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/disassembler.hpp" 34 #include "compiler/oopMap.hpp" 35 #include "memory/allocation.inline.hpp" 36 #include "opto/ad.hpp" 37 #include "opto/callnode.hpp" 38 #include "opto/cfgnode.hpp" 39 #include "opto/locknode.hpp" 40 #include "opto/machnode.hpp" 41 #include "opto/optoreg.hpp" 42 #include "opto/output.hpp" 43 #include "opto/regalloc.hpp" 44 #include "opto/runtime.hpp" 45 #include "opto/subnode.hpp" 46 #include "opto/type.hpp" 47 #include "runtime/handles.inline.hpp" 48 #include "utilities/xmlstream.hpp" 49 50 #ifndef PRODUCT 51 #define DEBUG_ARG(x) , x 52 #else 53 #define DEBUG_ARG(x) 54 #endif 55 56 // Convert Nodes to instruction bits and pass off to the VM 57 void Compile::Output() { 58 // RootNode goes 59 assert( _cfg->get_root_block()->number_of_nodes() == 0, "" ); 60 61 // The number of new nodes (mostly MachNop) is proportional to 62 // the number of java calls and inner loops which are aligned. 63 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 + 64 C->inner_loops()*(OptoLoopAlignment-1)), 65 "out of nodes before code generation" ) ) { 66 return; 67 } 68 // Make sure I can find the Start Node 69 Block *entry = _cfg->get_block(1); 70 Block *broot = _cfg->get_root_block(); 71 72 const StartNode *start = entry->head()->as_Start(); 73 74 // Replace StartNode with prolog 75 MachPrologNode *prolog = new MachPrologNode(); 76 entry->map_node(prolog, 0); 77 _cfg->map_node_to_block(prolog, entry); 78 _cfg->unmap_node_from_block(start); // start is no longer in any block 79 80 // Virtual methods need an unverified entry point 81 82 if( is_osr_compilation() ) { 83 if( PoisonOSREntry ) { 84 // TODO: Should use a ShouldNotReachHereNode... 85 _cfg->insert( broot, 0, new MachBreakpointNode() ); 86 } 87 } else { 88 if( _method && !_method->flags().is_static() ) { 89 // Insert unvalidated entry point 90 _cfg->insert( broot, 0, new MachUEPNode() ); 91 } 92 93 } 94 95 // Break before main entry point 96 if ((_method && C->directive()->BreakAtExecuteOption) || 97 (OptoBreakpoint && is_method_compilation()) || 98 (OptoBreakpointOSR && is_osr_compilation()) || 99 (OptoBreakpointC2R && !_method) ) { 100 // checking for _method means that OptoBreakpoint does not apply to 101 // runtime stubs or frame converters 102 _cfg->insert( entry, 1, new MachBreakpointNode() ); 103 } 104 105 // Insert epilogs before every return 106 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 107 Block* block = _cfg->get_block(i); 108 if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point? 109 Node* m = block->end(); 110 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) { 111 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); 112 block->add_inst(epilog); 113 _cfg->map_node_to_block(epilog, block); 114 } 115 } 116 } 117 118 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1); 119 blk_starts[0] = 0; 120 121 // Initialize code buffer and process short branches. 122 CodeBuffer* cb = init_buffer(blk_starts); 123 124 if (cb == NULL || failing()) { 125 return; 126 } 127 128 ScheduleAndBundle(); 129 130 #ifndef PRODUCT 131 if (trace_opto_output()) { 132 tty->print("\n---- After ScheduleAndBundle ----\n"); 133 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 134 tty->print("\nBB#%03d:\n", i); 135 Block* block = _cfg->get_block(i); 136 for (uint j = 0; j < block->number_of_nodes(); j++) { 137 Node* n = block->get_node(j); 138 OptoReg::Name reg = _regalloc->get_reg_first(n); 139 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : ""); 140 n->dump(); 141 } 142 } 143 } 144 #endif 145 146 if (failing()) { 147 return; 148 } 149 150 BuildOopMaps(); 151 152 if (failing()) { 153 return; 154 } 155 156 fill_buffer(cb, blk_starts); 157 } 158 159 bool Compile::need_stack_bang(int frame_size_in_bytes) const { 160 // Determine if we need to generate a stack overflow check. 161 // Do it if the method is not a stub function and 162 // has java calls or has frame size > vm_page_size/8. 163 // The debug VM checks that deoptimization doesn't trigger an 164 // unexpected stack overflow (compiled method stack banging should 165 // guarantee it doesn't happen) so we always need the stack bang in 166 // a debug VM. 167 return (UseStackBanging && stub_function() == NULL && 168 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3 169 DEBUG_ONLY(|| true))); 170 } 171 172 bool Compile::need_register_stack_bang() const { 173 // Determine if we need to generate a register stack overflow check. 174 // This is only used on architectures which have split register 175 // and memory stacks (ie. IA64). 176 // Bang if the method is not a stub function and has java calls 177 return (stub_function() == NULL && has_java_calls()); 178 } 179 180 181 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top 182 // of a loop. When aligning a loop we need to provide enough instructions 183 // in cpu's fetch buffer to feed decoders. The loop alignment could be 184 // avoided if we have enough instructions in fetch buffer at the head of a loop. 185 // By default, the size is set to 999999 by Block's constructor so that 186 // a loop will be aligned if the size is not reset here. 187 // 188 // Note: Mach instructions could contain several HW instructions 189 // so the size is estimated only. 190 // 191 void Compile::compute_loop_first_inst_sizes() { 192 // The next condition is used to gate the loop alignment optimization. 193 // Don't aligned a loop if there are enough instructions at the head of a loop 194 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad 195 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is 196 // equal to 11 bytes which is the largest address NOP instruction. 197 if (MaxLoopPad < OptoLoopAlignment - 1) { 198 uint last_block = _cfg->number_of_blocks() - 1; 199 for (uint i = 1; i <= last_block; i++) { 200 Block* block = _cfg->get_block(i); 201 // Check the first loop's block which requires an alignment. 202 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) { 203 uint sum_size = 0; 204 uint inst_cnt = NumberOfLoopInstrToAlign; 205 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 206 207 // Check subsequent fallthrough blocks if the loop's first 208 // block(s) does not have enough instructions. 209 Block *nb = block; 210 while(inst_cnt > 0 && 211 i < last_block && 212 !_cfg->get_block(i + 1)->has_loop_alignment() && 213 !nb->has_successor(block)) { 214 i++; 215 nb = _cfg->get_block(i); 216 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 217 } // while( inst_cnt > 0 && i < last_block ) 218 219 block->set_first_inst_size(sum_size); 220 } // f( b->head()->is_Loop() ) 221 } // for( i <= last_block ) 222 } // if( MaxLoopPad < OptoLoopAlignment-1 ) 223 } 224 225 // The architecture description provides short branch variants for some long 226 // branch instructions. Replace eligible long branches with short branches. 227 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) { 228 // Compute size of each block, method size, and relocation information size 229 uint nblocks = _cfg->number_of_blocks(); 230 231 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 232 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 233 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks); 234 235 // Collect worst case block paddings 236 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks); 237 memset(block_worst_case_pad, 0, nblocks * sizeof(int)); 238 239 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); ) 240 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); ) 241 242 bool has_short_branch_candidate = false; 243 244 // Initialize the sizes to 0 245 code_size = 0; // Size in bytes of generated code 246 stub_size = 0; // Size in bytes of all stub entries 247 // Size in bytes of all relocation entries, including those in local stubs. 248 // Start with 2-bytes of reloc info for the unvalidated entry point 249 reloc_size = 1; // Number of relocation entries 250 251 // Make three passes. The first computes pessimistic blk_starts, 252 // relative jmp_offset and reloc_size information. The second performs 253 // short branch substitution using the pessimistic sizing. The 254 // third inserts nops where needed. 255 256 // Step one, perform a pessimistic sizing pass. 257 uint last_call_adr = max_juint; 258 uint last_avoid_back_to_back_adr = max_juint; 259 uint nop_size = (new MachNopNode())->size(_regalloc); 260 for (uint i = 0; i < nblocks; i++) { // For all blocks 261 Block* block = _cfg->get_block(i); 262 263 // During short branch replacement, we store the relative (to blk_starts) 264 // offset of jump in jmp_offset, rather than the absolute offset of jump. 265 // This is so that we do not need to recompute sizes of all nodes when 266 // we compute correct blk_starts in our next sizing pass. 267 jmp_offset[i] = 0; 268 jmp_size[i] = 0; 269 jmp_nidx[i] = -1; 270 DEBUG_ONLY( jmp_target[i] = 0; ) 271 DEBUG_ONLY( jmp_rule[i] = 0; ) 272 273 // Sum all instruction sizes to compute block size 274 uint last_inst = block->number_of_nodes(); 275 uint blk_size = 0; 276 for (uint j = 0; j < last_inst; j++) { 277 Node* nj = block->get_node(j); 278 // Handle machine instruction nodes 279 if (nj->is_Mach()) { 280 MachNode *mach = nj->as_Mach(); 281 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 282 reloc_size += mach->reloc(); 283 if (mach->is_MachCall()) { 284 // add size information for trampoline stub 285 // class CallStubImpl is platform-specific and defined in the *.ad files. 286 stub_size += CallStubImpl::size_call_trampoline(); 287 reloc_size += CallStubImpl::reloc_call_trampoline(); 288 289 MachCallNode *mcall = mach->as_MachCall(); 290 // This destination address is NOT PC-relative 291 292 mcall->method_set((intptr_t)mcall->entry_point()); 293 294 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) { 295 stub_size += CompiledStaticCall::to_interp_stub_size(); 296 reloc_size += CompiledStaticCall::reloc_to_interp_stub(); 297 #if INCLUDE_AOT 298 stub_size += CompiledStaticCall::to_aot_stub_size(); 299 reloc_size += CompiledStaticCall::reloc_to_aot_stub(); 300 #endif 301 } 302 } else if (mach->is_MachSafePoint()) { 303 // If call/safepoint are adjacent, account for possible 304 // nop to disambiguate the two safepoints. 305 // ScheduleAndBundle() can rearrange nodes in a block, 306 // check for all offsets inside this block. 307 if (last_call_adr >= blk_starts[i]) { 308 blk_size += nop_size; 309 } 310 } 311 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 312 // Nop is inserted between "avoid back to back" instructions. 313 // ScheduleAndBundle() can rearrange nodes in a block, 314 // check for all offsets inside this block. 315 if (last_avoid_back_to_back_adr >= blk_starts[i]) { 316 blk_size += nop_size; 317 } 318 } 319 if (mach->may_be_short_branch()) { 320 if (!nj->is_MachBranch()) { 321 #ifndef PRODUCT 322 nj->dump(3); 323 #endif 324 Unimplemented(); 325 } 326 assert(jmp_nidx[i] == -1, "block should have only one branch"); 327 jmp_offset[i] = blk_size; 328 jmp_size[i] = nj->size(_regalloc); 329 jmp_nidx[i] = j; 330 has_short_branch_candidate = true; 331 } 332 } 333 blk_size += nj->size(_regalloc); 334 // Remember end of call offset 335 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) { 336 last_call_adr = blk_starts[i]+blk_size; 337 } 338 // Remember end of avoid_back_to_back offset 339 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 340 last_avoid_back_to_back_adr = blk_starts[i]+blk_size; 341 } 342 } 343 344 // When the next block starts a loop, we may insert pad NOP 345 // instructions. Since we cannot know our future alignment, 346 // assume the worst. 347 if (i < nblocks - 1) { 348 Block* nb = _cfg->get_block(i + 1); 349 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 350 if (max_loop_pad > 0) { 351 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 352 // Adjust last_call_adr and/or last_avoid_back_to_back_adr. 353 // If either is the last instruction in this block, bump by 354 // max_loop_pad in lock-step with blk_size, so sizing 355 // calculations in subsequent blocks still can conservatively 356 // detect that it may the last instruction in this block. 357 if (last_call_adr == blk_starts[i]+blk_size) { 358 last_call_adr += max_loop_pad; 359 } 360 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) { 361 last_avoid_back_to_back_adr += max_loop_pad; 362 } 363 blk_size += max_loop_pad; 364 block_worst_case_pad[i + 1] = max_loop_pad; 365 } 366 } 367 368 // Save block size; update total method size 369 blk_starts[i+1] = blk_starts[i]+blk_size; 370 } 371 372 // Step two, replace eligible long jumps. 373 bool progress = true; 374 uint last_may_be_short_branch_adr = max_juint; 375 while (has_short_branch_candidate && progress) { 376 progress = false; 377 has_short_branch_candidate = false; 378 int adjust_block_start = 0; 379 for (uint i = 0; i < nblocks; i++) { 380 Block* block = _cfg->get_block(i); 381 int idx = jmp_nidx[i]; 382 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach(); 383 if (mach != NULL && mach->may_be_short_branch()) { 384 #ifdef ASSERT 385 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity"); 386 int j; 387 // Find the branch; ignore trailing NOPs. 388 for (j = block->number_of_nodes()-1; j>=0; j--) { 389 Node* n = block->get_node(j); 390 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) 391 break; 392 } 393 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity"); 394 #endif 395 int br_size = jmp_size[i]; 396 int br_offs = blk_starts[i] + jmp_offset[i]; 397 398 // This requires the TRUE branch target be in succs[0] 399 uint bnum = block->non_connector_successor(0)->_pre_order; 400 int offset = blk_starts[bnum] - br_offs; 401 if (bnum > i) { // adjust following block's offset 402 offset -= adjust_block_start; 403 } 404 405 // This block can be a loop header, account for the padding 406 // in the previous block. 407 int block_padding = block_worst_case_pad[i]; 408 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top"); 409 // In the following code a nop could be inserted before 410 // the branch which will increase the backward distance. 411 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr); 412 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block"); 413 414 if (needs_padding && offset <= 0) 415 offset -= nop_size; 416 417 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 418 // We've got a winner. Replace this branch. 419 MachNode* replacement = mach->as_MachBranch()->short_branch_version(); 420 421 // Update the jmp_size. 422 int new_size = replacement->size(_regalloc); 423 int diff = br_size - new_size; 424 assert(diff >= (int)nop_size, "short_branch size should be smaller"); 425 // Conservatively take into account padding between 426 // avoid_back_to_back branches. Previous branch could be 427 // converted into avoid_back_to_back branch during next 428 // rounds. 429 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 430 jmp_offset[i] += nop_size; 431 diff -= nop_size; 432 } 433 adjust_block_start += diff; 434 block->map_node(replacement, idx); 435 mach->subsume_by(replacement, C); 436 mach = replacement; 437 progress = true; 438 439 jmp_size[i] = new_size; 440 DEBUG_ONLY( jmp_target[i] = bnum; ); 441 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 442 } else { 443 // The jump distance is not short, try again during next iteration. 444 has_short_branch_candidate = true; 445 } 446 } // (mach->may_be_short_branch()) 447 if (mach != NULL && (mach->may_be_short_branch() || 448 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) { 449 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i]; 450 } 451 blk_starts[i+1] -= adjust_block_start; 452 } 453 } 454 455 #ifdef ASSERT 456 for (uint i = 0; i < nblocks; i++) { // For all blocks 457 if (jmp_target[i] != 0) { 458 int br_size = jmp_size[i]; 459 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 460 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 461 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]); 462 } 463 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp"); 464 } 465 } 466 #endif 467 468 // Step 3, compute the offsets of all blocks, will be done in fill_buffer() 469 // after ScheduleAndBundle(). 470 471 // ------------------ 472 // Compute size for code buffer 473 code_size = blk_starts[nblocks]; 474 475 // Relocation records 476 reloc_size += 1; // Relo entry for exception handler 477 478 // Adjust reloc_size to number of record of relocation info 479 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 480 // a relocation index. 481 // The CodeBuffer will expand the locs array if this estimate is too low. 482 reloc_size *= 10 / sizeof(relocInfo); 483 } 484 485 //------------------------------FillLocArray----------------------------------- 486 // Create a bit of debug info and append it to the array. The mapping is from 487 // Java local or expression stack to constant, register or stack-slot. For 488 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 489 // entry has been taken care of and caller should skip it). 490 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 491 // This should never have accepted Bad before 492 assert(OptoReg::is_valid(regnum), "location must be valid"); 493 return (OptoReg::is_reg(regnum)) 494 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 495 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 496 } 497 498 499 ObjectValue* 500 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 501 for (int i = 0; i < objs->length(); i++) { 502 assert(objs->at(i)->is_object(), "corrupt object cache"); 503 ObjectValue* sv = (ObjectValue*) objs->at(i); 504 if (sv->id() == id) { 505 return sv; 506 } 507 } 508 // Otherwise.. 509 return NULL; 510 } 511 512 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 513 ObjectValue* sv ) { 514 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition"); 515 objs->append(sv); 516 } 517 518 519 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 520 GrowableArray<ScopeValue*> *array, 521 GrowableArray<ScopeValue*> *objs ) { 522 assert( local, "use _top instead of null" ); 523 if (array->length() != idx) { 524 assert(array->length() == idx + 1, "Unexpected array count"); 525 // Old functionality: 526 // return 527 // New functionality: 528 // Assert if the local is not top. In product mode let the new node 529 // override the old entry. 530 assert(local == top(), "LocArray collision"); 531 if (local == top()) { 532 return; 533 } 534 array->pop(); 535 } 536 const Type *t = local->bottom_type(); 537 538 // Is it a safepoint scalar object node? 539 if (local->is_SafePointScalarObject()) { 540 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 541 542 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx); 543 if (sv == NULL) { 544 ciKlass* cik = t->is_oopptr()->klass(); 545 assert(cik->is_instance_klass() || 546 cik->is_array_klass(), "Not supported allocation."); 547 sv = new ObjectValue(spobj->_idx, 548 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 549 Compile::set_sv_for_object_node(objs, sv); 550 551 uint first_ind = spobj->first_index(sfpt->jvms()); 552 for (uint i = 0; i < spobj->n_fields(); i++) { 553 Node* fld_node = sfpt->in(first_ind+i); 554 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 555 } 556 } 557 array->append(sv); 558 return; 559 } 560 561 // Grab the register number for the local 562 OptoReg::Name regnum = _regalloc->get_reg_first(local); 563 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 564 // Record the double as two float registers. 565 // The register mask for such a value always specifies two adjacent 566 // float registers, with the lower register number even. 567 // Normally, the allocation of high and low words to these registers 568 // is irrelevant, because nearly all operations on register pairs 569 // (e.g., StoreD) treat them as a single unit. 570 // Here, we assume in addition that the words in these two registers 571 // stored "naturally" (by operations like StoreD and double stores 572 // within the interpreter) such that the lower-numbered register 573 // is written to the lower memory address. This may seem like 574 // a machine dependency, but it is not--it is a requirement on 575 // the author of the <arch>.ad file to ensure that, for every 576 // even/odd double-register pair to which a double may be allocated, 577 // the word in the even single-register is stored to the first 578 // memory word. (Note that register numbers are completely 579 // arbitrary, and are not tied to any machine-level encodings.) 580 #ifdef _LP64 581 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 582 array->append(new ConstantIntValue((jint)0)); 583 array->append(new_loc_value( _regalloc, regnum, Location::dbl )); 584 } else if ( t->base() == Type::Long ) { 585 array->append(new ConstantIntValue((jint)0)); 586 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 587 } else if ( t->base() == Type::RawPtr ) { 588 // jsr/ret return address which must be restored into a the full 589 // width 64-bit stack slot. 590 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 591 } 592 #else //_LP64 593 #ifdef SPARC 594 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) { 595 // For SPARC we have to swap high and low words for 596 // long values stored in a single-register (g0-g7). 597 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 598 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 599 } else 600 #endif //SPARC 601 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 602 // Repack the double/long as two jints. 603 // The convention the interpreter uses is that the second local 604 // holds the first raw word of the native double representation. 605 // This is actually reasonable, since locals and stack arrays 606 // grow downwards in all implementations. 607 // (If, on some machine, the interpreter's Java locals or stack 608 // were to grow upwards, the embedded doubles would be word-swapped.) 609 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 610 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 611 } 612 #endif //_LP64 613 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 614 OptoReg::is_reg(regnum) ) { 615 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double() 616 ? Location::float_in_dbl : Location::normal )); 617 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 618 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long 619 ? Location::int_in_long : Location::normal )); 620 } else if( t->base() == Type::NarrowOop ) { 621 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop )); 622 } else { 623 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal )); 624 } 625 return; 626 } 627 628 // No register. It must be constant data. 629 switch (t->base()) { 630 case Type::Half: // Second half of a double 631 ShouldNotReachHere(); // Caller should skip 2nd halves 632 break; 633 case Type::AnyPtr: 634 array->append(new ConstantOopWriteValue(NULL)); 635 break; 636 case Type::AryPtr: 637 case Type::InstPtr: // fall through 638 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 639 break; 640 case Type::NarrowOop: 641 if (t == TypeNarrowOop::NULL_PTR) { 642 array->append(new ConstantOopWriteValue(NULL)); 643 } else { 644 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 645 } 646 break; 647 case Type::Int: 648 array->append(new ConstantIntValue(t->is_int()->get_con())); 649 break; 650 case Type::RawPtr: 651 // A return address (T_ADDRESS). 652 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 653 #ifdef _LP64 654 // Must be restored to the full-width 64-bit stack slot. 655 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 656 #else 657 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 658 #endif 659 break; 660 case Type::FloatCon: { 661 float f = t->is_float_constant()->getf(); 662 array->append(new ConstantIntValue(jint_cast(f))); 663 break; 664 } 665 case Type::DoubleCon: { 666 jdouble d = t->is_double_constant()->getd(); 667 #ifdef _LP64 668 array->append(new ConstantIntValue((jint)0)); 669 array->append(new ConstantDoubleValue(d)); 670 #else 671 // Repack the double as two jints. 672 // The convention the interpreter uses is that the second local 673 // holds the first raw word of the native double representation. 674 // This is actually reasonable, since locals and stack arrays 675 // grow downwards in all implementations. 676 // (If, on some machine, the interpreter's Java locals or stack 677 // were to grow upwards, the embedded doubles would be word-swapped.) 678 jlong_accessor acc; 679 acc.long_value = jlong_cast(d); 680 array->append(new ConstantIntValue(acc.words[1])); 681 array->append(new ConstantIntValue(acc.words[0])); 682 #endif 683 break; 684 } 685 case Type::Long: { 686 jlong d = t->is_long()->get_con(); 687 #ifdef _LP64 688 array->append(new ConstantIntValue((jint)0)); 689 array->append(new ConstantLongValue(d)); 690 #else 691 // Repack the long as two jints. 692 // The convention the interpreter uses is that the second local 693 // holds the first raw word of the native double representation. 694 // This is actually reasonable, since locals and stack arrays 695 // grow downwards in all implementations. 696 // (If, on some machine, the interpreter's Java locals or stack 697 // were to grow upwards, the embedded doubles would be word-swapped.) 698 jlong_accessor acc; 699 acc.long_value = d; 700 array->append(new ConstantIntValue(acc.words[1])); 701 array->append(new ConstantIntValue(acc.words[0])); 702 #endif 703 break; 704 } 705 case Type::Top: // Add an illegal value here 706 array->append(new LocationValue(Location())); 707 break; 708 default: 709 ShouldNotReachHere(); 710 break; 711 } 712 } 713 714 // Determine if this node starts a bundle 715 bool Compile::starts_bundle(const Node *n) const { 716 return (_node_bundling_limit > n->_idx && 717 _node_bundling_base[n->_idx].starts_bundle()); 718 } 719 720 //--------------------------Process_OopMap_Node-------------------------------- 721 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) { 722 723 // Handle special safepoint nodes for synchronization 724 MachSafePointNode *sfn = mach->as_MachSafePoint(); 725 MachCallNode *mcall; 726 727 int safepoint_pc_offset = current_offset; 728 bool is_method_handle_invoke = false; 729 bool return_oop = false; 730 731 // Add the safepoint in the DebugInfoRecorder 732 if( !mach->is_MachCall() ) { 733 mcall = NULL; 734 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 735 } else { 736 mcall = mach->as_MachCall(); 737 738 // Is the call a MethodHandle call? 739 if (mcall->is_MachCallJava()) { 740 if (mcall->as_MachCallJava()->_method_handle_invoke) { 741 assert(has_method_handle_invokes(), "must have been set during call generation"); 742 is_method_handle_invoke = true; 743 } 744 } 745 746 // Check if a call returns an object. 747 if (mcall->returns_pointer()) { 748 return_oop = true; 749 } 750 safepoint_pc_offset += mcall->ret_addr_offset(); 751 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 752 } 753 754 // Loop over the JVMState list to add scope information 755 // Do not skip safepoints with a NULL method, they need monitor info 756 JVMState* youngest_jvms = sfn->jvms(); 757 int max_depth = youngest_jvms->depth(); 758 759 // Allocate the object pool for scalar-replaced objects -- the map from 760 // small-integer keys (which can be recorded in the local and ostack 761 // arrays) to descriptions of the object state. 762 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 763 764 // Visit scopes from oldest to youngest. 765 for (int depth = 1; depth <= max_depth; depth++) { 766 JVMState* jvms = youngest_jvms->of_depth(depth); 767 int idx; 768 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 769 // Safepoints that do not have method() set only provide oop-map and monitor info 770 // to support GC; these do not support deoptimization. 771 int num_locs = (method == NULL) ? 0 : jvms->loc_size(); 772 int num_exps = (method == NULL) ? 0 : jvms->stk_size(); 773 int num_mon = jvms->nof_monitors(); 774 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), 775 "JVMS local count must match that of the method"); 776 777 // Add Local and Expression Stack Information 778 779 // Insert locals into the locarray 780 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 781 for( idx = 0; idx < num_locs; idx++ ) { 782 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 783 } 784 785 // Insert expression stack entries into the exparray 786 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 787 for( idx = 0; idx < num_exps; idx++ ) { 788 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 789 } 790 791 // Add in mappings of the monitors 792 assert( !method || 793 !method->is_synchronized() || 794 method->is_native() || 795 num_mon > 0 || 796 !GenerateSynchronizationCode, 797 "monitors must always exist for synchronized methods"); 798 799 // Build the growable array of ScopeValues for exp stack 800 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 801 802 // Loop over monitors and insert into array 803 for (idx = 0; idx < num_mon; idx++) { 804 // Grab the node that defines this monitor 805 Node* box_node = sfn->monitor_box(jvms, idx); 806 Node* obj_node = sfn->monitor_obj(jvms, idx); 807 808 // Create ScopeValue for object 809 ScopeValue *scval = NULL; 810 811 if (obj_node->is_SafePointScalarObject()) { 812 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 813 scval = Compile::sv_for_node_id(objs, spobj->_idx); 814 if (scval == NULL) { 815 const Type *t = spobj->bottom_type(); 816 ciKlass* cik = t->is_oopptr()->klass(); 817 assert(cik->is_instance_klass() || 818 cik->is_array_klass(), "Not supported allocation."); 819 ObjectValue* sv = new ObjectValue(spobj->_idx, 820 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 821 Compile::set_sv_for_object_node(objs, sv); 822 823 uint first_ind = spobj->first_index(youngest_jvms); 824 for (uint i = 0; i < spobj->n_fields(); i++) { 825 Node* fld_node = sfn->in(first_ind+i); 826 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 827 } 828 scval = sv; 829 } 830 } else if (!obj_node->is_Con()) { 831 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); 832 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 833 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop ); 834 } else { 835 scval = new_loc_value( _regalloc, obj_reg, Location::oop ); 836 } 837 } else { 838 const TypePtr *tp = obj_node->get_ptr_type(); 839 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding()); 840 } 841 842 OptoReg::Name box_reg = BoxLockNode::reg(box_node); 843 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)); 844 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated()); 845 monarray->append(new MonitorValue(scval, basic_lock, eliminated)); 846 } 847 848 // We dump the object pool first, since deoptimization reads it in first. 849 debug_info()->dump_object_pool(objs); 850 851 // Build first class objects to pass to scope 852 DebugToken *locvals = debug_info()->create_scope_values(locarray); 853 DebugToken *expvals = debug_info()->create_scope_values(exparray); 854 DebugToken *monvals = debug_info()->create_monitor_values(monarray); 855 856 // Make method available for all Safepoints 857 ciMethod* scope_method = method ? method : _method; 858 // Describe the scope here 859 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 860 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 861 // Now we can describe the scope. 862 methodHandle null_mh; 863 bool rethrow_exception = false; 864 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); 865 } // End jvms loop 866 867 // Mark the end of the scope set. 868 debug_info()->end_safepoint(safepoint_pc_offset); 869 } 870 871 872 873 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 874 class NonSafepointEmitter { 875 Compile* C; 876 JVMState* _pending_jvms; 877 int _pending_offset; 878 879 void emit_non_safepoint(); 880 881 public: 882 NonSafepointEmitter(Compile* compile) { 883 this->C = compile; 884 _pending_jvms = NULL; 885 _pending_offset = 0; 886 } 887 888 void observe_instruction(Node* n, int pc_offset) { 889 if (!C->debug_info()->recording_non_safepoints()) return; 890 891 Node_Notes* nn = C->node_notes_at(n->_idx); 892 if (nn == NULL || nn->jvms() == NULL) return; 893 if (_pending_jvms != NULL && 894 _pending_jvms->same_calls_as(nn->jvms())) { 895 // Repeated JVMS? Stretch it up here. 896 _pending_offset = pc_offset; 897 } else { 898 if (_pending_jvms != NULL && 899 _pending_offset < pc_offset) { 900 emit_non_safepoint(); 901 } 902 _pending_jvms = NULL; 903 if (pc_offset > C->debug_info()->last_pc_offset()) { 904 // This is the only way _pending_jvms can become non-NULL: 905 _pending_jvms = nn->jvms(); 906 _pending_offset = pc_offset; 907 } 908 } 909 } 910 911 // Stay out of the way of real safepoints: 912 void observe_safepoint(JVMState* jvms, int pc_offset) { 913 if (_pending_jvms != NULL && 914 !_pending_jvms->same_calls_as(jvms) && 915 _pending_offset < pc_offset) { 916 emit_non_safepoint(); 917 } 918 _pending_jvms = NULL; 919 } 920 921 void flush_at_end() { 922 if (_pending_jvms != NULL) { 923 emit_non_safepoint(); 924 } 925 _pending_jvms = NULL; 926 } 927 }; 928 929 void NonSafepointEmitter::emit_non_safepoint() { 930 JVMState* youngest_jvms = _pending_jvms; 931 int pc_offset = _pending_offset; 932 933 // Clear it now: 934 _pending_jvms = NULL; 935 936 DebugInformationRecorder* debug_info = C->debug_info(); 937 assert(debug_info->recording_non_safepoints(), "sanity"); 938 939 debug_info->add_non_safepoint(pc_offset); 940 int max_depth = youngest_jvms->depth(); 941 942 // Visit scopes from oldest to youngest. 943 for (int depth = 1; depth <= max_depth; depth++) { 944 JVMState* jvms = youngest_jvms->of_depth(depth); 945 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 946 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 947 methodHandle null_mh; 948 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute()); 949 } 950 951 // Mark the end of the scope set. 952 debug_info->end_non_safepoint(pc_offset); 953 } 954 955 //------------------------------init_buffer------------------------------------ 956 CodeBuffer* Compile::init_buffer(uint* blk_starts) { 957 958 // Set the initially allocated size 959 int code_req = initial_code_capacity; 960 int locs_req = initial_locs_capacity; 961 int stub_req = initial_stub_capacity; 962 int const_req = initial_const_capacity; 963 964 int pad_req = NativeCall::instruction_size; 965 // The extra spacing after the code is necessary on some platforms. 966 // Sometimes we need to patch in a jump after the last instruction, 967 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 968 969 // Compute the byte offset where we can store the deopt pc. 970 if (fixed_slots() != 0) { 971 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 972 } 973 974 // Compute prolog code size 975 _method_size = 0; 976 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; 977 #if defined(IA64) && !defined(AIX) 978 if (save_argument_registers()) { 979 // 4815101: this is a stub with implicit and unknown precision fp args. 980 // The usual spill mechanism can only generate stfd's in this case, which 981 // doesn't work if the fp reg to spill contains a single-precision denorm. 982 // Instead, we hack around the normal spill mechanism using stfspill's and 983 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate 984 // space here for the fp arg regs (f8-f15) we're going to thusly spill. 985 // 986 // If we ever implement 16-byte 'registers' == stack slots, we can 987 // get rid of this hack and have SpillCopy generate stfspill/ldffill 988 // instead of stfd/stfs/ldfd/ldfs. 989 _frame_slots += 8*(16/BytesPerInt); 990 } 991 #endif 992 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check"); 993 994 if (has_mach_constant_base_node()) { 995 uint add_size = 0; 996 // Fill the constant table. 997 // Note: This must happen before shorten_branches. 998 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 999 Block* b = _cfg->get_block(i); 1000 1001 for (uint j = 0; j < b->number_of_nodes(); j++) { 1002 Node* n = b->get_node(j); 1003 1004 // If the node is a MachConstantNode evaluate the constant 1005 // value section. 1006 if (n->is_MachConstant()) { 1007 MachConstantNode* machcon = n->as_MachConstant(); 1008 machcon->eval_constant(C); 1009 } else if (n->is_Mach()) { 1010 // On Power there are more nodes that issue constants. 1011 add_size += (n->as_Mach()->ins_num_consts() * 8); 1012 } 1013 } 1014 } 1015 1016 // Calculate the offsets of the constants and the size of the 1017 // constant table (including the padding to the next section). 1018 constant_table().calculate_offsets_and_size(); 1019 const_req = constant_table().size() + add_size; 1020 } 1021 1022 // Initialize the space for the BufferBlob used to find and verify 1023 // instruction size in MachNode::emit_size() 1024 init_scratch_buffer_blob(const_req); 1025 if (failing()) return NULL; // Out of memory 1026 1027 // Pre-compute the length of blocks and replace 1028 // long branches with short if machine supports it. 1029 shorten_branches(blk_starts, code_req, locs_req, stub_req); 1030 1031 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1032 // class HandlerImpl is platform-specific and defined in the *.ad files. 1033 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler 1034 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler 1035 stub_req += MAX_stubs_size; // ensure per-stub margin 1036 code_req += MAX_inst_size; // ensure per-instruction margin 1037 1038 if (StressCodeBuffers) 1039 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1040 1041 int total_req = 1042 const_req + 1043 code_req + 1044 pad_req + 1045 stub_req + 1046 exception_handler_req + 1047 deopt_handler_req; // deopt handler 1048 1049 if (has_method_handle_invokes()) 1050 total_req += deopt_handler_req; // deopt MH handler 1051 1052 CodeBuffer* cb = code_buffer(); 1053 cb->initialize(total_req, locs_req); 1054 1055 // Have we run out of code space? 1056 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1057 C->record_failure("CodeCache is full"); 1058 return NULL; 1059 } 1060 // Configure the code buffer. 1061 cb->initialize_consts_size(const_req); 1062 cb->initialize_stubs_size(stub_req); 1063 cb->initialize_oop_recorder(env()->oop_recorder()); 1064 1065 // fill in the nop array for bundling computations 1066 MachNode *_nop_list[Bundle::_nop_count]; 1067 Bundle::initialize_nops(_nop_list); 1068 1069 return cb; 1070 } 1071 1072 //------------------------------fill_buffer------------------------------------ 1073 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) { 1074 // blk_starts[] contains offsets calculated during short branches processing, 1075 // offsets should not be increased during following steps. 1076 1077 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 1078 // of a loop. It is used to determine the padding for loop alignment. 1079 compute_loop_first_inst_sizes(); 1080 1081 // Create oopmap set. 1082 _oop_map_set = new OopMapSet(); 1083 1084 // !!!!! This preserves old handling of oopmaps for now 1085 debug_info()->set_oopmaps(_oop_map_set); 1086 1087 uint nblocks = _cfg->number_of_blocks(); 1088 // Count and start of implicit null check instructions 1089 uint inct_cnt = 0; 1090 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1091 1092 // Count and start of calls 1093 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1094 1095 uint return_offset = 0; 1096 int nop_size = (new MachNopNode())->size(_regalloc); 1097 1098 int previous_offset = 0; 1099 int current_offset = 0; 1100 int last_call_offset = -1; 1101 int last_avoid_back_to_back_offset = -1; 1102 #ifdef ASSERT 1103 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); 1104 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 1105 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 1106 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); 1107 #endif 1108 1109 // Create an array of unused labels, one for each basic block, if printing is enabled 1110 #ifndef PRODUCT 1111 int *node_offsets = NULL; 1112 uint node_offset_limit = unique(); 1113 1114 if (print_assembly()) 1115 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1116 #endif 1117 1118 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily 1119 1120 // Emit the constant table. 1121 if (has_mach_constant_base_node()) { 1122 constant_table().emit(*cb); 1123 } 1124 1125 // Create an array of labels, one for each basic block 1126 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1); 1127 for (uint i=0; i <= nblocks; i++) { 1128 blk_labels[i].init(); 1129 } 1130 1131 // ------------------ 1132 // Now fill in the code buffer 1133 Node *delay_slot = NULL; 1134 1135 for (uint i = 0; i < nblocks; i++) { 1136 Block* block = _cfg->get_block(i); 1137 Node* head = block->head(); 1138 1139 // If this block needs to start aligned (i.e, can be reached other 1140 // than by falling-thru from the previous block), then force the 1141 // start of a new bundle. 1142 if (Pipeline::requires_bundling() && starts_bundle(head)) { 1143 cb->flush_bundle(true); 1144 } 1145 1146 #ifdef ASSERT 1147 if (!block->is_connector()) { 1148 stringStream st; 1149 block->dump_head(_cfg, &st); 1150 MacroAssembler(cb).block_comment(st.as_string()); 1151 } 1152 jmp_target[i] = 0; 1153 jmp_offset[i] = 0; 1154 jmp_size[i] = 0; 1155 jmp_rule[i] = 0; 1156 #endif 1157 int blk_offset = current_offset; 1158 1159 // Define the label at the beginning of the basic block 1160 MacroAssembler(cb).bind(blk_labels[block->_pre_order]); 1161 1162 uint last_inst = block->number_of_nodes(); 1163 1164 // Emit block normally, except for last instruction. 1165 // Emit means "dump code bits into code buffer". 1166 for (uint j = 0; j<last_inst; j++) { 1167 1168 // Get the node 1169 Node* n = block->get_node(j); 1170 1171 // See if delay slots are supported 1172 if (valid_bundle_info(n) && 1173 node_bundling(n)->used_in_unconditional_delay()) { 1174 assert(delay_slot == NULL, "no use of delay slot node"); 1175 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1176 1177 delay_slot = n; 1178 continue; 1179 } 1180 1181 // If this starts a new instruction group, then flush the current one 1182 // (but allow split bundles) 1183 if (Pipeline::requires_bundling() && starts_bundle(n)) 1184 cb->flush_bundle(false); 1185 1186 // Special handling for SafePoint/Call Nodes 1187 bool is_mcall = false; 1188 if (n->is_Mach()) { 1189 MachNode *mach = n->as_Mach(); 1190 is_mcall = n->is_MachCall(); 1191 bool is_sfn = n->is_MachSafePoint(); 1192 1193 // If this requires all previous instructions be flushed, then do so 1194 if (is_sfn || is_mcall || mach->alignment_required() != 1) { 1195 cb->flush_bundle(true); 1196 current_offset = cb->insts_size(); 1197 } 1198 1199 // A padding may be needed again since a previous instruction 1200 // could be moved to delay slot. 1201 1202 // align the instruction if necessary 1203 int padding = mach->compute_padding(current_offset); 1204 // Make sure safepoint node for polling is distinct from a call's 1205 // return by adding a nop if needed. 1206 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) { 1207 padding = nop_size; 1208 } 1209 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) && 1210 current_offset == last_avoid_back_to_back_offset) { 1211 // Avoid back to back some instructions. 1212 padding = nop_size; 1213 } 1214 1215 if (padding > 0) { 1216 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1217 int nops_cnt = padding / nop_size; 1218 MachNode *nop = new MachNopNode(nops_cnt); 1219 block->insert_node(nop, j++); 1220 last_inst++; 1221 _cfg->map_node_to_block(nop, block); 1222 // Ensure enough space. 1223 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1224 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1225 C->record_failure("CodeCache is full"); 1226 return; 1227 } 1228 nop->emit(*cb, _regalloc); 1229 cb->flush_bundle(true); 1230 current_offset = cb->insts_size(); 1231 } 1232 1233 // Remember the start of the last call in a basic block 1234 if (is_mcall) { 1235 MachCallNode *mcall = mach->as_MachCall(); 1236 1237 // This destination address is NOT PC-relative 1238 mcall->method_set((intptr_t)mcall->entry_point()); 1239 1240 // Save the return address 1241 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset(); 1242 1243 if (mcall->is_MachCallLeaf()) { 1244 is_mcall = false; 1245 is_sfn = false; 1246 } 1247 } 1248 1249 // sfn will be valid whenever mcall is valid now because of inheritance 1250 if (is_sfn || is_mcall) { 1251 1252 // Handle special safepoint nodes for synchronization 1253 if (!is_mcall) { 1254 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1255 // !!!!! Stubs only need an oopmap right now, so bail out 1256 if (sfn->jvms()->method() == NULL) { 1257 // Write the oopmap directly to the code blob??!! 1258 continue; 1259 } 1260 } // End synchronization 1261 1262 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1263 current_offset); 1264 Process_OopMap_Node(mach, current_offset); 1265 } // End if safepoint 1266 1267 // If this is a null check, then add the start of the previous instruction to the list 1268 else if( mach->is_MachNullCheck() ) { 1269 inct_starts[inct_cnt++] = previous_offset; 1270 } 1271 1272 // If this is a branch, then fill in the label with the target BB's label 1273 else if (mach->is_MachBranch()) { 1274 // This requires the TRUE branch target be in succs[0] 1275 uint block_num = block->non_connector_successor(0)->_pre_order; 1276 1277 // Try to replace long branch if delay slot is not used, 1278 // it is mostly for back branches since forward branch's 1279 // distance is not updated yet. 1280 bool delay_slot_is_used = valid_bundle_info(n) && 1281 node_bundling(n)->use_unconditional_delay(); 1282 if (!delay_slot_is_used && mach->may_be_short_branch()) { 1283 assert(delay_slot == NULL, "not expecting delay slot node"); 1284 int br_size = n->size(_regalloc); 1285 int offset = blk_starts[block_num] - current_offset; 1286 if (block_num >= i) { 1287 // Current and following block's offset are not 1288 // finalized yet, adjust distance by the difference 1289 // between calculated and final offsets of current block. 1290 offset -= (blk_starts[i] - blk_offset); 1291 } 1292 // In the following code a nop could be inserted before 1293 // the branch which will increase the backward distance. 1294 bool needs_padding = (current_offset == last_avoid_back_to_back_offset); 1295 if (needs_padding && offset <= 0) 1296 offset -= nop_size; 1297 1298 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 1299 // We've got a winner. Replace this branch. 1300 MachNode* replacement = mach->as_MachBranch()->short_branch_version(); 1301 1302 // Update the jmp_size. 1303 int new_size = replacement->size(_regalloc); 1304 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller"); 1305 // Insert padding between avoid_back_to_back branches. 1306 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 1307 MachNode *nop = new MachNopNode(); 1308 block->insert_node(nop, j++); 1309 _cfg->map_node_to_block(nop, block); 1310 last_inst++; 1311 nop->emit(*cb, _regalloc); 1312 cb->flush_bundle(true); 1313 current_offset = cb->insts_size(); 1314 } 1315 #ifdef ASSERT 1316 jmp_target[i] = block_num; 1317 jmp_offset[i] = current_offset - blk_offset; 1318 jmp_size[i] = new_size; 1319 jmp_rule[i] = mach->rule(); 1320 #endif 1321 block->map_node(replacement, j); 1322 mach->subsume_by(replacement, C); 1323 n = replacement; 1324 mach = replacement; 1325 } 1326 } 1327 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num ); 1328 } else if (mach->ideal_Opcode() == Op_Jump) { 1329 for (uint h = 0; h < block->_num_succs; h++) { 1330 Block* succs_block = block->_succs[h]; 1331 for (uint j = 1; j < succs_block->num_preds(); j++) { 1332 Node* jpn = succs_block->pred(j); 1333 if (jpn->is_JumpProj() && jpn->in(0) == mach) { 1334 uint block_num = succs_block->non_connector()->_pre_order; 1335 Label *blkLabel = &blk_labels[block_num]; 1336 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1337 } 1338 } 1339 } 1340 } 1341 #ifdef ASSERT 1342 // Check that oop-store precedes the card-mark 1343 else if (mach->ideal_Opcode() == Op_StoreCM) { 1344 uint storeCM_idx = j; 1345 int count = 0; 1346 for (uint prec = mach->req(); prec < mach->len(); prec++) { 1347 Node *oop_store = mach->in(prec); // Precedence edge 1348 if (oop_store == NULL) continue; 1349 count++; 1350 uint i4; 1351 for (i4 = 0; i4 < last_inst; ++i4) { 1352 if (block->get_node(i4) == oop_store) { 1353 break; 1354 } 1355 } 1356 // Note: This test can provide a false failure if other precedence 1357 // edges have been added to the storeCMNode. 1358 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1359 } 1360 assert(count > 0, "storeCM expects at least one precedence edge"); 1361 } 1362 #endif 1363 else if (!n->is_Proj()) { 1364 // Remember the beginning of the previous instruction, in case 1365 // it's followed by a flag-kill and a null-check. Happens on 1366 // Intel all the time, with add-to-memory kind of opcodes. 1367 previous_offset = current_offset; 1368 } 1369 1370 // Not an else-if! 1371 // If this is a trap based cmp then add its offset to the list. 1372 if (mach->is_TrapBasedCheckNode()) { 1373 inct_starts[inct_cnt++] = current_offset; 1374 } 1375 } 1376 1377 // Verify that there is sufficient space remaining 1378 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1379 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1380 C->record_failure("CodeCache is full"); 1381 return; 1382 } 1383 1384 // Save the offset for the listing 1385 #ifndef PRODUCT 1386 if (node_offsets && n->_idx < node_offset_limit) 1387 node_offsets[n->_idx] = cb->insts_size(); 1388 #endif 1389 1390 // "Normal" instruction case 1391 DEBUG_ONLY( uint instr_offset = cb->insts_size(); ) 1392 n->emit(*cb, _regalloc); 1393 current_offset = cb->insts_size(); 1394 1395 // Above we only verified that there is enough space in the instruction section. 1396 // However, the instruction may emit stubs that cause code buffer expansion. 1397 // Bail out here if expansion failed due to a lack of code cache space. 1398 if (failing()) { 1399 return; 1400 } 1401 1402 #ifdef ASSERT 1403 uint n_size = n->size(C->regalloc()); 1404 if (n_size < (current_offset-instr_offset)) { 1405 MachNode* mach = n->as_Mach(); 1406 n->dump(); 1407 mach->dump_format(C->regalloc(), tty); 1408 tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset); 1409 Disassembler::decode(cb->insts_begin() + instr_offset, cb->insts_begin() + current_offset + 1, tty); 1410 tty->print_cr(" ------------------- "); 1411 BufferBlob* blob = this->scratch_buffer_blob(); 1412 address blob_begin = blob->content_begin(); 1413 Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty); 1414 assert(false, "wrong size of mach node"); 1415 } 1416 #endif 1417 non_safepoints.observe_instruction(n, current_offset); 1418 1419 // mcall is last "call" that can be a safepoint 1420 // record it so we can see if a poll will directly follow it 1421 // in which case we'll need a pad to make the PcDesc sites unique 1422 // see 5010568. This can be slightly inaccurate but conservative 1423 // in the case that return address is not actually at current_offset. 1424 // This is a small price to pay. 1425 1426 if (is_mcall) { 1427 last_call_offset = current_offset; 1428 } 1429 1430 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 1431 // Avoid back to back some instructions. 1432 last_avoid_back_to_back_offset = current_offset; 1433 } 1434 1435 // See if this instruction has a delay slot 1436 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1437 guarantee(delay_slot != NULL, "expecting delay slot node"); 1438 1439 // Back up 1 instruction 1440 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size()); 1441 1442 // Save the offset for the listing 1443 #ifndef PRODUCT 1444 if (node_offsets && delay_slot->_idx < node_offset_limit) 1445 node_offsets[delay_slot->_idx] = cb->insts_size(); 1446 #endif 1447 1448 // Support a SafePoint in the delay slot 1449 if (delay_slot->is_MachSafePoint()) { 1450 MachNode *mach = delay_slot->as_Mach(); 1451 // !!!!! Stubs only need an oopmap right now, so bail out 1452 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) { 1453 // Write the oopmap directly to the code blob??!! 1454 delay_slot = NULL; 1455 continue; 1456 } 1457 1458 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1459 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1460 adjusted_offset); 1461 // Generate an OopMap entry 1462 Process_OopMap_Node(mach, adjusted_offset); 1463 } 1464 1465 // Insert the delay slot instruction 1466 delay_slot->emit(*cb, _regalloc); 1467 1468 // Don't reuse it 1469 delay_slot = NULL; 1470 } 1471 1472 } // End for all instructions in block 1473 1474 // If the next block is the top of a loop, pad this block out to align 1475 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1476 if (i < nblocks-1) { 1477 Block *nb = _cfg->get_block(i + 1); 1478 int padding = nb->alignment_padding(current_offset); 1479 if( padding > 0 ) { 1480 MachNode *nop = new MachNopNode(padding / nop_size); 1481 block->insert_node(nop, block->number_of_nodes()); 1482 _cfg->map_node_to_block(nop, block); 1483 nop->emit(*cb, _regalloc); 1484 current_offset = cb->insts_size(); 1485 } 1486 } 1487 // Verify that the distance for generated before forward 1488 // short branches is still valid. 1489 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size"); 1490 1491 // Save new block start offset 1492 blk_starts[i] = blk_offset; 1493 } // End of for all blocks 1494 blk_starts[nblocks] = current_offset; 1495 1496 non_safepoints.flush_at_end(); 1497 1498 // Offset too large? 1499 if (failing()) return; 1500 1501 // Define a pseudo-label at the end of the code 1502 MacroAssembler(cb).bind( blk_labels[nblocks] ); 1503 1504 // Compute the size of the first block 1505 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1506 1507 #ifdef ASSERT 1508 for (uint i = 0; i < nblocks; i++) { // For all blocks 1509 if (jmp_target[i] != 0) { 1510 int br_size = jmp_size[i]; 1511 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 1512 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 1513 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]); 1514 assert(false, "Displacement too large for short jmp"); 1515 } 1516 } 1517 } 1518 #endif 1519 1520 #ifndef PRODUCT 1521 // Information on the size of the method, without the extraneous code 1522 Scheduling::increment_method_size(cb->insts_size()); 1523 #endif 1524 1525 // ------------------ 1526 // Fill in exception table entries. 1527 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1528 1529 // Only java methods have exception handlers and deopt handlers 1530 // class HandlerImpl is platform-specific and defined in the *.ad files. 1531 if (_method) { 1532 // Emit the exception handler code. 1533 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb)); 1534 if (failing()) { 1535 return; // CodeBuffer::expand failed 1536 } 1537 // Emit the deopt handler code. 1538 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb)); 1539 1540 // Emit the MethodHandle deopt handler code (if required). 1541 if (has_method_handle_invokes() && !failing()) { 1542 // We can use the same code as for the normal deopt handler, we 1543 // just need a different entry point address. 1544 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb)); 1545 } 1546 } 1547 1548 // One last check for failed CodeBuffer::expand: 1549 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1550 C->record_failure("CodeCache is full"); 1551 return; 1552 } 1553 1554 #ifndef PRODUCT 1555 // Dump the assembly code, including basic-block numbers 1556 if (print_assembly()) { 1557 ttyLocker ttyl; // keep the following output all in one block 1558 if (!VMThread::should_terminate()) { // test this under the tty lock 1559 // This output goes directly to the tty, not the compiler log. 1560 // To enable tools to match it up with the compilation activity, 1561 // be sure to tag this tty output with the compile ID. 1562 if (xtty != NULL) { 1563 xtty->head("opto_assembly compile_id='%d'%s", compile_id(), 1564 is_osr_compilation() ? " compile_kind='osr'" : 1565 ""); 1566 } 1567 if (method() != NULL) { 1568 method()->print_metadata(); 1569 } else if (stub_name() != NULL) { 1570 tty->print_cr("Generating RuntimeStub - %s", stub_name()); 1571 } 1572 dump_asm(node_offsets, node_offset_limit); 1573 if (xtty != NULL) { 1574 // print_metadata and dump_asm above may safepoint which makes us loose the ttylock. 1575 // Retake lock too make sure the end tag is coherent, and that xmlStream->pop_tag is done 1576 // thread safe 1577 ttyLocker ttyl2; 1578 xtty->tail("opto_assembly"); 1579 } 1580 } 1581 } 1582 #endif 1583 1584 } 1585 1586 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1587 _inc_table.set_size(cnt); 1588 1589 uint inct_cnt = 0; 1590 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 1591 Block* block = _cfg->get_block(i); 1592 Node *n = NULL; 1593 int j; 1594 1595 // Find the branch; ignore trailing NOPs. 1596 for (j = block->number_of_nodes() - 1; j >= 0; j--) { 1597 n = block->get_node(j); 1598 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) { 1599 break; 1600 } 1601 } 1602 1603 // If we didn't find anything, continue 1604 if (j < 0) { 1605 continue; 1606 } 1607 1608 // Compute ExceptionHandlerTable subtable entry and add it 1609 // (skip empty blocks) 1610 if (n->is_Catch()) { 1611 1612 // Get the offset of the return from the call 1613 uint call_return = call_returns[block->_pre_order]; 1614 #ifdef ASSERT 1615 assert( call_return > 0, "no call seen for this basic block" ); 1616 while (block->get_node(--j)->is_MachProj()) ; 1617 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call"); 1618 #endif 1619 // last instruction is a CatchNode, find it's CatchProjNodes 1620 int nof_succs = block->_num_succs; 1621 // allocate space 1622 GrowableArray<intptr_t> handler_bcis(nof_succs); 1623 GrowableArray<intptr_t> handler_pcos(nof_succs); 1624 // iterate through all successors 1625 for (int j = 0; j < nof_succs; j++) { 1626 Block* s = block->_succs[j]; 1627 bool found_p = false; 1628 for (uint k = 1; k < s->num_preds(); k++) { 1629 Node* pk = s->pred(k); 1630 if (pk->is_CatchProj() && pk->in(0) == n) { 1631 const CatchProjNode* p = pk->as_CatchProj(); 1632 found_p = true; 1633 // add the corresponding handler bci & pco information 1634 if (p->_con != CatchProjNode::fall_through_index) { 1635 // p leads to an exception handler (and is not fall through) 1636 assert(s == _cfg->get_block(s->_pre_order), "bad numbering"); 1637 // no duplicates, please 1638 if (!handler_bcis.contains(p->handler_bci())) { 1639 uint block_num = s->non_connector()->_pre_order; 1640 handler_bcis.append(p->handler_bci()); 1641 handler_pcos.append(blk_labels[block_num].loc_pos()); 1642 } 1643 } 1644 } 1645 } 1646 assert(found_p, "no matching predecessor found"); 1647 // Note: Due to empty block removal, one block may have 1648 // several CatchProj inputs, from the same Catch. 1649 } 1650 1651 // Set the offset of the return from the call 1652 assert(handler_bcis.find(-1) != -1, "must have default handler"); 1653 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); 1654 continue; 1655 } 1656 1657 // Handle implicit null exception table updates 1658 if (n->is_MachNullCheck()) { 1659 uint block_num = block->non_connector_successor(0)->_pre_order; 1660 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1661 continue; 1662 } 1663 // Handle implicit exception table updates: trap instructions. 1664 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) { 1665 uint block_num = block->non_connector_successor(0)->_pre_order; 1666 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1667 continue; 1668 } 1669 } // End of for all blocks fill in exception table entries 1670 } 1671 1672 // Static Variables 1673 #ifndef PRODUCT 1674 uint Scheduling::_total_nop_size = 0; 1675 uint Scheduling::_total_method_size = 0; 1676 uint Scheduling::_total_branches = 0; 1677 uint Scheduling::_total_unconditional_delays = 0; 1678 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1679 #endif 1680 1681 // Initializer for class Scheduling 1682 1683 Scheduling::Scheduling(Arena *arena, Compile &compile) 1684 : _arena(arena), 1685 _cfg(compile.cfg()), 1686 _regalloc(compile.regalloc()), 1687 _reg_node(arena), 1688 _bundle_instr_count(0), 1689 _bundle_cycle_number(0), 1690 _scheduled(arena), 1691 _available(arena), 1692 _next_node(NULL), 1693 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]), 1694 _pinch_free_list(arena) 1695 #ifndef PRODUCT 1696 , _branches(0) 1697 , _unconditional_delays(0) 1698 #endif 1699 { 1700 // Create a MachNopNode 1701 _nop = new MachNopNode(); 1702 1703 // Now that the nops are in the array, save the count 1704 // (but allow entries for the nops) 1705 _node_bundling_limit = compile.unique(); 1706 uint node_max = _regalloc->node_regs_max_index(); 1707 1708 compile.set_node_bundling_limit(_node_bundling_limit); 1709 1710 // This one is persistent within the Compile class 1711 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1712 1713 // Allocate space for fixed-size arrays 1714 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1715 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1716 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1717 1718 // Clear the arrays 1719 memset(_node_bundling_base, 0, node_max * sizeof(Bundle)); 1720 memset(_node_latency, 0, node_max * sizeof(unsigned short)); 1721 memset(_uses, 0, node_max * sizeof(short)); 1722 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 1723 1724 // Clear the bundling information 1725 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements)); 1726 1727 // Get the last node 1728 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1); 1729 1730 _next_node = block->get_node(block->number_of_nodes() - 1); 1731 } 1732 1733 #ifndef PRODUCT 1734 // Scheduling destructor 1735 Scheduling::~Scheduling() { 1736 _total_branches += _branches; 1737 _total_unconditional_delays += _unconditional_delays; 1738 } 1739 #endif 1740 1741 // Step ahead "i" cycles 1742 void Scheduling::step(uint i) { 1743 1744 Bundle *bundle = node_bundling(_next_node); 1745 bundle->set_starts_bundle(); 1746 1747 // Update the bundle record, but leave the flags information alone 1748 if (_bundle_instr_count > 0) { 1749 bundle->set_instr_count(_bundle_instr_count); 1750 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1751 } 1752 1753 // Update the state information 1754 _bundle_instr_count = 0; 1755 _bundle_cycle_number += i; 1756 _bundle_use.step(i); 1757 } 1758 1759 void Scheduling::step_and_clear() { 1760 Bundle *bundle = node_bundling(_next_node); 1761 bundle->set_starts_bundle(); 1762 1763 // Update the bundle record 1764 if (_bundle_instr_count > 0) { 1765 bundle->set_instr_count(_bundle_instr_count); 1766 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1767 1768 _bundle_cycle_number += 1; 1769 } 1770 1771 // Clear the bundling information 1772 _bundle_instr_count = 0; 1773 _bundle_use.reset(); 1774 1775 memcpy(_bundle_use_elements, 1776 Pipeline_Use::elaborated_elements, 1777 sizeof(Pipeline_Use::elaborated_elements)); 1778 } 1779 1780 // Perform instruction scheduling and bundling over the sequence of 1781 // instructions in backwards order. 1782 void Compile::ScheduleAndBundle() { 1783 1784 // Don't optimize this if it isn't a method 1785 if (!_method) 1786 return; 1787 1788 // Don't optimize this if scheduling is disabled 1789 if (!do_scheduling()) 1790 return; 1791 1792 // Scheduling code works only with pairs (16 bytes) maximum. 1793 if (max_vector_size() > 16) 1794 return; 1795 1796 TracePhase tp("isched", &timers[_t_instrSched]); 1797 1798 // Create a data structure for all the scheduling information 1799 Scheduling scheduling(Thread::current()->resource_area(), *this); 1800 1801 // Walk backwards over each basic block, computing the needed alignment 1802 // Walk over all the basic blocks 1803 scheduling.DoScheduling(); 1804 } 1805 1806 // Compute the latency of all the instructions. This is fairly simple, 1807 // because we already have a legal ordering. Walk over the instructions 1808 // from first to last, and compute the latency of the instruction based 1809 // on the latency of the preceding instruction(s). 1810 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { 1811 #ifndef PRODUCT 1812 if (_cfg->C->trace_opto_output()) 1813 tty->print("# -> ComputeLocalLatenciesForward\n"); 1814 #endif 1815 1816 // Walk over all the schedulable instructions 1817 for( uint j=_bb_start; j < _bb_end; j++ ) { 1818 1819 // This is a kludge, forcing all latency calculations to start at 1. 1820 // Used to allow latency 0 to force an instruction to the beginning 1821 // of the bb 1822 uint latency = 1; 1823 Node *use = bb->get_node(j); 1824 uint nlen = use->len(); 1825 1826 // Walk over all the inputs 1827 for ( uint k=0; k < nlen; k++ ) { 1828 Node *def = use->in(k); 1829 if (!def) 1830 continue; 1831 1832 uint l = _node_latency[def->_idx] + use->latency(k); 1833 if (latency < l) 1834 latency = l; 1835 } 1836 1837 _node_latency[use->_idx] = latency; 1838 1839 #ifndef PRODUCT 1840 if (_cfg->C->trace_opto_output()) { 1841 tty->print("# latency %4d: ", latency); 1842 use->dump(); 1843 } 1844 #endif 1845 } 1846 1847 #ifndef PRODUCT 1848 if (_cfg->C->trace_opto_output()) 1849 tty->print("# <- ComputeLocalLatenciesForward\n"); 1850 #endif 1851 1852 } // end ComputeLocalLatenciesForward 1853 1854 // See if this node fits into the present instruction bundle 1855 bool Scheduling::NodeFitsInBundle(Node *n) { 1856 uint n_idx = n->_idx; 1857 1858 // If this is the unconditional delay instruction, then it fits 1859 if (n == _unconditional_delay_slot) { 1860 #ifndef PRODUCT 1861 if (_cfg->C->trace_opto_output()) 1862 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 1863 #endif 1864 return (true); 1865 } 1866 1867 // If the node cannot be scheduled this cycle, skip it 1868 if (_current_latency[n_idx] > _bundle_cycle_number) { 1869 #ifndef PRODUCT 1870 if (_cfg->C->trace_opto_output()) 1871 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 1872 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 1873 #endif 1874 return (false); 1875 } 1876 1877 const Pipeline *node_pipeline = n->pipeline(); 1878 1879 uint instruction_count = node_pipeline->instructionCount(); 1880 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1881 instruction_count = 0; 1882 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1883 instruction_count++; 1884 1885 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 1886 #ifndef PRODUCT 1887 if (_cfg->C->trace_opto_output()) 1888 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 1889 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 1890 #endif 1891 return (false); 1892 } 1893 1894 // Don't allow non-machine nodes to be handled this way 1895 if (!n->is_Mach() && instruction_count == 0) 1896 return (false); 1897 1898 // See if there is any overlap 1899 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 1900 1901 if (delay > 0) { 1902 #ifndef PRODUCT 1903 if (_cfg->C->trace_opto_output()) 1904 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 1905 #endif 1906 return false; 1907 } 1908 1909 #ifndef PRODUCT 1910 if (_cfg->C->trace_opto_output()) 1911 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 1912 #endif 1913 1914 return true; 1915 } 1916 1917 Node * Scheduling::ChooseNodeToBundle() { 1918 uint siz = _available.size(); 1919 1920 if (siz == 0) { 1921 1922 #ifndef PRODUCT 1923 if (_cfg->C->trace_opto_output()) 1924 tty->print("# ChooseNodeToBundle: NULL\n"); 1925 #endif 1926 return (NULL); 1927 } 1928 1929 // Fast path, if only 1 instruction in the bundle 1930 if (siz == 1) { 1931 #ifndef PRODUCT 1932 if (_cfg->C->trace_opto_output()) { 1933 tty->print("# ChooseNodeToBundle (only 1): "); 1934 _available[0]->dump(); 1935 } 1936 #endif 1937 return (_available[0]); 1938 } 1939 1940 // Don't bother, if the bundle is already full 1941 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 1942 for ( uint i = 0; i < siz; i++ ) { 1943 Node *n = _available[i]; 1944 1945 // Skip projections, we'll handle them another way 1946 if (n->is_Proj()) 1947 continue; 1948 1949 // This presupposed that instructions are inserted into the 1950 // available list in a legality order; i.e. instructions that 1951 // must be inserted first are at the head of the list 1952 if (NodeFitsInBundle(n)) { 1953 #ifndef PRODUCT 1954 if (_cfg->C->trace_opto_output()) { 1955 tty->print("# ChooseNodeToBundle: "); 1956 n->dump(); 1957 } 1958 #endif 1959 return (n); 1960 } 1961 } 1962 } 1963 1964 // Nothing fits in this bundle, choose the highest priority 1965 #ifndef PRODUCT 1966 if (_cfg->C->trace_opto_output()) { 1967 tty->print("# ChooseNodeToBundle: "); 1968 _available[0]->dump(); 1969 } 1970 #endif 1971 1972 return _available[0]; 1973 } 1974 1975 void Scheduling::AddNodeToAvailableList(Node *n) { 1976 assert( !n->is_Proj(), "projections never directly made available" ); 1977 #ifndef PRODUCT 1978 if (_cfg->C->trace_opto_output()) { 1979 tty->print("# AddNodeToAvailableList: "); 1980 n->dump(); 1981 } 1982 #endif 1983 1984 int latency = _current_latency[n->_idx]; 1985 1986 // Insert in latency order (insertion sort) 1987 uint i; 1988 for ( i=0; i < _available.size(); i++ ) 1989 if (_current_latency[_available[i]->_idx] > latency) 1990 break; 1991 1992 // Special Check for compares following branches 1993 if( n->is_Mach() && _scheduled.size() > 0 ) { 1994 int op = n->as_Mach()->ideal_Opcode(); 1995 Node *last = _scheduled[0]; 1996 if( last->is_MachIf() && last->in(1) == n && 1997 ( op == Op_CmpI || 1998 op == Op_CmpU || 1999 op == Op_CmpUL || 2000 op == Op_CmpP || 2001 op == Op_CmpF || 2002 op == Op_CmpD || 2003 op == Op_CmpL ) ) { 2004 2005 // Recalculate position, moving to front of same latency 2006 for ( i=0 ; i < _available.size(); i++ ) 2007 if (_current_latency[_available[i]->_idx] >= latency) 2008 break; 2009 } 2010 } 2011 2012 // Insert the node in the available list 2013 _available.insert(i, n); 2014 2015 #ifndef PRODUCT 2016 if (_cfg->C->trace_opto_output()) 2017 dump_available(); 2018 #endif 2019 } 2020 2021 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 2022 for ( uint i=0; i < n->len(); i++ ) { 2023 Node *def = n->in(i); 2024 if (!def) continue; 2025 if( def->is_Proj() ) // If this is a machine projection, then 2026 def = def->in(0); // propagate usage thru to the base instruction 2027 2028 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local 2029 continue; 2030 } 2031 2032 // Compute the latency 2033 uint l = _bundle_cycle_number + n->latency(i); 2034 if (_current_latency[def->_idx] < l) 2035 _current_latency[def->_idx] = l; 2036 2037 // If this does not have uses then schedule it 2038 if ((--_uses[def->_idx]) == 0) 2039 AddNodeToAvailableList(def); 2040 } 2041 } 2042 2043 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 2044 #ifndef PRODUCT 2045 if (_cfg->C->trace_opto_output()) { 2046 tty->print("# AddNodeToBundle: "); 2047 n->dump(); 2048 } 2049 #endif 2050 2051 // Remove this from the available list 2052 uint i; 2053 for (i = 0; i < _available.size(); i++) 2054 if (_available[i] == n) 2055 break; 2056 assert(i < _available.size(), "entry in _available list not found"); 2057 _available.remove(i); 2058 2059 // See if this fits in the current bundle 2060 const Pipeline *node_pipeline = n->pipeline(); 2061 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 2062 2063 // Check for instructions to be placed in the delay slot. We 2064 // do this before we actually schedule the current instruction, 2065 // because the delay slot follows the current instruction. 2066 if (Pipeline::_branch_has_delay_slot && 2067 node_pipeline->hasBranchDelay() && 2068 !_unconditional_delay_slot) { 2069 2070 uint siz = _available.size(); 2071 2072 // Conditional branches can support an instruction that 2073 // is unconditionally executed and not dependent by the 2074 // branch, OR a conditionally executed instruction if 2075 // the branch is taken. In practice, this means that 2076 // the first instruction at the branch target is 2077 // copied to the delay slot, and the branch goes to 2078 // the instruction after that at the branch target 2079 if ( n->is_MachBranch() ) { 2080 2081 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 2082 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 2083 2084 #ifndef PRODUCT 2085 _branches++; 2086 #endif 2087 2088 // At least 1 instruction is on the available list 2089 // that is not dependent on the branch 2090 for (uint i = 0; i < siz; i++) { 2091 Node *d = _available[i]; 2092 const Pipeline *avail_pipeline = d->pipeline(); 2093 2094 // Don't allow safepoints in the branch shadow, that will 2095 // cause a number of difficulties 2096 if ( avail_pipeline->instructionCount() == 1 && 2097 !avail_pipeline->hasMultipleBundles() && 2098 !avail_pipeline->hasBranchDelay() && 2099 Pipeline::instr_has_unit_size() && 2100 d->size(_regalloc) == Pipeline::instr_unit_size() && 2101 NodeFitsInBundle(d) && 2102 !node_bundling(d)->used_in_delay()) { 2103 2104 if (d->is_Mach() && !d->is_MachSafePoint()) { 2105 // A node that fits in the delay slot was found, so we need to 2106 // set the appropriate bits in the bundle pipeline information so 2107 // that it correctly indicates resource usage. Later, when we 2108 // attempt to add this instruction to the bundle, we will skip 2109 // setting the resource usage. 2110 _unconditional_delay_slot = d; 2111 node_bundling(n)->set_use_unconditional_delay(); 2112 node_bundling(d)->set_used_in_unconditional_delay(); 2113 _bundle_use.add_usage(avail_pipeline->resourceUse()); 2114 _current_latency[d->_idx] = _bundle_cycle_number; 2115 _next_node = d; 2116 ++_bundle_instr_count; 2117 #ifndef PRODUCT 2118 _unconditional_delays++; 2119 #endif 2120 break; 2121 } 2122 } 2123 } 2124 } 2125 2126 // No delay slot, add a nop to the usage 2127 if (!_unconditional_delay_slot) { 2128 // See if adding an instruction in the delay slot will overflow 2129 // the bundle. 2130 if (!NodeFitsInBundle(_nop)) { 2131 #ifndef PRODUCT 2132 if (_cfg->C->trace_opto_output()) 2133 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2134 #endif 2135 step(1); 2136 } 2137 2138 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2139 _next_node = _nop; 2140 ++_bundle_instr_count; 2141 } 2142 2143 // See if the instruction in the delay slot requires a 2144 // step of the bundles 2145 if (!NodeFitsInBundle(n)) { 2146 #ifndef PRODUCT 2147 if (_cfg->C->trace_opto_output()) 2148 tty->print("# *** STEP(branch won't fit) ***\n"); 2149 #endif 2150 // Update the state information 2151 _bundle_instr_count = 0; 2152 _bundle_cycle_number += 1; 2153 _bundle_use.step(1); 2154 } 2155 } 2156 2157 // Get the number of instructions 2158 uint instruction_count = node_pipeline->instructionCount(); 2159 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2160 instruction_count = 0; 2161 2162 // Compute the latency information 2163 uint delay = 0; 2164 2165 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2166 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2167 if (relative_latency < 0) 2168 relative_latency = 0; 2169 2170 delay = _bundle_use.full_latency(relative_latency, node_usage); 2171 2172 // Does not fit in this bundle, start a new one 2173 if (delay > 0) { 2174 step(delay); 2175 2176 #ifndef PRODUCT 2177 if (_cfg->C->trace_opto_output()) 2178 tty->print("# *** STEP(%d) ***\n", delay); 2179 #endif 2180 } 2181 } 2182 2183 // If this was placed in the delay slot, ignore it 2184 if (n != _unconditional_delay_slot) { 2185 2186 if (delay == 0) { 2187 if (node_pipeline->hasMultipleBundles()) { 2188 #ifndef PRODUCT 2189 if (_cfg->C->trace_opto_output()) 2190 tty->print("# *** STEP(multiple instructions) ***\n"); 2191 #endif 2192 step(1); 2193 } 2194 2195 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2196 #ifndef PRODUCT 2197 if (_cfg->C->trace_opto_output()) 2198 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2199 instruction_count + _bundle_instr_count, 2200 Pipeline::_max_instrs_per_cycle); 2201 #endif 2202 step(1); 2203 } 2204 } 2205 2206 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2207 _bundle_instr_count++; 2208 2209 // Set the node's latency 2210 _current_latency[n->_idx] = _bundle_cycle_number; 2211 2212 // Now merge the functional unit information 2213 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2214 _bundle_use.add_usage(node_usage); 2215 2216 // Increment the number of instructions in this bundle 2217 _bundle_instr_count += instruction_count; 2218 2219 // Remember this node for later 2220 if (n->is_Mach()) 2221 _next_node = n; 2222 } 2223 2224 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2225 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2226 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2227 // into the block. All other scheduled nodes get put in the schedule here. 2228 int op = n->Opcode(); 2229 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2230 (op != Op_Node && // Not an unused antidepedence node and 2231 // not an unallocated boxlock 2232 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2233 2234 // Push any trailing projections 2235 if( bb->get_node(bb->number_of_nodes()-1) != n ) { 2236 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2237 Node *foi = n->fast_out(i); 2238 if( foi->is_Proj() ) 2239 _scheduled.push(foi); 2240 } 2241 } 2242 2243 // Put the instruction in the schedule list 2244 _scheduled.push(n); 2245 } 2246 2247 #ifndef PRODUCT 2248 if (_cfg->C->trace_opto_output()) 2249 dump_available(); 2250 #endif 2251 2252 // Walk all the definitions, decrementing use counts, and 2253 // if a definition has a 0 use count, place it in the available list. 2254 DecrementUseCounts(n,bb); 2255 } 2256 2257 // This method sets the use count within a basic block. We will ignore all 2258 // uses outside the current basic block. As we are doing a backwards walk, 2259 // any node we reach that has a use count of 0 may be scheduled. This also 2260 // avoids the problem of cyclic references from phi nodes, as long as phi 2261 // nodes are at the front of the basic block. This method also initializes 2262 // the available list to the set of instructions that have no uses within this 2263 // basic block. 2264 void Scheduling::ComputeUseCount(const Block *bb) { 2265 #ifndef PRODUCT 2266 if (_cfg->C->trace_opto_output()) 2267 tty->print("# -> ComputeUseCount\n"); 2268 #endif 2269 2270 // Clear the list of available and scheduled instructions, just in case 2271 _available.clear(); 2272 _scheduled.clear(); 2273 2274 // No delay slot specified 2275 _unconditional_delay_slot = NULL; 2276 2277 #ifdef ASSERT 2278 for( uint i=0; i < bb->number_of_nodes(); i++ ) 2279 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" ); 2280 #endif 2281 2282 // Force the _uses count to never go to zero for unscheduable pieces 2283 // of the block 2284 for( uint k = 0; k < _bb_start; k++ ) 2285 _uses[bb->get_node(k)->_idx] = 1; 2286 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ ) 2287 _uses[bb->get_node(l)->_idx] = 1; 2288 2289 // Iterate backwards over the instructions in the block. Don't count the 2290 // branch projections at end or the block header instructions. 2291 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2292 Node *n = bb->get_node(j); 2293 if( n->is_Proj() ) continue; // Projections handled another way 2294 2295 // Account for all uses 2296 for ( uint k = 0; k < n->len(); k++ ) { 2297 Node *inp = n->in(k); 2298 if (!inp) continue; 2299 assert(inp != n, "no cycles allowed" ); 2300 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use? 2301 if (inp->is_Proj()) { // Skip through Proj's 2302 inp = inp->in(0); 2303 } 2304 ++_uses[inp->_idx]; // Count 1 block-local use 2305 } 2306 } 2307 2308 // If this instruction has a 0 use count, then it is available 2309 if (!_uses[n->_idx]) { 2310 _current_latency[n->_idx] = _bundle_cycle_number; 2311 AddNodeToAvailableList(n); 2312 } 2313 2314 #ifndef PRODUCT 2315 if (_cfg->C->trace_opto_output()) { 2316 tty->print("# uses: %3d: ", _uses[n->_idx]); 2317 n->dump(); 2318 } 2319 #endif 2320 } 2321 2322 #ifndef PRODUCT 2323 if (_cfg->C->trace_opto_output()) 2324 tty->print("# <- ComputeUseCount\n"); 2325 #endif 2326 } 2327 2328 // This routine performs scheduling on each basic block in reverse order, 2329 // using instruction latencies and taking into account function unit 2330 // availability. 2331 void Scheduling::DoScheduling() { 2332 #ifndef PRODUCT 2333 if (_cfg->C->trace_opto_output()) 2334 tty->print("# -> DoScheduling\n"); 2335 #endif 2336 2337 Block *succ_bb = NULL; 2338 Block *bb; 2339 2340 // Walk over all the basic blocks in reverse order 2341 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) { 2342 bb = _cfg->get_block(i); 2343 2344 #ifndef PRODUCT 2345 if (_cfg->C->trace_opto_output()) { 2346 tty->print("# Schedule BB#%03d (initial)\n", i); 2347 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2348 bb->get_node(j)->dump(); 2349 } 2350 } 2351 #endif 2352 2353 // On the head node, skip processing 2354 if (bb == _cfg->get_root_block()) { 2355 continue; 2356 } 2357 2358 // Skip empty, connector blocks 2359 if (bb->is_connector()) 2360 continue; 2361 2362 // If the following block is not the sole successor of 2363 // this one, then reset the pipeline information 2364 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2365 #ifndef PRODUCT 2366 if (_cfg->C->trace_opto_output()) { 2367 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2368 _next_node->_idx, _bundle_instr_count); 2369 } 2370 #endif 2371 step_and_clear(); 2372 } 2373 2374 // Leave untouched the starting instruction, any Phis, a CreateEx node 2375 // or Top. bb->get_node(_bb_start) is the first schedulable instruction. 2376 _bb_end = bb->number_of_nodes()-1; 2377 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2378 Node *n = bb->get_node(_bb_start); 2379 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2380 // Also, MachIdealNodes do not get scheduled 2381 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2382 MachNode *mach = n->as_Mach(); 2383 int iop = mach->ideal_Opcode(); 2384 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2385 if( iop == Op_Con ) continue; // Do not schedule Top 2386 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2387 mach->pipeline() == MachNode::pipeline_class() && 2388 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc 2389 continue; 2390 break; // Funny loop structure to be sure... 2391 } 2392 // Compute last "interesting" instruction in block - last instruction we 2393 // might schedule. _bb_end points just after last schedulable inst. We 2394 // normally schedule conditional branches (despite them being forced last 2395 // in the block), because they have delay slots we can fill. Calls all 2396 // have their delay slots filled in the template expansions, so we don't 2397 // bother scheduling them. 2398 Node *last = bb->get_node(_bb_end); 2399 // Ignore trailing NOPs. 2400 while (_bb_end > 0 && last->is_Mach() && 2401 last->as_Mach()->ideal_Opcode() == Op_Con) { 2402 last = bb->get_node(--_bb_end); 2403 } 2404 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, ""); 2405 if( last->is_Catch() || 2406 (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2407 // There might be a prior call. Skip it. 2408 while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj()); 2409 } else if( last->is_MachNullCheck() ) { 2410 // Backup so the last null-checked memory instruction is 2411 // outside the schedulable range. Skip over the nullcheck, 2412 // projection, and the memory nodes. 2413 Node *mem = last->in(1); 2414 do { 2415 _bb_end--; 2416 } while (mem != bb->get_node(_bb_end)); 2417 } else { 2418 // Set _bb_end to point after last schedulable inst. 2419 _bb_end++; 2420 } 2421 2422 assert( _bb_start <= _bb_end, "inverted block ends" ); 2423 2424 // Compute the register antidependencies for the basic block 2425 ComputeRegisterAntidependencies(bb); 2426 if (_cfg->C->failing()) return; // too many D-U pinch points 2427 2428 // Compute intra-bb latencies for the nodes 2429 ComputeLocalLatenciesForward(bb); 2430 2431 // Compute the usage within the block, and set the list of all nodes 2432 // in the block that have no uses within the block. 2433 ComputeUseCount(bb); 2434 2435 // Schedule the remaining instructions in the block 2436 while ( _available.size() > 0 ) { 2437 Node *n = ChooseNodeToBundle(); 2438 guarantee(n != NULL, "no nodes available"); 2439 AddNodeToBundle(n,bb); 2440 } 2441 2442 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2443 #ifdef ASSERT 2444 for( uint l = _bb_start; l < _bb_end; l++ ) { 2445 Node *n = bb->get_node(l); 2446 uint m; 2447 for( m = 0; m < _bb_end-_bb_start; m++ ) 2448 if( _scheduled[m] == n ) 2449 break; 2450 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2451 } 2452 #endif 2453 2454 // Now copy the instructions (in reverse order) back to the block 2455 for ( uint k = _bb_start; k < _bb_end; k++ ) 2456 bb->map_node(_scheduled[_bb_end-k-1], k); 2457 2458 #ifndef PRODUCT 2459 if (_cfg->C->trace_opto_output()) { 2460 tty->print("# Schedule BB#%03d (final)\n", i); 2461 uint current = 0; 2462 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2463 Node *n = bb->get_node(j); 2464 if( valid_bundle_info(n) ) { 2465 Bundle *bundle = node_bundling(n); 2466 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2467 tty->print("*** Bundle: "); 2468 bundle->dump(); 2469 } 2470 n->dump(); 2471 } 2472 } 2473 } 2474 #endif 2475 #ifdef ASSERT 2476 verify_good_schedule(bb,"after block local scheduling"); 2477 #endif 2478 } 2479 2480 #ifndef PRODUCT 2481 if (_cfg->C->trace_opto_output()) 2482 tty->print("# <- DoScheduling\n"); 2483 #endif 2484 2485 // Record final node-bundling array location 2486 _regalloc->C->set_node_bundling_base(_node_bundling_base); 2487 2488 } // end DoScheduling 2489 2490 // Verify that no live-range used in the block is killed in the block by a 2491 // wrong DEF. This doesn't verify live-ranges that span blocks. 2492 2493 // Check for edge existence. Used to avoid adding redundant precedence edges. 2494 static bool edge_from_to( Node *from, Node *to ) { 2495 for( uint i=0; i<from->len(); i++ ) 2496 if( from->in(i) == to ) 2497 return true; 2498 return false; 2499 } 2500 2501 #ifdef ASSERT 2502 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2503 // Check for bad kills 2504 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2505 Node *prior_use = _reg_node[def]; 2506 if( prior_use && !edge_from_to(prior_use,n) ) { 2507 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2508 n->dump(); 2509 tty->print_cr("..."); 2510 prior_use->dump(); 2511 assert(edge_from_to(prior_use,n), "%s", msg); 2512 } 2513 _reg_node.map(def,NULL); // Kill live USEs 2514 } 2515 } 2516 2517 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2518 2519 // Zap to something reasonable for the verify code 2520 _reg_node.clear(); 2521 2522 // Walk over the block backwards. Check to make sure each DEF doesn't 2523 // kill a live value (other than the one it's supposed to). Add each 2524 // USE to the live set. 2525 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) { 2526 Node *n = b->get_node(i); 2527 int n_op = n->Opcode(); 2528 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2529 // Fat-proj kills a slew of registers 2530 RegMask rm = n->out_RegMask();// Make local copy 2531 while( rm.is_NotEmpty() ) { 2532 OptoReg::Name kill = rm.find_first_elem(); 2533 rm.Remove(kill); 2534 verify_do_def( n, kill, msg ); 2535 } 2536 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2537 // Get DEF'd registers the normal way 2538 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2539 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2540 } 2541 2542 // Now make all USEs live 2543 for( uint i=1; i<n->req(); i++ ) { 2544 Node *def = n->in(i); 2545 assert(def != 0, "input edge required"); 2546 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2547 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2548 if( OptoReg::is_valid(reg_lo) ) { 2549 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg); 2550 _reg_node.map(reg_lo,n); 2551 } 2552 if( OptoReg::is_valid(reg_hi) ) { 2553 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg); 2554 _reg_node.map(reg_hi,n); 2555 } 2556 } 2557 2558 } 2559 2560 // Zap to something reasonable for the Antidependence code 2561 _reg_node.clear(); 2562 } 2563 #endif 2564 2565 // Conditionally add precedence edges. Avoid putting edges on Projs. 2566 static void add_prec_edge_from_to( Node *from, Node *to ) { 2567 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2568 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2569 from = from->in(0); 2570 } 2571 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2572 !edge_from_to( from, to ) ) // Avoid duplicate edge 2573 from->add_prec(to); 2574 } 2575 2576 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2577 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2578 return; 2579 2580 Node *pinch = _reg_node[def_reg]; // Get pinch point 2581 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet? 2582 is_def ) { // Check for a true def (not a kill) 2583 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2584 return; 2585 } 2586 2587 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2588 debug_only( def = (Node*)((intptr_t)0xdeadbeef); ) 2589 2590 // After some number of kills there _may_ be a later def 2591 Node *later_def = NULL; 2592 2593 // Finding a kill requires a real pinch-point. 2594 // Check for not already having a pinch-point. 2595 // Pinch points are Op_Node's. 2596 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2597 later_def = pinch; // Must be def/kill as optimistic pinch-point 2598 if ( _pinch_free_list.size() > 0) { 2599 pinch = _pinch_free_list.pop(); 2600 } else { 2601 pinch = new Node(1); // Pinch point to-be 2602 } 2603 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2604 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2605 return; 2606 } 2607 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init) 2608 _reg_node.map(def_reg,pinch); // Record pinch-point 2609 //_regalloc->set_bad(pinch->_idx); // Already initialized this way. 2610 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2611 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call 2612 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2613 later_def = NULL; // and no later def 2614 } 2615 pinch->set_req(0,later_def); // Hook later def so we can find it 2616 } else { // Else have valid pinch point 2617 if( pinch->in(0) ) // If there is a later-def 2618 later_def = pinch->in(0); // Get it 2619 } 2620 2621 // Add output-dependence edge from later def to kill 2622 if( later_def ) // If there is some original def 2623 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2624 2625 // See if current kill is also a use, and so is forced to be the pinch-point. 2626 if( pinch->Opcode() == Op_Node ) { 2627 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2628 for( uint i=1; i<uses->req(); i++ ) { 2629 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2630 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2631 // Yes, found a use/kill pinch-point 2632 pinch->set_req(0,NULL); // 2633 pinch->replace_by(kill); // Move anti-dep edges up 2634 pinch = kill; 2635 _reg_node.map(def_reg,pinch); 2636 return; 2637 } 2638 } 2639 } 2640 2641 // Add edge from kill to pinch-point 2642 add_prec_edge_from_to(kill,pinch); 2643 } 2644 2645 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2646 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2647 return; 2648 Node *pinch = _reg_node[use_reg]; // Get pinch point 2649 // Check for no later def_reg/kill in block 2650 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b && 2651 // Use has to be block-local as well 2652 _cfg->get_block_for_node(use) == b) { 2653 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2654 pinch->req() == 1 ) { // pinch not yet in block? 2655 pinch->del_req(0); // yank pointer to later-def, also set flag 2656 // Insert the pinch-point in the block just after the last use 2657 b->insert_node(pinch, b->find_node(use) + 1); 2658 _bb_end++; // Increase size scheduled region in block 2659 } 2660 2661 add_prec_edge_from_to(pinch,use); 2662 } 2663 } 2664 2665 // We insert antidependences between the reads and following write of 2666 // allocated registers to prevent illegal code motion. Hopefully, the 2667 // number of added references should be fairly small, especially as we 2668 // are only adding references within the current basic block. 2669 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2670 2671 #ifdef ASSERT 2672 verify_good_schedule(b,"before block local scheduling"); 2673 #endif 2674 2675 // A valid schedule, for each register independently, is an endless cycle 2676 // of: a def, then some uses (connected to the def by true dependencies), 2677 // then some kills (defs with no uses), finally the cycle repeats with a new 2678 // def. The uses are allowed to float relative to each other, as are the 2679 // kills. No use is allowed to slide past a kill (or def). This requires 2680 // antidependencies between all uses of a single def and all kills that 2681 // follow, up to the next def. More edges are redundant, because later defs 2682 // & kills are already serialized with true or antidependencies. To keep 2683 // the edge count down, we add a 'pinch point' node if there's more than 2684 // one use or more than one kill/def. 2685 2686 // We add dependencies in one bottom-up pass. 2687 2688 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2689 2690 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2691 // register. If not, we record the DEF/KILL in _reg_node, the 2692 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2693 // "pinch point", a new Node that's in the graph but not in the block. 2694 // We put edges from the prior and current DEF/KILLs to the pinch point. 2695 // We put the pinch point in _reg_node. If there's already a pinch point 2696 // we merely add an edge from the current DEF/KILL to the pinch point. 2697 2698 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2699 // put an edge from the pinch point to the USE. 2700 2701 // To be expedient, the _reg_node array is pre-allocated for the whole 2702 // compilation. _reg_node is lazily initialized; it either contains a NULL, 2703 // or a valid def/kill/pinch-point, or a leftover node from some prior 2704 // block. Leftover node from some prior block is treated like a NULL (no 2705 // prior def, so no anti-dependence needed). Valid def is distinguished by 2706 // it being in the current block. 2707 bool fat_proj_seen = false; 2708 uint last_safept = _bb_end-1; 2709 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL; 2710 Node* last_safept_node = end_node; 2711 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2712 Node *n = b->get_node(i); 2713 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2714 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) { 2715 // Fat-proj kills a slew of registers 2716 // This can add edges to 'n' and obscure whether or not it was a def, 2717 // hence the is_def flag. 2718 fat_proj_seen = true; 2719 RegMask rm = n->out_RegMask();// Make local copy 2720 while( rm.is_NotEmpty() ) { 2721 OptoReg::Name kill = rm.find_first_elem(); 2722 rm.Remove(kill); 2723 anti_do_def( b, n, kill, is_def ); 2724 } 2725 } else { 2726 // Get DEF'd registers the normal way 2727 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2728 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2729 } 2730 2731 // Kill projections on a branch should appear to occur on the 2732 // branch, not afterwards, so grab the masks from the projections 2733 // and process them. 2734 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) { 2735 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2736 Node* use = n->fast_out(i); 2737 if (use->is_Proj()) { 2738 RegMask rm = use->out_RegMask();// Make local copy 2739 while( rm.is_NotEmpty() ) { 2740 OptoReg::Name kill = rm.find_first_elem(); 2741 rm.Remove(kill); 2742 anti_do_def( b, n, kill, false ); 2743 } 2744 } 2745 } 2746 } 2747 2748 // Check each register used by this instruction for a following DEF/KILL 2749 // that must occur afterward and requires an anti-dependence edge. 2750 for( uint j=0; j<n->req(); j++ ) { 2751 Node *def = n->in(j); 2752 if( def ) { 2753 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2754 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2755 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2756 } 2757 } 2758 // Do not allow defs of new derived values to float above GC 2759 // points unless the base is definitely available at the GC point. 2760 2761 Node *m = b->get_node(i); 2762 2763 // Add precedence edge from following safepoint to use of derived pointer 2764 if( last_safept_node != end_node && 2765 m != last_safept_node) { 2766 for (uint k = 1; k < m->req(); k++) { 2767 const Type *t = m->in(k)->bottom_type(); 2768 if( t->isa_oop_ptr() && 2769 t->is_ptr()->offset() != 0 ) { 2770 last_safept_node->add_prec( m ); 2771 break; 2772 } 2773 } 2774 } 2775 2776 if( n->jvms() ) { // Precedence edge from derived to safept 2777 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2778 if( b->get_node(last_safept) != last_safept_node ) { 2779 last_safept = b->find_node(last_safept_node); 2780 } 2781 for( uint j=last_safept; j > i; j-- ) { 2782 Node *mach = b->get_node(j); 2783 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2784 mach->add_prec( n ); 2785 } 2786 last_safept = i; 2787 last_safept_node = m; 2788 } 2789 } 2790 2791 if (fat_proj_seen) { 2792 // Garbage collect pinch nodes that were not consumed. 2793 // They are usually created by a fat kill MachProj for a call. 2794 garbage_collect_pinch_nodes(); 2795 } 2796 } 2797 2798 // Garbage collect pinch nodes for reuse by other blocks. 2799 // 2800 // The block scheduler's insertion of anti-dependence 2801 // edges creates many pinch nodes when the block contains 2802 // 2 or more Calls. A pinch node is used to prevent a 2803 // combinatorial explosion of edges. If a set of kills for a 2804 // register is anti-dependent on a set of uses (or defs), rather 2805 // than adding an edge in the graph between each pair of kill 2806 // and use (or def), a pinch is inserted between them: 2807 // 2808 // use1 use2 use3 2809 // \ | / 2810 // \ | / 2811 // pinch 2812 // / | \ 2813 // / | \ 2814 // kill1 kill2 kill3 2815 // 2816 // One pinch node is created per register killed when 2817 // the second call is encountered during a backwards pass 2818 // over the block. Most of these pinch nodes are never 2819 // wired into the graph because the register is never 2820 // used or def'ed in the block. 2821 // 2822 void Scheduling::garbage_collect_pinch_nodes() { 2823 #ifndef PRODUCT 2824 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 2825 #endif 2826 int trace_cnt = 0; 2827 for (uint k = 0; k < _reg_node.Size(); k++) { 2828 Node* pinch = _reg_node[k]; 2829 if ((pinch != NULL) && pinch->Opcode() == Op_Node && 2830 // no predecence input edges 2831 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { 2832 cleanup_pinch(pinch); 2833 _pinch_free_list.push(pinch); 2834 _reg_node.map(k, NULL); 2835 #ifndef PRODUCT 2836 if (_cfg->C->trace_opto_output()) { 2837 trace_cnt++; 2838 if (trace_cnt > 40) { 2839 tty->print("\n"); 2840 trace_cnt = 0; 2841 } 2842 tty->print(" %d", pinch->_idx); 2843 } 2844 #endif 2845 } 2846 } 2847 #ifndef PRODUCT 2848 if (_cfg->C->trace_opto_output()) tty->print("\n"); 2849 #endif 2850 } 2851 2852 // Clean up a pinch node for reuse. 2853 void Scheduling::cleanup_pinch( Node *pinch ) { 2854 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 2855 2856 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 2857 Node* use = pinch->last_out(i); 2858 uint uses_found = 0; 2859 for (uint j = use->req(); j < use->len(); j++) { 2860 if (use->in(j) == pinch) { 2861 use->rm_prec(j); 2862 uses_found++; 2863 } 2864 } 2865 assert(uses_found > 0, "must be a precedence edge"); 2866 i -= uses_found; // we deleted 1 or more copies of this edge 2867 } 2868 // May have a later_def entry 2869 pinch->set_req(0, NULL); 2870 } 2871 2872 #ifndef PRODUCT 2873 2874 void Scheduling::dump_available() const { 2875 tty->print("#Availist "); 2876 for (uint i = 0; i < _available.size(); i++) 2877 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 2878 tty->cr(); 2879 } 2880 2881 // Print Scheduling Statistics 2882 void Scheduling::print_statistics() { 2883 // Print the size added by nops for bundling 2884 tty->print("Nops added %d bytes to total of %d bytes", 2885 _total_nop_size, _total_method_size); 2886 if (_total_method_size > 0) 2887 tty->print(", for %.2f%%", 2888 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 2889 tty->print("\n"); 2890 2891 // Print the number of branch shadows filled 2892 if (Pipeline::_branch_has_delay_slot) { 2893 tty->print("Of %d branches, %d had unconditional delay slots filled", 2894 _total_branches, _total_unconditional_delays); 2895 if (_total_branches > 0) 2896 tty->print(", for %.2f%%", 2897 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 2898 tty->print("\n"); 2899 } 2900 2901 uint total_instructions = 0, total_bundles = 0; 2902 2903 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 2904 uint bundle_count = _total_instructions_per_bundle[i]; 2905 total_instructions += bundle_count * i; 2906 total_bundles += bundle_count; 2907 } 2908 2909 if (total_bundles > 0) 2910 tty->print("Average ILP (excluding nops) is %.2f\n", 2911 ((double)total_instructions) / ((double)total_bundles)); 2912 } 2913 #endif