bb_end--; 2502 } while (mem != bb->get_node(_bb_end)); 2503 } else { 2504 // Set _bb_end to point after last schedulable inst. 2505 _bb_end++; 2506 } 2507 2508 assert( _bb_start <= _bb_end, "inverted block ends" ); 2509 2510 // Compute the register antidependencies for the basic block 2511 ComputeRegisterAntidependencies(bb); 2512 if (_cfg->C->failing()) return; // too many D-U pinch points 2513 2514 // Compute intra-bb latencies for the nodes 2515 ComputeLocalLatenciesForward(bb); 2516 2517 // Compute the usage within the block, and set the list of all nodes 2518 // in the block that have no uses within the block. 2519 ComputeUseCount(bb); 2520 2521 // Schedule the remaining instructions in the block 2522 while ( _available.size() > 0 ) { 2523 Node *n = ChooseNodeToBundle(); 2524 guarantee(n != NULL, "no nodes available"); 2525 AddNodeToBundle(n,bb); 2526 } 2527 2528 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2529 #ifdef ASSERT 2530 for( uint l = _bb_start; l < _bb_end; l++ ) { 2531 Node *n = bb->get_node(l); 2532 uint m; 2533 for( m = 0; m < _bb_end-_bb_start; m++ ) 2534 if( _scheduled[m] == n ) 2535 break; 2536 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2537 } 2538 #endif 2539 2540 // Now copy the instructions (in reverse order) back to the block 2541 for ( uint k = _bb_start; k < _bb_end; k++ ) 2542 bb->map_node(_scheduled[_bb_end-k-1], k); 2543 2544 #ifndef PRODUCT 2545 if (_cfg->C->trace_opto_output()) { 2546 tty->print("# Schedule BB#%03d (final)\n", i); 2547 uint current = 0; 2548 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2549 Node *n = bb->get_node(j); 2550 if( valid_bundle_info(n) ) { 2551 Bundle *bundle = node_bundling(n); 2552 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2553 tty->print("*** Bundle: "); 2554 bundle->dump(); 2555 } 2556 n->dump(); 2557 } 2558 } 2559 } 2560 #endif 2561 #ifdef ASSERT 2562 verify_good_schedule(bb,"after block local scheduling"); 2563 #endif 2564 } 2565 2566 #ifndef PRODUCT 2567 if (_cfg->C->trace_opto_output()) 2568 tty->print("# <- DoScheduling\n"); 2569 #endif 2570 2571 // Record final node-bundling array location 2572 _regalloc->C->set_node_bundling_base(_node_bundling_base); 2573 2574 } // end DoScheduling 2575 2576 // Verify that no live-range used in the block is killed in the block by a 2577 // wrong DEF. This doesn't verify live-ranges that span blocks. 2578 2579 // Check for edge existence. Used to avoid adding redundant precedence edges. 2580 static bool edge_from_to( Node *from, Node *to ) { 2581 for( uint i=0; i<from->len(); i++ ) 2582 if( from->in(i) == to ) 2583 return true; 2584 return false; 2585 } 2586 2587 #ifdef ASSERT 2588 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2589 // Check for bad kills 2590 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2591 Node *prior_use = _reg_node[def]; 2592 if( prior_use && !edge_from_to(prior_use,n) ) { 2593 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2594 n->dump(); 2595 tty->print_cr("..."); 2596 prior_use->dump(); 2597 assert(edge_from_to(prior_use,n),msg); 2598 } 2599 _reg_node.map(def,NULL); // Kill live USEs 2600 } 2601 } 2602 2603 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2604 2605 // Zap to something reasonable for the verify code 2606 _reg_node.clear(); 2607 2608 // Walk over the block backwards. Check to make sure each DEF doesn't 2609 // kill a live value (other than the one it's supposed to). Add each 2610 // USE to the live set. 2611 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) { 2612 Node *n = b->get_node(i); 2613 int n_op = n->Opcode(); 2614 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2615 // Fat-proj kills a slew of registers 2616 RegMask rm = n->out_RegMask();// Make local copy 2617 while( rm.is_NotEmpty() ) { 2618 OptoReg::Name kill = rm.find_first_elem(); 2619 rm.Remove(kill); 2620 verify_do_def( n, kill, msg ); 2621 } 2622 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2623 // Get DEF'd registers the normal way 2624 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2625 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2626 } 2627 2628 // Now make all USEs live 2629 for( uint i=1; i<n->req(); i++ ) { 2630 Node *def = n->in(i); 2631 assert(def != 0, "input edge required"); 2632 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2633 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2634 if( OptoReg::is_valid(reg_lo) ) { 2635 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg); 2636 _reg_node.map(reg_lo,n); 2637 } 2638 if( OptoReg::is_valid(reg_hi) ) { 2639 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg); 2640 _reg_node.map(reg_hi,n); 2641 } 2642 } 2643 2644 } 2645 2646 // Zap to something reasonable for the Antidependence code 2647 _reg_node.clear(); 2648 } 2649 #endif 2650 2651 // Conditionally add precedence edges. Avoid putting edges on Projs. 2652 static void add_prec_edge_from_to( Node *from, Node *to ) { 2653 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2654 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2655 from = from->in(0); 2656 } 2657 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2658 !edge_from_to( from, to ) ) // Avoid duplicate edge 2659 from->add_prec(to); 2660 } 2661 2662 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2663 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2664 return; 2665 2666 Node *pinch = _reg_node[def_reg]; // Get pinch point 2667 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet? 2668 is_def ) { // Check for a true def (not a kill) 2669 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2670 return; 2671 } 2672 2673 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2674 debug_only( def = (Node*)0xdeadbeef; ) 2675 2676 // After some number of kills there _may_ be a later def 2677 Node *later_def = NULL; 2678 2679 // Finding a kill requires a real pinch-point. 2680 // Check for not already having a pinch-point. 2681 // Pinch points are Op_Node's. 2682 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2683 later_def = pinch; // Must be def/kill as optimistic pinch-point 2684 if ( _pinch_free_list.size() > 0) { 2685 pinch = _pinch_free_list.pop(); 2686 } else { 2687 pinch = new (_cfg->C) Node(1); // Pinch point to-be 2688 } 2689 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2690 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2691 return; 2692 } 2693 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init) 2694 _reg_node.map(def_reg,pinch); // Record pinch-point 2695 //_regalloc->set_bad(pinch->_idx); // Already initialized this way. 2696 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2697 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call 2698 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2699 later_def = NULL; // and no later def 2700 } 2701 pinch->set_req(0,later_def); // Hook later def so we can find it 2702 } else { // Else have valid pinch point 2703 if( pinch->in(0) ) // If there is a later-def 2704 later_def = pinch->in(0); // Get it 2705 } 2706 2707 // Add output-dependence edge from later def to kill 2708 if( later_def ) // If there is some original def 2709 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2710 2711 // See if current kill is also a use, and so is forced to be the pinch-point. 2712 if( pinch->Opcode() == Op_Node ) { 2713 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2714 for( uint i=1; i<uses->req(); i++ ) { 2715 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2716 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2717 // Yes, found a use/kill pinch-point 2718 pinch->set_req(0,NULL); // 2719 pinch->replace_by(kill); // Move anti-dep edges up 2720 pinch = kill; 2721 _reg_node.map(def_reg,pinch); 2722 return; 2723 } 2724 } 2725 } 2726 2727 // Add edge from kill to pinch-point 2728 add_prec_edge_from_to(kill,pinch); 2729 } 2730 2731 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2732 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2733 return; 2734 Node *pinch = _reg_node[use_reg]; // Get pinch point 2735 // Check for no later def_reg/kill in block 2736 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b && 2737 // Use has to be block-local as well 2738 _cfg->get_block_for_node(use) == b) { 2739 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2740 pinch->req() == 1 ) { // pinch not yet in block? 2741 pinch->del_req(0); // yank pointer to later-def, also set flag 2742 // Insert the pinch-point in the block just after the last use 2743 b->insert_node(pinch, b->find_node(use) + 1); 2744 _bb_end++; // Increase size scheduled region in block 2745 } 2746 2747 add_prec_edge_from_to(pinch,use); 2748 } 2749 } 2750 2751 // We insert antidependences between the reads and following write of 2752 // allocated registers to prevent illegal code motion. Hopefully, the 2753 // number of added references should be fairly small, especially as we 2754 // are only adding references within the current basic block. 2755 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2756 2757 #ifdef ASSERT 2758 verify_good_schedule(b,"before block local scheduling"); 2759 #endif 2760 2761 // A valid schedule, for each register independently, is an endless cycle 2762 // of: a def, then some uses (connected to the def by true dependencies), 2763 // then some kills (defs with no uses), finally the cycle repeats with a new 2764 // def. The uses are allowed to float relative to each other, as are the 2765 // kills. No use is allowed to slide past a kill (or def). This requires 2766 // antidependencies between all uses of a single def and all kills that 2767 // follow, up to the next def. More edges are redundant, because later defs 2768 // & kills are already serialized with true or antidependencies. To keep 2769 // the edge count down, we add a 'pinch point' node if there's more than 2770 // one use or more than one kill/def. 2771 2772 // We add dependencies in one bottom-up pass. 2773 2774 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2775 2776 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2777 // register. If not, we record the DEF/KILL in _reg_node, the 2778 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2779 // "pinch point", a new Node that's in the graph but not in the block. 2780 // We put edges from the prior and current DEF/KILLs to the pinch point. 2781 // We put the pinch point in _reg_node. If there's already a pinch point 2782 // we merely add an edge from the current DEF/KILL to the pinch point. 2783 2784 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2785 // put an edge from the pinch point to the USE. 2786 2787 // To be expedient, the _reg_node array is pre-allocated for the whole 2788 // compilation. _reg_node is lazily initialized; it either contains a NULL, 2789 // or a valid def/kill/pinch-point, or a leftover node from some prior 2790 // block. Leftover node from some prior block is treated like a NULL (no 2791 // prior def, so no anti-dependence needed). Valid def is distinguished by 2792 // it being in the current block. 2793 bool fat_proj_seen = false; 2794 uint last_safept = _bb_end-1; 2795 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL; 2796 Node* last_safept_node = end_node; 2797 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2798 Node *n = b->get_node(i); 2799 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2800 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) { 2801 // Fat-proj kills a slew of registers 2802 // This can add edges to 'n' and obscure whether or not it was a def, 2803 // hence the is_def flag. 2804 fat_proj_seen = true; 2805 RegMask rm = n->out_RegMask();// Make local copy 2806 while( rm.is_NotEmpty() ) { 2807 OptoReg::Name kill = rm.find_first_elem(); 2808 rm.Remove(kill); 2809 anti_do_def( b, n, kill, is_def ); 2810 } 2811 } else { 2812 // Get DEF'd registers the normal way 2813 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2814 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2815 } 2816 2817 // Kill projections on a branch should appear to occur on the 2818 // branch, not afterwards, so grab the masks from the projections 2819 // and process them. 2820 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) { 2821 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2822 Node* use = n->fast_out(i); 2823 if (use->is_Proj()) { 2824 RegMask rm = use->out_RegMask();// Make local copy 2825 while( rm.is_NotEmpty() ) { 2826 OptoReg::Name kill = rm.find_first_elem(); 2827 rm.Remove(kill); 2828 anti_do_def( b, n, kill, false ); 2829 } 2830 } 2831 } 2832 } 2833 2834 // Check each register used by this instruction for a following DEF/KILL 2835 // that must occur afterward and requires an anti-dependence edge. 2836 for( uint j=0; j<n->req(); j++ ) { 2837 Node *def = n->in(j); 2838 if( def ) { 2839 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2840 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2841 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2842 } 2843 } 2844 // Do not allow defs of new derived values to float above GC 2845 // points unless the base is definitely available at the GC point. 2846 2847 Node *m = b->get_node(i); 2848 2849 // Add precedence edge from following safepoint to use of derived pointer 2850 if( last_safept_node != end_node && 2851 m != last_safept_node) { 2852 for (uint k = 1; k < m->req(); k++) { 2853 const Type *t = m->in(k)->bottom_type(); 2854 if( t->isa_oop_ptr() && 2855 t->is_ptr()->offset() != 0 ) { 2856 last_safept_node->add_prec( m ); 2857 break; 2858 } 2859 } 2860 } 2861 2862 if( n->jvms() ) { // Precedence edge from derived to safept 2863 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2864 if( b->get_node(last_safept) != last_safept_node ) { 2865 last_safept = b->find_node(last_safept_node); 2866 } 2867 for( uint j=last_safept; j > i; j-- ) { 2868 Node *mach = b->get_node(j); 2869 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2870 mach->add_prec( n ); 2871 } 2872 last_safept = i; 2873 last_safept_node = m; 2874 } 2875 } 2876 2877 if (fat_proj_seen) { 2878 // Garbage collect pinch nodes that were not consumed. 2879 // They are usually created by a fat kill MachProj for a call. 2880 garbage_collect_pinch_nodes(); 2881 } 2882 } 2883 2884 // Garbage collect pinch nodes for reuse by other blocks. 2885 // 2886 // The block scheduler's insertion of anti-dependence 2887 // edges creates many pinch nodes when the block contains 2888 // 2 or more Calls. A pinch node is used to prevent a 2889 // combinatorial explosion of edges. If a set of kills for a 2890 // register is anti-dependent on a set of uses (or defs), rather 2891 // than adding an edge in the graph between each pair of kill 2892 // and use (or def), a pinch is inserted between them: 2893 // 2894 // use1 use2 use3 2895 // \ | / 2896 // \ | / 2897 // pinch 2898 // / | \ 2899 // / | \ 2900 // kill1 kill2 kill3 2901 // 2902 // One pinch node is created per register killed when 2903 // the second call is encountered during a backwards pass 2904 // over the block. Most of these pinch nodes are never 2905 // wired into the graph because the register is never 2906 // used or def'ed in the block. 2907 // 2908 void Scheduling::garbage_collect_pinch_nodes() { 2909 #ifndef PRODUCT 2910 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 2911 #endif 2912 int trace_cnt = 0; 2913 for (uint k = 0; k < _reg_node.Size(); k++) { 2914 Node* pinch = _reg_node[k]; 2915 if ((pinch != NULL) && pinch->Opcode() == Op_Node && 2916 // no predecence input edges 2917 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { 2918 cleanup_pinch(pinch); 2919 _pinch_free_list.push(pinch); 2920 _reg_node.map(k, NULL); 2921 #ifndef PRODUCT 2922 if (_cfg->C->trace_opto_output()) { 2923 trace_cnt++; 2924 if (trace_cnt > 40) { 2925 tty->print("\n"); 2926 trace_cnt = 0; 2927 } 2928 tty->print(" %d", pinch->_idx); 2929 } 2930 #endif 2931 } 2932 } 2933 #ifndef PRODUCT 2934 if (_cfg->C->trace_opto_output()) tty->print("\n"); 2935 #endif 2936 } 2937 2938 // Clean up a pinch node for reuse. 2939 void Scheduling::cleanup_pinch( Node *pinch ) { 2940 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 2941 2942 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 2943 Node* use = pinch->last_out(i); 2944 uint uses_found = 0; 2945 for (uint j = use->req(); j < use->len(); j++) { 2946 if (use->in(j) == pinch) { 2947 use->rm_prec(j); 2948 uses_found++; 2949 } 2950 } 2951 assert(uses_found > 0, "must be a precedence edge"); 2952 i -= uses_found; // we deleted 1 or more copies of this edge 2953 } 2954 // May have a later_def entry 2955 pinch->set_req(0, NULL); 2956 } 2957 2958 #ifndef PRODUCT 2959 2960 void Scheduling::dump_available() const { 2961 tty->print("#Availist "); 2962 for (uint i = 0; i < _available.size(); i++) 2963 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 2964 tty->cr(); 2965 } 2966 2967 // Print Scheduling Statistics 2968 void Scheduling::print_statistics() { 2969 // Print the size added by nops for bundling 2970 tty->print("Nops added %d bytes to total of %d bytes", 2971 _total_nop_size, _total_method_size); 2972 if (_total_method_size > 0) 2973 tty->print(", for %.2f%%", 2974 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 2975 tty->print("\n"); 2976 2977 // Print the number of branch shadows filled 2978 if (Pipeline::_branch_has_delay_slot) { 2979 tty->print("Of %d branches, %d had unconditional delay slots filled", 2980 _total_branches, _total_unconditional_delays); 2981 if (_total_branches > 0) 2982 tty->print(", for %.2f%%", 2983 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 2984 tty->print("\n"); 2985 } 2986 2987 uint total_instructions = 0, total_bundles = 0; 2988 2989 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 2990 uint bundle_count = _total_instructions_per_bundle[i]; 2991 total_instructions += bundle_count * i; 2992 total_bundles += bundle_count; 2993 } 2994 2995 if (total_bundles > 0) 2996 tty->print("Average ILP (excluding nops) is %.2f\n", 2997 ((double)total_instructions) / ((double)total_bundles)); 2998 } 2999 #endif