1 /* 2 * Copyright (c) 2002, 2014, 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 "code/vmreg.inline.hpp" 27 #include "compiler/oopMap.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/callnode.hpp" 30 #include "opto/compile.hpp" 31 #include "opto/machnode.hpp" 32 #include "opto/matcher.hpp" 33 #include "opto/phase.hpp" 34 #include "opto/regalloc.hpp" 35 #include "opto/rootnode.hpp" 36 37 // The functions in this file builds OopMaps after all scheduling is done. 38 // 39 // OopMaps contain a list of all registers and stack-slots containing oops (so 40 // they can be updated by GC). OopMaps also contain a list of derived-pointer 41 // base-pointer pairs. When the base is moved, the derived pointer moves to 42 // follow it. Finally, any registers holding callee-save values are also 43 // recorded. These might contain oops, but only the caller knows. 44 // 45 // BuildOopMaps implements a simple forward reaching-defs solution. At each 46 // GC point we'll have the reaching-def Nodes. If the reaching Nodes are 47 // typed as pointers (no offset), then they are oops. Pointers+offsets are 48 // derived pointers, and bases can be found from them. Finally, we'll also 49 // track reaching callee-save values. Note that a copy of a callee-save value 50 // "kills" it's source, so that only 1 copy of a callee-save value is alive at 51 // a time. 52 // 53 // We run a simple bitvector liveness pass to help trim out dead oops. Due to 54 // irreducible loops, we can have a reaching def of an oop that only reaches 55 // along one path and no way to know if it's valid or not on the other path. 56 // The bitvectors are quite dense and the liveness pass is fast. 57 // 58 // At GC points, we consult this information to build OopMaps. All reaching 59 // defs typed as oops are added to the OopMap. Only 1 instance of a 60 // callee-save register can be recorded. For derived pointers, we'll have to 61 // find and record the register holding the base. 62 // 63 // The reaching def's is a simple 1-pass worklist approach. I tried a clever 64 // breadth-first approach but it was worse (showed O(n^2) in the 65 // pick-next-block code). 66 // 67 // The relevant data is kept in a struct of arrays (it could just as well be 68 // an array of structs, but the struct-of-arrays is generally a little more 69 // efficient). The arrays are indexed by register number (including 70 // stack-slots as registers) and so is bounded by 200 to 300 elements in 71 // practice. One array will map to a reaching def Node (or NULL for 72 // conflict/dead). The other array will map to a callee-saved register or 73 // OptoReg::Bad for not-callee-saved. 74 75 76 // Structure to pass around 77 struct OopFlow : public ResourceObj { 78 short *_callees; // Array mapping register to callee-saved 79 Node **_defs; // array mapping register to reaching def 80 // or NULL if dead/conflict 81 // OopFlow structs, when not being actively modified, describe the _end_ of 82 // this block. 83 Block *_b; // Block for this struct 84 OopFlow *_next; // Next free OopFlow 85 // or NULL if dead/conflict 86 Compile* C; 87 88 OopFlow( short *callees, Node **defs, Compile* c ) : _callees(callees), _defs(defs), 89 _b(NULL), _next(NULL), C(c) { } 90 91 // Given reaching-defs for this block start, compute it for this block end 92 void compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash ); 93 94 // Merge these two OopFlows into the 'this' pointer. 95 void merge( OopFlow *flow, int max_reg ); 96 97 // Copy a 'flow' over an existing flow 98 void clone( OopFlow *flow, int max_size); 99 100 // Make a new OopFlow from scratch 101 static OopFlow *make( Arena *A, int max_size, Compile* C ); 102 103 // Build an oopmap from the current flow info 104 OopMap *build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live ); 105 }; 106 107 // Given reaching-defs for this block start, compute it for this block end 108 void OopFlow::compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash ) { 109 110 for( uint i=0; i<_b->number_of_nodes(); i++ ) { 111 Node *n = _b->get_node(i); 112 113 if( n->jvms() ) { // Build an OopMap here? 114 JVMState *jvms = n->jvms(); 115 // no map needed for leaf calls 116 if( n->is_MachSafePoint() && !n->is_MachCallLeaf() ) { 117 int *live = (int*) (*safehash)[n]; 118 assert( live, "must find live" ); 119 n->as_MachSafePoint()->set_oop_map( build_oop_map(n,max_reg,regalloc, live) ); 120 } 121 } 122 123 // Assign new reaching def's. 124 // Note that I padded the _defs and _callees arrays so it's legal 125 // to index at _defs[OptoReg::Bad]. 126 OptoReg::Name first = regalloc->get_reg_first(n); 127 OptoReg::Name second = regalloc->get_reg_second(n); 128 _defs[first] = n; 129 _defs[second] = n; 130 131 // Pass callee-save info around copies 132 int idx = n->is_Copy(); 133 if( idx ) { // Copies move callee-save info 134 OptoReg::Name old_first = regalloc->get_reg_first(n->in(idx)); 135 OptoReg::Name old_second = regalloc->get_reg_second(n->in(idx)); 136 int tmp_first = _callees[old_first]; 137 int tmp_second = _callees[old_second]; 138 _callees[old_first] = OptoReg::Bad; // callee-save is moved, dead in old location 139 _callees[old_second] = OptoReg::Bad; 140 _callees[first] = tmp_first; 141 _callees[second] = tmp_second; 142 } else if( n->is_Phi() ) { // Phis do not mod callee-saves 143 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(1))], "" ); 144 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(1))], "" ); 145 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(n->req()-1))], "" ); 146 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(n->req()-1))], "" ); 147 } else { 148 _callees[first] = OptoReg::Bad; // No longer holding a callee-save value 149 _callees[second] = OptoReg::Bad; 150 151 // Find base case for callee saves 152 if( n->is_Proj() && n->in(0)->is_Start() ) { 153 if( OptoReg::is_reg(first) && 154 regalloc->_matcher.is_save_on_entry(first) ) 155 _callees[first] = first; 156 if( OptoReg::is_reg(second) && 157 regalloc->_matcher.is_save_on_entry(second) ) 158 _callees[second] = second; 159 } 160 } 161 } 162 } 163 164 // Merge the given flow into the 'this' flow 165 void OopFlow::merge( OopFlow *flow, int max_reg ) { 166 assert( _b == NULL, "merging into a happy flow" ); 167 assert( flow->_b, "this flow is still alive" ); 168 assert( flow != this, "no self flow" ); 169 170 // Do the merge. If there are any differences, drop to 'bottom' which 171 // is OptoReg::Bad or NULL depending. 172 for( int i=0; i<max_reg; i++ ) { 173 // Merge the callee-save's 174 if( _callees[i] != flow->_callees[i] ) 175 _callees[i] = OptoReg::Bad; 176 // Merge the reaching defs 177 if( _defs[i] != flow->_defs[i] ) 178 _defs[i] = NULL; 179 } 180 181 } 182 183 void OopFlow::clone( OopFlow *flow, int max_size ) { 184 _b = flow->_b; 185 memcpy( _callees, flow->_callees, sizeof(short)*max_size); 186 memcpy( _defs , flow->_defs , sizeof(Node*)*max_size); 187 } 188 189 OopFlow *OopFlow::make( Arena *A, int max_size, Compile* C ) { 190 short *callees = NEW_ARENA_ARRAY(A,short,max_size+1); 191 Node **defs = NEW_ARENA_ARRAY(A,Node*,max_size+1); 192 debug_only( memset(defs,0,(max_size+1)*sizeof(Node*)) ); 193 OopFlow *flow = new (A) OopFlow(callees+1, defs+1, C); 194 assert( &flow->_callees[OptoReg::Bad] == callees, "Ok to index at OptoReg::Bad" ); 195 assert( &flow->_defs [OptoReg::Bad] == defs , "Ok to index at OptoReg::Bad" ); 196 return flow; 197 } 198 199 static int get_live_bit( int *live, int reg ) { 200 return live[reg>>LogBitsPerInt] & (1<<(reg&(BitsPerInt-1))); } 201 static void set_live_bit( int *live, int reg ) { 202 live[reg>>LogBitsPerInt] |= (1<<(reg&(BitsPerInt-1))); } 203 static void clr_live_bit( int *live, int reg ) { 204 live[reg>>LogBitsPerInt] &= ~(1<<(reg&(BitsPerInt-1))); } 205 206 // Build an oopmap from the current flow info 207 OopMap *OopFlow::build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live ) { 208 int framesize = regalloc->_framesize; 209 int max_inarg_slot = OptoReg::reg2stack(regalloc->_matcher._new_SP); 210 debug_only( char *dup_check = NEW_RESOURCE_ARRAY(char,OptoReg::stack0()); 211 memset(dup_check,0,OptoReg::stack0()) ); 212 213 OopMap *omap = new OopMap( framesize, max_inarg_slot ); 214 MachCallNode *mcall = n->is_MachCall() ? n->as_MachCall() : NULL; 215 JVMState* jvms = n->jvms(); 216 217 // For all registers do... 218 for( int reg=0; reg<max_reg; reg++ ) { 219 if( get_live_bit(live,reg) == 0 ) 220 continue; // Ignore if not live 221 222 // %%% C2 can use 2 OptoRegs when the physical register is only one 64bit 223 // register in that case we'll get an non-concrete register for the second 224 // half. We only need to tell the map the register once! 225 // 226 // However for the moment we disable this change and leave things as they 227 // were. 228 229 VMReg r = OptoReg::as_VMReg(OptoReg::Name(reg), framesize, max_inarg_slot); 230 231 if (false && r->is_reg() && !r->is_concrete()) { 232 continue; 233 } 234 235 // See if dead (no reaching def). 236 Node *def = _defs[reg]; // Get reaching def 237 assert( def, "since live better have reaching def" ); 238 239 // Classify the reaching def as oop, derived, callee-save, dead, or other 240 const Type *t = def->bottom_type(); 241 if( t->isa_oop_ptr() ) { // Oop or derived? 242 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" ); 243 #ifdef _LP64 244 // 64-bit pointers record oop-ishness on 2 aligned adjacent registers. 245 // Make sure both are record from the same reaching def, but do not 246 // put both into the oopmap. 247 if( (reg&1) == 1 ) { // High half of oop-pair? 248 assert( _defs[reg-1] == _defs[reg], "both halves from same reaching def" ); 249 continue; // Do not record high parts in oopmap 250 } 251 #endif 252 253 // Check for a legal reg name in the oopMap and bailout if it is not. 254 if (!omap->legal_vm_reg_name(r)) { 255 regalloc->C->record_method_not_compilable("illegal oopMap register name"); 256 continue; 257 } 258 if( t->is_ptr()->_offset == 0 ) { // Not derived? 259 if( mcall ) { 260 // Outgoing argument GC mask responsibility belongs to the callee, 261 // not the caller. Inspect the inputs to the call, to see if 262 // this live-range is one of them. 263 uint cnt = mcall->tf()->domain()->cnt(); 264 uint j; 265 for( j = TypeFunc::Parms; j < cnt; j++) 266 if( mcall->in(j) == def ) 267 break; // reaching def is an argument oop 268 if( j < cnt ) // arg oops dont go in GC map 269 continue; // Continue on to the next register 270 } 271 omap->set_oop(r); 272 } else { // Else it's derived. 273 // Find the base of the derived value. 274 uint i; 275 // Fast, common case, scan 276 for( i = jvms->oopoff(); i < n->req(); i+=2 ) 277 if( n->in(i) == def ) break; // Common case 278 if( i == n->req() ) { // Missed, try a more generous scan 279 // Scan again, but this time peek through copies 280 for( i = jvms->oopoff(); i < n->req(); i+=2 ) { 281 Node *m = n->in(i); // Get initial derived value 282 while( 1 ) { 283 Node *d = def; // Get initial reaching def 284 while( 1 ) { // Follow copies of reaching def to end 285 if( m == d ) goto found; // breaks 3 loops 286 int idx = d->is_Copy(); 287 if( !idx ) break; 288 d = d->in(idx); // Link through copy 289 } 290 int idx = m->is_Copy(); 291 if( !idx ) break; 292 m = m->in(idx); 293 } 294 } 295 guarantee( 0, "must find derived/base pair" ); 296 } 297 found: ; 298 Node *base = n->in(i+1); // Base is other half of pair 299 int breg = regalloc->get_reg_first(base); 300 VMReg b = OptoReg::as_VMReg(OptoReg::Name(breg), framesize, max_inarg_slot); 301 302 // I record liveness at safepoints BEFORE I make the inputs 303 // live. This is because argument oops are NOT live at a 304 // safepoint (or at least they cannot appear in the oopmap). 305 // Thus bases of base/derived pairs might not be in the 306 // liveness data but they need to appear in the oopmap. 307 if( get_live_bit(live,breg) == 0 ) {// Not live? 308 // Flag it, so next derived pointer won't re-insert into oopmap 309 set_live_bit(live,breg); 310 // Already missed our turn? 311 if( breg < reg ) { 312 if (b->is_stack() || b->is_concrete() || true ) { 313 omap->set_oop( b); 314 } 315 } 316 } 317 if (b->is_stack() || b->is_concrete() || true ) { 318 omap->set_derived_oop( r, b); 319 } 320 } 321 322 } else if( t->isa_narrowoop() ) { 323 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" ); 324 // Check for a legal reg name in the oopMap and bailout if it is not. 325 if (!omap->legal_vm_reg_name(r)) { 326 regalloc->C->record_method_not_compilable("illegal oopMap register name"); 327 continue; 328 } 329 if( mcall ) { 330 // Outgoing argument GC mask responsibility belongs to the callee, 331 // not the caller. Inspect the inputs to the call, to see if 332 // this live-range is one of them. 333 uint cnt = mcall->tf()->domain()->cnt(); 334 uint j; 335 for( j = TypeFunc::Parms; j < cnt; j++) 336 if( mcall->in(j) == def ) 337 break; // reaching def is an argument oop 338 if( j < cnt ) // arg oops dont go in GC map 339 continue; // Continue on to the next register 340 } 341 omap->set_narrowoop(r); 342 } else if( OptoReg::is_valid(_callees[reg])) { // callee-save? 343 // It's a callee-save value 344 assert( dup_check[_callees[reg]]==0, "trying to callee save same reg twice" ); 345 debug_only( dup_check[_callees[reg]]=1; ) 346 VMReg callee = OptoReg::as_VMReg(OptoReg::Name(_callees[reg])); 347 if ( callee->is_concrete() || true ) { 348 omap->set_callee_saved( r, callee); 349 } 350 351 } else { 352 // Other - some reaching non-oop value 353 omap->set_value( r); 354 #ifdef ASSERT 355 if( t->isa_rawptr() && C->cfg()->_raw_oops.member(def) ) { 356 def->dump(); 357 n->dump(); 358 assert(false, "there should be a oop in OopMap instead of a live raw oop at safepoint"); 359 } 360 #endif 361 } 362 363 } 364 365 #ifdef ASSERT 366 /* Nice, Intel-only assert 367 int cnt_callee_saves=0; 368 int reg2 = 0; 369 while (OptoReg::is_reg(reg2)) { 370 if( dup_check[reg2] != 0) cnt_callee_saves++; 371 assert( cnt_callee_saves==3 || cnt_callee_saves==5, "missed some callee-save" ); 372 reg2++; 373 } 374 */ 375 #endif 376 377 #ifdef ASSERT 378 for( OopMapStream oms1(omap, OopMapValue::derived_oop_value); !oms1.is_done(); oms1.next()) { 379 OopMapValue omv1 = oms1.current(); 380 bool found = false; 381 for( OopMapStream oms2(omap,OopMapValue::oop_value); !oms2.is_done(); oms2.next()) { 382 if( omv1.content_reg() == oms2.current().reg() ) { 383 found = true; 384 break; 385 } 386 } 387 assert( found, "derived with no base in oopmap" ); 388 } 389 #endif 390 391 return omap; 392 } 393 394 // Compute backwards liveness on registers 395 static void do_liveness(PhaseRegAlloc* regalloc, PhaseCFG* cfg, Block_List* worklist, int max_reg_ints, Arena* A, Dict* safehash) { 396 int* live = NEW_ARENA_ARRAY(A, int, (cfg->number_of_blocks() + 1) * max_reg_ints); 397 int* tmp_live = &live[cfg->number_of_blocks() * max_reg_ints]; 398 Node* root = cfg->get_root_node(); 399 // On CISC platforms, get the node representing the stack pointer that regalloc 400 // used for spills 401 Node *fp = NodeSentinel; 402 if (UseCISCSpill && root->req() > 1) { 403 fp = root->in(1)->in(TypeFunc::FramePtr); 404 } 405 memset(live, 0, cfg->number_of_blocks() * (max_reg_ints << LogBytesPerInt)); 406 // Push preds onto worklist 407 for (uint i = 1; i < root->req(); i++) { 408 Block* block = cfg->get_block_for_node(root->in(i)); 409 worklist->push(block); 410 } 411 412 // ZKM.jar includes tiny infinite loops which are unreached from below. 413 // If we missed any blocks, we'll retry here after pushing all missed 414 // blocks on the worklist. Normally this outer loop never trips more 415 // than once. 416 while (1) { 417 418 while( worklist->size() ) { // Standard worklist algorithm 419 Block *b = worklist->rpop(); 420 421 // Copy first successor into my tmp_live space 422 int s0num = b->_succs[0]->_pre_order; 423 int *t = &live[s0num*max_reg_ints]; 424 for( int i=0; i<max_reg_ints; i++ ) 425 tmp_live[i] = t[i]; 426 427 // OR in the remaining live registers 428 for( uint j=1; j<b->_num_succs; j++ ) { 429 uint sjnum = b->_succs[j]->_pre_order; 430 int *t = &live[sjnum*max_reg_ints]; 431 for( int i=0; i<max_reg_ints; i++ ) 432 tmp_live[i] |= t[i]; 433 } 434 435 // Now walk tmp_live up the block backwards, computing live 436 for( int k=b->number_of_nodes()-1; k>=0; k-- ) { 437 Node *n = b->get_node(k); 438 // KILL def'd bits 439 int first = regalloc->get_reg_first(n); 440 int second = regalloc->get_reg_second(n); 441 if( OptoReg::is_valid(first) ) clr_live_bit(tmp_live,first); 442 if( OptoReg::is_valid(second) ) clr_live_bit(tmp_live,second); 443 444 MachNode *m = n->is_Mach() ? n->as_Mach() : NULL; 445 446 // Check if m is potentially a CISC alternate instruction (i.e, possibly 447 // synthesized by RegAlloc from a conventional instruction and a 448 // spilled input) 449 bool is_cisc_alternate = false; 450 if (UseCISCSpill && m) { 451 is_cisc_alternate = m->is_cisc_alternate(); 452 } 453 454 // GEN use'd bits 455 for( uint l=1; l<n->req(); l++ ) { 456 Node *def = n->in(l); 457 assert(def != 0, "input edge required"); 458 int first = regalloc->get_reg_first(def); 459 int second = regalloc->get_reg_second(def); 460 if( OptoReg::is_valid(first) ) set_live_bit(tmp_live,first); 461 if( OptoReg::is_valid(second) ) set_live_bit(tmp_live,second); 462 // If we use the stack pointer in a cisc-alternative instruction, 463 // check for use as a memory operand. Then reconstruct the RegName 464 // for this stack location, and set the appropriate bit in the 465 // live vector 4987749. 466 if (is_cisc_alternate && def == fp) { 467 const TypePtr *adr_type = NULL; 468 intptr_t offset; 469 const Node* base = m->get_base_and_disp(offset, adr_type); 470 if (base == NodeSentinel) { 471 // Machnode has multiple memory inputs. We are unable to reason 472 // with these, but are presuming (with trepidation) that not any of 473 // them are oops. This can be fixed by making get_base_and_disp() 474 // look at a specific input instead of all inputs. 475 assert(!def->bottom_type()->isa_oop_ptr(), "expecting non-oop mem input"); 476 } else if (base != fp || offset == Type::OffsetBot) { 477 // Do nothing: the fp operand is either not from a memory use 478 // (base == NULL) OR the fp is used in a non-memory context 479 // (base is some other register) OR the offset is not constant, 480 // so it is not a stack slot. 481 } else { 482 assert(offset >= 0, "unexpected negative offset"); 483 offset -= (offset % jintSize); // count the whole word 484 int stack_reg = regalloc->offset2reg(offset); 485 if (OptoReg::is_stack(stack_reg)) { 486 set_live_bit(tmp_live, stack_reg); 487 } else { 488 assert(false, "stack_reg not on stack?"); 489 } 490 } 491 } 492 } 493 494 if( n->jvms() ) { // Record liveness at safepoint 495 496 // This placement of this stanza means inputs to calls are 497 // considered live at the callsite's OopMap. Argument oops are 498 // hence live, but NOT included in the oopmap. See cutout in 499 // build_oop_map. Debug oops are live (and in OopMap). 500 int *n_live = NEW_ARENA_ARRAY(A, int, max_reg_ints); 501 for( int l=0; l<max_reg_ints; l++ ) 502 n_live[l] = tmp_live[l]; 503 safehash->Insert(n,n_live); 504 } 505 506 } 507 508 // Now at block top, see if we have any changes. If so, propagate 509 // to prior blocks. 510 int *old_live = &live[b->_pre_order*max_reg_ints]; 511 int l; 512 for( l=0; l<max_reg_ints; l++ ) 513 if( tmp_live[l] != old_live[l] ) 514 break; 515 if( l<max_reg_ints ) { // Change! 516 // Copy in new value 517 for( l=0; l<max_reg_ints; l++ ) 518 old_live[l] = tmp_live[l]; 519 // Push preds onto worklist 520 for (l = 1; l < (int)b->num_preds(); l++) { 521 Block* block = cfg->get_block_for_node(b->pred(l)); 522 worklist->push(block); 523 } 524 } 525 } 526 527 // Scan for any missing safepoints. Happens to infinite loops 528 // ala ZKM.jar 529 uint i; 530 for (i = 1; i < cfg->number_of_blocks(); i++) { 531 Block* block = cfg->get_block(i); 532 uint j; 533 for (j = 1; j < block->number_of_nodes(); j++) { 534 if (block->get_node(j)->jvms() && (*safehash)[block->get_node(j)] == NULL) { 535 break; 536 } 537 } 538 if (j < block->number_of_nodes()) { 539 break; 540 } 541 } 542 if (i == cfg->number_of_blocks()) { 543 break; // Got 'em all 544 } 545 546 if (PrintOpto && Verbose) { 547 tty->print_cr("retripping live calc"); 548 } 549 550 // Force the issue (expensively): recheck everybody 551 for (i = 1; i < cfg->number_of_blocks(); i++) { 552 worklist->push(cfg->get_block(i)); 553 } 554 } 555 } 556 557 // Collect GC mask info - where are all the OOPs? 558 void Compile::BuildOopMaps() { 559 TracePhase tp("bldOopMaps", &timers[_t_buildOopMaps]); 560 // Can't resource-mark because I need to leave all those OopMaps around, 561 // or else I need to resource-mark some arena other than the default. 562 // ResourceMark rm; // Reclaim all OopFlows when done 563 int max_reg = _regalloc->_max_reg; // Current array extent 564 565 Arena *A = Thread::current()->resource_area(); 566 Block_List worklist; // Worklist of pending blocks 567 568 int max_reg_ints = round_to(max_reg, BitsPerInt)>>LogBitsPerInt; 569 Dict *safehash = NULL; // Used for assert only 570 // Compute a backwards liveness per register. Needs a bitarray of 571 // #blocks x (#registers, rounded up to ints) 572 safehash = new Dict(cmpkey,hashkey,A); 573 do_liveness( _regalloc, _cfg, &worklist, max_reg_ints, A, safehash ); 574 OopFlow *free_list = NULL; // Free, unused 575 576 // Array mapping blocks to completed oopflows 577 OopFlow **flows = NEW_ARENA_ARRAY(A, OopFlow*, _cfg->number_of_blocks()); 578 memset( flows, 0, _cfg->number_of_blocks() * sizeof(OopFlow*) ); 579 580 581 // Do the first block 'by hand' to prime the worklist 582 Block *entry = _cfg->get_block(1); 583 OopFlow *rootflow = OopFlow::make(A,max_reg,this); 584 // Initialize to 'bottom' (not 'top') 585 memset( rootflow->_callees, OptoReg::Bad, max_reg*sizeof(short) ); 586 memset( rootflow->_defs , 0, max_reg*sizeof(Node*) ); 587 flows[entry->_pre_order] = rootflow; 588 589 // Do the first block 'by hand' to prime the worklist 590 rootflow->_b = entry; 591 rootflow->compute_reach( _regalloc, max_reg, safehash ); 592 for( uint i=0; i<entry->_num_succs; i++ ) 593 worklist.push(entry->_succs[i]); 594 595 // Now worklist contains blocks which have some, but perhaps not all, 596 // predecessors visited. 597 while( worklist.size() ) { 598 // Scan for a block with all predecessors visited, or any randoms slob 599 // otherwise. All-preds-visited order allows me to recycle OopFlow 600 // structures rapidly and cut down on the memory footprint. 601 // Note: not all predecessors might be visited yet (must happen for 602 // irreducible loops). This is OK, since every live value must have the 603 // SAME reaching def for the block, so any reaching def is OK. 604 uint i; 605 606 Block *b = worklist.pop(); 607 // Ignore root block 608 if (b == _cfg->get_root_block()) { 609 continue; 610 } 611 // Block is already done? Happens if block has several predecessors, 612 // he can get on the worklist more than once. 613 if( flows[b->_pre_order] ) continue; 614 615 // If this block has a visited predecessor AND that predecessor has this 616 // last block as his only undone child, we can move the OopFlow from the 617 // pred to this block. Otherwise we have to grab a new OopFlow. 618 OopFlow *flow = NULL; // Flag for finding optimized flow 619 Block *pred = (Block*)0xdeadbeef; 620 // Scan this block's preds to find a done predecessor 621 for (uint j = 1; j < b->num_preds(); j++) { 622 Block* p = _cfg->get_block_for_node(b->pred(j)); 623 OopFlow *p_flow = flows[p->_pre_order]; 624 if( p_flow ) { // Predecessor is done 625 assert( p_flow->_b == p, "cross check" ); 626 pred = p; // Record some predecessor 627 // If all successors of p are done except for 'b', then we can carry 628 // p_flow forward to 'b' without copying, otherwise we have to draw 629 // from the free_list and clone data. 630 uint k; 631 for( k=0; k<p->_num_succs; k++ ) 632 if( !flows[p->_succs[k]->_pre_order] && 633 p->_succs[k] != b ) 634 break; 635 636 // Either carry-forward the now-unused OopFlow for b's use 637 // or draw a new one from the free list 638 if( k==p->_num_succs ) { 639 flow = p_flow; 640 break; // Found an ideal pred, use him 641 } 642 } 643 } 644 645 if( flow ) { 646 // We have an OopFlow that's the last-use of a predecessor. 647 // Carry it forward. 648 } else { // Draw a new OopFlow from the freelist 649 if( !free_list ) 650 free_list = OopFlow::make(A,max_reg,C); 651 flow = free_list; 652 assert( flow->_b == NULL, "oopFlow is not free" ); 653 free_list = flow->_next; 654 flow->_next = NULL; 655 656 // Copy/clone over the data 657 flow->clone(flows[pred->_pre_order], max_reg); 658 } 659 660 // Mark flow for block. Blocks can only be flowed over once, 661 // because after the first time they are guarded from entering 662 // this code again. 663 assert( flow->_b == pred, "have some prior flow" ); 664 flow->_b = NULL; 665 666 // Now push flow forward 667 flows[b->_pre_order] = flow;// Mark flow for this block 668 flow->_b = b; 669 flow->compute_reach( _regalloc, max_reg, safehash ); 670 671 // Now push children onto worklist 672 for( i=0; i<b->_num_succs; i++ ) 673 worklist.push(b->_succs[i]); 674 675 } 676 }