1 /* 2 * Copyright (c) 1998, 2018, 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 "asm/macroAssembler.inline.hpp" 28 #include "code/compiledIC.hpp" 29 #include "code/debugInfo.hpp" 30 #include "code/debugInfoRec.hpp" 31 #include "compiler/compileBroker.hpp" 32 #include "compiler/compilerDirectives.hpp" 33 #include "compiler/oopMap.hpp" 34 #include "memory/allocation.inline.hpp" 35 #include "opto/ad.hpp" 36 #include "opto/callnode.hpp" 37 #include "opto/cfgnode.hpp" 38 #include "opto/locknode.hpp" 39 #include "opto/machnode.hpp" 40 #include "opto/optoreg.hpp" 41 #include "opto/output.hpp" 42 #include "opto/regalloc.hpp" 43 #include "opto/runtime.hpp" 44 #include "opto/subnode.hpp" 45 #include "opto/type.hpp" 46 #include "runtime/handles.inline.hpp" 47 #include "utilities/xmlstream.hpp" 48 49 #ifndef PRODUCT 50 #define DEBUG_ARG(x) , x 51 #else 52 #define DEBUG_ARG(x) 53 #endif 54 55 // Convert Nodes to instruction bits and pass off to the VM 56 void Compile::Output() { 57 // RootNode goes 58 assert( _cfg->get_root_block()->number_of_nodes() == 0, "" ); 59 60 // The number of new nodes (mostly MachNop) is proportional to 61 // the number of java calls and inner loops which are aligned. 62 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 + 63 C->inner_loops()*(OptoLoopAlignment-1)), 64 "out of nodes before code generation" ) ) { 65 return; 66 } 67 // Make sure I can find the Start Node 68 Block *entry = _cfg->get_block(1); 69 Block *broot = _cfg->get_root_block(); 70 71 const StartNode *start = entry->head()->as_Start(); 72 73 // Replace StartNode with prolog 74 MachPrologNode *prolog = new MachPrologNode(); 75 entry->map_node(prolog, 0); 76 _cfg->map_node_to_block(prolog, entry); 77 _cfg->unmap_node_from_block(start); // start is no longer in any block 78 79 // Virtual methods need an unverified entry point 80 81 if( is_osr_compilation() ) { 82 if( PoisonOSREntry ) { 83 // TODO: Should use a ShouldNotReachHereNode... 84 _cfg->insert( broot, 0, new MachBreakpointNode() ); 85 } 86 } else { 87 if( _method && !_method->flags().is_static() ) { 88 // Insert unvalidated entry point 89 _cfg->insert( broot, 0, new MachUEPNode() ); 90 } 91 92 } 93 94 // Break before main entry point 95 if ((_method && C->directive()->BreakAtExecuteOption) || 96 (OptoBreakpoint && is_method_compilation()) || 97 (OptoBreakpointOSR && is_osr_compilation()) || 98 (OptoBreakpointC2R && !_method) ) { 99 // checking for _method means that OptoBreakpoint does not apply to 100 // runtime stubs or frame converters 101 _cfg->insert( entry, 1, new MachBreakpointNode() ); 102 } 103 104 // Insert epilogs before every return 105 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 106 Block* block = _cfg->get_block(i); 107 if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point? 108 Node* m = block->end(); 109 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) { 110 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); 111 block->add_inst(epilog); 112 _cfg->map_node_to_block(epilog, block); 113 } 114 } 115 } 116 117 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1); 118 blk_starts[0] = 0; 119 120 // Initialize code buffer and process short branches. 121 CodeBuffer* cb = init_buffer(blk_starts); 122 123 if (cb == NULL || failing()) { 124 return; 125 } 126 127 ScheduleAndBundle(); 128 129 #ifndef PRODUCT 130 if (trace_opto_output()) { 131 tty->print("\n---- After ScheduleAndBundle ----\n"); 132 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 133 tty->print("\nBB#%03d:\n", i); 134 Block* block = _cfg->get_block(i); 135 for (uint j = 0; j < block->number_of_nodes(); j++) { 136 Node* n = block->get_node(j); 137 OptoReg::Name reg = _regalloc->get_reg_first(n); 138 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : ""); 139 n->dump(); 140 } 141 } 142 } 143 #endif 144 145 if (failing()) { 146 return; 147 } 148 149 BuildOopMaps(); 150 151 if (failing()) { 152 return; 153 } 154 155 fill_buffer(cb, blk_starts); 156 } 157 158 bool Compile::need_stack_bang(int frame_size_in_bytes) const { 159 // Determine if we need to generate a stack overflow check. 160 // Do it if the method is not a stub function and 161 // has java calls or has frame size > vm_page_size/8. 162 // The debug VM checks that deoptimization doesn't trigger an 163 // unexpected stack overflow (compiled method stack banging should 164 // guarantee it doesn't happen) so we always need the stack bang in 165 // a debug VM. 166 return (UseStackBanging && stub_function() == NULL && 167 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3 168 DEBUG_ONLY(|| true))); 169 } 170 171 bool Compile::need_register_stack_bang() const { 172 // Determine if we need to generate a register stack overflow check. 173 // This is only used on architectures which have split register 174 // and memory stacks (ie. IA64). 175 // Bang if the method is not a stub function and has java calls 176 return (stub_function() == NULL && has_java_calls()); 177 } 178 179 180 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top 181 // of a loop. When aligning a loop we need to provide enough instructions 182 // in cpu's fetch buffer to feed decoders. The loop alignment could be 183 // avoided if we have enough instructions in fetch buffer at the head of a loop. 184 // By default, the size is set to 999999 by Block's constructor so that 185 // a loop will be aligned if the size is not reset here. 186 // 187 // Note: Mach instructions could contain several HW instructions 188 // so the size is estimated only. 189 // 190 void Compile::compute_loop_first_inst_sizes() { 191 // The next condition is used to gate the loop alignment optimization. 192 // Don't aligned a loop if there are enough instructions at the head of a loop 193 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad 194 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is 195 // equal to 11 bytes which is the largest address NOP instruction. 196 if (MaxLoopPad < OptoLoopAlignment - 1) { 197 uint last_block = _cfg->number_of_blocks() - 1; 198 for (uint i = 1; i <= last_block; i++) { 199 Block* block = _cfg->get_block(i); 200 // Check the first loop's block which requires an alignment. 201 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) { 202 uint sum_size = 0; 203 uint inst_cnt = NumberOfLoopInstrToAlign; 204 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 205 206 // Check subsequent fallthrough blocks if the loop's first 207 // block(s) does not have enough instructions. 208 Block *nb = block; 209 while(inst_cnt > 0 && 210 i < last_block && 211 !_cfg->get_block(i + 1)->has_loop_alignment() && 212 !nb->has_successor(block)) { 213 i++; 214 nb = _cfg->get_block(i); 215 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 216 } // while( inst_cnt > 0 && i < last_block ) 217 218 block->set_first_inst_size(sum_size); 219 } // f( b->head()->is_Loop() ) 220 } // for( i <= last_block ) 221 } // if( MaxLoopPad < OptoLoopAlignment-1 ) 222 } 223 224 // The architecture description provides short branch variants for some long 225 // branch instructions. Replace eligible long branches with short branches. 226 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) { 227 // Compute size of each block, method size, and relocation information size 228 uint nblocks = _cfg->number_of_blocks(); 229 230 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 231 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 232 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks); 233 234 // Collect worst case block paddings 235 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks); 236 memset(block_worst_case_pad, 0, nblocks * sizeof(int)); 237 238 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); ) 239 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); ) 240 241 bool has_short_branch_candidate = false; 242 243 // Initialize the sizes to 0 244 code_size = 0; // Size in bytes of generated code 245 stub_size = 0; // Size in bytes of all stub entries 246 // Size in bytes of all relocation entries, including those in local stubs. 247 // Start with 2-bytes of reloc info for the unvalidated entry point 248 reloc_size = 1; // Number of relocation entries 249 250 // Make three passes. The first computes pessimistic blk_starts, 251 // relative jmp_offset and reloc_size information. The second performs 252 // short branch substitution using the pessimistic sizing. The 253 // third inserts nops where needed. 254 255 // Step one, perform a pessimistic sizing pass. 256 uint last_call_adr = max_juint; 257 uint last_avoid_back_to_back_adr = max_juint; 258 uint nop_size = (new MachNopNode())->size(_regalloc); 259 for (uint i = 0; i < nblocks; i++) { // For all blocks 260 Block* block = _cfg->get_block(i); 261 262 // During short branch replacement, we store the relative (to blk_starts) 263 // offset of jump in jmp_offset, rather than the absolute offset of jump. 264 // This is so that we do not need to recompute sizes of all nodes when 265 // we compute correct blk_starts in our next sizing pass. 266 jmp_offset[i] = 0; 267 jmp_size[i] = 0; 268 jmp_nidx[i] = -1; 269 DEBUG_ONLY( jmp_target[i] = 0; ) 270 DEBUG_ONLY( jmp_rule[i] = 0; ) 271 272 // Sum all instruction sizes to compute block size 273 uint last_inst = block->number_of_nodes(); 274 uint blk_size = 0; 275 for (uint j = 0; j < last_inst; j++) { 276 Node* nj = block->get_node(j); 277 // Handle machine instruction nodes 278 if (nj->is_Mach()) { 279 MachNode *mach = nj->as_Mach(); 280 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 281 reloc_size += mach->reloc(); 282 if (mach->is_MachCall()) { 283 // add size information for trampoline stub 284 // class CallStubImpl is platform-specific and defined in the *.ad files. 285 stub_size += CallStubImpl::size_call_trampoline(); 286 reloc_size += CallStubImpl::reloc_call_trampoline(); 287 288 MachCallNode *mcall = mach->as_MachCall(); 289 // This destination address is NOT PC-relative 290 291 mcall->method_set((intptr_t)mcall->entry_point()); 292 293 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) { 294 stub_size += CompiledStaticCall::to_interp_stub_size(); 295 reloc_size += CompiledStaticCall::reloc_to_interp_stub(); 296 #if INCLUDE_AOT 297 stub_size += CompiledStaticCall::to_aot_stub_size(); 298 reloc_size += CompiledStaticCall::reloc_to_aot_stub(); 299 #endif 300 } 301 } else if (mach->is_MachSafePoint()) { 302 // If call/safepoint are adjacent, account for possible 303 // nop to disambiguate the two safepoints. 304 // ScheduleAndBundle() can rearrange nodes in a block, 305 // check for all offsets inside this block. 306 if (last_call_adr >= blk_starts[i]) { 307 blk_size += nop_size; 308 } 309 } 310 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 311 // Nop is inserted between "avoid back to back" instructions. 312 // ScheduleAndBundle() can rearrange nodes in a block, 313 // check for all offsets inside this block. 314 if (last_avoid_back_to_back_adr >= blk_starts[i]) { 315 blk_size += nop_size; 316 } 317 } 318 if (mach->may_be_short_branch()) { 319 if (!nj->is_MachBranch()) { 320 #ifndef PRODUCT 321 nj->dump(3); 322 #endif 323 Unimplemented(); 324 } 325 assert(jmp_nidx[i] == -1, "block should have only one branch"); 326 jmp_offset[i] = blk_size; 327 jmp_size[i] = nj->size(_regalloc); 328 jmp_nidx[i] = j; 329 has_short_branch_candidate = true; 330 } 331 } 332 blk_size += nj->size(_regalloc); 333 // Remember end of call offset 334 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) { 335 last_call_adr = blk_starts[i]+blk_size; 336 } 337 // Remember end of avoid_back_to_back offset 338 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 339 last_avoid_back_to_back_adr = blk_starts[i]+blk_size; 340 } 341 } 342 343 // When the next block starts a loop, we may insert pad NOP 344 // instructions. Since we cannot know our future alignment, 345 // assume the worst. 346 if (i < nblocks - 1) { 347 Block* nb = _cfg->get_block(i + 1); 348 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 349 if (max_loop_pad > 0) { 350 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 351 // Adjust last_call_adr and/or last_avoid_back_to_back_adr. 352 // If either is the last instruction in this block, bump by 353 // max_loop_pad in lock-step with blk_size, so sizing 354 // calculations in subsequent blocks still can conservatively 355 // detect that it may the last instruction in this block. 356 if (last_call_adr == blk_starts[i]+blk_size) { 357 last_call_adr += max_loop_pad; 358 } 359 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) { 360 last_avoid_back_to_back_adr += max_loop_pad; 361 } 362 blk_size += max_loop_pad; 363 block_worst_case_pad[i + 1] = max_loop_pad; 364 } 365 } 366 367 // Save block size; update total method size 368 blk_starts[i+1] = blk_starts[i]+blk_size; 369 } 370 371 // Step two, replace eligible long jumps. 372 bool progress = true; 373 uint last_may_be_short_branch_adr = max_juint; 374 while (has_short_branch_candidate && progress) { 375 progress = false; 376 has_short_branch_candidate = false; 377 int adjust_block_start = 0; 378 for (uint i = 0; i < nblocks; i++) { 379 Block* block = _cfg->get_block(i); 380 int idx = jmp_nidx[i]; 381 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach(); 382 if (mach != NULL && mach->may_be_short_branch()) { 383 #ifdef ASSERT 384 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity"); 385 int j; 386 // Find the branch; ignore trailing NOPs. 387 for (j = block->number_of_nodes()-1; j>=0; j--) { 388 Node* n = block->get_node(j); 389 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) 390 break; 391 } 392 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity"); 393 #endif 394 int br_size = jmp_size[i]; 395 int br_offs = blk_starts[i] + jmp_offset[i]; 396 397 // This requires the TRUE branch target be in succs[0] 398 uint bnum = block->non_connector_successor(0)->_pre_order; 399 int offset = blk_starts[bnum] - br_offs; 400 if (bnum > i) { // adjust following block's offset 401 offset -= adjust_block_start; 402 } 403 404 // This block can be a loop header, account for the padding 405 // in the previous block. 406 int block_padding = block_worst_case_pad[i]; 407 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top"); 408 // In the following code a nop could be inserted before 409 // the branch which will increase the backward distance. 410 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr); 411 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block"); 412 413 if (needs_padding && offset <= 0) 414 offset -= nop_size; 415 416 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 417 // We've got a winner. Replace this branch. 418 MachNode* replacement = mach->as_MachBranch()->short_branch_version(); 419 420 // Update the jmp_size. 421 int new_size = replacement->size(_regalloc); 422 int diff = br_size - new_size; 423 assert(diff >= (int)nop_size, "short_branch size should be smaller"); 424 // Conservatively take into account padding between 425 // avoid_back_to_back branches. Previous branch could be 426 // converted into avoid_back_to_back branch during next 427 // rounds. 428 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 429 jmp_offset[i] += nop_size; 430 diff -= nop_size; 431 } 432 adjust_block_start += diff; 433 block->map_node(replacement, idx); 434 mach->subsume_by(replacement, C); 435 mach = replacement; 436 progress = true; 437 438 jmp_size[i] = new_size; 439 DEBUG_ONLY( jmp_target[i] = bnum; ); 440 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 441 } else { 442 // The jump distance is not short, try again during next iteration. 443 has_short_branch_candidate = true; 444 } 445 } // (mach->may_be_short_branch()) 446 if (mach != NULL && (mach->may_be_short_branch() || 447 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) { 448 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i]; 449 } 450 blk_starts[i+1] -= adjust_block_start; 451 } 452 } 453 454 #ifdef ASSERT 455 for (uint i = 0; i < nblocks; i++) { // For all blocks 456 if (jmp_target[i] != 0) { 457 int br_size = jmp_size[i]; 458 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 459 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 460 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]); 461 } 462 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp"); 463 } 464 } 465 #endif 466 467 // Step 3, compute the offsets of all blocks, will be done in fill_buffer() 468 // after ScheduleAndBundle(). 469 470 // ------------------ 471 // Compute size for code buffer 472 code_size = blk_starts[nblocks]; 473 474 // Relocation records 475 reloc_size += 1; // Relo entry for exception handler 476 477 // Adjust reloc_size to number of record of relocation info 478 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 479 // a relocation index. 480 // The CodeBuffer will expand the locs array if this estimate is too low. 481 reloc_size *= 10 / sizeof(relocInfo); 482 } 483 484 //------------------------------FillLocArray----------------------------------- 485 // Create a bit of debug info and append it to the array. The mapping is from 486 // Java local or expression stack to constant, register or stack-slot. For 487 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 488 // entry has been taken care of and caller should skip it). 489 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 490 // This should never have accepted Bad before 491 assert(OptoReg::is_valid(regnum), "location must be valid"); 492 return (OptoReg::is_reg(regnum)) 493 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 494 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 495 } 496 497 498 ObjectValue* 499 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 500 for (int i = 0; i < objs->length(); i++) { 501 assert(objs->at(i)->is_object(), "corrupt object cache"); 502 ObjectValue* sv = (ObjectValue*) objs->at(i); 503 if (sv->id() == id) { 504 return sv; 505 } 506 } 507 // Otherwise.. 508 return NULL; 509 } 510 511 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 512 ObjectValue* sv ) { 513 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition"); 514 objs->append(sv); 515 } 516 517 518 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 519 GrowableArray<ScopeValue*> *array, 520 GrowableArray<ScopeValue*> *objs ) { 521 assert( local, "use _top instead of null" ); 522 if (array->length() != idx) { 523 assert(array->length() == idx + 1, "Unexpected array count"); 524 // Old functionality: 525 // return 526 // New functionality: 527 // Assert if the local is not top. In product mode let the new node 528 // override the old entry. 529 assert(local == top(), "LocArray collision"); 530 if (local == top()) { 531 return; 532 } 533 array->pop(); 534 } 535 const Type *t = local->bottom_type(); 536 537 // Is it a safepoint scalar object node? 538 if (local->is_SafePointScalarObject()) { 539 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 540 541 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx); 542 if (sv == NULL) { 543 ciKlass* cik = t->is_oopptr()->klass(); 544 assert(cik->is_instance_klass() || 545 cik->is_array_klass(), "Not supported allocation."); 546 sv = new ObjectValue(spobj->_idx, 547 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 548 Compile::set_sv_for_object_node(objs, sv); 549 550 uint first_ind = spobj->first_index(sfpt->jvms()); 551 for (uint i = 0; i < spobj->n_fields(); i++) { 552 Node* fld_node = sfpt->in(first_ind+i); 553 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 554 } 555 } 556 array->append(sv); 557 return; 558 } 559 560 // Grab the register number for the local 561 OptoReg::Name regnum = _regalloc->get_reg_first(local); 562 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 563 // Record the double as two float registers. 564 // The register mask for such a value always specifies two adjacent 565 // float registers, with the lower register number even. 566 // Normally, the allocation of high and low words to these registers 567 // is irrelevant, because nearly all operations on register pairs 568 // (e.g., StoreD) treat them as a single unit. 569 // Here, we assume in addition that the words in these two registers 570 // stored "naturally" (by operations like StoreD and double stores 571 // within the interpreter) such that the lower-numbered register 572 // is written to the lower memory address. This may seem like 573 // a machine dependency, but it is not--it is a requirement on 574 // the author of the <arch>.ad file to ensure that, for every 575 // even/odd double-register pair to which a double may be allocated, 576 // the word in the even single-register is stored to the first 577 // memory word. (Note that register numbers are completely 578 // arbitrary, and are not tied to any machine-level encodings.) 579 #ifdef _LP64 580 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 581 array->append(new ConstantIntValue((jint)0)); 582 array->append(new_loc_value( _regalloc, regnum, Location::dbl )); 583 } else if ( t->base() == Type::Long ) { 584 array->append(new ConstantIntValue((jint)0)); 585 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 586 } else if ( t->base() == Type::RawPtr ) { 587 // jsr/ret return address which must be restored into a the full 588 // width 64-bit stack slot. 589 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 590 } 591 #else //_LP64 592 #ifdef SPARC 593 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) { 594 // For SPARC we have to swap high and low words for 595 // long values stored in a single-register (g0-g7). 596 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 597 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 598 } else 599 #endif //SPARC 600 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 601 // Repack the double/long as two jints. 602 // The convention the interpreter uses is that the second local 603 // holds the first raw word of the native double representation. 604 // This is actually reasonable, since locals and stack arrays 605 // grow downwards in all implementations. 606 // (If, on some machine, the interpreter's Java locals or stack 607 // were to grow upwards, the embedded doubles would be word-swapped.) 608 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 609 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 610 } 611 #endif //_LP64 612 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 613 OptoReg::is_reg(regnum) ) { 614 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double() 615 ? Location::float_in_dbl : Location::normal )); 616 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 617 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long 618 ? Location::int_in_long : Location::normal )); 619 } else if( t->base() == Type::NarrowOop ) { 620 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop )); 621 } else { 622 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal )); 623 } 624 return; 625 } 626 627 // No register. It must be constant data. 628 switch (t->base()) { 629 case Type::Half: // Second half of a double 630 ShouldNotReachHere(); // Caller should skip 2nd halves 631 break; 632 case Type::AnyPtr: 633 array->append(new ConstantOopWriteValue(NULL)); 634 break; 635 case Type::AryPtr: 636 case Type::InstPtr: // fall through 637 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 638 break; 639 case Type::NarrowOop: 640 if (t == TypeNarrowOop::NULL_PTR) { 641 array->append(new ConstantOopWriteValue(NULL)); 642 } else { 643 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 644 } 645 break; 646 case Type::Int: 647 array->append(new ConstantIntValue(t->is_int()->get_con())); 648 break; 649 case Type::RawPtr: 650 // A return address (T_ADDRESS). 651 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 652 #ifdef _LP64 653 // Must be restored to the full-width 64-bit stack slot. 654 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 655 #else 656 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 657 #endif 658 break; 659 case Type::FloatCon: { 660 float f = t->is_float_constant()->getf(); 661 array->append(new ConstantIntValue(jint_cast(f))); 662 break; 663 } 664 case Type::DoubleCon: { 665 jdouble d = t->is_double_constant()->getd(); 666 #ifdef _LP64 667 array->append(new ConstantIntValue((jint)0)); 668 array->append(new ConstantDoubleValue(d)); 669 #else 670 // Repack the double as two jints. 671 // The convention the interpreter uses is that the second local 672 // holds the first raw word of the native double representation. 673 // This is actually reasonable, since locals and stack arrays 674 // grow downwards in all implementations. 675 // (If, on some machine, the interpreter's Java locals or stack 676 // were to grow upwards, the embedded doubles would be word-swapped.) 677 jlong_accessor acc; 678 acc.long_value = jlong_cast(d); 679 array->append(new ConstantIntValue(acc.words[1])); 680 array->append(new ConstantIntValue(acc.words[0])); 681 #endif 682 break; 683 } 684 case Type::Long: { 685 jlong d = t->is_long()->get_con(); 686 #ifdef _LP64 687 array->append(new ConstantIntValue((jint)0)); 688 array->append(new ConstantLongValue(d)); 689 #else 690 // Repack the long as two jints. 691 // The convention the interpreter uses is that the second local 692 // holds the first raw word of the native double representation. 693 // This is actually reasonable, since locals and stack arrays 694 // grow downwards in all implementations. 695 // (If, on some machine, the interpreter's Java locals or stack 696 // were to grow upwards, the embedded doubles would be word-swapped.) 697 jlong_accessor acc; 698 acc.long_value = d; 699 array->append(new ConstantIntValue(acc.words[1])); 700 array->append(new ConstantIntValue(acc.words[0])); 701 #endif 702 break; 703 } 704 case Type::Top: // Add an illegal value here 705 array->append(new LocationValue(Location())); 706 break; 707 default: 708 ShouldNotReachHere(); 709 break; 710 } 711 } 712 713 // Determine if this node starts a bundle 714 bool Compile::starts_bundle(const Node *n) const { 715 return (_node_bundling_limit > n->_idx && 716 _node_bundling_base[n->_idx].starts_bundle()); 717 } 718 719 //--------------------------Process_OopMap_Node-------------------------------- 720 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) { 721 722 // Handle special safepoint nodes for synchronization 723 MachSafePointNode *sfn = mach->as_MachSafePoint(); 724 MachCallNode *mcall; 725 726 int safepoint_pc_offset = current_offset; 727 bool is_method_handle_invoke = false; 728 bool return_oop = false; 729 730 // Add the safepoint in the DebugInfoRecorder 731 if( !mach->is_MachCall() ) { 732 mcall = NULL; 733 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 734 } else { 735 mcall = mach->as_MachCall(); 736 737 // Is the call a MethodHandle call? 738 if (mcall->is_MachCallJava()) { 739 if (mcall->as_MachCallJava()->_method_handle_invoke) { 740 assert(has_method_handle_invokes(), "must have been set during call generation"); 741 is_method_handle_invoke = true; 742 } 743 } 744 745 // Check if a call returns an object. 746 if (mcall->returns_pointer()) { 747 return_oop = true; 748 } 749 safepoint_pc_offset += mcall->ret_addr_offset(); 750 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 751 } 752 753 // Loop over the JVMState list to add scope information 754 // Do not skip safepoints with a NULL method, they need monitor info 755 JVMState* youngest_jvms = sfn->jvms(); 756 int max_depth = youngest_jvms->depth(); 757 758 // Allocate the object pool for scalar-replaced objects -- the map from 759 // small-integer keys (which can be recorded in the local and ostack 760 // arrays) to descriptions of the object state. 761 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 762 763 // Visit scopes from oldest to youngest. 764 for (int depth = 1; depth <= max_depth; depth++) { 765 JVMState* jvms = youngest_jvms->of_depth(depth); 766 int idx; 767 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 768 // Safepoints that do not have method() set only provide oop-map and monitor info 769 // to support GC; these do not support deoptimization. 770 int num_locs = (method == NULL) ? 0 : jvms->loc_size(); 771 int num_exps = (method == NULL) ? 0 : jvms->stk_size(); 772 int num_mon = jvms->nof_monitors(); 773 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), 774 "JVMS local count must match that of the method"); 775 776 // Add Local and Expression Stack Information 777 778 // Insert locals into the locarray 779 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 780 for( idx = 0; idx < num_locs; idx++ ) { 781 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 782 } 783 784 // Insert expression stack entries into the exparray 785 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 786 for( idx = 0; idx < num_exps; idx++ ) { 787 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 788 } 789 790 // Add in mappings of the monitors 791 assert( !method || 792 !method->is_synchronized() || 793 method->is_native() || 794 num_mon > 0 || 795 !GenerateSynchronizationCode, 796 "monitors must always exist for synchronized methods"); 797 798 // Build the growable array of ScopeValues for exp stack 799 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 800 801 // Loop over monitors and insert into array 802 for (idx = 0; idx < num_mon; idx++) { 803 // Grab the node that defines this monitor 804 Node* box_node = sfn->monitor_box(jvms, idx); 805 Node* obj_node = sfn->monitor_obj(jvms, idx); 806 807 // Create ScopeValue for object 808 ScopeValue *scval = NULL; 809 810 if (obj_node->is_SafePointScalarObject()) { 811 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 812 scval = Compile::sv_for_node_id(objs, spobj->_idx); 813 if (scval == NULL) { 814 const Type *t = spobj->bottom_type(); 815 ciKlass* cik = t->is_oopptr()->klass(); 816 assert(cik->is_instance_klass() || 817 cik->is_array_klass(), "Not supported allocation."); 818 ObjectValue* sv = new ObjectValue(spobj->_idx, 819 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 820 Compile::set_sv_for_object_node(objs, sv); 821 822 uint first_ind = spobj->first_index(youngest_jvms); 823 for (uint i = 0; i < spobj->n_fields(); i++) { 824 Node* fld_node = sfn->in(first_ind+i); 825 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 826 } 827 scval = sv; 828 } 829 } else if (!obj_node->is_Con()) { 830 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); 831 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 832 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop ); 833 } else { 834 scval = new_loc_value( _regalloc, obj_reg, Location::oop ); 835 } 836 } else { 837 const TypePtr *tp = obj_node->get_ptr_type(); 838 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding()); 839 } 840 841 OptoReg::Name box_reg = BoxLockNode::reg(box_node); 842 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)); 843 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated()); 844 monarray->append(new MonitorValue(scval, basic_lock, eliminated)); 845 } 846 847 // We dump the object pool first, since deoptimization reads it in first. 848 debug_info()->dump_object_pool(objs); 849 850 // Build first class objects to pass to scope 851 DebugToken *locvals = debug_info()->create_scope_values(locarray); 852 DebugToken *expvals = debug_info()->create_scope_values(exparray); 853 DebugToken *monvals = debug_info()->create_monitor_values(monarray); 854 855 // Make method available for all Safepoints 856 ciMethod* scope_method = method ? method : _method; 857 // Describe the scope here 858 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 859 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 860 // Now we can describe the scope. 861 methodHandle null_mh; 862 bool rethrow_exception = false; 863 debug_info()->describe_scope(safepoint_pc_offset, null_mh, scope_method, jvms->bci(), jvms->should_reexecute(), rethrow_exception, is_method_handle_invoke, return_oop, locvals, expvals, monvals); 864 } // End jvms loop 865 866 // Mark the end of the scope set. 867 debug_info()->end_safepoint(safepoint_pc_offset); 868 } 869 870 871 872 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 873 class NonSafepointEmitter { 874 Compile* C; 875 JVMState* _pending_jvms; 876 int _pending_offset; 877 878 void emit_non_safepoint(); 879 880 public: 881 NonSafepointEmitter(Compile* compile) { 882 this->C = compile; 883 _pending_jvms = NULL; 884 _pending_offset = 0; 885 } 886 887 void observe_instruction(Node* n, int pc_offset) { 888 if (!C->debug_info()->recording_non_safepoints()) return; 889 890 Node_Notes* nn = C->node_notes_at(n->_idx); 891 if (nn == NULL || nn->jvms() == NULL) return; 892 if (_pending_jvms != NULL && 893 _pending_jvms->same_calls_as(nn->jvms())) { 894 // Repeated JVMS? Stretch it up here. 895 _pending_offset = pc_offset; 896 } else { 897 if (_pending_jvms != NULL && 898 _pending_offset < pc_offset) { 899 emit_non_safepoint(); 900 } 901 _pending_jvms = NULL; 902 if (pc_offset > C->debug_info()->last_pc_offset()) { 903 // This is the only way _pending_jvms can become non-NULL: 904 _pending_jvms = nn->jvms(); 905 _pending_offset = pc_offset; 906 } 907 } 908 } 909 910 // Stay out of the way of real safepoints: 911 void observe_safepoint(JVMState* jvms, int pc_offset) { 912 if (_pending_jvms != NULL && 913 !_pending_jvms->same_calls_as(jvms) && 914 _pending_offset < pc_offset) { 915 emit_non_safepoint(); 916 } 917 _pending_jvms = NULL; 918 } 919 920 void flush_at_end() { 921 if (_pending_jvms != NULL) { 922 emit_non_safepoint(); 923 } 924 _pending_jvms = NULL; 925 } 926 }; 927 928 void NonSafepointEmitter::emit_non_safepoint() { 929 JVMState* youngest_jvms = _pending_jvms; 930 int pc_offset = _pending_offset; 931 932 // Clear it now: 933 _pending_jvms = NULL; 934 935 DebugInformationRecorder* debug_info = C->debug_info(); 936 assert(debug_info->recording_non_safepoints(), "sanity"); 937 938 debug_info->add_non_safepoint(pc_offset); 939 int max_depth = youngest_jvms->depth(); 940 941 // Visit scopes from oldest to youngest. 942 for (int depth = 1; depth <= max_depth; depth++) { 943 JVMState* jvms = youngest_jvms->of_depth(depth); 944 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 945 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 946 methodHandle null_mh; 947 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute()); 948 } 949 950 // Mark the end of the scope set. 951 debug_info->end_non_safepoint(pc_offset); 952 } 953 954 //------------------------------init_buffer------------------------------------ 955 CodeBuffer* Compile::init_buffer(uint* blk_starts) { 956 957 // Set the initially allocated size 958 int code_req = initial_code_capacity; 959 int locs_req = initial_locs_capacity; 960 int stub_req = initial_stub_capacity; 961 int const_req = initial_const_capacity; 962 963 int pad_req = NativeCall::instruction_size; 964 // The extra spacing after the code is necessary on some platforms. 965 // Sometimes we need to patch in a jump after the last instruction, 966 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 967 968 // Compute the byte offset where we can store the deopt pc. 969 if (fixed_slots() != 0) { 970 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 971 } 972 973 // Compute prolog code size 974 _method_size = 0; 975 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; 976 #if defined(IA64) && !defined(AIX) 977 if (save_argument_registers()) { 978 // 4815101: this is a stub with implicit and unknown precision fp args. 979 // The usual spill mechanism can only generate stfd's in this case, which 980 // doesn't work if the fp reg to spill contains a single-precision denorm. 981 // Instead, we hack around the normal spill mechanism using stfspill's and 982 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate 983 // space here for the fp arg regs (f8-f15) we're going to thusly spill. 984 // 985 // If we ever implement 16-byte 'registers' == stack slots, we can 986 // get rid of this hack and have SpillCopy generate stfspill/ldffill 987 // instead of stfd/stfs/ldfd/ldfs. 988 _frame_slots += 8*(16/BytesPerInt); 989 } 990 #endif 991 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check"); 992 993 if (has_mach_constant_base_node()) { 994 uint add_size = 0; 995 // Fill the constant table. 996 // Note: This must happen before shorten_branches. 997 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 998 Block* b = _cfg->get_block(i); 999 1000 for (uint j = 0; j < b->number_of_nodes(); j++) { 1001 Node* n = b->get_node(j); 1002 1003 // If the node is a MachConstantNode evaluate the constant 1004 // value section. 1005 if (n->is_MachConstant()) { 1006 MachConstantNode* machcon = n->as_MachConstant(); 1007 machcon->eval_constant(C); 1008 } else if (n->is_Mach()) { 1009 // On Power there are more nodes that issue constants. 1010 add_size += (n->as_Mach()->ins_num_consts() * 8); 1011 } 1012 } 1013 } 1014 1015 // Calculate the offsets of the constants and the size of the 1016 // constant table (including the padding to the next section). 1017 constant_table().calculate_offsets_and_size(); 1018 const_req = constant_table().size() + add_size; 1019 } 1020 1021 // Initialize the space for the BufferBlob used to find and verify 1022 // instruction size in MachNode::emit_size() 1023 init_scratch_buffer_blob(const_req); 1024 if (failing()) return NULL; // Out of memory 1025 1026 // Pre-compute the length of blocks and replace 1027 // long branches with short if machine supports it. 1028 shorten_branches(blk_starts, code_req, locs_req, stub_req); 1029 1030 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1031 // class HandlerImpl is platform-specific and defined in the *.ad files. 1032 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler 1033 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler 1034 stub_req += MAX_stubs_size; // ensure per-stub margin 1035 code_req += MAX_inst_size; // ensure per-instruction margin 1036 1037 if (StressCodeBuffers) 1038 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1039 1040 int total_req = 1041 const_req + 1042 code_req + 1043 pad_req + 1044 stub_req + 1045 exception_handler_req + 1046 deopt_handler_req; // deopt handler 1047 1048 if (has_method_handle_invokes()) 1049 total_req += deopt_handler_req; // deopt MH handler 1050 1051 CodeBuffer* cb = code_buffer(); 1052 cb->initialize(total_req, locs_req); 1053 1054 // Have we run out of code space? 1055 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1056 C->record_failure("CodeCache is full"); 1057 return NULL; 1058 } 1059 // Configure the code buffer. 1060 cb->initialize_consts_size(const_req); 1061 cb->initialize_stubs_size(stub_req); 1062 cb->initialize_oop_recorder(env()->oop_recorder()); 1063 1064 // fill in the nop array for bundling computations 1065 MachNode *_nop_list[Bundle::_nop_count]; 1066 Bundle::initialize_nops(_nop_list); 1067 1068 return cb; 1069 } 1070 1071 //------------------------------fill_buffer------------------------------------ 1072 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) { 1073 // blk_starts[] contains offsets calculated during short branches processing, 1074 // offsets should not be increased during following steps. 1075 1076 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 1077 // of a loop. It is used to determine the padding for loop alignment. 1078 compute_loop_first_inst_sizes(); 1079 1080 // Create oopmap set. 1081 _oop_map_set = new OopMapSet(); 1082 1083 // !!!!! This preserves old handling of oopmaps for now 1084 debug_info()->set_oopmaps(_oop_map_set); 1085 1086 uint nblocks = _cfg->number_of_blocks(); 1087 // Count and start of implicit null check instructions 1088 uint inct_cnt = 0; 1089 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1090 1091 // Count and start of calls 1092 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1093 1094 uint return_offset = 0; 1095 int nop_size = (new MachNopNode())->size(_regalloc); 1096 1097 int previous_offset = 0; 1098 int current_offset = 0; 1099 int last_call_offset = -1; 1100 int last_avoid_back_to_back_offset = -1; 1101 #ifdef ASSERT 1102 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); 1103 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 1104 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 1105 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); 1106 #endif 1107 1108 // Create an array of unused labels, one for each basic block, if printing is enabled 1109 #ifndef PRODUCT 1110 int *node_offsets = NULL; 1111 uint node_offset_limit = unique(); 1112 1113 if (print_assembly()) 1114 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1115 #endif 1116 1117 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily 1118 1119 // Emit the constant table. 1120 if (has_mach_constant_base_node()) { 1121 constant_table().emit(*cb); 1122 } 1123 1124 // Create an array of labels, one for each basic block 1125 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1); 1126 for (uint i=0; i <= nblocks; i++) { 1127 blk_labels[i].init(); 1128 } 1129 1130 // ------------------ 1131 // Now fill in the code buffer 1132 Node *delay_slot = NULL; 1133 1134 for (uint i = 0; i < nblocks; i++) { 1135 Block* block = _cfg->get_block(i); 1136 Node* head = block->head(); 1137 1138 // If this block needs to start aligned (i.e, can be reached other 1139 // than by falling-thru from the previous block), then force the 1140 // start of a new bundle. 1141 if (Pipeline::requires_bundling() && starts_bundle(head)) { 1142 cb->flush_bundle(true); 1143 } 1144 1145 #ifdef ASSERT 1146 if (!block->is_connector()) { 1147 stringStream st; 1148 block->dump_head(_cfg, &st); 1149 MacroAssembler(cb).block_comment(st.as_string()); 1150 } 1151 jmp_target[i] = 0; 1152 jmp_offset[i] = 0; 1153 jmp_size[i] = 0; 1154 jmp_rule[i] = 0; 1155 #endif 1156 int blk_offset = current_offset; 1157 1158 // Define the label at the beginning of the basic block 1159 MacroAssembler(cb).bind(blk_labels[block->_pre_order]); 1160 1161 uint last_inst = block->number_of_nodes(); 1162 1163 // Emit block normally, except for last instruction. 1164 // Emit means "dump code bits into code buffer". 1165 for (uint j = 0; j<last_inst; j++) { 1166 1167 // Get the node 1168 Node* n = block->get_node(j); 1169 1170 // See if delay slots are supported 1171 if (valid_bundle_info(n) && 1172 node_bundling(n)->used_in_unconditional_delay()) { 1173 assert(delay_slot == NULL, "no use of delay slot node"); 1174 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1175 1176 delay_slot = n; 1177 continue; 1178 } 1179 1180 // If this starts a new instruction group, then flush the current one 1181 // (but allow split bundles) 1182 if (Pipeline::requires_bundling() && starts_bundle(n)) 1183 cb->flush_bundle(false); 1184 1185 // Special handling for SafePoint/Call Nodes 1186 bool is_mcall = false; 1187 if (n->is_Mach()) { 1188 MachNode *mach = n->as_Mach(); 1189 is_mcall = n->is_MachCall(); 1190 bool is_sfn = n->is_MachSafePoint(); 1191 1192 // If this requires all previous instructions be flushed, then do so 1193 if (is_sfn || is_mcall || mach->alignment_required() != 1) { 1194 cb->flush_bundle(true); 1195 current_offset = cb->insts_size(); 1196 } 1197 1198 // A padding may be needed again since a previous instruction 1199 // could be moved to delay slot. 1200 1201 // align the instruction if necessary 1202 int padding = mach->compute_padding(current_offset); 1203 // Make sure safepoint node for polling is distinct from a call's 1204 // return by adding a nop if needed. 1205 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) { 1206 padding = nop_size; 1207 } 1208 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) && 1209 current_offset == last_avoid_back_to_back_offset) { 1210 // Avoid back to back some instructions. 1211 padding = nop_size; 1212 } 1213 1214 if (padding > 0) { 1215 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1216 int nops_cnt = padding / nop_size; 1217 MachNode *nop = new MachNopNode(nops_cnt); 1218 block->insert_node(nop, j++); 1219 last_inst++; 1220 _cfg->map_node_to_block(nop, block); 1221 // Ensure enough space. 1222 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1223 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1224 C->record_failure("CodeCache is full"); 1225 return; 1226 } 1227 nop->emit(*cb, _regalloc); 1228 cb->flush_bundle(true); 1229 current_offset = cb->insts_size(); 1230 } 1231 1232 // Remember the start of the last call in a basic block 1233 if (is_mcall) { 1234 MachCallNode *mcall = mach->as_MachCall(); 1235 1236 // This destination address is NOT PC-relative 1237 mcall->method_set((intptr_t)mcall->entry_point()); 1238 1239 // Save the return address 1240 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset(); 1241 1242 if (mcall->is_MachCallLeaf()) { 1243 is_mcall = false; 1244 is_sfn = false; 1245 } 1246 } 1247 1248 // sfn will be valid whenever mcall is valid now because of inheritance 1249 if (is_sfn || is_mcall) { 1250 1251 // Handle special safepoint nodes for synchronization 1252 if (!is_mcall) { 1253 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1254 // !!!!! Stubs only need an oopmap right now, so bail out 1255 if (sfn->jvms()->method() == NULL) { 1256 // Write the oopmap directly to the code blob??!! 1257 continue; 1258 } 1259 } // End synchronization 1260 1261 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1262 current_offset); 1263 Process_OopMap_Node(mach, current_offset); 1264 } // End if safepoint 1265 1266 // If this is a null check, then add the start of the previous instruction to the list 1267 else if( mach->is_MachNullCheck() ) { 1268 inct_starts[inct_cnt++] = previous_offset; 1269 } 1270 1271 // If this is a branch, then fill in the label with the target BB's label 1272 else if (mach->is_MachBranch()) { 1273 // This requires the TRUE branch target be in succs[0] 1274 uint block_num = block->non_connector_successor(0)->_pre_order; 1275 1276 // Try to replace long branch if delay slot is not used, 1277 // it is mostly for back branches since forward branch's 1278 // distance is not updated yet. 1279 bool delay_slot_is_used = valid_bundle_info(n) && 1280 node_bundling(n)->use_unconditional_delay(); 1281 if (!delay_slot_is_used && mach->may_be_short_branch()) { 1282 assert(delay_slot == NULL, "not expecting delay slot node"); 1283 int br_size = n->size(_regalloc); 1284 int offset = blk_starts[block_num] - current_offset; 1285 if (block_num >= i) { 1286 // Current and following block's offset are not 1287 // finalized yet, adjust distance by the difference 1288 // between calculated and final offsets of current block. 1289 offset -= (blk_starts[i] - blk_offset); 1290 } 1291 // In the following code a nop could be inserted before 1292 // the branch which will increase the backward distance. 1293 bool needs_padding = (current_offset == last_avoid_back_to_back_offset); 1294 if (needs_padding && offset <= 0) 1295 offset -= nop_size; 1296 1297 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 1298 // We've got a winner. Replace this branch. 1299 MachNode* replacement = mach->as_MachBranch()->short_branch_version(); 1300 1301 // Update the jmp_size. 1302 int new_size = replacement->size(_regalloc); 1303 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller"); 1304 // Insert padding between avoid_back_to_back branches. 1305 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 1306 MachNode *nop = new MachNopNode(); 1307 block->insert_node(nop, j++); 1308 _cfg->map_node_to_block(nop, block); 1309 last_inst++; 1310 nop->emit(*cb, _regalloc); 1311 cb->flush_bundle(true); 1312 current_offset = cb->insts_size(); 1313 } 1314 #ifdef ASSERT 1315 jmp_target[i] = block_num; 1316 jmp_offset[i] = current_offset - blk_offset; 1317 jmp_size[i] = new_size; 1318 jmp_rule[i] = mach->rule(); 1319 #endif 1320 block->map_node(replacement, j); 1321 mach->subsume_by(replacement, C); 1322 n = replacement; 1323 mach = replacement; 1324 } 1325 } 1326 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num ); 1327 } else if (mach->ideal_Opcode() == Op_Jump) { 1328 for (uint h = 0; h < block->_num_succs; h++) { 1329 Block* succs_block = block->_succs[h]; 1330 for (uint j = 1; j < succs_block->num_preds(); j++) { 1331 Node* jpn = succs_block->pred(j); 1332 if (jpn->is_JumpProj() && jpn->in(0) == mach) { 1333 uint block_num = succs_block->non_connector()->_pre_order; 1334 Label *blkLabel = &blk_labels[block_num]; 1335 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1336 } 1337 } 1338 } 1339 } 1340 #ifdef ASSERT 1341 // Check that oop-store precedes the card-mark 1342 else if (mach->ideal_Opcode() == Op_StoreCM) { 1343 uint storeCM_idx = j; 1344 int count = 0; 1345 for (uint prec = mach->req(); prec < mach->len(); prec++) { 1346 Node *oop_store = mach->in(prec); // Precedence edge 1347 if (oop_store == NULL) continue; 1348 count++; 1349 uint i4; 1350 for (i4 = 0; i4 < last_inst; ++i4) { 1351 if (block->get_node(i4) == oop_store) { 1352 break; 1353 } 1354 } 1355 // Note: This test can provide a false failure if other precedence 1356 // edges have been added to the storeCMNode. 1357 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1358 } 1359 assert(count > 0, "storeCM expects at least one precedence edge"); 1360 } 1361 #endif 1362 else if (!n->is_Proj()) { 1363 // Remember the beginning of the previous instruction, in case 1364 // it's followed by a flag-kill and a null-check. Happens on 1365 // Intel all the time, with add-to-memory kind of opcodes. 1366 previous_offset = current_offset; 1367 } 1368 1369 // Not an else-if! 1370 // If this is a trap based cmp then add its offset to the list. 1371 if (mach->is_TrapBasedCheckNode()) { 1372 inct_starts[inct_cnt++] = current_offset; 1373 } 1374 } 1375 1376 // Verify that there is sufficient space remaining 1377 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1378 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1379 C->record_failure("CodeCache is full"); 1380 return; 1381 } 1382 1383 // Save the offset for the listing 1384 #ifndef PRODUCT 1385 if (node_offsets && n->_idx < node_offset_limit) 1386 node_offsets[n->_idx] = cb->insts_size(); 1387 #endif 1388 1389 // "Normal" instruction case 1390 DEBUG_ONLY( uint instr_offset = cb->insts_size(); ) 1391 n->emit(*cb, _regalloc); 1392 current_offset = cb->insts_size(); 1393 1394 // Above we only verified that there is enough space in the instruction section. 1395 // However, the instruction may emit stubs that cause code buffer expansion. 1396 // Bail out here if expansion failed due to a lack of code cache space. 1397 if (failing()) { 1398 return; 1399 } 1400 1401 #ifdef ASSERT 1402 if (n->size(_regalloc) < (current_offset-instr_offset)) { 1403 n->dump(); 1404 assert(false, "wrong size of mach node"); 1405 } 1406 #endif 1407 non_safepoints.observe_instruction(n, current_offset); 1408 1409 // mcall is last "call" that can be a safepoint 1410 // record it so we can see if a poll will directly follow it 1411 // in which case we'll need a pad to make the PcDesc sites unique 1412 // see 5010568. This can be slightly inaccurate but conservative 1413 // in the case that return address is not actually at current_offset. 1414 // This is a small price to pay. 1415 1416 if (is_mcall) { 1417 last_call_offset = current_offset; 1418 } 1419 1420 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 1421 // Avoid back to back some instructions. 1422 last_avoid_back_to_back_offset = current_offset; 1423 } 1424 1425 // See if this instruction has a delay slot 1426 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1427 guarantee(delay_slot != NULL, "expecting delay slot node"); 1428 1429 // Back up 1 instruction 1430 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size()); 1431 1432 // Save the offset for the listing 1433 #ifndef PRODUCT 1434 if (node_offsets && delay_slot->_idx < node_offset_limit) 1435 node_offsets[delay_slot->_idx] = cb->insts_size(); 1436 #endif 1437 1438 // Support a SafePoint in the delay slot 1439 if (delay_slot->is_MachSafePoint()) { 1440 MachNode *mach = delay_slot->as_Mach(); 1441 // !!!!! Stubs only need an oopmap right now, so bail out 1442 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) { 1443 // Write the oopmap directly to the code blob??!! 1444 delay_slot = NULL; 1445 continue; 1446 } 1447 1448 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1449 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1450 adjusted_offset); 1451 // Generate an OopMap entry 1452 Process_OopMap_Node(mach, adjusted_offset); 1453 } 1454 1455 // Insert the delay slot instruction 1456 delay_slot->emit(*cb, _regalloc); 1457 1458 // Don't reuse it 1459 delay_slot = NULL; 1460 } 1461 1462 } // End for all instructions in block 1463 1464 // If the next block is the top of a loop, pad this block out to align 1465 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1466 if (i < nblocks-1) { 1467 Block *nb = _cfg->get_block(i + 1); 1468 int padding = nb->alignment_padding(current_offset); 1469 if( padding > 0 ) { 1470 MachNode *nop = new MachNopNode(padding / nop_size); 1471 block->insert_node(nop, block->number_of_nodes()); 1472 _cfg->map_node_to_block(nop, block); 1473 nop->emit(*cb, _regalloc); 1474 current_offset = cb->insts_size(); 1475 } 1476 } 1477 // Verify that the distance for generated before forward 1478 // short branches is still valid. 1479 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size"); 1480 1481 // Save new block start offset 1482 blk_starts[i] = blk_offset; 1483 } // End of for all blocks 1484 blk_starts[nblocks] = current_offset; 1485 1486 non_safepoints.flush_at_end(); 1487 1488 // Offset too large? 1489 if (failing()) return; 1490 1491 // Define a pseudo-label at the end of the code 1492 MacroAssembler(cb).bind( blk_labels[nblocks] ); 1493 1494 // Compute the size of the first block 1495 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1496 1497 #ifdef ASSERT 1498 for (uint i = 0; i < nblocks; i++) { // For all blocks 1499 if (jmp_target[i] != 0) { 1500 int br_size = jmp_size[i]; 1501 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 1502 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 1503 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]); 1504 assert(false, "Displacement too large for short jmp"); 1505 } 1506 } 1507 } 1508 #endif 1509 1510 #ifndef PRODUCT 1511 // Information on the size of the method, without the extraneous code 1512 Scheduling::increment_method_size(cb->insts_size()); 1513 #endif 1514 1515 // ------------------ 1516 // Fill in exception table entries. 1517 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1518 1519 // Only java methods have exception handlers and deopt handlers 1520 // class HandlerImpl is platform-specific and defined in the *.ad files. 1521 if (_method) { 1522 // Emit the exception handler code. 1523 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb)); 1524 if (failing()) { 1525 return; // CodeBuffer::expand failed 1526 } 1527 // Emit the deopt handler code. 1528 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb)); 1529 1530 // Emit the MethodHandle deopt handler code (if required). 1531 if (has_method_handle_invokes() && !failing()) { 1532 // We can use the same code as for the normal deopt handler, we 1533 // just need a different entry point address. 1534 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb)); 1535 } 1536 } 1537 1538 // One last check for failed CodeBuffer::expand: 1539 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1540 C->record_failure("CodeCache is full"); 1541 return; 1542 } 1543 1544 #ifndef PRODUCT 1545 // Dump the assembly code, including basic-block numbers 1546 if (print_assembly()) { 1547 ttyLocker ttyl; // keep the following output all in one block 1548 if (!VMThread::should_terminate()) { // test this under the tty lock 1549 // This output goes directly to the tty, not the compiler log. 1550 // To enable tools to match it up with the compilation activity, 1551 // be sure to tag this tty output with the compile ID. 1552 if (xtty != NULL) { 1553 xtty->head("opto_assembly compile_id='%d'%s", compile_id(), 1554 is_osr_compilation() ? " compile_kind='osr'" : 1555 ""); 1556 } 1557 if (method() != NULL) { 1558 method()->print_metadata(); 1559 } else if (stub_name() != NULL) { 1560 tty->print_cr("Generating RuntimeStub - %s", stub_name()); 1561 } 1562 dump_asm(node_offsets, node_offset_limit); 1563 if (xtty != NULL) { 1564 // print_metadata and dump_asm above may safepoint which makes us loose the ttylock. 1565 // Retake lock too make sure the end tag is coherent, and that xmlStream->pop_tag is done 1566 // thread safe 1567 ttyLocker ttyl2; 1568 xtty->tail("opto_assembly"); 1569 } 1570 } 1571 } 1572 #endif 1573 1574 } 1575 1576 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1577 _inc_table.set_size(cnt); 1578 1579 uint inct_cnt = 0; 1580 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 1581 Block* block = _cfg->get_block(i); 1582 Node *n = NULL; 1583 int j; 1584 1585 // Find the branch; ignore trailing NOPs. 1586 for (j = block->number_of_nodes() - 1; j >= 0; j--) { 1587 n = block->get_node(j); 1588 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) { 1589 break; 1590 } 1591 } 1592 1593 // If we didn't find anything, continue 1594 if (j < 0) { 1595 continue; 1596 } 1597 1598 // Compute ExceptionHandlerTable subtable entry and add it 1599 // (skip empty blocks) 1600 if (n->is_Catch()) { 1601 1602 // Get the offset of the return from the call 1603 uint call_return = call_returns[block->_pre_order]; 1604 #ifdef ASSERT 1605 assert( call_return > 0, "no call seen for this basic block" ); 1606 while (block->get_node(--j)->is_MachProj()) ; 1607 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call"); 1608 #endif 1609 // last instruction is a CatchNode, find it's CatchProjNodes 1610 int nof_succs = block->_num_succs; 1611 // allocate space 1612 GrowableArray<intptr_t> handler_bcis(nof_succs); 1613 GrowableArray<intptr_t> handler_pcos(nof_succs); 1614 // iterate through all successors 1615 for (int j = 0; j < nof_succs; j++) { 1616 Block* s = block->_succs[j]; 1617 bool found_p = false; 1618 for (uint k = 1; k < s->num_preds(); k++) { 1619 Node* pk = s->pred(k); 1620 if (pk->is_CatchProj() && pk->in(0) == n) { 1621 const CatchProjNode* p = pk->as_CatchProj(); 1622 found_p = true; 1623 // add the corresponding handler bci & pco information 1624 if (p->_con != CatchProjNode::fall_through_index) { 1625 // p leads to an exception handler (and is not fall through) 1626 assert(s == _cfg->get_block(s->_pre_order), "bad numbering"); 1627 // no duplicates, please 1628 if (!handler_bcis.contains(p->handler_bci())) { 1629 uint block_num = s->non_connector()->_pre_order; 1630 handler_bcis.append(p->handler_bci()); 1631 handler_pcos.append(blk_labels[block_num].loc_pos()); 1632 } 1633 } 1634 } 1635 } 1636 assert(found_p, "no matching predecessor found"); 1637 // Note: Due to empty block removal, one block may have 1638 // several CatchProj inputs, from the same Catch. 1639 } 1640 1641 // Set the offset of the return from the call 1642 assert(handler_bcis.find(-1) != -1, "must have default handler"); 1643 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); 1644 continue; 1645 } 1646 1647 // Handle implicit null exception table updates 1648 if (n->is_MachNullCheck()) { 1649 uint block_num = block->non_connector_successor(0)->_pre_order; 1650 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1651 continue; 1652 } 1653 // Handle implicit exception table updates: trap instructions. 1654 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) { 1655 uint block_num = block->non_connector_successor(0)->_pre_order; 1656 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1657 continue; 1658 } 1659 } // End of for all blocks fill in exception table entries 1660 } 1661 1662 // Static Variables 1663 #ifndef PRODUCT 1664 uint Scheduling::_total_nop_size = 0; 1665 uint Scheduling::_total_method_size = 0; 1666 uint Scheduling::_total_branches = 0; 1667 uint Scheduling::_total_unconditional_delays = 0; 1668 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1669 #endif 1670 1671 // Initializer for class Scheduling 1672 1673 Scheduling::Scheduling(Arena *arena, Compile &compile) 1674 : _arena(arena), 1675 _cfg(compile.cfg()), 1676 _regalloc(compile.regalloc()), 1677 _scheduled(arena), 1678 _available(arena), 1679 _reg_node(arena), 1680 _pinch_free_list(arena), 1681 _next_node(NULL), 1682 _bundle_instr_count(0), 1683 _bundle_cycle_number(0), 1684 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]) 1685 #ifndef PRODUCT 1686 , _branches(0) 1687 , _unconditional_delays(0) 1688 #endif 1689 { 1690 // Create a MachNopNode 1691 _nop = new MachNopNode(); 1692 1693 // Now that the nops are in the array, save the count 1694 // (but allow entries for the nops) 1695 _node_bundling_limit = compile.unique(); 1696 uint node_max = _regalloc->node_regs_max_index(); 1697 1698 compile.set_node_bundling_limit(_node_bundling_limit); 1699 1700 // This one is persistent within the Compile class 1701 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1702 1703 // Allocate space for fixed-size arrays 1704 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1705 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1706 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1707 1708 // Clear the arrays 1709 for (uint i = 0; i < node_max; i++) { 1710 ::new (&_node_bundling_base[i]) Bundle(); 1711 } 1712 memset(_node_latency, 0, node_max * sizeof(unsigned short)); 1713 memset(_uses, 0, node_max * sizeof(short)); 1714 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 1715 1716 // Clear the bundling information 1717 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements)); 1718 1719 // Get the last node 1720 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1); 1721 1722 _next_node = block->get_node(block->number_of_nodes() - 1); 1723 } 1724 1725 #ifndef PRODUCT 1726 // Scheduling destructor 1727 Scheduling::~Scheduling() { 1728 _total_branches += _branches; 1729 _total_unconditional_delays += _unconditional_delays; 1730 } 1731 #endif 1732 1733 // Step ahead "i" cycles 1734 void Scheduling::step(uint i) { 1735 1736 Bundle *bundle = node_bundling(_next_node); 1737 bundle->set_starts_bundle(); 1738 1739 // Update the bundle record, but leave the flags information alone 1740 if (_bundle_instr_count > 0) { 1741 bundle->set_instr_count(_bundle_instr_count); 1742 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1743 } 1744 1745 // Update the state information 1746 _bundle_instr_count = 0; 1747 _bundle_cycle_number += i; 1748 _bundle_use.step(i); 1749 } 1750 1751 void Scheduling::step_and_clear() { 1752 Bundle *bundle = node_bundling(_next_node); 1753 bundle->set_starts_bundle(); 1754 1755 // Update the bundle record 1756 if (_bundle_instr_count > 0) { 1757 bundle->set_instr_count(_bundle_instr_count); 1758 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1759 1760 _bundle_cycle_number += 1; 1761 } 1762 1763 // Clear the bundling information 1764 _bundle_instr_count = 0; 1765 _bundle_use.reset(); 1766 1767 memcpy(_bundle_use_elements, 1768 Pipeline_Use::elaborated_elements, 1769 sizeof(Pipeline_Use::elaborated_elements)); 1770 } 1771 1772 // Perform instruction scheduling and bundling over the sequence of 1773 // instructions in backwards order. 1774 void Compile::ScheduleAndBundle() { 1775 1776 // Don't optimize this if it isn't a method 1777 if (!_method) 1778 return; 1779 1780 // Don't optimize this if scheduling is disabled 1781 if (!do_scheduling()) 1782 return; 1783 1784 // Scheduling code works only with pairs (16 bytes) maximum. 1785 if (max_vector_size() > 16) 1786 return; 1787 1788 TracePhase tp("isched", &timers[_t_instrSched]); 1789 1790 // Create a data structure for all the scheduling information 1791 Scheduling scheduling(Thread::current()->resource_area(), *this); 1792 1793 // Walk backwards over each basic block, computing the needed alignment 1794 // Walk over all the basic blocks 1795 scheduling.DoScheduling(); 1796 } 1797 1798 // Compute the latency of all the instructions. This is fairly simple, 1799 // because we already have a legal ordering. Walk over the instructions 1800 // from first to last, and compute the latency of the instruction based 1801 // on the latency of the preceding instruction(s). 1802 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { 1803 #ifndef PRODUCT 1804 if (_cfg->C->trace_opto_output()) 1805 tty->print("# -> ComputeLocalLatenciesForward\n"); 1806 #endif 1807 1808 // Walk over all the schedulable instructions 1809 for( uint j=_bb_start; j < _bb_end; j++ ) { 1810 1811 // This is a kludge, forcing all latency calculations to start at 1. 1812 // Used to allow latency 0 to force an instruction to the beginning 1813 // of the bb 1814 uint latency = 1; 1815 Node *use = bb->get_node(j); 1816 uint nlen = use->len(); 1817 1818 // Walk over all the inputs 1819 for ( uint k=0; k < nlen; k++ ) { 1820 Node *def = use->in(k); 1821 if (!def) 1822 continue; 1823 1824 uint l = _node_latency[def->_idx] + use->latency(k); 1825 if (latency < l) 1826 latency = l; 1827 } 1828 1829 _node_latency[use->_idx] = latency; 1830 1831 #ifndef PRODUCT 1832 if (_cfg->C->trace_opto_output()) { 1833 tty->print("# latency %4d: ", latency); 1834 use->dump(); 1835 } 1836 #endif 1837 } 1838 1839 #ifndef PRODUCT 1840 if (_cfg->C->trace_opto_output()) 1841 tty->print("# <- ComputeLocalLatenciesForward\n"); 1842 #endif 1843 1844 } // end ComputeLocalLatenciesForward 1845 1846 // See if this node fits into the present instruction bundle 1847 bool Scheduling::NodeFitsInBundle(Node *n) { 1848 uint n_idx = n->_idx; 1849 1850 // If this is the unconditional delay instruction, then it fits 1851 if (n == _unconditional_delay_slot) { 1852 #ifndef PRODUCT 1853 if (_cfg->C->trace_opto_output()) 1854 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 1855 #endif 1856 return (true); 1857 } 1858 1859 // If the node cannot be scheduled this cycle, skip it 1860 if (_current_latency[n_idx] > _bundle_cycle_number) { 1861 #ifndef PRODUCT 1862 if (_cfg->C->trace_opto_output()) 1863 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 1864 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 1865 #endif 1866 return (false); 1867 } 1868 1869 const Pipeline *node_pipeline = n->pipeline(); 1870 1871 uint instruction_count = node_pipeline->instructionCount(); 1872 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1873 instruction_count = 0; 1874 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1875 instruction_count++; 1876 1877 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 1878 #ifndef PRODUCT 1879 if (_cfg->C->trace_opto_output()) 1880 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 1881 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 1882 #endif 1883 return (false); 1884 } 1885 1886 // Don't allow non-machine nodes to be handled this way 1887 if (!n->is_Mach() && instruction_count == 0) 1888 return (false); 1889 1890 // See if there is any overlap 1891 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 1892 1893 if (delay > 0) { 1894 #ifndef PRODUCT 1895 if (_cfg->C->trace_opto_output()) 1896 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 1897 #endif 1898 return false; 1899 } 1900 1901 #ifndef PRODUCT 1902 if (_cfg->C->trace_opto_output()) 1903 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 1904 #endif 1905 1906 return true; 1907 } 1908 1909 Node * Scheduling::ChooseNodeToBundle() { 1910 uint siz = _available.size(); 1911 1912 if (siz == 0) { 1913 1914 #ifndef PRODUCT 1915 if (_cfg->C->trace_opto_output()) 1916 tty->print("# ChooseNodeToBundle: NULL\n"); 1917 #endif 1918 return (NULL); 1919 } 1920 1921 // Fast path, if only 1 instruction in the bundle 1922 if (siz == 1) { 1923 #ifndef PRODUCT 1924 if (_cfg->C->trace_opto_output()) { 1925 tty->print("# ChooseNodeToBundle (only 1): "); 1926 _available[0]->dump(); 1927 } 1928 #endif 1929 return (_available[0]); 1930 } 1931 1932 // Don't bother, if the bundle is already full 1933 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 1934 for ( uint i = 0; i < siz; i++ ) { 1935 Node *n = _available[i]; 1936 1937 // Skip projections, we'll handle them another way 1938 if (n->is_Proj()) 1939 continue; 1940 1941 // This presupposed that instructions are inserted into the 1942 // available list in a legality order; i.e. instructions that 1943 // must be inserted first are at the head of the list 1944 if (NodeFitsInBundle(n)) { 1945 #ifndef PRODUCT 1946 if (_cfg->C->trace_opto_output()) { 1947 tty->print("# ChooseNodeToBundle: "); 1948 n->dump(); 1949 } 1950 #endif 1951 return (n); 1952 } 1953 } 1954 } 1955 1956 // Nothing fits in this bundle, choose the highest priority 1957 #ifndef PRODUCT 1958 if (_cfg->C->trace_opto_output()) { 1959 tty->print("# ChooseNodeToBundle: "); 1960 _available[0]->dump(); 1961 } 1962 #endif 1963 1964 return _available[0]; 1965 } 1966 1967 void Scheduling::AddNodeToAvailableList(Node *n) { 1968 assert( !n->is_Proj(), "projections never directly made available" ); 1969 #ifndef PRODUCT 1970 if (_cfg->C->trace_opto_output()) { 1971 tty->print("# AddNodeToAvailableList: "); 1972 n->dump(); 1973 } 1974 #endif 1975 1976 int latency = _current_latency[n->_idx]; 1977 1978 // Insert in latency order (insertion sort) 1979 uint i; 1980 for ( i=0; i < _available.size(); i++ ) 1981 if (_current_latency[_available[i]->_idx] > latency) 1982 break; 1983 1984 // Special Check for compares following branches 1985 if( n->is_Mach() && _scheduled.size() > 0 ) { 1986 int op = n->as_Mach()->ideal_Opcode(); 1987 Node *last = _scheduled[0]; 1988 if( last->is_MachIf() && last->in(1) == n && 1989 ( op == Op_CmpI || 1990 op == Op_CmpU || 1991 op == Op_CmpUL || 1992 op == Op_CmpP || 1993 op == Op_CmpF || 1994 op == Op_CmpD || 1995 op == Op_CmpL ) ) { 1996 1997 // Recalculate position, moving to front of same latency 1998 for ( i=0 ; i < _available.size(); i++ ) 1999 if (_current_latency[_available[i]->_idx] >= latency) 2000 break; 2001 } 2002 } 2003 2004 // Insert the node in the available list 2005 _available.insert(i, n); 2006 2007 #ifndef PRODUCT 2008 if (_cfg->C->trace_opto_output()) 2009 dump_available(); 2010 #endif 2011 } 2012 2013 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 2014 for ( uint i=0; i < n->len(); i++ ) { 2015 Node *def = n->in(i); 2016 if (!def) continue; 2017 if( def->is_Proj() ) // If this is a machine projection, then 2018 def = def->in(0); // propagate usage thru to the base instruction 2019 2020 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local 2021 continue; 2022 } 2023 2024 // Compute the latency 2025 uint l = _bundle_cycle_number + n->latency(i); 2026 if (_current_latency[def->_idx] < l) 2027 _current_latency[def->_idx] = l; 2028 2029 // If this does not have uses then schedule it 2030 if ((--_uses[def->_idx]) == 0) 2031 AddNodeToAvailableList(def); 2032 } 2033 } 2034 2035 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 2036 #ifndef PRODUCT 2037 if (_cfg->C->trace_opto_output()) { 2038 tty->print("# AddNodeToBundle: "); 2039 n->dump(); 2040 } 2041 #endif 2042 2043 // Remove this from the available list 2044 uint i; 2045 for (i = 0; i < _available.size(); i++) 2046 if (_available[i] == n) 2047 break; 2048 assert(i < _available.size(), "entry in _available list not found"); 2049 _available.remove(i); 2050 2051 // See if this fits in the current bundle 2052 const Pipeline *node_pipeline = n->pipeline(); 2053 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 2054 2055 // Check for instructions to be placed in the delay slot. We 2056 // do this before we actually schedule the current instruction, 2057 // because the delay slot follows the current instruction. 2058 if (Pipeline::_branch_has_delay_slot && 2059 node_pipeline->hasBranchDelay() && 2060 !_unconditional_delay_slot) { 2061 2062 uint siz = _available.size(); 2063 2064 // Conditional branches can support an instruction that 2065 // is unconditionally executed and not dependent by the 2066 // branch, OR a conditionally executed instruction if 2067 // the branch is taken. In practice, this means that 2068 // the first instruction at the branch target is 2069 // copied to the delay slot, and the branch goes to 2070 // the instruction after that at the branch target 2071 if ( n->is_MachBranch() ) { 2072 2073 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 2074 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 2075 2076 #ifndef PRODUCT 2077 _branches++; 2078 #endif 2079 2080 // At least 1 instruction is on the available list 2081 // that is not dependent on the branch 2082 for (uint i = 0; i < siz; i++) { 2083 Node *d = _available[i]; 2084 const Pipeline *avail_pipeline = d->pipeline(); 2085 2086 // Don't allow safepoints in the branch shadow, that will 2087 // cause a number of difficulties 2088 if ( avail_pipeline->instructionCount() == 1 && 2089 !avail_pipeline->hasMultipleBundles() && 2090 !avail_pipeline->hasBranchDelay() && 2091 Pipeline::instr_has_unit_size() && 2092 d->size(_regalloc) == Pipeline::instr_unit_size() && 2093 NodeFitsInBundle(d) && 2094 !node_bundling(d)->used_in_delay()) { 2095 2096 if (d->is_Mach() && !d->is_MachSafePoint()) { 2097 // A node that fits in the delay slot was found, so we need to 2098 // set the appropriate bits in the bundle pipeline information so 2099 // that it correctly indicates resource usage. Later, when we 2100 // attempt to add this instruction to the bundle, we will skip 2101 // setting the resource usage. 2102 _unconditional_delay_slot = d; 2103 node_bundling(n)->set_use_unconditional_delay(); 2104 node_bundling(d)->set_used_in_unconditional_delay(); 2105 _bundle_use.add_usage(avail_pipeline->resourceUse()); 2106 _current_latency[d->_idx] = _bundle_cycle_number; 2107 _next_node = d; 2108 ++_bundle_instr_count; 2109 #ifndef PRODUCT 2110 _unconditional_delays++; 2111 #endif 2112 break; 2113 } 2114 } 2115 } 2116 } 2117 2118 // No delay slot, add a nop to the usage 2119 if (!_unconditional_delay_slot) { 2120 // See if adding an instruction in the delay slot will overflow 2121 // the bundle. 2122 if (!NodeFitsInBundle(_nop)) { 2123 #ifndef PRODUCT 2124 if (_cfg->C->trace_opto_output()) 2125 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2126 #endif 2127 step(1); 2128 } 2129 2130 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2131 _next_node = _nop; 2132 ++_bundle_instr_count; 2133 } 2134 2135 // See if the instruction in the delay slot requires a 2136 // step of the bundles 2137 if (!NodeFitsInBundle(n)) { 2138 #ifndef PRODUCT 2139 if (_cfg->C->trace_opto_output()) 2140 tty->print("# *** STEP(branch won't fit) ***\n"); 2141 #endif 2142 // Update the state information 2143 _bundle_instr_count = 0; 2144 _bundle_cycle_number += 1; 2145 _bundle_use.step(1); 2146 } 2147 } 2148 2149 // Get the number of instructions 2150 uint instruction_count = node_pipeline->instructionCount(); 2151 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2152 instruction_count = 0; 2153 2154 // Compute the latency information 2155 uint delay = 0; 2156 2157 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2158 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2159 if (relative_latency < 0) 2160 relative_latency = 0; 2161 2162 delay = _bundle_use.full_latency(relative_latency, node_usage); 2163 2164 // Does not fit in this bundle, start a new one 2165 if (delay > 0) { 2166 step(delay); 2167 2168 #ifndef PRODUCT 2169 if (_cfg->C->trace_opto_output()) 2170 tty->print("# *** STEP(%d) ***\n", delay); 2171 #endif 2172 } 2173 } 2174 2175 // If this was placed in the delay slot, ignore it 2176 if (n != _unconditional_delay_slot) { 2177 2178 if (delay == 0) { 2179 if (node_pipeline->hasMultipleBundles()) { 2180 #ifndef PRODUCT 2181 if (_cfg->C->trace_opto_output()) 2182 tty->print("# *** STEP(multiple instructions) ***\n"); 2183 #endif 2184 step(1); 2185 } 2186 2187 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2188 #ifndef PRODUCT 2189 if (_cfg->C->trace_opto_output()) 2190 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2191 instruction_count + _bundle_instr_count, 2192 Pipeline::_max_instrs_per_cycle); 2193 #endif 2194 step(1); 2195 } 2196 } 2197 2198 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2199 _bundle_instr_count++; 2200 2201 // Set the node's latency 2202 _current_latency[n->_idx] = _bundle_cycle_number; 2203 2204 // Now merge the functional unit information 2205 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2206 _bundle_use.add_usage(node_usage); 2207 2208 // Increment the number of instructions in this bundle 2209 _bundle_instr_count += instruction_count; 2210 2211 // Remember this node for later 2212 if (n->is_Mach()) 2213 _next_node = n; 2214 } 2215 2216 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2217 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2218 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2219 // into the block. All other scheduled nodes get put in the schedule here. 2220 int op = n->Opcode(); 2221 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2222 (op != Op_Node && // Not an unused antidepedence node and 2223 // not an unallocated boxlock 2224 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2225 2226 // Push any trailing projections 2227 if( bb->get_node(bb->number_of_nodes()-1) != n ) { 2228 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2229 Node *foi = n->fast_out(i); 2230 if( foi->is_Proj() ) 2231 _scheduled.push(foi); 2232 } 2233 } 2234 2235 // Put the instruction in the schedule list 2236 _scheduled.push(n); 2237 } 2238 2239 #ifndef PRODUCT 2240 if (_cfg->C->trace_opto_output()) 2241 dump_available(); 2242 #endif 2243 2244 // Walk all the definitions, decrementing use counts, and 2245 // if a definition has a 0 use count, place it in the available list. 2246 DecrementUseCounts(n,bb); 2247 } 2248 2249 // This method sets the use count within a basic block. We will ignore all 2250 // uses outside the current basic block. As we are doing a backwards walk, 2251 // any node we reach that has a use count of 0 may be scheduled. This also 2252 // avoids the problem of cyclic references from phi nodes, as long as phi 2253 // nodes are at the front of the basic block. This method also initializes 2254 // the available list to the set of instructions that have no uses within this 2255 // basic block. 2256 void Scheduling::ComputeUseCount(const Block *bb) { 2257 #ifndef PRODUCT 2258 if (_cfg->C->trace_opto_output()) 2259 tty->print("# -> ComputeUseCount\n"); 2260 #endif 2261 2262 // Clear the list of available and scheduled instructions, just in case 2263 _available.clear(); 2264 _scheduled.clear(); 2265 2266 // No delay slot specified 2267 _unconditional_delay_slot = NULL; 2268 2269 #ifdef ASSERT 2270 for( uint i=0; i < bb->number_of_nodes(); i++ ) 2271 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" ); 2272 #endif 2273 2274 // Force the _uses count to never go to zero for unscheduable pieces 2275 // of the block 2276 for( uint k = 0; k < _bb_start; k++ ) 2277 _uses[bb->get_node(k)->_idx] = 1; 2278 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ ) 2279 _uses[bb->get_node(l)->_idx] = 1; 2280 2281 // Iterate backwards over the instructions in the block. Don't count the 2282 // branch projections at end or the block header instructions. 2283 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2284 Node *n = bb->get_node(j); 2285 if( n->is_Proj() ) continue; // Projections handled another way 2286 2287 // Account for all uses 2288 for ( uint k = 0; k < n->len(); k++ ) { 2289 Node *inp = n->in(k); 2290 if (!inp) continue; 2291 assert(inp != n, "no cycles allowed" ); 2292 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use? 2293 if (inp->is_Proj()) { // Skip through Proj's 2294 inp = inp->in(0); 2295 } 2296 ++_uses[inp->_idx]; // Count 1 block-local use 2297 } 2298 } 2299 2300 // If this instruction has a 0 use count, then it is available 2301 if (!_uses[n->_idx]) { 2302 _current_latency[n->_idx] = _bundle_cycle_number; 2303 AddNodeToAvailableList(n); 2304 } 2305 2306 #ifndef PRODUCT 2307 if (_cfg->C->trace_opto_output()) { 2308 tty->print("# uses: %3d: ", _uses[n->_idx]); 2309 n->dump(); 2310 } 2311 #endif 2312 } 2313 2314 #ifndef PRODUCT 2315 if (_cfg->C->trace_opto_output()) 2316 tty->print("# <- ComputeUseCount\n"); 2317 #endif 2318 } 2319 2320 // This routine performs scheduling on each basic block in reverse order, 2321 // using instruction latencies and taking into account function unit 2322 // availability. 2323 void Scheduling::DoScheduling() { 2324 #ifndef PRODUCT 2325 if (_cfg->C->trace_opto_output()) 2326 tty->print("# -> DoScheduling\n"); 2327 #endif 2328 2329 Block *succ_bb = NULL; 2330 Block *bb; 2331 2332 // Walk over all the basic blocks in reverse order 2333 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) { 2334 bb = _cfg->get_block(i); 2335 2336 #ifndef PRODUCT 2337 if (_cfg->C->trace_opto_output()) { 2338 tty->print("# Schedule BB#%03d (initial)\n", i); 2339 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2340 bb->get_node(j)->dump(); 2341 } 2342 } 2343 #endif 2344 2345 // On the head node, skip processing 2346 if (bb == _cfg->get_root_block()) { 2347 continue; 2348 } 2349 2350 // Skip empty, connector blocks 2351 if (bb->is_connector()) 2352 continue; 2353 2354 // If the following block is not the sole successor of 2355 // this one, then reset the pipeline information 2356 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2357 #ifndef PRODUCT 2358 if (_cfg->C->trace_opto_output()) { 2359 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2360 _next_node->_idx, _bundle_instr_count); 2361 } 2362 #endif 2363 step_and_clear(); 2364 } 2365 2366 // Leave untouched the starting instruction, any Phis, a CreateEx node 2367 // or Top. bb->get_node(_bb_start) is the first schedulable instruction. 2368 _bb_end = bb->number_of_nodes()-1; 2369 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2370 Node *n = bb->get_node(_bb_start); 2371 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2372 // Also, MachIdealNodes do not get scheduled 2373 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2374 MachNode *mach = n->as_Mach(); 2375 int iop = mach->ideal_Opcode(); 2376 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2377 if( iop == Op_Con ) continue; // Do not schedule Top 2378 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2379 mach->pipeline() == MachNode::pipeline_class() && 2380 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc 2381 continue; 2382 break; // Funny loop structure to be sure... 2383 } 2384 // Compute last "interesting" instruction in block - last instruction we 2385 // might schedule. _bb_end points just after last schedulable inst. We 2386 // normally schedule conditional branches (despite them being forced last 2387 // in the block), because they have delay slots we can fill. Calls all 2388 // have their delay slots filled in the template expansions, so we don't 2389 // bother scheduling them. 2390 Node *last = bb->get_node(_bb_end); 2391 // Ignore trailing NOPs. 2392 while (_bb_end > 0 && last->is_Mach() && 2393 last->as_Mach()->ideal_Opcode() == Op_Con) { 2394 last = bb->get_node(--_bb_end); 2395 } 2396 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, ""); 2397 if( last->is_Catch() || 2398 // Exclude unreachable path case when Halt node is in a separate block. 2399 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2400 // There must be a prior call. Skip it. 2401 while( !bb->get_node(--_bb_end)->is_MachCall() ) { 2402 assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" ); 2403 } 2404 } else if( last->is_MachNullCheck() ) { 2405 // Backup so the last null-checked memory instruction is 2406 // outside the schedulable range. Skip over the nullcheck, 2407 // projection, and the memory nodes. 2408 Node *mem = last->in(1); 2409 do { 2410 _bb_end--; 2411 } while (mem != bb->get_node(_bb_end)); 2412 } else { 2413 // Set _bb_end to point after last schedulable inst. 2414 _bb_end++; 2415 } 2416 2417 assert( _bb_start <= _bb_end, "inverted block ends" ); 2418 2419 // Compute the register antidependencies for the basic block 2420 ComputeRegisterAntidependencies(bb); 2421 if (_cfg->C->failing()) return; // too many D-U pinch points 2422 2423 // Compute intra-bb latencies for the nodes 2424 ComputeLocalLatenciesForward(bb); 2425 2426 // Compute the usage within the block, and set the list of all nodes 2427 // in the block that have no uses within the block. 2428 ComputeUseCount(bb); 2429 2430 // Schedule the remaining instructions in the block 2431 while ( _available.size() > 0 ) { 2432 Node *n = ChooseNodeToBundle(); 2433 guarantee(n != NULL, "no nodes available"); 2434 AddNodeToBundle(n,bb); 2435 } 2436 2437 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2438 #ifdef ASSERT 2439 for( uint l = _bb_start; l < _bb_end; l++ ) { 2440 Node *n = bb->get_node(l); 2441 uint m; 2442 for( m = 0; m < _bb_end-_bb_start; m++ ) 2443 if( _scheduled[m] == n ) 2444 break; 2445 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2446 } 2447 #endif 2448 2449 // Now copy the instructions (in reverse order) back to the block 2450 for ( uint k = _bb_start; k < _bb_end; k++ ) 2451 bb->map_node(_scheduled[_bb_end-k-1], k); 2452 2453 #ifndef PRODUCT 2454 if (_cfg->C->trace_opto_output()) { 2455 tty->print("# Schedule BB#%03d (final)\n", i); 2456 uint current = 0; 2457 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2458 Node *n = bb->get_node(j); 2459 if( valid_bundle_info(n) ) { 2460 Bundle *bundle = node_bundling(n); 2461 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2462 tty->print("*** Bundle: "); 2463 bundle->dump(); 2464 } 2465 n->dump(); 2466 } 2467 } 2468 } 2469 #endif 2470 #ifdef ASSERT 2471 verify_good_schedule(bb,"after block local scheduling"); 2472 #endif 2473 } 2474 2475 #ifndef PRODUCT 2476 if (_cfg->C->trace_opto_output()) 2477 tty->print("# <- DoScheduling\n"); 2478 #endif 2479 2480 // Record final node-bundling array location 2481 _regalloc->C->set_node_bundling_base(_node_bundling_base); 2482 2483 } // end DoScheduling 2484 2485 // Verify that no live-range used in the block is killed in the block by a 2486 // wrong DEF. This doesn't verify live-ranges that span blocks. 2487 2488 // Check for edge existence. Used to avoid adding redundant precedence edges. 2489 static bool edge_from_to( Node *from, Node *to ) { 2490 for( uint i=0; i<from->len(); i++ ) 2491 if( from->in(i) == to ) 2492 return true; 2493 return false; 2494 } 2495 2496 #ifdef ASSERT 2497 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2498 // Check for bad kills 2499 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2500 Node *prior_use = _reg_node[def]; 2501 if( prior_use && !edge_from_to(prior_use,n) ) { 2502 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2503 n->dump(); 2504 tty->print_cr("..."); 2505 prior_use->dump(); 2506 assert(edge_from_to(prior_use,n), "%s", msg); 2507 } 2508 _reg_node.map(def,NULL); // Kill live USEs 2509 } 2510 } 2511 2512 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2513 2514 // Zap to something reasonable for the verify code 2515 _reg_node.clear(); 2516 2517 // Walk over the block backwards. Check to make sure each DEF doesn't 2518 // kill a live value (other than the one it's supposed to). Add each 2519 // USE to the live set. 2520 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) { 2521 Node *n = b->get_node(i); 2522 int n_op = n->Opcode(); 2523 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2524 // Fat-proj kills a slew of registers 2525 RegMask rm = n->out_RegMask();// Make local copy 2526 while( rm.is_NotEmpty() ) { 2527 OptoReg::Name kill = rm.find_first_elem(); 2528 rm.Remove(kill); 2529 verify_do_def( n, kill, msg ); 2530 } 2531 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2532 // Get DEF'd registers the normal way 2533 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2534 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2535 } 2536 2537 // Now make all USEs live 2538 for( uint i=1; i<n->req(); i++ ) { 2539 Node *def = n->in(i); 2540 assert(def != 0, "input edge required"); 2541 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2542 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2543 if( OptoReg::is_valid(reg_lo) ) { 2544 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg); 2545 _reg_node.map(reg_lo,n); 2546 } 2547 if( OptoReg::is_valid(reg_hi) ) { 2548 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg); 2549 _reg_node.map(reg_hi,n); 2550 } 2551 } 2552 2553 } 2554 2555 // Zap to something reasonable for the Antidependence code 2556 _reg_node.clear(); 2557 } 2558 #endif 2559 2560 // Conditionally add precedence edges. Avoid putting edges on Projs. 2561 static void add_prec_edge_from_to( Node *from, Node *to ) { 2562 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2563 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2564 from = from->in(0); 2565 } 2566 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2567 !edge_from_to( from, to ) ) // Avoid duplicate edge 2568 from->add_prec(to); 2569 } 2570 2571 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2572 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2573 return; 2574 2575 Node *pinch = _reg_node[def_reg]; // Get pinch point 2576 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet? 2577 is_def ) { // Check for a true def (not a kill) 2578 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2579 return; 2580 } 2581 2582 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2583 debug_only( def = (Node*)((intptr_t)0xdeadbeef); ) 2584 2585 // After some number of kills there _may_ be a later def 2586 Node *later_def = NULL; 2587 2588 // Finding a kill requires a real pinch-point. 2589 // Check for not already having a pinch-point. 2590 // Pinch points are Op_Node's. 2591 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2592 later_def = pinch; // Must be def/kill as optimistic pinch-point 2593 if ( _pinch_free_list.size() > 0) { 2594 pinch = _pinch_free_list.pop(); 2595 } else { 2596 pinch = new Node(1); // Pinch point to-be 2597 } 2598 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2599 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2600 return; 2601 } 2602 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init) 2603 _reg_node.map(def_reg,pinch); // Record pinch-point 2604 //_regalloc->set_bad(pinch->_idx); // Already initialized this way. 2605 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2606 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call 2607 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2608 later_def = NULL; // and no later def 2609 } 2610 pinch->set_req(0,later_def); // Hook later def so we can find it 2611 } else { // Else have valid pinch point 2612 if( pinch->in(0) ) // If there is a later-def 2613 later_def = pinch->in(0); // Get it 2614 } 2615 2616 // Add output-dependence edge from later def to kill 2617 if( later_def ) // If there is some original def 2618 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2619 2620 // See if current kill is also a use, and so is forced to be the pinch-point. 2621 if( pinch->Opcode() == Op_Node ) { 2622 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2623 for( uint i=1; i<uses->req(); i++ ) { 2624 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2625 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2626 // Yes, found a use/kill pinch-point 2627 pinch->set_req(0,NULL); // 2628 pinch->replace_by(kill); // Move anti-dep edges up 2629 pinch = kill; 2630 _reg_node.map(def_reg,pinch); 2631 return; 2632 } 2633 } 2634 } 2635 2636 // Add edge from kill to pinch-point 2637 add_prec_edge_from_to(kill,pinch); 2638 } 2639 2640 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2641 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2642 return; 2643 Node *pinch = _reg_node[use_reg]; // Get pinch point 2644 // Check for no later def_reg/kill in block 2645 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b && 2646 // Use has to be block-local as well 2647 _cfg->get_block_for_node(use) == b) { 2648 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2649 pinch->req() == 1 ) { // pinch not yet in block? 2650 pinch->del_req(0); // yank pointer to later-def, also set flag 2651 // Insert the pinch-point in the block just after the last use 2652 b->insert_node(pinch, b->find_node(use) + 1); 2653 _bb_end++; // Increase size scheduled region in block 2654 } 2655 2656 add_prec_edge_from_to(pinch,use); 2657 } 2658 } 2659 2660 // We insert antidependences between the reads and following write of 2661 // allocated registers to prevent illegal code motion. Hopefully, the 2662 // number of added references should be fairly small, especially as we 2663 // are only adding references within the current basic block. 2664 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2665 2666 #ifdef ASSERT 2667 verify_good_schedule(b,"before block local scheduling"); 2668 #endif 2669 2670 // A valid schedule, for each register independently, is an endless cycle 2671 // of: a def, then some uses (connected to the def by true dependencies), 2672 // then some kills (defs with no uses), finally the cycle repeats with a new 2673 // def. The uses are allowed to float relative to each other, as are the 2674 // kills. No use is allowed to slide past a kill (or def). This requires 2675 // antidependencies between all uses of a single def and all kills that 2676 // follow, up to the next def. More edges are redundant, because later defs 2677 // & kills are already serialized with true or antidependencies. To keep 2678 // the edge count down, we add a 'pinch point' node if there's more than 2679 // one use or more than one kill/def. 2680 2681 // We add dependencies in one bottom-up pass. 2682 2683 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2684 2685 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2686 // register. If not, we record the DEF/KILL in _reg_node, the 2687 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2688 // "pinch point", a new Node that's in the graph but not in the block. 2689 // We put edges from the prior and current DEF/KILLs to the pinch point. 2690 // We put the pinch point in _reg_node. If there's already a pinch point 2691 // we merely add an edge from the current DEF/KILL to the pinch point. 2692 2693 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2694 // put an edge from the pinch point to the USE. 2695 2696 // To be expedient, the _reg_node array is pre-allocated for the whole 2697 // compilation. _reg_node is lazily initialized; it either contains a NULL, 2698 // or a valid def/kill/pinch-point, or a leftover node from some prior 2699 // block. Leftover node from some prior block is treated like a NULL (no 2700 // prior def, so no anti-dependence needed). Valid def is distinguished by 2701 // it being in the current block. 2702 bool fat_proj_seen = false; 2703 uint last_safept = _bb_end-1; 2704 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL; 2705 Node* last_safept_node = end_node; 2706 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2707 Node *n = b->get_node(i); 2708 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2709 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) { 2710 // Fat-proj kills a slew of registers 2711 // This can add edges to 'n' and obscure whether or not it was a def, 2712 // hence the is_def flag. 2713 fat_proj_seen = true; 2714 RegMask rm = n->out_RegMask();// Make local copy 2715 while( rm.is_NotEmpty() ) { 2716 OptoReg::Name kill = rm.find_first_elem(); 2717 rm.Remove(kill); 2718 anti_do_def( b, n, kill, is_def ); 2719 } 2720 } else { 2721 // Get DEF'd registers the normal way 2722 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2723 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2724 } 2725 2726 // Kill projections on a branch should appear to occur on the 2727 // branch, not afterwards, so grab the masks from the projections 2728 // and process them. 2729 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) { 2730 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2731 Node* use = n->fast_out(i); 2732 if (use->is_Proj()) { 2733 RegMask rm = use->out_RegMask();// Make local copy 2734 while( rm.is_NotEmpty() ) { 2735 OptoReg::Name kill = rm.find_first_elem(); 2736 rm.Remove(kill); 2737 anti_do_def( b, n, kill, false ); 2738 } 2739 } 2740 } 2741 } 2742 2743 // Check each register used by this instruction for a following DEF/KILL 2744 // that must occur afterward and requires an anti-dependence edge. 2745 for( uint j=0; j<n->req(); j++ ) { 2746 Node *def = n->in(j); 2747 if( def ) { 2748 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2749 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2750 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2751 } 2752 } 2753 // Do not allow defs of new derived values to float above GC 2754 // points unless the base is definitely available at the GC point. 2755 2756 Node *m = b->get_node(i); 2757 2758 // Add precedence edge from following safepoint to use of derived pointer 2759 if( last_safept_node != end_node && 2760 m != last_safept_node) { 2761 for (uint k = 1; k < m->req(); k++) { 2762 const Type *t = m->in(k)->bottom_type(); 2763 if( t->isa_oop_ptr() && 2764 t->is_ptr()->offset() != 0 ) { 2765 last_safept_node->add_prec( m ); 2766 break; 2767 } 2768 } 2769 } 2770 2771 if( n->jvms() ) { // Precedence edge from derived to safept 2772 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2773 if( b->get_node(last_safept) != last_safept_node ) { 2774 last_safept = b->find_node(last_safept_node); 2775 } 2776 for( uint j=last_safept; j > i; j-- ) { 2777 Node *mach = b->get_node(j); 2778 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2779 mach->add_prec( n ); 2780 } 2781 last_safept = i; 2782 last_safept_node = m; 2783 } 2784 } 2785 2786 if (fat_proj_seen) { 2787 // Garbage collect pinch nodes that were not consumed. 2788 // They are usually created by a fat kill MachProj for a call. 2789 garbage_collect_pinch_nodes(); 2790 } 2791 } 2792 2793 // Garbage collect pinch nodes for reuse by other blocks. 2794 // 2795 // The block scheduler's insertion of anti-dependence 2796 // edges creates many pinch nodes when the block contains 2797 // 2 or more Calls. A pinch node is used to prevent a 2798 // combinatorial explosion of edges. If a set of kills for a 2799 // register is anti-dependent on a set of uses (or defs), rather 2800 // than adding an edge in the graph between each pair of kill 2801 // and use (or def), a pinch is inserted between them: 2802 // 2803 // use1 use2 use3 2804 // \ | / 2805 // \ | / 2806 // pinch 2807 // / | \ 2808 // / | \ 2809 // kill1 kill2 kill3 2810 // 2811 // One pinch node is created per register killed when 2812 // the second call is encountered during a backwards pass 2813 // over the block. Most of these pinch nodes are never 2814 // wired into the graph because the register is never 2815 // used or def'ed in the block. 2816 // 2817 void Scheduling::garbage_collect_pinch_nodes() { 2818 #ifndef PRODUCT 2819 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 2820 #endif 2821 int trace_cnt = 0; 2822 for (uint k = 0; k < _reg_node.Size(); k++) { 2823 Node* pinch = _reg_node[k]; 2824 if ((pinch != NULL) && pinch->Opcode() == Op_Node && 2825 // no predecence input edges 2826 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { 2827 cleanup_pinch(pinch); 2828 _pinch_free_list.push(pinch); 2829 _reg_node.map(k, NULL); 2830 #ifndef PRODUCT 2831 if (_cfg->C->trace_opto_output()) { 2832 trace_cnt++; 2833 if (trace_cnt > 40) { 2834 tty->print("\n"); 2835 trace_cnt = 0; 2836 } 2837 tty->print(" %d", pinch->_idx); 2838 } 2839 #endif 2840 } 2841 } 2842 #ifndef PRODUCT 2843 if (_cfg->C->trace_opto_output()) tty->print("\n"); 2844 #endif 2845 } 2846 2847 // Clean up a pinch node for reuse. 2848 void Scheduling::cleanup_pinch( Node *pinch ) { 2849 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 2850 2851 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 2852 Node* use = pinch->last_out(i); 2853 uint uses_found = 0; 2854 for (uint j = use->req(); j < use->len(); j++) { 2855 if (use->in(j) == pinch) { 2856 use->rm_prec(j); 2857 uses_found++; 2858 } 2859 } 2860 assert(uses_found > 0, "must be a precedence edge"); 2861 i -= uses_found; // we deleted 1 or more copies of this edge 2862 } 2863 // May have a later_def entry 2864 pinch->set_req(0, NULL); 2865 } 2866 2867 #ifndef PRODUCT 2868 2869 void Scheduling::dump_available() const { 2870 tty->print("#Availist "); 2871 for (uint i = 0; i < _available.size(); i++) 2872 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 2873 tty->cr(); 2874 } 2875 2876 // Print Scheduling Statistics 2877 void Scheduling::print_statistics() { 2878 // Print the size added by nops for bundling 2879 tty->print("Nops added %d bytes to total of %d bytes", 2880 _total_nop_size, _total_method_size); 2881 if (_total_method_size > 0) 2882 tty->print(", for %.2f%%", 2883 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 2884 tty->print("\n"); 2885 2886 // Print the number of branch shadows filled 2887 if (Pipeline::_branch_has_delay_slot) { 2888 tty->print("Of %d branches, %d had unconditional delay slots filled", 2889 _total_branches, _total_unconditional_delays); 2890 if (_total_branches > 0) 2891 tty->print(", for %.2f%%", 2892 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 2893 tty->print("\n"); 2894 } 2895 2896 uint total_instructions = 0, total_bundles = 0; 2897 2898 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 2899 uint bundle_count = _total_instructions_per_bundle[i]; 2900 total_instructions += bundle_count * i; 2901 total_bundles += bundle_count; 2902 } 2903 2904 if (total_bundles > 0) 2905 tty->print("Average ILP (excluding nops) is %.2f\n", 2906 ((double)total_instructions) / ((double)total_bundles)); 2907 } 2908 #endif