1 /* 2 * Copyright (c) 1998, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/assembler.inline.hpp" 27 #include "code/compiledIC.hpp" 28 #include "code/debugInfo.hpp" 29 #include "code/debugInfoRec.hpp" 30 #include "compiler/compileBroker.hpp" 31 #include "compiler/compilerDirectives.hpp" 32 #include "compiler/oopMap.hpp" 33 #include "memory/allocation.inline.hpp" 34 #include "opto/ad.hpp" 35 #include "opto/callnode.hpp" 36 #include "opto/cfgnode.hpp" 37 #include "opto/locknode.hpp" 38 #include "opto/machnode.hpp" 39 #include "opto/optoreg.hpp" 40 #include "opto/output.hpp" 41 #include "opto/regalloc.hpp" 42 #include "opto/runtime.hpp" 43 #include "opto/subnode.hpp" 44 #include "opto/type.hpp" 45 #include "runtime/handles.inline.hpp" 46 #include "utilities/xmlstream.hpp" 47 48 #ifndef PRODUCT 49 #define DEBUG_ARG(x) , x 50 #else 51 #define DEBUG_ARG(x) 52 #endif 53 54 // Convert Nodes to instruction bits and pass off to the VM 55 void Compile::Output() { 56 // RootNode goes 57 assert( _cfg->get_root_block()->number_of_nodes() == 0, "" ); 58 59 // The number of new nodes (mostly MachNop) is proportional to 60 // the number of java calls and inner loops which are aligned. 61 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 + 62 C->inner_loops()*(OptoLoopAlignment-1)), 63 "out of nodes before code generation" ) ) { 64 return; 65 } 66 // Make sure I can find the Start Node 67 Block *entry = _cfg->get_block(1); 68 Block *broot = _cfg->get_root_block(); 69 70 const StartNode *start = entry->head()->as_Start(); 71 72 // Replace StartNode with prolog 73 MachPrologNode *prolog = new MachPrologNode(); 74 entry->map_node(prolog, 0); 75 _cfg->map_node_to_block(prolog, entry); 76 _cfg->unmap_node_from_block(start); // start is no longer in any block 77 78 // Virtual methods need an unverified entry point 79 80 if( is_osr_compilation() ) { 81 if( PoisonOSREntry ) { 82 // TODO: Should use a ShouldNotReachHereNode... 83 _cfg->insert( broot, 0, new MachBreakpointNode() ); 84 } 85 } else { 86 if( _method && !_method->flags().is_static() ) { 87 // Insert unvalidated entry point 88 _cfg->insert( broot, 0, new MachUEPNode() ); 89 } 90 91 } 92 93 // Break before main entry point 94 if ((_method && C->directive()->BreakAtExecuteOption) || 95 (OptoBreakpoint && is_method_compilation()) || 96 (OptoBreakpointOSR && is_osr_compilation()) || 97 (OptoBreakpointC2R && !_method) ) { 98 // checking for _method means that OptoBreakpoint does not apply to 99 // runtime stubs or frame converters 100 _cfg->insert( entry, 1, new MachBreakpointNode() ); 101 } 102 103 // Insert epilogs before every return 104 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 105 Block* block = _cfg->get_block(i); 106 if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point? 107 Node* m = block->end(); 108 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) { 109 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); 110 block->add_inst(epilog); 111 _cfg->map_node_to_block(epilog, block); 112 } 113 } 114 } 115 116 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1); 117 blk_starts[0] = 0; 118 119 // Initialize code buffer and process short branches. 120 CodeBuffer* cb = init_buffer(blk_starts); 121 122 if (cb == NULL || failing()) { 123 return; 124 } 125 126 ScheduleAndBundle(); 127 128 #ifndef PRODUCT 129 if (trace_opto_output()) { 130 tty->print("\n---- After ScheduleAndBundle ----\n"); 131 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 132 tty->print("\nBB#%03d:\n", i); 133 Block* block = _cfg->get_block(i); 134 for (uint j = 0; j < block->number_of_nodes(); j++) { 135 Node* n = block->get_node(j); 136 OptoReg::Name reg = _regalloc->get_reg_first(n); 137 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : ""); 138 n->dump(); 139 } 140 } 141 } 142 #endif 143 144 if (failing()) { 145 return; 146 } 147 148 BuildOopMaps(); 149 150 if (failing()) { 151 return; 152 } 153 154 fill_buffer(cb, blk_starts); 155 } 156 157 bool Compile::need_stack_bang(int frame_size_in_bytes) const { 158 // Determine if we need to generate a stack overflow check. 159 // Do it if the method is not a stub function and 160 // has java calls or has frame size > vm_page_size/8. 161 // The debug VM checks that deoptimization doesn't trigger an 162 // unexpected stack overflow (compiled method stack banging should 163 // guarantee it doesn't happen) so we always need the stack bang in 164 // a debug VM. 165 return (UseStackBanging && stub_function() == NULL && 166 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3 167 DEBUG_ONLY(|| true))); 168 } 169 170 bool Compile::need_register_stack_bang() const { 171 // Determine if we need to generate a register stack overflow check. 172 // This is only used on architectures which have split register 173 // and memory stacks (ie. IA64). 174 // Bang if the method is not a stub function and has java calls 175 return (stub_function() == NULL && has_java_calls()); 176 } 177 178 179 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top 180 // of a loop. When aligning a loop we need to provide enough instructions 181 // in cpu's fetch buffer to feed decoders. The loop alignment could be 182 // avoided if we have enough instructions in fetch buffer at the head of a loop. 183 // By default, the size is set to 999999 by Block's constructor so that 184 // a loop will be aligned if the size is not reset here. 185 // 186 // Note: Mach instructions could contain several HW instructions 187 // so the size is estimated only. 188 // 189 void Compile::compute_loop_first_inst_sizes() { 190 // The next condition is used to gate the loop alignment optimization. 191 // Don't aligned a loop if there are enough instructions at the head of a loop 192 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad 193 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is 194 // equal to 11 bytes which is the largest address NOP instruction. 195 if (MaxLoopPad < OptoLoopAlignment - 1) { 196 uint last_block = _cfg->number_of_blocks() - 1; 197 for (uint i = 1; i <= last_block; i++) { 198 Block* block = _cfg->get_block(i); 199 // Check the first loop's block which requires an alignment. 200 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) { 201 uint sum_size = 0; 202 uint inst_cnt = NumberOfLoopInstrToAlign; 203 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 204 205 // Check subsequent fallthrough blocks if the loop's first 206 // block(s) does not have enough instructions. 207 Block *nb = block; 208 while(inst_cnt > 0 && 209 i < last_block && 210 !_cfg->get_block(i + 1)->has_loop_alignment() && 211 !nb->has_successor(block)) { 212 i++; 213 nb = _cfg->get_block(i); 214 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 215 } // while( inst_cnt > 0 && i < last_block ) 216 217 block->set_first_inst_size(sum_size); 218 } // f( b->head()->is_Loop() ) 219 } // for( i <= last_block ) 220 } // if( MaxLoopPad < OptoLoopAlignment-1 ) 221 } 222 223 // The architecture description provides short branch variants for some long 224 // branch instructions. Replace eligible long branches with short branches. 225 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) { 226 // Compute size of each block, method size, and relocation information size 227 uint nblocks = _cfg->number_of_blocks(); 228 229 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 230 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 231 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks); 232 233 // Collect worst case block paddings 234 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks); 235 memset(block_worst_case_pad, 0, nblocks * sizeof(int)); 236 237 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); ) 238 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); ) 239 240 bool has_short_branch_candidate = false; 241 242 // Initialize the sizes to 0 243 code_size = 0; // Size in bytes of generated code 244 stub_size = 0; // Size in bytes of all stub entries 245 // Size in bytes of all relocation entries, including those in local stubs. 246 // Start with 2-bytes of reloc info for the unvalidated entry point 247 reloc_size = 1; // Number of relocation entries 248 249 // Make three passes. The first computes pessimistic blk_starts, 250 // relative jmp_offset and reloc_size information. The second performs 251 // short branch substitution using the pessimistic sizing. The 252 // third inserts nops where needed. 253 254 // Step one, perform a pessimistic sizing pass. 255 uint last_call_adr = max_juint; 256 uint last_avoid_back_to_back_adr = max_juint; 257 uint nop_size = (new MachNopNode())->size(_regalloc); 258 for (uint i = 0; i < nblocks; i++) { // For all blocks 259 Block* block = _cfg->get_block(i); 260 261 // During short branch replacement, we store the relative (to blk_starts) 262 // offset of jump in jmp_offset, rather than the absolute offset of jump. 263 // This is so that we do not need to recompute sizes of all nodes when 264 // we compute correct blk_starts in our next sizing pass. 265 jmp_offset[i] = 0; 266 jmp_size[i] = 0; 267 jmp_nidx[i] = -1; 268 DEBUG_ONLY( jmp_target[i] = 0; ) 269 DEBUG_ONLY( jmp_rule[i] = 0; ) 270 271 // Sum all instruction sizes to compute block size 272 uint last_inst = block->number_of_nodes(); 273 uint blk_size = 0; 274 for (uint j = 0; j < last_inst; j++) { 275 Node* nj = block->get_node(j); 276 // Handle machine instruction nodes 277 if (nj->is_Mach()) { 278 MachNode *mach = nj->as_Mach(); 279 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 280 reloc_size += mach->reloc(); 281 if (mach->is_MachCall()) { 282 // add size information for trampoline stub 283 // class CallStubImpl is platform-specific and defined in the *.ad files. 284 stub_size += CallStubImpl::size_call_trampoline(); 285 reloc_size += CallStubImpl::reloc_call_trampoline(); 286 287 MachCallNode *mcall = mach->as_MachCall(); 288 // This destination address is NOT PC-relative 289 290 mcall->method_set((intptr_t)mcall->entry_point()); 291 292 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) { 293 stub_size += CompiledStaticCall::to_interp_stub_size(); 294 reloc_size += CompiledStaticCall::reloc_to_interp_stub(); 295 } 296 } else if (mach->is_MachSafePoint()) { 297 // If call/safepoint are adjacent, account for possible 298 // nop to disambiguate the two safepoints. 299 // ScheduleAndBundle() can rearrange nodes in a block, 300 // check for all offsets inside this block. 301 if (last_call_adr >= blk_starts[i]) { 302 blk_size += nop_size; 303 } 304 } 305 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 306 // Nop is inserted between "avoid back to back" instructions. 307 // ScheduleAndBundle() can rearrange nodes in a block, 308 // check for all offsets inside this block. 309 if (last_avoid_back_to_back_adr >= blk_starts[i]) { 310 blk_size += nop_size; 311 } 312 } 313 if (mach->may_be_short_branch()) { 314 if (!nj->is_MachBranch()) { 315 #ifndef PRODUCT 316 nj->dump(3); 317 #endif 318 Unimplemented(); 319 } 320 assert(jmp_nidx[i] == -1, "block should have only one branch"); 321 jmp_offset[i] = blk_size; 322 jmp_size[i] = nj->size(_regalloc); 323 jmp_nidx[i] = j; 324 has_short_branch_candidate = true; 325 } 326 } 327 blk_size += nj->size(_regalloc); 328 // Remember end of call offset 329 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) { 330 last_call_adr = blk_starts[i]+blk_size; 331 } 332 // Remember end of avoid_back_to_back offset 333 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 334 last_avoid_back_to_back_adr = blk_starts[i]+blk_size; 335 } 336 } 337 338 // When the next block starts a loop, we may insert pad NOP 339 // instructions. Since we cannot know our future alignment, 340 // assume the worst. 341 if (i < nblocks - 1) { 342 Block* nb = _cfg->get_block(i + 1); 343 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 344 if (max_loop_pad > 0) { 345 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 346 // Adjust last_call_adr and/or last_avoid_back_to_back_adr. 347 // If either is the last instruction in this block, bump by 348 // max_loop_pad in lock-step with blk_size, so sizing 349 // calculations in subsequent blocks still can conservatively 350 // detect that it may the last instruction in this block. 351 if (last_call_adr == blk_starts[i]+blk_size) { 352 last_call_adr += max_loop_pad; 353 } 354 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) { 355 last_avoid_back_to_back_adr += max_loop_pad; 356 } 357 blk_size += max_loop_pad; 358 block_worst_case_pad[i + 1] = max_loop_pad; 359 } 360 } 361 362 // Save block size; update total method size 363 blk_starts[i+1] = blk_starts[i]+blk_size; 364 } 365 366 // Step two, replace eligible long jumps. 367 bool progress = true; 368 uint last_may_be_short_branch_adr = max_juint; 369 while (has_short_branch_candidate && progress) { 370 progress = false; 371 has_short_branch_candidate = false; 372 int adjust_block_start = 0; 373 for (uint i = 0; i < nblocks; i++) { 374 Block* block = _cfg->get_block(i); 375 int idx = jmp_nidx[i]; 376 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach(); 377 if (mach != NULL && mach->may_be_short_branch()) { 378 #ifdef ASSERT 379 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity"); 380 int j; 381 // Find the branch; ignore trailing NOPs. 382 for (j = block->number_of_nodes()-1; j>=0; j--) { 383 Node* n = block->get_node(j); 384 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) 385 break; 386 } 387 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity"); 388 #endif 389 int br_size = jmp_size[i]; 390 int br_offs = blk_starts[i] + jmp_offset[i]; 391 392 // This requires the TRUE branch target be in succs[0] 393 uint bnum = block->non_connector_successor(0)->_pre_order; 394 int offset = blk_starts[bnum] - br_offs; 395 if (bnum > i) { // adjust following block's offset 396 offset -= adjust_block_start; 397 } 398 399 // This block can be a loop header, account for the padding 400 // in the previous block. 401 int block_padding = block_worst_case_pad[i]; 402 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top"); 403 // In the following code a nop could be inserted before 404 // the branch which will increase the backward distance. 405 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr); 406 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block"); 407 408 if (needs_padding && offset <= 0) 409 offset -= nop_size; 410 411 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 412 // We've got a winner. Replace this branch. 413 MachNode* replacement = mach->as_MachBranch()->short_branch_version(); 414 415 // Update the jmp_size. 416 int new_size = replacement->size(_regalloc); 417 int diff = br_size - new_size; 418 assert(diff >= (int)nop_size, "short_branch size should be smaller"); 419 // Conservatively take into account padding between 420 // avoid_back_to_back branches. Previous branch could be 421 // converted into avoid_back_to_back branch during next 422 // rounds. 423 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 424 jmp_offset[i] += nop_size; 425 diff -= nop_size; 426 } 427 adjust_block_start += diff; 428 block->map_node(replacement, idx); 429 mach->subsume_by(replacement, C); 430 mach = replacement; 431 progress = true; 432 433 jmp_size[i] = new_size; 434 DEBUG_ONLY( jmp_target[i] = bnum; ); 435 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 436 } else { 437 // The jump distance is not short, try again during next iteration. 438 has_short_branch_candidate = true; 439 } 440 } // (mach->may_be_short_branch()) 441 if (mach != NULL && (mach->may_be_short_branch() || 442 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) { 443 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i]; 444 } 445 blk_starts[i+1] -= adjust_block_start; 446 } 447 } 448 449 #ifdef ASSERT 450 for (uint i = 0; i < nblocks; i++) { // For all blocks 451 if (jmp_target[i] != 0) { 452 int br_size = jmp_size[i]; 453 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 454 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 455 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]); 456 } 457 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp"); 458 } 459 } 460 #endif 461 462 // Step 3, compute the offsets of all blocks, will be done in fill_buffer() 463 // after ScheduleAndBundle(). 464 465 // ------------------ 466 // Compute size for code buffer 467 code_size = blk_starts[nblocks]; 468 469 // Relocation records 470 reloc_size += 1; // Relo entry for exception handler 471 472 // Adjust reloc_size to number of record of relocation info 473 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 474 // a relocation index. 475 // The CodeBuffer will expand the locs array if this estimate is too low. 476 reloc_size *= 10 / sizeof(relocInfo); 477 } 478 479 //------------------------------FillLocArray----------------------------------- 480 // Create a bit of debug info and append it to the array. The mapping is from 481 // Java local or expression stack to constant, register or stack-slot. For 482 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 483 // entry has been taken care of and caller should skip it). 484 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 485 // This should never have accepted Bad before 486 assert(OptoReg::is_valid(regnum), "location must be valid"); 487 return (OptoReg::is_reg(regnum)) 488 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 489 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 490 } 491 492 493 ObjectValue* 494 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 495 for (int i = 0; i < objs->length(); i++) { 496 assert(objs->at(i)->is_object(), "corrupt object cache"); 497 ObjectValue* sv = (ObjectValue*) objs->at(i); 498 if (sv->id() == id) { 499 return sv; 500 } 501 } 502 // Otherwise.. 503 return NULL; 504 } 505 506 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 507 ObjectValue* sv ) { 508 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition"); 509 objs->append(sv); 510 } 511 512 513 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 514 GrowableArray<ScopeValue*> *array, 515 GrowableArray<ScopeValue*> *objs ) { 516 assert( local, "use _top instead of null" ); 517 if (array->length() != idx) { 518 assert(array->length() == idx + 1, "Unexpected array count"); 519 // Old functionality: 520 // return 521 // New functionality: 522 // Assert if the local is not top. In product mode let the new node 523 // override the old entry. 524 assert(local == top(), "LocArray collision"); 525 if (local == top()) { 526 return; 527 } 528 array->pop(); 529 } 530 const Type *t = local->bottom_type(); 531 532 // Is it a safepoint scalar object node? 533 if (local->is_SafePointScalarObject()) { 534 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 535 536 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx); 537 if (sv == NULL) { 538 ciKlass* cik = t->is_oopptr()->klass(); 539 assert(cik->is_instance_klass() || 540 cik->is_array_klass(), "Not supported allocation."); 541 sv = new ObjectValue(spobj->_idx, 542 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 543 Compile::set_sv_for_object_node(objs, sv); 544 545 uint first_ind = spobj->first_index(sfpt->jvms()); 546 for (uint i = 0; i < spobj->n_fields(); i++) { 547 Node* fld_node = sfpt->in(first_ind+i); 548 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 549 } 550 } 551 array->append(sv); 552 return; 553 } 554 555 // Grab the register number for the local 556 OptoReg::Name regnum = _regalloc->get_reg_first(local); 557 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 558 // Record the double as two float registers. 559 // The register mask for such a value always specifies two adjacent 560 // float registers, with the lower register number even. 561 // Normally, the allocation of high and low words to these registers 562 // is irrelevant, because nearly all operations on register pairs 563 // (e.g., StoreD) treat them as a single unit. 564 // Here, we assume in addition that the words in these two registers 565 // stored "naturally" (by operations like StoreD and double stores 566 // within the interpreter) such that the lower-numbered register 567 // is written to the lower memory address. This may seem like 568 // a machine dependency, but it is not--it is a requirement on 569 // the author of the <arch>.ad file to ensure that, for every 570 // even/odd double-register pair to which a double may be allocated, 571 // the word in the even single-register is stored to the first 572 // memory word. (Note that register numbers are completely 573 // arbitrary, and are not tied to any machine-level encodings.) 574 #ifdef _LP64 575 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 576 array->append(new ConstantIntValue(0)); 577 array->append(new_loc_value( _regalloc, regnum, Location::dbl )); 578 } else if ( t->base() == Type::Long ) { 579 array->append(new ConstantIntValue(0)); 580 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 581 } else if ( t->base() == Type::RawPtr ) { 582 // jsr/ret return address which must be restored into a the full 583 // width 64-bit stack slot. 584 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 585 } 586 #else //_LP64 587 #ifdef SPARC 588 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) { 589 // For SPARC we have to swap high and low words for 590 // long values stored in a single-register (g0-g7). 591 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 592 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 593 } else 594 #endif //SPARC 595 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 596 // Repack the double/long as two jints. 597 // The convention the interpreter uses is that the second local 598 // holds the first raw word of the native double representation. 599 // This is actually reasonable, since locals and stack arrays 600 // grow downwards in all implementations. 601 // (If, on some machine, the interpreter's Java locals or stack 602 // were to grow upwards, the embedded doubles would be word-swapped.) 603 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 604 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 605 } 606 #endif //_LP64 607 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 608 OptoReg::is_reg(regnum) ) { 609 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double() 610 ? Location::float_in_dbl : Location::normal )); 611 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 612 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long 613 ? Location::int_in_long : Location::normal )); 614 } else if( t->base() == Type::NarrowOop ) { 615 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop )); 616 } else { 617 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal )); 618 } 619 return; 620 } 621 622 // No register. It must be constant data. 623 switch (t->base()) { 624 case Type::Half: // Second half of a double 625 ShouldNotReachHere(); // Caller should skip 2nd halves 626 break; 627 case Type::AnyPtr: 628 array->append(new ConstantOopWriteValue(NULL)); 629 break; 630 case Type::AryPtr: 631 case Type::InstPtr: // fall through 632 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 633 break; 634 case Type::NarrowOop: 635 if (t == TypeNarrowOop::NULL_PTR) { 636 array->append(new ConstantOopWriteValue(NULL)); 637 } else { 638 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 639 } 640 break; 641 case Type::Int: 642 array->append(new ConstantIntValue(t->is_int()->get_con())); 643 break; 644 case Type::RawPtr: 645 // A return address (T_ADDRESS). 646 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 647 #ifdef _LP64 648 // Must be restored to the full-width 64-bit stack slot. 649 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 650 #else 651 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 652 #endif 653 break; 654 case Type::FloatCon: { 655 float f = t->is_float_constant()->getf(); 656 array->append(new ConstantIntValue(jint_cast(f))); 657 break; 658 } 659 case Type::DoubleCon: { 660 jdouble d = t->is_double_constant()->getd(); 661 #ifdef _LP64 662 array->append(new ConstantIntValue(0)); 663 array->append(new ConstantDoubleValue(d)); 664 #else 665 // Repack the double as two jints. 666 // The convention the interpreter uses is that the second local 667 // holds the first raw word of the native double representation. 668 // This is actually reasonable, since locals and stack arrays 669 // grow downwards in all implementations. 670 // (If, on some machine, the interpreter's Java locals or stack 671 // were to grow upwards, the embedded doubles would be word-swapped.) 672 jlong_accessor acc; 673 acc.long_value = jlong_cast(d); 674 array->append(new ConstantIntValue(acc.words[1])); 675 array->append(new ConstantIntValue(acc.words[0])); 676 #endif 677 break; 678 } 679 case Type::Long: { 680 jlong d = t->is_long()->get_con(); 681 #ifdef _LP64 682 array->append(new ConstantIntValue(0)); 683 array->append(new ConstantLongValue(d)); 684 #else 685 // Repack the long as two jints. 686 // The convention the interpreter uses is that the second local 687 // holds the first raw word of the native double representation. 688 // This is actually reasonable, since locals and stack arrays 689 // grow downwards in all implementations. 690 // (If, on some machine, the interpreter's Java locals or stack 691 // were to grow upwards, the embedded doubles would be word-swapped.) 692 jlong_accessor acc; 693 acc.long_value = d; 694 array->append(new ConstantIntValue(acc.words[1])); 695 array->append(new ConstantIntValue(acc.words[0])); 696 #endif 697 break; 698 } 699 case Type::Top: // Add an illegal value here 700 array->append(new LocationValue(Location())); 701 break; 702 default: 703 ShouldNotReachHere(); 704 break; 705 } 706 } 707 708 // Determine if this node starts a bundle 709 bool Compile::starts_bundle(const Node *n) const { 710 return (_node_bundling_limit > n->_idx && 711 _node_bundling_base[n->_idx].starts_bundle()); 712 } 713 714 //--------------------------Process_OopMap_Node-------------------------------- 715 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) { 716 717 // Handle special safepoint nodes for synchronization 718 MachSafePointNode *sfn = mach->as_MachSafePoint(); 719 MachCallNode *mcall; 720 721 int safepoint_pc_offset = current_offset; 722 bool is_method_handle_invoke = false; 723 bool return_oop = false; 724 725 // Add the safepoint in the DebugInfoRecorder 726 if( !mach->is_MachCall() ) { 727 mcall = NULL; 728 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 729 } else { 730 mcall = mach->as_MachCall(); 731 732 // Is the call a MethodHandle call? 733 if (mcall->is_MachCallJava()) { 734 if (mcall->as_MachCallJava()->_method_handle_invoke) { 735 assert(has_method_handle_invokes(), "must have been set during call generation"); 736 is_method_handle_invoke = true; 737 } 738 } 739 740 // Check if a call returns an object. 741 if (mcall->returns_pointer()) { 742 return_oop = true; 743 } 744 safepoint_pc_offset += mcall->ret_addr_offset(); 745 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 746 } 747 748 // Loop over the JVMState list to add scope information 749 // Do not skip safepoints with a NULL method, they need monitor info 750 JVMState* youngest_jvms = sfn->jvms(); 751 int max_depth = youngest_jvms->depth(); 752 753 // Allocate the object pool for scalar-replaced objects -- the map from 754 // small-integer keys (which can be recorded in the local and ostack 755 // arrays) to descriptions of the object state. 756 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 757 758 // Visit scopes from oldest to youngest. 759 for (int depth = 1; depth <= max_depth; depth++) { 760 JVMState* jvms = youngest_jvms->of_depth(depth); 761 int idx; 762 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 763 // Safepoints that do not have method() set only provide oop-map and monitor info 764 // to support GC; these do not support deoptimization. 765 int num_locs = (method == NULL) ? 0 : jvms->loc_size(); 766 int num_exps = (method == NULL) ? 0 : jvms->stk_size(); 767 int num_mon = jvms->nof_monitors(); 768 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), 769 "JVMS local count must match that of the method"); 770 771 // Add Local and Expression Stack Information 772 773 // Insert locals into the locarray 774 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 775 for( idx = 0; idx < num_locs; idx++ ) { 776 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 777 } 778 779 // Insert expression stack entries into the exparray 780 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 781 for( idx = 0; idx < num_exps; idx++ ) { 782 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 783 } 784 785 // Add in mappings of the monitors 786 assert( !method || 787 !method->is_synchronized() || 788 method->is_native() || 789 num_mon > 0 || 790 !GenerateSynchronizationCode, 791 "monitors must always exist for synchronized methods"); 792 793 // Build the growable array of ScopeValues for exp stack 794 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 795 796 // Loop over monitors and insert into array 797 for (idx = 0; idx < num_mon; idx++) { 798 // Grab the node that defines this monitor 799 Node* box_node = sfn->monitor_box(jvms, idx); 800 Node* obj_node = sfn->monitor_obj(jvms, idx); 801 802 // Create ScopeValue for object 803 ScopeValue *scval = NULL; 804 805 if (obj_node->is_SafePointScalarObject()) { 806 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 807 scval = Compile::sv_for_node_id(objs, spobj->_idx); 808 if (scval == NULL) { 809 const Type *t = spobj->bottom_type(); 810 ciKlass* cik = t->is_oopptr()->klass(); 811 assert(cik->is_instance_klass() || 812 cik->is_array_klass(), "Not supported allocation."); 813 ObjectValue* sv = new ObjectValue(spobj->_idx, 814 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 815 Compile::set_sv_for_object_node(objs, sv); 816 817 uint first_ind = spobj->first_index(youngest_jvms); 818 for (uint i = 0; i < spobj->n_fields(); i++) { 819 Node* fld_node = sfn->in(first_ind+i); 820 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 821 } 822 scval = sv; 823 } 824 } else if (!obj_node->is_Con()) { 825 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); 826 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 827 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop ); 828 } else { 829 scval = new_loc_value( _regalloc, obj_reg, Location::oop ); 830 } 831 } else { 832 const TypePtr *tp = obj_node->get_ptr_type(); 833 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding()); 834 } 835 836 OptoReg::Name box_reg = BoxLockNode::reg(box_node); 837 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)); 838 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated()); 839 monarray->append(new MonitorValue(scval, basic_lock, eliminated)); 840 } 841 842 // We dump the object pool first, since deoptimization reads it in first. 843 debug_info()->dump_object_pool(objs); 844 845 // Build first class objects to pass to scope 846 DebugToken *locvals = debug_info()->create_scope_values(locarray); 847 DebugToken *expvals = debug_info()->create_scope_values(exparray); 848 DebugToken *monvals = debug_info()->create_monitor_values(monarray); 849 850 // Make method available for all Safepoints 851 ciMethod* scope_method = method ? method : _method; 852 // Describe the scope here 853 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 854 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 855 // Now we can describe the scope. 856 methodHandle null_mh; 857 bool rethrow_exception = false; 858 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); 859 } // End jvms loop 860 861 // Mark the end of the scope set. 862 debug_info()->end_safepoint(safepoint_pc_offset); 863 } 864 865 866 867 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 868 class NonSafepointEmitter { 869 Compile* C; 870 JVMState* _pending_jvms; 871 int _pending_offset; 872 873 void emit_non_safepoint(); 874 875 public: 876 NonSafepointEmitter(Compile* compile) { 877 this->C = compile; 878 _pending_jvms = NULL; 879 _pending_offset = 0; 880 } 881 882 void observe_instruction(Node* n, int pc_offset) { 883 if (!C->debug_info()->recording_non_safepoints()) return; 884 885 Node_Notes* nn = C->node_notes_at(n->_idx); 886 if (nn == NULL || nn->jvms() == NULL) return; 887 if (_pending_jvms != NULL && 888 _pending_jvms->same_calls_as(nn->jvms())) { 889 // Repeated JVMS? Stretch it up here. 890 _pending_offset = pc_offset; 891 } else { 892 if (_pending_jvms != NULL && 893 _pending_offset < pc_offset) { 894 emit_non_safepoint(); 895 } 896 _pending_jvms = NULL; 897 if (pc_offset > C->debug_info()->last_pc_offset()) { 898 // This is the only way _pending_jvms can become non-NULL: 899 _pending_jvms = nn->jvms(); 900 _pending_offset = pc_offset; 901 } 902 } 903 } 904 905 // Stay out of the way of real safepoints: 906 void observe_safepoint(JVMState* jvms, int pc_offset) { 907 if (_pending_jvms != NULL && 908 !_pending_jvms->same_calls_as(jvms) && 909 _pending_offset < pc_offset) { 910 emit_non_safepoint(); 911 } 912 _pending_jvms = NULL; 913 } 914 915 void flush_at_end() { 916 if (_pending_jvms != NULL) { 917 emit_non_safepoint(); 918 } 919 _pending_jvms = NULL; 920 } 921 }; 922 923 void NonSafepointEmitter::emit_non_safepoint() { 924 JVMState* youngest_jvms = _pending_jvms; 925 int pc_offset = _pending_offset; 926 927 // Clear it now: 928 _pending_jvms = NULL; 929 930 DebugInformationRecorder* debug_info = C->debug_info(); 931 assert(debug_info->recording_non_safepoints(), "sanity"); 932 933 debug_info->add_non_safepoint(pc_offset); 934 int max_depth = youngest_jvms->depth(); 935 936 // Visit scopes from oldest to youngest. 937 for (int depth = 1; depth <= max_depth; depth++) { 938 JVMState* jvms = youngest_jvms->of_depth(depth); 939 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 940 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 941 methodHandle null_mh; 942 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute()); 943 } 944 945 // Mark the end of the scope set. 946 debug_info->end_non_safepoint(pc_offset); 947 } 948 949 //------------------------------init_buffer------------------------------------ 950 CodeBuffer* Compile::init_buffer(uint* blk_starts) { 951 952 // Set the initially allocated size 953 int code_req = initial_code_capacity; 954 int locs_req = initial_locs_capacity; 955 int stub_req = initial_stub_capacity; 956 int const_req = initial_const_capacity; 957 958 int pad_req = NativeCall::instruction_size; 959 // The extra spacing after the code is necessary on some platforms. 960 // Sometimes we need to patch in a jump after the last instruction, 961 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 962 963 // Compute the byte offset where we can store the deopt pc. 964 if (fixed_slots() != 0) { 965 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 966 } 967 968 // Compute prolog code size 969 _method_size = 0; 970 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; 971 #if defined(IA64) && !defined(AIX) 972 if (save_argument_registers()) { 973 // 4815101: this is a stub with implicit and unknown precision fp args. 974 // The usual spill mechanism can only generate stfd's in this case, which 975 // doesn't work if the fp reg to spill contains a single-precision denorm. 976 // Instead, we hack around the normal spill mechanism using stfspill's and 977 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate 978 // space here for the fp arg regs (f8-f15) we're going to thusly spill. 979 // 980 // If we ever implement 16-byte 'registers' == stack slots, we can 981 // get rid of this hack and have SpillCopy generate stfspill/ldffill 982 // instead of stfd/stfs/ldfd/ldfs. 983 _frame_slots += 8*(16/BytesPerInt); 984 } 985 #endif 986 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check"); 987 988 if (has_mach_constant_base_node()) { 989 uint add_size = 0; 990 // Fill the constant table. 991 // Note: This must happen before shorten_branches. 992 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 993 Block* b = _cfg->get_block(i); 994 995 for (uint j = 0; j < b->number_of_nodes(); j++) { 996 Node* n = b->get_node(j); 997 998 // If the node is a MachConstantNode evaluate the constant 999 // value section. 1000 if (n->is_MachConstant()) { 1001 MachConstantNode* machcon = n->as_MachConstant(); 1002 machcon->eval_constant(C); 1003 } else if (n->is_Mach()) { 1004 // On Power there are more nodes that issue constants. 1005 add_size += (n->as_Mach()->ins_num_consts() * 8); 1006 } 1007 } 1008 } 1009 1010 // Calculate the offsets of the constants and the size of the 1011 // constant table (including the padding to the next section). 1012 constant_table().calculate_offsets_and_size(); 1013 const_req = constant_table().size() + add_size; 1014 } 1015 1016 // Initialize the space for the BufferBlob used to find and verify 1017 // instruction size in MachNode::emit_size() 1018 init_scratch_buffer_blob(const_req); 1019 if (failing()) return NULL; // Out of memory 1020 1021 // Pre-compute the length of blocks and replace 1022 // long branches with short if machine supports it. 1023 shorten_branches(blk_starts, code_req, locs_req, stub_req); 1024 1025 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1026 // class HandlerImpl is platform-specific and defined in the *.ad files. 1027 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler 1028 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler 1029 stub_req += MAX_stubs_size; // ensure per-stub margin 1030 code_req += MAX_inst_size; // ensure per-instruction margin 1031 1032 if (StressCodeBuffers) 1033 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1034 1035 int total_req = 1036 const_req + 1037 code_req + 1038 pad_req + 1039 stub_req + 1040 exception_handler_req + 1041 deopt_handler_req; // deopt handler 1042 1043 if (has_method_handle_invokes()) 1044 total_req += deopt_handler_req; // deopt MH handler 1045 1046 CodeBuffer* cb = code_buffer(); 1047 cb->initialize(total_req, locs_req); 1048 1049 // Have we run out of code space? 1050 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1051 C->record_failure("CodeCache is full"); 1052 return NULL; 1053 } 1054 // Configure the code buffer. 1055 cb->initialize_consts_size(const_req); 1056 cb->initialize_stubs_size(stub_req); 1057 cb->initialize_oop_recorder(env()->oop_recorder()); 1058 1059 // fill in the nop array for bundling computations 1060 MachNode *_nop_list[Bundle::_nop_count]; 1061 Bundle::initialize_nops(_nop_list); 1062 1063 return cb; 1064 } 1065 1066 //------------------------------fill_buffer------------------------------------ 1067 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) { 1068 // blk_starts[] contains offsets calculated during short branches processing, 1069 // offsets should not be increased during following steps. 1070 1071 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 1072 // of a loop. It is used to determine the padding for loop alignment. 1073 compute_loop_first_inst_sizes(); 1074 1075 // Create oopmap set. 1076 _oop_map_set = new OopMapSet(); 1077 1078 // !!!!! This preserves old handling of oopmaps for now 1079 debug_info()->set_oopmaps(_oop_map_set); 1080 1081 uint nblocks = _cfg->number_of_blocks(); 1082 // Count and start of implicit null check instructions 1083 uint inct_cnt = 0; 1084 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1085 1086 // Count and start of calls 1087 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1088 1089 uint return_offset = 0; 1090 int nop_size = (new MachNopNode())->size(_regalloc); 1091 1092 int previous_offset = 0; 1093 int current_offset = 0; 1094 int last_call_offset = -1; 1095 int last_avoid_back_to_back_offset = -1; 1096 #ifdef ASSERT 1097 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); 1098 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 1099 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 1100 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); 1101 #endif 1102 1103 // Create an array of unused labels, one for each basic block, if printing is enabled 1104 #ifndef PRODUCT 1105 int *node_offsets = NULL; 1106 uint node_offset_limit = unique(); 1107 1108 if (print_assembly()) 1109 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1110 #endif 1111 1112 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily 1113 1114 // Emit the constant table. 1115 if (has_mach_constant_base_node()) { 1116 constant_table().emit(*cb); 1117 } 1118 1119 // Create an array of labels, one for each basic block 1120 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1); 1121 for (uint i=0; i <= nblocks; i++) { 1122 blk_labels[i].init(); 1123 } 1124 1125 // ------------------ 1126 // Now fill in the code buffer 1127 Node *delay_slot = NULL; 1128 1129 for (uint i = 0; i < nblocks; i++) { 1130 Block* block = _cfg->get_block(i); 1131 Node* head = block->head(); 1132 1133 // If this block needs to start aligned (i.e, can be reached other 1134 // than by falling-thru from the previous block), then force the 1135 // start of a new bundle. 1136 if (Pipeline::requires_bundling() && starts_bundle(head)) { 1137 cb->flush_bundle(true); 1138 } 1139 1140 #ifdef ASSERT 1141 if (!block->is_connector()) { 1142 stringStream st; 1143 block->dump_head(_cfg, &st); 1144 MacroAssembler(cb).block_comment(st.as_string()); 1145 } 1146 jmp_target[i] = 0; 1147 jmp_offset[i] = 0; 1148 jmp_size[i] = 0; 1149 jmp_rule[i] = 0; 1150 #endif 1151 int blk_offset = current_offset; 1152 1153 // Define the label at the beginning of the basic block 1154 MacroAssembler(cb).bind(blk_labels[block->_pre_order]); 1155 1156 uint last_inst = block->number_of_nodes(); 1157 1158 // Emit block normally, except for last instruction. 1159 // Emit means "dump code bits into code buffer". 1160 for (uint j = 0; j<last_inst; j++) { 1161 1162 // Get the node 1163 Node* n = block->get_node(j); 1164 1165 // See if delay slots are supported 1166 if (valid_bundle_info(n) && 1167 node_bundling(n)->used_in_unconditional_delay()) { 1168 assert(delay_slot == NULL, "no use of delay slot node"); 1169 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1170 1171 delay_slot = n; 1172 continue; 1173 } 1174 1175 // If this starts a new instruction group, then flush the current one 1176 // (but allow split bundles) 1177 if (Pipeline::requires_bundling() && starts_bundle(n)) 1178 cb->flush_bundle(false); 1179 1180 // Special handling for SafePoint/Call Nodes 1181 bool is_mcall = false; 1182 if (n->is_Mach()) { 1183 MachNode *mach = n->as_Mach(); 1184 is_mcall = n->is_MachCall(); 1185 bool is_sfn = n->is_MachSafePoint(); 1186 1187 // If this requires all previous instructions be flushed, then do so 1188 if (is_sfn || is_mcall || mach->alignment_required() != 1) { 1189 cb->flush_bundle(true); 1190 current_offset = cb->insts_size(); 1191 } 1192 1193 // A padding may be needed again since a previous instruction 1194 // could be moved to delay slot. 1195 1196 // align the instruction if necessary 1197 int padding = mach->compute_padding(current_offset); 1198 // Make sure safepoint node for polling is distinct from a call's 1199 // return by adding a nop if needed. 1200 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) { 1201 padding = nop_size; 1202 } 1203 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) && 1204 current_offset == last_avoid_back_to_back_offset) { 1205 // Avoid back to back some instructions. 1206 padding = nop_size; 1207 } 1208 1209 if (padding > 0) { 1210 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1211 int nops_cnt = padding / nop_size; 1212 MachNode *nop = new MachNopNode(nops_cnt); 1213 block->insert_node(nop, j++); 1214 last_inst++; 1215 _cfg->map_node_to_block(nop, block); 1216 // Ensure enough space. 1217 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1218 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1219 C->record_failure("CodeCache is full"); 1220 return; 1221 } 1222 nop->emit(*cb, _regalloc); 1223 cb->flush_bundle(true); 1224 current_offset = cb->insts_size(); 1225 } 1226 1227 // Remember the start of the last call in a basic block 1228 if (is_mcall) { 1229 MachCallNode *mcall = mach->as_MachCall(); 1230 1231 // This destination address is NOT PC-relative 1232 mcall->method_set((intptr_t)mcall->entry_point()); 1233 1234 // Save the return address 1235 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset(); 1236 1237 if (mcall->is_MachCallLeaf()) { 1238 is_mcall = false; 1239 is_sfn = false; 1240 } 1241 } 1242 1243 // sfn will be valid whenever mcall is valid now because of inheritance 1244 if (is_sfn || is_mcall) { 1245 1246 // Handle special safepoint nodes for synchronization 1247 if (!is_mcall) { 1248 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1249 // !!!!! Stubs only need an oopmap right now, so bail out 1250 if (sfn->jvms()->method() == NULL) { 1251 // Write the oopmap directly to the code blob??!! 1252 continue; 1253 } 1254 } // End synchronization 1255 1256 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1257 current_offset); 1258 Process_OopMap_Node(mach, current_offset); 1259 } // End if safepoint 1260 1261 // If this is a null check, then add the start of the previous instruction to the list 1262 else if( mach->is_MachNullCheck() ) { 1263 inct_starts[inct_cnt++] = previous_offset; 1264 } 1265 1266 // If this is a branch, then fill in the label with the target BB's label 1267 else if (mach->is_MachBranch()) { 1268 // This requires the TRUE branch target be in succs[0] 1269 uint block_num = block->non_connector_successor(0)->_pre_order; 1270 1271 // Try to replace long branch if delay slot is not used, 1272 // it is mostly for back branches since forward branch's 1273 // distance is not updated yet. 1274 bool delay_slot_is_used = valid_bundle_info(n) && 1275 node_bundling(n)->use_unconditional_delay(); 1276 if (!delay_slot_is_used && mach->may_be_short_branch()) { 1277 assert(delay_slot == NULL, "not expecting delay slot node"); 1278 int br_size = n->size(_regalloc); 1279 int offset = blk_starts[block_num] - current_offset; 1280 if (block_num >= i) { 1281 // Current and following block's offset are not 1282 // finalized yet, adjust distance by the difference 1283 // between calculated and final offsets of current block. 1284 offset -= (blk_starts[i] - blk_offset); 1285 } 1286 // In the following code a nop could be inserted before 1287 // the branch which will increase the backward distance. 1288 bool needs_padding = (current_offset == last_avoid_back_to_back_offset); 1289 if (needs_padding && offset <= 0) 1290 offset -= nop_size; 1291 1292 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 1293 // We've got a winner. Replace this branch. 1294 MachNode* replacement = mach->as_MachBranch()->short_branch_version(); 1295 1296 // Update the jmp_size. 1297 int new_size = replacement->size(_regalloc); 1298 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller"); 1299 // Insert padding between avoid_back_to_back branches. 1300 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 1301 MachNode *nop = new MachNopNode(); 1302 block->insert_node(nop, j++); 1303 _cfg->map_node_to_block(nop, block); 1304 last_inst++; 1305 nop->emit(*cb, _regalloc); 1306 cb->flush_bundle(true); 1307 current_offset = cb->insts_size(); 1308 } 1309 #ifdef ASSERT 1310 jmp_target[i] = block_num; 1311 jmp_offset[i] = current_offset - blk_offset; 1312 jmp_size[i] = new_size; 1313 jmp_rule[i] = mach->rule(); 1314 #endif 1315 block->map_node(replacement, j); 1316 mach->subsume_by(replacement, C); 1317 n = replacement; 1318 mach = replacement; 1319 } 1320 } 1321 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num ); 1322 } else if (mach->ideal_Opcode() == Op_Jump) { 1323 for (uint h = 0; h < block->_num_succs; h++) { 1324 Block* succs_block = block->_succs[h]; 1325 for (uint j = 1; j < succs_block->num_preds(); j++) { 1326 Node* jpn = succs_block->pred(j); 1327 if (jpn->is_JumpProj() && jpn->in(0) == mach) { 1328 uint block_num = succs_block->non_connector()->_pre_order; 1329 Label *blkLabel = &blk_labels[block_num]; 1330 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1331 } 1332 } 1333 } 1334 } 1335 #ifdef ASSERT 1336 // Check that oop-store precedes the card-mark 1337 else if (mach->ideal_Opcode() == Op_StoreCM) { 1338 uint storeCM_idx = j; 1339 int count = 0; 1340 for (uint prec = mach->req(); prec < mach->len(); prec++) { 1341 Node *oop_store = mach->in(prec); // Precedence edge 1342 if (oop_store == NULL) continue; 1343 count++; 1344 uint i4; 1345 for (i4 = 0; i4 < last_inst; ++i4) { 1346 if (block->get_node(i4) == oop_store) { 1347 break; 1348 } 1349 } 1350 // Note: This test can provide a false failure if other precedence 1351 // edges have been added to the storeCMNode. 1352 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1353 } 1354 assert(count > 0, "storeCM expects at least one precedence edge"); 1355 } 1356 #endif 1357 else if (!n->is_Proj()) { 1358 // Remember the beginning of the previous instruction, in case 1359 // it's followed by a flag-kill and a null-check. Happens on 1360 // Intel all the time, with add-to-memory kind of opcodes. 1361 previous_offset = current_offset; 1362 } 1363 1364 // Not an else-if! 1365 // If this is a trap based cmp then add its offset to the list. 1366 if (mach->is_TrapBasedCheckNode()) { 1367 inct_starts[inct_cnt++] = current_offset; 1368 } 1369 } 1370 1371 // Verify that there is sufficient space remaining 1372 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1373 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1374 C->record_failure("CodeCache is full"); 1375 return; 1376 } 1377 1378 // Save the offset for the listing 1379 #ifndef PRODUCT 1380 if (node_offsets && n->_idx < node_offset_limit) 1381 node_offsets[n->_idx] = cb->insts_size(); 1382 #endif 1383 1384 // "Normal" instruction case 1385 DEBUG_ONLY( uint instr_offset = cb->insts_size(); ) 1386 n->emit(*cb, _regalloc); 1387 current_offset = cb->insts_size(); 1388 1389 // Above we only verified that there is enough space in the instruction section. 1390 // However, the instruction may emit stubs that cause code buffer expansion. 1391 // Bail out here if expansion failed due to a lack of code cache space. 1392 if (failing()) { 1393 return; 1394 } 1395 1396 #ifdef ASSERT 1397 if (n->size(_regalloc) < (current_offset-instr_offset)) { 1398 n->dump(); 1399 assert(false, "wrong size of mach node"); 1400 } 1401 #endif 1402 non_safepoints.observe_instruction(n, current_offset); 1403 1404 // mcall is last "call" that can be a safepoint 1405 // record it so we can see if a poll will directly follow it 1406 // in which case we'll need a pad to make the PcDesc sites unique 1407 // see 5010568. This can be slightly inaccurate but conservative 1408 // in the case that return address is not actually at current_offset. 1409 // This is a small price to pay. 1410 1411 if (is_mcall) { 1412 last_call_offset = current_offset; 1413 } 1414 1415 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 1416 // Avoid back to back some instructions. 1417 last_avoid_back_to_back_offset = current_offset; 1418 } 1419 1420 // See if this instruction has a delay slot 1421 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1422 assert(delay_slot != NULL, "expecting delay slot node"); 1423 1424 // Back up 1 instruction 1425 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size()); 1426 1427 // Save the offset for the listing 1428 #ifndef PRODUCT 1429 if (node_offsets && delay_slot->_idx < node_offset_limit) 1430 node_offsets[delay_slot->_idx] = cb->insts_size(); 1431 #endif 1432 1433 // Support a SafePoint in the delay slot 1434 if (delay_slot->is_MachSafePoint()) { 1435 MachNode *mach = delay_slot->as_Mach(); 1436 // !!!!! Stubs only need an oopmap right now, so bail out 1437 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) { 1438 // Write the oopmap directly to the code blob??!! 1439 delay_slot = NULL; 1440 continue; 1441 } 1442 1443 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1444 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1445 adjusted_offset); 1446 // Generate an OopMap entry 1447 Process_OopMap_Node(mach, adjusted_offset); 1448 } 1449 1450 // Insert the delay slot instruction 1451 delay_slot->emit(*cb, _regalloc); 1452 1453 // Don't reuse it 1454 delay_slot = NULL; 1455 } 1456 1457 } // End for all instructions in block 1458 1459 // If the next block is the top of a loop, pad this block out to align 1460 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1461 if (i < nblocks-1) { 1462 Block *nb = _cfg->get_block(i + 1); 1463 int padding = nb->alignment_padding(current_offset); 1464 if( padding > 0 ) { 1465 MachNode *nop = new MachNopNode(padding / nop_size); 1466 block->insert_node(nop, block->number_of_nodes()); 1467 _cfg->map_node_to_block(nop, block); 1468 nop->emit(*cb, _regalloc); 1469 current_offset = cb->insts_size(); 1470 } 1471 } 1472 // Verify that the distance for generated before forward 1473 // short branches is still valid. 1474 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size"); 1475 1476 // Save new block start offset 1477 blk_starts[i] = blk_offset; 1478 } // End of for all blocks 1479 blk_starts[nblocks] = current_offset; 1480 1481 non_safepoints.flush_at_end(); 1482 1483 // Offset too large? 1484 if (failing()) return; 1485 1486 // Define a pseudo-label at the end of the code 1487 MacroAssembler(cb).bind( blk_labels[nblocks] ); 1488 1489 // Compute the size of the first block 1490 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1491 1492 #ifdef ASSERT 1493 for (uint i = 0; i < nblocks; i++) { // For all blocks 1494 if (jmp_target[i] != 0) { 1495 int br_size = jmp_size[i]; 1496 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 1497 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 1498 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]); 1499 assert(false, "Displacement too large for short jmp"); 1500 } 1501 } 1502 } 1503 #endif 1504 1505 #ifndef PRODUCT 1506 // Information on the size of the method, without the extraneous code 1507 Scheduling::increment_method_size(cb->insts_size()); 1508 #endif 1509 1510 // ------------------ 1511 // Fill in exception table entries. 1512 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1513 1514 // Only java methods have exception handlers and deopt handlers 1515 // class HandlerImpl is platform-specific and defined in the *.ad files. 1516 if (_method) { 1517 // Emit the exception handler code. 1518 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb)); 1519 if (failing()) { 1520 return; // CodeBuffer::expand failed 1521 } 1522 // Emit the deopt handler code. 1523 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb)); 1524 1525 // Emit the MethodHandle deopt handler code (if required). 1526 if (has_method_handle_invokes() && !failing()) { 1527 // We can use the same code as for the normal deopt handler, we 1528 // just need a different entry point address. 1529 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb)); 1530 } 1531 } 1532 1533 // One last check for failed CodeBuffer::expand: 1534 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1535 C->record_failure("CodeCache is full"); 1536 return; 1537 } 1538 1539 #ifndef PRODUCT 1540 // Dump the assembly code, including basic-block numbers 1541 if (print_assembly()) { 1542 ttyLocker ttyl; // keep the following output all in one block 1543 if (!VMThread::should_terminate()) { // test this under the tty lock 1544 // This output goes directly to the tty, not the compiler log. 1545 // To enable tools to match it up with the compilation activity, 1546 // be sure to tag this tty output with the compile ID. 1547 if (xtty != NULL) { 1548 xtty->head("opto_assembly compile_id='%d'%s", compile_id(), 1549 is_osr_compilation() ? " compile_kind='osr'" : 1550 ""); 1551 } 1552 if (method() != NULL) { 1553 method()->print_metadata(); 1554 } 1555 dump_asm(node_offsets, node_offset_limit); 1556 if (xtty != NULL) { 1557 // print_metadata and dump_asm above may safepoint which makes us loose the ttylock. 1558 // Retake lock too make sure the end tag is coherent, and that xmlStream->pop_tag is done 1559 // thread safe 1560 ttyLocker ttyl2; 1561 xtty->tail("opto_assembly"); 1562 } 1563 } 1564 } 1565 #endif 1566 1567 } 1568 1569 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1570 _inc_table.set_size(cnt); 1571 1572 uint inct_cnt = 0; 1573 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 1574 Block* block = _cfg->get_block(i); 1575 Node *n = NULL; 1576 int j; 1577 1578 // Find the branch; ignore trailing NOPs. 1579 for (j = block->number_of_nodes() - 1; j >= 0; j--) { 1580 n = block->get_node(j); 1581 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) { 1582 break; 1583 } 1584 } 1585 1586 // If we didn't find anything, continue 1587 if (j < 0) { 1588 continue; 1589 } 1590 1591 // Compute ExceptionHandlerTable subtable entry and add it 1592 // (skip empty blocks) 1593 if (n->is_Catch()) { 1594 1595 // Get the offset of the return from the call 1596 uint call_return = call_returns[block->_pre_order]; 1597 #ifdef ASSERT 1598 assert( call_return > 0, "no call seen for this basic block" ); 1599 while (block->get_node(--j)->is_MachProj()) ; 1600 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call"); 1601 #endif 1602 // last instruction is a CatchNode, find it's CatchProjNodes 1603 int nof_succs = block->_num_succs; 1604 // allocate space 1605 GrowableArray<intptr_t> handler_bcis(nof_succs); 1606 GrowableArray<intptr_t> handler_pcos(nof_succs); 1607 // iterate through all successors 1608 for (int j = 0; j < nof_succs; j++) { 1609 Block* s = block->_succs[j]; 1610 bool found_p = false; 1611 for (uint k = 1; k < s->num_preds(); k++) { 1612 Node* pk = s->pred(k); 1613 if (pk->is_CatchProj() && pk->in(0) == n) { 1614 const CatchProjNode* p = pk->as_CatchProj(); 1615 found_p = true; 1616 // add the corresponding handler bci & pco information 1617 if (p->_con != CatchProjNode::fall_through_index) { 1618 // p leads to an exception handler (and is not fall through) 1619 assert(s == _cfg->get_block(s->_pre_order), "bad numbering"); 1620 // no duplicates, please 1621 if (!handler_bcis.contains(p->handler_bci())) { 1622 uint block_num = s->non_connector()->_pre_order; 1623 handler_bcis.append(p->handler_bci()); 1624 handler_pcos.append(blk_labels[block_num].loc_pos()); 1625 } 1626 } 1627 } 1628 } 1629 assert(found_p, "no matching predecessor found"); 1630 // Note: Due to empty block removal, one block may have 1631 // several CatchProj inputs, from the same Catch. 1632 } 1633 1634 // Set the offset of the return from the call 1635 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); 1636 continue; 1637 } 1638 1639 // Handle implicit null exception table updates 1640 if (n->is_MachNullCheck()) { 1641 uint block_num = block->non_connector_successor(0)->_pre_order; 1642 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1643 continue; 1644 } 1645 // Handle implicit exception table updates: trap instructions. 1646 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) { 1647 uint block_num = block->non_connector_successor(0)->_pre_order; 1648 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1649 continue; 1650 } 1651 } // End of for all blocks fill in exception table entries 1652 } 1653 1654 // Static Variables 1655 #ifndef PRODUCT 1656 uint Scheduling::_total_nop_size = 0; 1657 uint Scheduling::_total_method_size = 0; 1658 uint Scheduling::_total_branches = 0; 1659 uint Scheduling::_total_unconditional_delays = 0; 1660 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1661 #endif 1662 1663 // Initializer for class Scheduling 1664 1665 Scheduling::Scheduling(Arena *arena, Compile &compile) 1666 : _arena(arena), 1667 _cfg(compile.cfg()), 1668 _regalloc(compile.regalloc()), 1669 _reg_node(arena), 1670 _bundle_instr_count(0), 1671 _bundle_cycle_number(0), 1672 _scheduled(arena), 1673 _available(arena), 1674 _next_node(NULL), 1675 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]), 1676 _pinch_free_list(arena) 1677 #ifndef PRODUCT 1678 , _branches(0) 1679 , _unconditional_delays(0) 1680 #endif 1681 { 1682 // Create a MachNopNode 1683 _nop = new MachNopNode(); 1684 1685 // Now that the nops are in the array, save the count 1686 // (but allow entries for the nops) 1687 _node_bundling_limit = compile.unique(); 1688 uint node_max = _regalloc->node_regs_max_index(); 1689 1690 compile.set_node_bundling_limit(_node_bundling_limit); 1691 1692 // This one is persistent within the Compile class 1693 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1694 1695 // Allocate space for fixed-size arrays 1696 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1697 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1698 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1699 1700 // Clear the arrays 1701 memset(_node_bundling_base, 0, node_max * sizeof(Bundle)); 1702 memset(_node_latency, 0, node_max * sizeof(unsigned short)); 1703 memset(_uses, 0, node_max * sizeof(short)); 1704 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 1705 1706 // Clear the bundling information 1707 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements)); 1708 1709 // Get the last node 1710 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1); 1711 1712 _next_node = block->get_node(block->number_of_nodes() - 1); 1713 } 1714 1715 #ifndef PRODUCT 1716 // Scheduling destructor 1717 Scheduling::~Scheduling() { 1718 _total_branches += _branches; 1719 _total_unconditional_delays += _unconditional_delays; 1720 } 1721 #endif 1722 1723 // Step ahead "i" cycles 1724 void Scheduling::step(uint i) { 1725 1726 Bundle *bundle = node_bundling(_next_node); 1727 bundle->set_starts_bundle(); 1728 1729 // Update the bundle record, but leave the flags information alone 1730 if (_bundle_instr_count > 0) { 1731 bundle->set_instr_count(_bundle_instr_count); 1732 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1733 } 1734 1735 // Update the state information 1736 _bundle_instr_count = 0; 1737 _bundle_cycle_number += i; 1738 _bundle_use.step(i); 1739 } 1740 1741 void Scheduling::step_and_clear() { 1742 Bundle *bundle = node_bundling(_next_node); 1743 bundle->set_starts_bundle(); 1744 1745 // Update the bundle record 1746 if (_bundle_instr_count > 0) { 1747 bundle->set_instr_count(_bundle_instr_count); 1748 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1749 1750 _bundle_cycle_number += 1; 1751 } 1752 1753 // Clear the bundling information 1754 _bundle_instr_count = 0; 1755 _bundle_use.reset(); 1756 1757 memcpy(_bundle_use_elements, 1758 Pipeline_Use::elaborated_elements, 1759 sizeof(Pipeline_Use::elaborated_elements)); 1760 } 1761 1762 // Perform instruction scheduling and bundling over the sequence of 1763 // instructions in backwards order. 1764 void Compile::ScheduleAndBundle() { 1765 1766 // Don't optimize this if it isn't a method 1767 if (!_method) 1768 return; 1769 1770 // Don't optimize this if scheduling is disabled 1771 if (!do_scheduling()) 1772 return; 1773 1774 // Scheduling code works only with pairs (16 bytes) maximum. 1775 if (max_vector_size() > 16) 1776 return; 1777 1778 TracePhase tp("isched", &timers[_t_instrSched]); 1779 1780 // Create a data structure for all the scheduling information 1781 Scheduling scheduling(Thread::current()->resource_area(), *this); 1782 1783 // Walk backwards over each basic block, computing the needed alignment 1784 // Walk over all the basic blocks 1785 scheduling.DoScheduling(); 1786 } 1787 1788 // Compute the latency of all the instructions. This is fairly simple, 1789 // because we already have a legal ordering. Walk over the instructions 1790 // from first to last, and compute the latency of the instruction based 1791 // on the latency of the preceding instruction(s). 1792 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { 1793 #ifndef PRODUCT 1794 if (_cfg->C->trace_opto_output()) 1795 tty->print("# -> ComputeLocalLatenciesForward\n"); 1796 #endif 1797 1798 // Walk over all the schedulable instructions 1799 for( uint j=_bb_start; j < _bb_end; j++ ) { 1800 1801 // This is a kludge, forcing all latency calculations to start at 1. 1802 // Used to allow latency 0 to force an instruction to the beginning 1803 // of the bb 1804 uint latency = 1; 1805 Node *use = bb->get_node(j); 1806 uint nlen = use->len(); 1807 1808 // Walk over all the inputs 1809 for ( uint k=0; k < nlen; k++ ) { 1810 Node *def = use->in(k); 1811 if (!def) 1812 continue; 1813 1814 uint l = _node_latency[def->_idx] + use->latency(k); 1815 if (latency < l) 1816 latency = l; 1817 } 1818 1819 _node_latency[use->_idx] = latency; 1820 1821 #ifndef PRODUCT 1822 if (_cfg->C->trace_opto_output()) { 1823 tty->print("# latency %4d: ", latency); 1824 use->dump(); 1825 } 1826 #endif 1827 } 1828 1829 #ifndef PRODUCT 1830 if (_cfg->C->trace_opto_output()) 1831 tty->print("# <- ComputeLocalLatenciesForward\n"); 1832 #endif 1833 1834 } // end ComputeLocalLatenciesForward 1835 1836 // See if this node fits into the present instruction bundle 1837 bool Scheduling::NodeFitsInBundle(Node *n) { 1838 uint n_idx = n->_idx; 1839 1840 // If this is the unconditional delay instruction, then it fits 1841 if (n == _unconditional_delay_slot) { 1842 #ifndef PRODUCT 1843 if (_cfg->C->trace_opto_output()) 1844 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 1845 #endif 1846 return (true); 1847 } 1848 1849 // If the node cannot be scheduled this cycle, skip it 1850 if (_current_latency[n_idx] > _bundle_cycle_number) { 1851 #ifndef PRODUCT 1852 if (_cfg->C->trace_opto_output()) 1853 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 1854 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 1855 #endif 1856 return (false); 1857 } 1858 1859 const Pipeline *node_pipeline = n->pipeline(); 1860 1861 uint instruction_count = node_pipeline->instructionCount(); 1862 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1863 instruction_count = 0; 1864 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1865 instruction_count++; 1866 1867 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 1868 #ifndef PRODUCT 1869 if (_cfg->C->trace_opto_output()) 1870 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 1871 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 1872 #endif 1873 return (false); 1874 } 1875 1876 // Don't allow non-machine nodes to be handled this way 1877 if (!n->is_Mach() && instruction_count == 0) 1878 return (false); 1879 1880 // See if there is any overlap 1881 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 1882 1883 if (delay > 0) { 1884 #ifndef PRODUCT 1885 if (_cfg->C->trace_opto_output()) 1886 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 1887 #endif 1888 return false; 1889 } 1890 1891 #ifndef PRODUCT 1892 if (_cfg->C->trace_opto_output()) 1893 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 1894 #endif 1895 1896 return true; 1897 } 1898 1899 Node * Scheduling::ChooseNodeToBundle() { 1900 uint siz = _available.size(); 1901 1902 if (siz == 0) { 1903 1904 #ifndef PRODUCT 1905 if (_cfg->C->trace_opto_output()) 1906 tty->print("# ChooseNodeToBundle: NULL\n"); 1907 #endif 1908 return (NULL); 1909 } 1910 1911 // Fast path, if only 1 instruction in the bundle 1912 if (siz == 1) { 1913 #ifndef PRODUCT 1914 if (_cfg->C->trace_opto_output()) { 1915 tty->print("# ChooseNodeToBundle (only 1): "); 1916 _available[0]->dump(); 1917 } 1918 #endif 1919 return (_available[0]); 1920 } 1921 1922 // Don't bother, if the bundle is already full 1923 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 1924 for ( uint i = 0; i < siz; i++ ) { 1925 Node *n = _available[i]; 1926 1927 // Skip projections, we'll handle them another way 1928 if (n->is_Proj()) 1929 continue; 1930 1931 // This presupposed that instructions are inserted into the 1932 // available list in a legality order; i.e. instructions that 1933 // must be inserted first are at the head of the list 1934 if (NodeFitsInBundle(n)) { 1935 #ifndef PRODUCT 1936 if (_cfg->C->trace_opto_output()) { 1937 tty->print("# ChooseNodeToBundle: "); 1938 n->dump(); 1939 } 1940 #endif 1941 return (n); 1942 } 1943 } 1944 } 1945 1946 // Nothing fits in this bundle, choose the highest priority 1947 #ifndef PRODUCT 1948 if (_cfg->C->trace_opto_output()) { 1949 tty->print("# ChooseNodeToBundle: "); 1950 _available[0]->dump(); 1951 } 1952 #endif 1953 1954 return _available[0]; 1955 } 1956 1957 void Scheduling::AddNodeToAvailableList(Node *n) { 1958 assert( !n->is_Proj(), "projections never directly made available" ); 1959 #ifndef PRODUCT 1960 if (_cfg->C->trace_opto_output()) { 1961 tty->print("# AddNodeToAvailableList: "); 1962 n->dump(); 1963 } 1964 #endif 1965 1966 int latency = _current_latency[n->_idx]; 1967 1968 // Insert in latency order (insertion sort) 1969 uint i; 1970 for ( i=0; i < _available.size(); i++ ) 1971 if (_current_latency[_available[i]->_idx] > latency) 1972 break; 1973 1974 // Special Check for compares following branches 1975 if( n->is_Mach() && _scheduled.size() > 0 ) { 1976 int op = n->as_Mach()->ideal_Opcode(); 1977 Node *last = _scheduled[0]; 1978 if( last->is_MachIf() && last->in(1) == n && 1979 ( op == Op_CmpI || 1980 op == Op_CmpU || 1981 op == Op_CmpP || 1982 op == Op_CmpF || 1983 op == Op_CmpD || 1984 op == Op_CmpL ) ) { 1985 1986 // Recalculate position, moving to front of same latency 1987 for ( i=0 ; i < _available.size(); i++ ) 1988 if (_current_latency[_available[i]->_idx] >= latency) 1989 break; 1990 } 1991 } 1992 1993 // Insert the node in the available list 1994 _available.insert(i, n); 1995 1996 #ifndef PRODUCT 1997 if (_cfg->C->trace_opto_output()) 1998 dump_available(); 1999 #endif 2000 } 2001 2002 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 2003 for ( uint i=0; i < n->len(); i++ ) { 2004 Node *def = n->in(i); 2005 if (!def) continue; 2006 if( def->is_Proj() ) // If this is a machine projection, then 2007 def = def->in(0); // propagate usage thru to the base instruction 2008 2009 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local 2010 continue; 2011 } 2012 2013 // Compute the latency 2014 uint l = _bundle_cycle_number + n->latency(i); 2015 if (_current_latency[def->_idx] < l) 2016 _current_latency[def->_idx] = l; 2017 2018 // If this does not have uses then schedule it 2019 if ((--_uses[def->_idx]) == 0) 2020 AddNodeToAvailableList(def); 2021 } 2022 } 2023 2024 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 2025 #ifndef PRODUCT 2026 if (_cfg->C->trace_opto_output()) { 2027 tty->print("# AddNodeToBundle: "); 2028 n->dump(); 2029 } 2030 #endif 2031 2032 // Remove this from the available list 2033 uint i; 2034 for (i = 0; i < _available.size(); i++) 2035 if (_available[i] == n) 2036 break; 2037 assert(i < _available.size(), "entry in _available list not found"); 2038 _available.remove(i); 2039 2040 // See if this fits in the current bundle 2041 const Pipeline *node_pipeline = n->pipeline(); 2042 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 2043 2044 // Check for instructions to be placed in the delay slot. We 2045 // do this before we actually schedule the current instruction, 2046 // because the delay slot follows the current instruction. 2047 if (Pipeline::_branch_has_delay_slot && 2048 node_pipeline->hasBranchDelay() && 2049 !_unconditional_delay_slot) { 2050 2051 uint siz = _available.size(); 2052 2053 // Conditional branches can support an instruction that 2054 // is unconditionally executed and not dependent by the 2055 // branch, OR a conditionally executed instruction if 2056 // the branch is taken. In practice, this means that 2057 // the first instruction at the branch target is 2058 // copied to the delay slot, and the branch goes to 2059 // the instruction after that at the branch target 2060 if ( n->is_MachBranch() ) { 2061 2062 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 2063 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 2064 2065 #ifndef PRODUCT 2066 _branches++; 2067 #endif 2068 2069 // At least 1 instruction is on the available list 2070 // that is not dependent on the branch 2071 for (uint i = 0; i < siz; i++) { 2072 Node *d = _available[i]; 2073 const Pipeline *avail_pipeline = d->pipeline(); 2074 2075 // Don't allow safepoints in the branch shadow, that will 2076 // cause a number of difficulties 2077 if ( avail_pipeline->instructionCount() == 1 && 2078 !avail_pipeline->hasMultipleBundles() && 2079 !avail_pipeline->hasBranchDelay() && 2080 Pipeline::instr_has_unit_size() && 2081 d->size(_regalloc) == Pipeline::instr_unit_size() && 2082 NodeFitsInBundle(d) && 2083 !node_bundling(d)->used_in_delay()) { 2084 2085 if (d->is_Mach() && !d->is_MachSafePoint()) { 2086 // A node that fits in the delay slot was found, so we need to 2087 // set the appropriate bits in the bundle pipeline information so 2088 // that it correctly indicates resource usage. Later, when we 2089 // attempt to add this instruction to the bundle, we will skip 2090 // setting the resource usage. 2091 _unconditional_delay_slot = d; 2092 node_bundling(n)->set_use_unconditional_delay(); 2093 node_bundling(d)->set_used_in_unconditional_delay(); 2094 _bundle_use.add_usage(avail_pipeline->resourceUse()); 2095 _current_latency[d->_idx] = _bundle_cycle_number; 2096 _next_node = d; 2097 ++_bundle_instr_count; 2098 #ifndef PRODUCT 2099 _unconditional_delays++; 2100 #endif 2101 break; 2102 } 2103 } 2104 } 2105 } 2106 2107 // No delay slot, add a nop to the usage 2108 if (!_unconditional_delay_slot) { 2109 // See if adding an instruction in the delay slot will overflow 2110 // the bundle. 2111 if (!NodeFitsInBundle(_nop)) { 2112 #ifndef PRODUCT 2113 if (_cfg->C->trace_opto_output()) 2114 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2115 #endif 2116 step(1); 2117 } 2118 2119 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2120 _next_node = _nop; 2121 ++_bundle_instr_count; 2122 } 2123 2124 // See if the instruction in the delay slot requires a 2125 // step of the bundles 2126 if (!NodeFitsInBundle(n)) { 2127 #ifndef PRODUCT 2128 if (_cfg->C->trace_opto_output()) 2129 tty->print("# *** STEP(branch won't fit) ***\n"); 2130 #endif 2131 // Update the state information 2132 _bundle_instr_count = 0; 2133 _bundle_cycle_number += 1; 2134 _bundle_use.step(1); 2135 } 2136 } 2137 2138 // Get the number of instructions 2139 uint instruction_count = node_pipeline->instructionCount(); 2140 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2141 instruction_count = 0; 2142 2143 // Compute the latency information 2144 uint delay = 0; 2145 2146 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2147 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2148 if (relative_latency < 0) 2149 relative_latency = 0; 2150 2151 delay = _bundle_use.full_latency(relative_latency, node_usage); 2152 2153 // Does not fit in this bundle, start a new one 2154 if (delay > 0) { 2155 step(delay); 2156 2157 #ifndef PRODUCT 2158 if (_cfg->C->trace_opto_output()) 2159 tty->print("# *** STEP(%d) ***\n", delay); 2160 #endif 2161 } 2162 } 2163 2164 // If this was placed in the delay slot, ignore it 2165 if (n != _unconditional_delay_slot) { 2166 2167 if (delay == 0) { 2168 if (node_pipeline->hasMultipleBundles()) { 2169 #ifndef PRODUCT 2170 if (_cfg->C->trace_opto_output()) 2171 tty->print("# *** STEP(multiple instructions) ***\n"); 2172 #endif 2173 step(1); 2174 } 2175 2176 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2177 #ifndef PRODUCT 2178 if (_cfg->C->trace_opto_output()) 2179 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2180 instruction_count + _bundle_instr_count, 2181 Pipeline::_max_instrs_per_cycle); 2182 #endif 2183 step(1); 2184 } 2185 } 2186 2187 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2188 _bundle_instr_count++; 2189 2190 // Set the node's latency 2191 _current_latency[n->_idx] = _bundle_cycle_number; 2192 2193 // Now merge the functional unit information 2194 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2195 _bundle_use.add_usage(node_usage); 2196 2197 // Increment the number of instructions in this bundle 2198 _bundle_instr_count += instruction_count; 2199 2200 // Remember this node for later 2201 if (n->is_Mach()) 2202 _next_node = n; 2203 } 2204 2205 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2206 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2207 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2208 // into the block. All other scheduled nodes get put in the schedule here. 2209 int op = n->Opcode(); 2210 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2211 (op != Op_Node && // Not an unused antidepedence node and 2212 // not an unallocated boxlock 2213 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2214 2215 // Push any trailing projections 2216 if( bb->get_node(bb->number_of_nodes()-1) != n ) { 2217 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2218 Node *foi = n->fast_out(i); 2219 if( foi->is_Proj() ) 2220 _scheduled.push(foi); 2221 } 2222 } 2223 2224 // Put the instruction in the schedule list 2225 _scheduled.push(n); 2226 } 2227 2228 #ifndef PRODUCT 2229 if (_cfg->C->trace_opto_output()) 2230 dump_available(); 2231 #endif 2232 2233 // Walk all the definitions, decrementing use counts, and 2234 // if a definition has a 0 use count, place it in the available list. 2235 DecrementUseCounts(n,bb); 2236 } 2237 2238 // This method sets the use count within a basic block. We will ignore all 2239 // uses outside the current basic block. As we are doing a backwards walk, 2240 // any node we reach that has a use count of 0 may be scheduled. This also 2241 // avoids the problem of cyclic references from phi nodes, as long as phi 2242 // nodes are at the front of the basic block. This method also initializes 2243 // the available list to the set of instructions that have no uses within this 2244 // basic block. 2245 void Scheduling::ComputeUseCount(const Block *bb) { 2246 #ifndef PRODUCT 2247 if (_cfg->C->trace_opto_output()) 2248 tty->print("# -> ComputeUseCount\n"); 2249 #endif 2250 2251 // Clear the list of available and scheduled instructions, just in case 2252 _available.clear(); 2253 _scheduled.clear(); 2254 2255 // No delay slot specified 2256 _unconditional_delay_slot = NULL; 2257 2258 #ifdef ASSERT 2259 for( uint i=0; i < bb->number_of_nodes(); i++ ) 2260 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" ); 2261 #endif 2262 2263 // Force the _uses count to never go to zero for unscheduable pieces 2264 // of the block 2265 for( uint k = 0; k < _bb_start; k++ ) 2266 _uses[bb->get_node(k)->_idx] = 1; 2267 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ ) 2268 _uses[bb->get_node(l)->_idx] = 1; 2269 2270 // Iterate backwards over the instructions in the block. Don't count the 2271 // branch projections at end or the block header instructions. 2272 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2273 Node *n = bb->get_node(j); 2274 if( n->is_Proj() ) continue; // Projections handled another way 2275 2276 // Account for all uses 2277 for ( uint k = 0; k < n->len(); k++ ) { 2278 Node *inp = n->in(k); 2279 if (!inp) continue; 2280 assert(inp != n, "no cycles allowed" ); 2281 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use? 2282 if (inp->is_Proj()) { // Skip through Proj's 2283 inp = inp->in(0); 2284 } 2285 ++_uses[inp->_idx]; // Count 1 block-local use 2286 } 2287 } 2288 2289 // If this instruction has a 0 use count, then it is available 2290 if (!_uses[n->_idx]) { 2291 _current_latency[n->_idx] = _bundle_cycle_number; 2292 AddNodeToAvailableList(n); 2293 } 2294 2295 #ifndef PRODUCT 2296 if (_cfg->C->trace_opto_output()) { 2297 tty->print("# uses: %3d: ", _uses[n->_idx]); 2298 n->dump(); 2299 } 2300 #endif 2301 } 2302 2303 #ifndef PRODUCT 2304 if (_cfg->C->trace_opto_output()) 2305 tty->print("# <- ComputeUseCount\n"); 2306 #endif 2307 } 2308 2309 // This routine performs scheduling on each basic block in reverse order, 2310 // using instruction latencies and taking into account function unit 2311 // availability. 2312 void Scheduling::DoScheduling() { 2313 #ifndef PRODUCT 2314 if (_cfg->C->trace_opto_output()) 2315 tty->print("# -> DoScheduling\n"); 2316 #endif 2317 2318 Block *succ_bb = NULL; 2319 Block *bb; 2320 2321 // Walk over all the basic blocks in reverse order 2322 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) { 2323 bb = _cfg->get_block(i); 2324 2325 #ifndef PRODUCT 2326 if (_cfg->C->trace_opto_output()) { 2327 tty->print("# Schedule BB#%03d (initial)\n", i); 2328 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2329 bb->get_node(j)->dump(); 2330 } 2331 } 2332 #endif 2333 2334 // On the head node, skip processing 2335 if (bb == _cfg->get_root_block()) { 2336 continue; 2337 } 2338 2339 // Skip empty, connector blocks 2340 if (bb->is_connector()) 2341 continue; 2342 2343 // If the following block is not the sole successor of 2344 // this one, then reset the pipeline information 2345 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2346 #ifndef PRODUCT 2347 if (_cfg->C->trace_opto_output()) { 2348 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2349 _next_node->_idx, _bundle_instr_count); 2350 } 2351 #endif 2352 step_and_clear(); 2353 } 2354 2355 // Leave untouched the starting instruction, any Phis, a CreateEx node 2356 // or Top. bb->get_node(_bb_start) is the first schedulable instruction. 2357 _bb_end = bb->number_of_nodes()-1; 2358 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2359 Node *n = bb->get_node(_bb_start); 2360 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2361 // Also, MachIdealNodes do not get scheduled 2362 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2363 MachNode *mach = n->as_Mach(); 2364 int iop = mach->ideal_Opcode(); 2365 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2366 if( iop == Op_Con ) continue; // Do not schedule Top 2367 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2368 mach->pipeline() == MachNode::pipeline_class() && 2369 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc 2370 continue; 2371 break; // Funny loop structure to be sure... 2372 } 2373 // Compute last "interesting" instruction in block - last instruction we 2374 // might schedule. _bb_end points just after last schedulable inst. We 2375 // normally schedule conditional branches (despite them being forced last 2376 // in the block), because they have delay slots we can fill. Calls all 2377 // have their delay slots filled in the template expansions, so we don't 2378 // bother scheduling them. 2379 Node *last = bb->get_node(_bb_end); 2380 // Ignore trailing NOPs. 2381 while (_bb_end > 0 && last->is_Mach() && 2382 last->as_Mach()->ideal_Opcode() == Op_Con) { 2383 last = bb->get_node(--_bb_end); 2384 } 2385 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, ""); 2386 if( last->is_Catch() || 2387 // Exclude unreachable path case when Halt node is in a separate block. 2388 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2389 // There must be a prior call. Skip it. 2390 while( !bb->get_node(--_bb_end)->is_MachCall() ) { 2391 assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" ); 2392 } 2393 } else if( last->is_MachNullCheck() ) { 2394 // Backup so the last null-checked memory instruction is 2395 // outside the schedulable range. Skip over the nullcheck, 2396 // projection, and the memory nodes. 2397 Node *mem = last->in(1); 2398 do { 2399 _bb_end--; 2400 } while (mem != bb->get_node(_bb_end)); 2401 } else { 2402 // Set _bb_end to point after last schedulable inst. 2403 _bb_end++; 2404 } 2405 2406 assert( _bb_start <= _bb_end, "inverted block ends" ); 2407 2408 // Compute the register antidependencies for the basic block 2409 ComputeRegisterAntidependencies(bb); 2410 if (_cfg->C->failing()) return; // too many D-U pinch points 2411 2412 // Compute intra-bb latencies for the nodes 2413 ComputeLocalLatenciesForward(bb); 2414 2415 // Compute the usage within the block, and set the list of all nodes 2416 // in the block that have no uses within the block. 2417 ComputeUseCount(bb); 2418 2419 // Schedule the remaining instructions in the block 2420 while ( _available.size() > 0 ) { 2421 Node *n = ChooseNodeToBundle(); 2422 guarantee(n != NULL, "no nodes available"); 2423 AddNodeToBundle(n,bb); 2424 } 2425 2426 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2427 #ifdef ASSERT 2428 for( uint l = _bb_start; l < _bb_end; l++ ) { 2429 Node *n = bb->get_node(l); 2430 uint m; 2431 for( m = 0; m < _bb_end-_bb_start; m++ ) 2432 if( _scheduled[m] == n ) 2433 break; 2434 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2435 } 2436 #endif 2437 2438 // Now copy the instructions (in reverse order) back to the block 2439 for ( uint k = _bb_start; k < _bb_end; k++ ) 2440 bb->map_node(_scheduled[_bb_end-k-1], k); 2441 2442 #ifndef PRODUCT 2443 if (_cfg->C->trace_opto_output()) { 2444 tty->print("# Schedule BB#%03d (final)\n", i); 2445 uint current = 0; 2446 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2447 Node *n = bb->get_node(j); 2448 if( valid_bundle_info(n) ) { 2449 Bundle *bundle = node_bundling(n); 2450 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2451 tty->print("*** Bundle: "); 2452 bundle->dump(); 2453 } 2454 n->dump(); 2455 } 2456 } 2457 } 2458 #endif 2459 #ifdef ASSERT 2460 verify_good_schedule(bb,"after block local scheduling"); 2461 #endif 2462 } 2463 2464 #ifndef PRODUCT 2465 if (_cfg->C->trace_opto_output()) 2466 tty->print("# <- DoScheduling\n"); 2467 #endif 2468 2469 // Record final node-bundling array location 2470 _regalloc->C->set_node_bundling_base(_node_bundling_base); 2471 2472 } // end DoScheduling 2473 2474 // Verify that no live-range used in the block is killed in the block by a 2475 // wrong DEF. This doesn't verify live-ranges that span blocks. 2476 2477 // Check for edge existence. Used to avoid adding redundant precedence edges. 2478 static bool edge_from_to( Node *from, Node *to ) { 2479 for( uint i=0; i<from->len(); i++ ) 2480 if( from->in(i) == to ) 2481 return true; 2482 return false; 2483 } 2484 2485 #ifdef ASSERT 2486 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2487 // Check for bad kills 2488 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2489 Node *prior_use = _reg_node[def]; 2490 if( prior_use && !edge_from_to(prior_use,n) ) { 2491 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2492 n->dump(); 2493 tty->print_cr("..."); 2494 prior_use->dump(); 2495 assert(edge_from_to(prior_use,n), "%s", msg); 2496 } 2497 _reg_node.map(def,NULL); // Kill live USEs 2498 } 2499 } 2500 2501 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2502 2503 // Zap to something reasonable for the verify code 2504 _reg_node.clear(); 2505 2506 // Walk over the block backwards. Check to make sure each DEF doesn't 2507 // kill a live value (other than the one it's supposed to). Add each 2508 // USE to the live set. 2509 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) { 2510 Node *n = b->get_node(i); 2511 int n_op = n->Opcode(); 2512 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2513 // Fat-proj kills a slew of registers 2514 RegMask rm = n->out_RegMask();// Make local copy 2515 while( rm.is_NotEmpty() ) { 2516 OptoReg::Name kill = rm.find_first_elem(); 2517 rm.Remove(kill); 2518 verify_do_def( n, kill, msg ); 2519 } 2520 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2521 // Get DEF'd registers the normal way 2522 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2523 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2524 } 2525 2526 // Now make all USEs live 2527 for( uint i=1; i<n->req(); i++ ) { 2528 Node *def = n->in(i); 2529 assert(def != 0, "input edge required"); 2530 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2531 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2532 if( OptoReg::is_valid(reg_lo) ) { 2533 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg); 2534 _reg_node.map(reg_lo,n); 2535 } 2536 if( OptoReg::is_valid(reg_hi) ) { 2537 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg); 2538 _reg_node.map(reg_hi,n); 2539 } 2540 } 2541 2542 } 2543 2544 // Zap to something reasonable for the Antidependence code 2545 _reg_node.clear(); 2546 } 2547 #endif 2548 2549 // Conditionally add precedence edges. Avoid putting edges on Projs. 2550 static void add_prec_edge_from_to( Node *from, Node *to ) { 2551 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2552 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2553 from = from->in(0); 2554 } 2555 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2556 !edge_from_to( from, to ) ) // Avoid duplicate edge 2557 from->add_prec(to); 2558 } 2559 2560 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2561 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2562 return; 2563 2564 Node *pinch = _reg_node[def_reg]; // Get pinch point 2565 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet? 2566 is_def ) { // Check for a true def (not a kill) 2567 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2568 return; 2569 } 2570 2571 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2572 debug_only( def = (Node*)0xdeadbeef; ) 2573 2574 // After some number of kills there _may_ be a later def 2575 Node *later_def = NULL; 2576 2577 // Finding a kill requires a real pinch-point. 2578 // Check for not already having a pinch-point. 2579 // Pinch points are Op_Node's. 2580 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2581 later_def = pinch; // Must be def/kill as optimistic pinch-point 2582 if ( _pinch_free_list.size() > 0) { 2583 pinch = _pinch_free_list.pop(); 2584 } else { 2585 pinch = new Node(1); // Pinch point to-be 2586 } 2587 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2588 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2589 return; 2590 } 2591 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init) 2592 _reg_node.map(def_reg,pinch); // Record pinch-point 2593 //_regalloc->set_bad(pinch->_idx); // Already initialized this way. 2594 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2595 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call 2596 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2597 later_def = NULL; // and no later def 2598 } 2599 pinch->set_req(0,later_def); // Hook later def so we can find it 2600 } else { // Else have valid pinch point 2601 if( pinch->in(0) ) // If there is a later-def 2602 later_def = pinch->in(0); // Get it 2603 } 2604 2605 // Add output-dependence edge from later def to kill 2606 if( later_def ) // If there is some original def 2607 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2608 2609 // See if current kill is also a use, and so is forced to be the pinch-point. 2610 if( pinch->Opcode() == Op_Node ) { 2611 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2612 for( uint i=1; i<uses->req(); i++ ) { 2613 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2614 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2615 // Yes, found a use/kill pinch-point 2616 pinch->set_req(0,NULL); // 2617 pinch->replace_by(kill); // Move anti-dep edges up 2618 pinch = kill; 2619 _reg_node.map(def_reg,pinch); 2620 return; 2621 } 2622 } 2623 } 2624 2625 // Add edge from kill to pinch-point 2626 add_prec_edge_from_to(kill,pinch); 2627 } 2628 2629 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2630 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2631 return; 2632 Node *pinch = _reg_node[use_reg]; // Get pinch point 2633 // Check for no later def_reg/kill in block 2634 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b && 2635 // Use has to be block-local as well 2636 _cfg->get_block_for_node(use) == b) { 2637 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2638 pinch->req() == 1 ) { // pinch not yet in block? 2639 pinch->del_req(0); // yank pointer to later-def, also set flag 2640 // Insert the pinch-point in the block just after the last use 2641 b->insert_node(pinch, b->find_node(use) + 1); 2642 _bb_end++; // Increase size scheduled region in block 2643 } 2644 2645 add_prec_edge_from_to(pinch,use); 2646 } 2647 } 2648 2649 // We insert antidependences between the reads and following write of 2650 // allocated registers to prevent illegal code motion. Hopefully, the 2651 // number of added references should be fairly small, especially as we 2652 // are only adding references within the current basic block. 2653 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2654 2655 #ifdef ASSERT 2656 verify_good_schedule(b,"before block local scheduling"); 2657 #endif 2658 2659 // A valid schedule, for each register independently, is an endless cycle 2660 // of: a def, then some uses (connected to the def by true dependencies), 2661 // then some kills (defs with no uses), finally the cycle repeats with a new 2662 // def. The uses are allowed to float relative to each other, as are the 2663 // kills. No use is allowed to slide past a kill (or def). This requires 2664 // antidependencies between all uses of a single def and all kills that 2665 // follow, up to the next def. More edges are redundant, because later defs 2666 // & kills are already serialized with true or antidependencies. To keep 2667 // the edge count down, we add a 'pinch point' node if there's more than 2668 // one use or more than one kill/def. 2669 2670 // We add dependencies in one bottom-up pass. 2671 2672 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2673 2674 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2675 // register. If not, we record the DEF/KILL in _reg_node, the 2676 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2677 // "pinch point", a new Node that's in the graph but not in the block. 2678 // We put edges from the prior and current DEF/KILLs to the pinch point. 2679 // We put the pinch point in _reg_node. If there's already a pinch point 2680 // we merely add an edge from the current DEF/KILL to the pinch point. 2681 2682 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2683 // put an edge from the pinch point to the USE. 2684 2685 // To be expedient, the _reg_node array is pre-allocated for the whole 2686 // compilation. _reg_node is lazily initialized; it either contains a NULL, 2687 // or a valid def/kill/pinch-point, or a leftover node from some prior 2688 // block. Leftover node from some prior block is treated like a NULL (no 2689 // prior def, so no anti-dependence needed). Valid def is distinguished by 2690 // it being in the current block. 2691 bool fat_proj_seen = false; 2692 uint last_safept = _bb_end-1; 2693 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL; 2694 Node* last_safept_node = end_node; 2695 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2696 Node *n = b->get_node(i); 2697 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2698 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) { 2699 // Fat-proj kills a slew of registers 2700 // This can add edges to 'n' and obscure whether or not it was a def, 2701 // hence the is_def flag. 2702 fat_proj_seen = true; 2703 RegMask rm = n->out_RegMask();// Make local copy 2704 while( rm.is_NotEmpty() ) { 2705 OptoReg::Name kill = rm.find_first_elem(); 2706 rm.Remove(kill); 2707 anti_do_def( b, n, kill, is_def ); 2708 } 2709 } else { 2710 // Get DEF'd registers the normal way 2711 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2712 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2713 } 2714 2715 // Kill projections on a branch should appear to occur on the 2716 // branch, not afterwards, so grab the masks from the projections 2717 // and process them. 2718 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) { 2719 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2720 Node* use = n->fast_out(i); 2721 if (use->is_Proj()) { 2722 RegMask rm = use->out_RegMask();// Make local copy 2723 while( rm.is_NotEmpty() ) { 2724 OptoReg::Name kill = rm.find_first_elem(); 2725 rm.Remove(kill); 2726 anti_do_def( b, n, kill, false ); 2727 } 2728 } 2729 } 2730 } 2731 2732 // Check each register used by this instruction for a following DEF/KILL 2733 // that must occur afterward and requires an anti-dependence edge. 2734 for( uint j=0; j<n->req(); j++ ) { 2735 Node *def = n->in(j); 2736 if( def ) { 2737 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2738 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2739 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2740 } 2741 } 2742 // Do not allow defs of new derived values to float above GC 2743 // points unless the base is definitely available at the GC point. 2744 2745 Node *m = b->get_node(i); 2746 2747 // Add precedence edge from following safepoint to use of derived pointer 2748 if( last_safept_node != end_node && 2749 m != last_safept_node) { 2750 for (uint k = 1; k < m->req(); k++) { 2751 const Type *t = m->in(k)->bottom_type(); 2752 if( t->isa_oop_ptr() && 2753 t->is_ptr()->offset() != 0 ) { 2754 last_safept_node->add_prec( m ); 2755 break; 2756 } 2757 } 2758 } 2759 2760 if( n->jvms() ) { // Precedence edge from derived to safept 2761 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2762 if( b->get_node(last_safept) != last_safept_node ) { 2763 last_safept = b->find_node(last_safept_node); 2764 } 2765 for( uint j=last_safept; j > i; j-- ) { 2766 Node *mach = b->get_node(j); 2767 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2768 mach->add_prec( n ); 2769 } 2770 last_safept = i; 2771 last_safept_node = m; 2772 } 2773 } 2774 2775 if (fat_proj_seen) { 2776 // Garbage collect pinch nodes that were not consumed. 2777 // They are usually created by a fat kill MachProj for a call. 2778 garbage_collect_pinch_nodes(); 2779 } 2780 } 2781 2782 // Garbage collect pinch nodes for reuse by other blocks. 2783 // 2784 // The block scheduler's insertion of anti-dependence 2785 // edges creates many pinch nodes when the block contains 2786 // 2 or more Calls. A pinch node is used to prevent a 2787 // combinatorial explosion of edges. If a set of kills for a 2788 // register is anti-dependent on a set of uses (or defs), rather 2789 // than adding an edge in the graph between each pair of kill 2790 // and use (or def), a pinch is inserted between them: 2791 // 2792 // use1 use2 use3 2793 // \ | / 2794 // \ | / 2795 // pinch 2796 // / | \ 2797 // / | \ 2798 // kill1 kill2 kill3 2799 // 2800 // One pinch node is created per register killed when 2801 // the second call is encountered during a backwards pass 2802 // over the block. Most of these pinch nodes are never 2803 // wired into the graph because the register is never 2804 // used or def'ed in the block. 2805 // 2806 void Scheduling::garbage_collect_pinch_nodes() { 2807 #ifndef PRODUCT 2808 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 2809 #endif 2810 int trace_cnt = 0; 2811 for (uint k = 0; k < _reg_node.Size(); k++) { 2812 Node* pinch = _reg_node[k]; 2813 if ((pinch != NULL) && pinch->Opcode() == Op_Node && 2814 // no predecence input edges 2815 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { 2816 cleanup_pinch(pinch); 2817 _pinch_free_list.push(pinch); 2818 _reg_node.map(k, NULL); 2819 #ifndef PRODUCT 2820 if (_cfg->C->trace_opto_output()) { 2821 trace_cnt++; 2822 if (trace_cnt > 40) { 2823 tty->print("\n"); 2824 trace_cnt = 0; 2825 } 2826 tty->print(" %d", pinch->_idx); 2827 } 2828 #endif 2829 } 2830 } 2831 #ifndef PRODUCT 2832 if (_cfg->C->trace_opto_output()) tty->print("\n"); 2833 #endif 2834 } 2835 2836 // Clean up a pinch node for reuse. 2837 void Scheduling::cleanup_pinch( Node *pinch ) { 2838 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 2839 2840 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 2841 Node* use = pinch->last_out(i); 2842 uint uses_found = 0; 2843 for (uint j = use->req(); j < use->len(); j++) { 2844 if (use->in(j) == pinch) { 2845 use->rm_prec(j); 2846 uses_found++; 2847 } 2848 } 2849 assert(uses_found > 0, "must be a precedence edge"); 2850 i -= uses_found; // we deleted 1 or more copies of this edge 2851 } 2852 // May have a later_def entry 2853 pinch->set_req(0, NULL); 2854 } 2855 2856 #ifndef PRODUCT 2857 2858 void Scheduling::dump_available() const { 2859 tty->print("#Availist "); 2860 for (uint i = 0; i < _available.size(); i++) 2861 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 2862 tty->cr(); 2863 } 2864 2865 // Print Scheduling Statistics 2866 void Scheduling::print_statistics() { 2867 // Print the size added by nops for bundling 2868 tty->print("Nops added %d bytes to total of %d bytes", 2869 _total_nop_size, _total_method_size); 2870 if (_total_method_size > 0) 2871 tty->print(", for %.2f%%", 2872 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 2873 tty->print("\n"); 2874 2875 // Print the number of branch shadows filled 2876 if (Pipeline::_branch_has_delay_slot) { 2877 tty->print("Of %d branches, %d had unconditional delay slots filled", 2878 _total_branches, _total_unconditional_delays); 2879 if (_total_branches > 0) 2880 tty->print(", for %.2f%%", 2881 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 2882 tty->print("\n"); 2883 } 2884 2885 uint total_instructions = 0, total_bundles = 0; 2886 2887 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 2888 uint bundle_count = _total_instructions_per_bundle[i]; 2889 total_instructions += bundle_count * i; 2890 total_bundles += bundle_count; 2891 } 2892 2893 if (total_bundles > 0) 2894 tty->print("Average ILP (excluding nops) is %.2f\n", 2895 ((double)total_instructions) / ((double)total_bundles)); 2896 } 2897 #endif