1 /* 2 * Copyright 1997-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 #include "incls/_precompiled.incl" 26 #include "incls/_matcher.cpp.incl" 27 28 OptoReg::Name OptoReg::c_frame_pointer; 29 30 31 32 const int Matcher::base2reg[Type::lastype] = { 33 Node::NotAMachineReg,0,0, Op_RegI, Op_RegL, 0, Op_RegN, 34 Node::NotAMachineReg, Node::NotAMachineReg, /* tuple, array */ 35 Op_RegP, Op_RegP, Op_RegP, Op_RegP, Op_RegP, Op_RegP, /* the pointers */ 36 0, 0/*abio*/, 37 Op_RegP /* Return address */, 0, /* the memories */ 38 Op_RegF, Op_RegF, Op_RegF, Op_RegD, Op_RegD, Op_RegD, 39 0 /*bottom*/ 40 }; 41 42 const RegMask *Matcher::idealreg2regmask[_last_machine_leaf]; 43 RegMask Matcher::mreg2regmask[_last_Mach_Reg]; 44 RegMask Matcher::STACK_ONLY_mask; 45 RegMask Matcher::c_frame_ptr_mask; 46 const uint Matcher::_begin_rematerialize = _BEGIN_REMATERIALIZE; 47 const uint Matcher::_end_rematerialize = _END_REMATERIALIZE; 48 49 //---------------------------Matcher------------------------------------------- 50 Matcher::Matcher( Node_List &proj_list ) : 51 PhaseTransform( Phase::Ins_Select ), 52 #ifdef ASSERT 53 _old2new_map(C->comp_arena()), 54 _new2old_map(C->comp_arena()), 55 #endif 56 _shared_nodes(C->comp_arena()), 57 _reduceOp(reduceOp), _leftOp(leftOp), _rightOp(rightOp), 58 _swallowed(swallowed), 59 _begin_inst_chain_rule(_BEGIN_INST_CHAIN_RULE), 60 _end_inst_chain_rule(_END_INST_CHAIN_RULE), 61 _must_clone(must_clone), _proj_list(proj_list), 62 _register_save_policy(register_save_policy), 63 _c_reg_save_policy(c_reg_save_policy), 64 _register_save_type(register_save_type), 65 _ruleName(ruleName), 66 _allocation_started(false), 67 _states_arena(Chunk::medium_size), 68 _visited(&_states_arena), 69 _shared(&_states_arena), 70 _dontcare(&_states_arena) { 71 C->set_matcher(this); 72 73 idealreg2spillmask[Op_RegI] = NULL; 74 idealreg2spillmask[Op_RegN] = NULL; 75 idealreg2spillmask[Op_RegL] = NULL; 76 idealreg2spillmask[Op_RegF] = NULL; 77 idealreg2spillmask[Op_RegD] = NULL; 78 idealreg2spillmask[Op_RegP] = NULL; 79 80 idealreg2debugmask[Op_RegI] = NULL; 81 idealreg2debugmask[Op_RegN] = NULL; 82 idealreg2debugmask[Op_RegL] = NULL; 83 idealreg2debugmask[Op_RegF] = NULL; 84 idealreg2debugmask[Op_RegD] = NULL; 85 idealreg2debugmask[Op_RegP] = NULL; 86 debug_only(_mem_node = NULL;) // Ideal memory node consumed by mach node 87 } 88 89 //------------------------------warp_incoming_stk_arg------------------------ 90 // This warps a VMReg into an OptoReg::Name 91 OptoReg::Name Matcher::warp_incoming_stk_arg( VMReg reg ) { 92 OptoReg::Name warped; 93 if( reg->is_stack() ) { // Stack slot argument? 94 warped = OptoReg::add(_old_SP, reg->reg2stack() ); 95 warped = OptoReg::add(warped, C->out_preserve_stack_slots()); 96 if( warped >= _in_arg_limit ) 97 _in_arg_limit = OptoReg::add(warped, 1); // Bump max stack slot seen 98 if (!RegMask::can_represent(warped)) { 99 // the compiler cannot represent this method's calling sequence 100 C->record_method_not_compilable_all_tiers("unsupported incoming calling sequence"); 101 return OptoReg::Bad; 102 } 103 return warped; 104 } 105 return OptoReg::as_OptoReg(reg); 106 } 107 108 //---------------------------compute_old_SP------------------------------------ 109 OptoReg::Name Compile::compute_old_SP() { 110 int fixed = fixed_slots(); 111 int preserve = in_preserve_stack_slots(); 112 return OptoReg::stack2reg(round_to(fixed + preserve, Matcher::stack_alignment_in_slots())); 113 } 114 115 116 117 #ifdef ASSERT 118 void Matcher::verify_new_nodes_only(Node* xroot) { 119 // Make sure that the new graph only references new nodes 120 ResourceMark rm; 121 Unique_Node_List worklist; 122 VectorSet visited(Thread::current()->resource_area()); 123 worklist.push(xroot); 124 while (worklist.size() > 0) { 125 Node* n = worklist.pop(); 126 visited <<= n->_idx; 127 assert(C->node_arena()->contains(n), "dead node"); 128 for (uint j = 0; j < n->req(); j++) { 129 Node* in = n->in(j); 130 if (in != NULL) { 131 assert(C->node_arena()->contains(in), "dead node"); 132 if (!visited.test(in->_idx)) { 133 worklist.push(in); 134 } 135 } 136 } 137 } 138 } 139 #endif 140 141 142 //---------------------------match--------------------------------------------- 143 void Matcher::match( ) { 144 // One-time initialization of some register masks. 145 init_spill_mask( C->root()->in(1) ); 146 _return_addr_mask = return_addr(); 147 #ifdef _LP64 148 // Pointers take 2 slots in 64-bit land 149 _return_addr_mask.Insert(OptoReg::add(return_addr(),1)); 150 #endif 151 152 // Map a Java-signature return type into return register-value 153 // machine registers for 0, 1 and 2 returned values. 154 const TypeTuple *range = C->tf()->range(); 155 if( range->cnt() > TypeFunc::Parms ) { // If not a void function 156 // Get ideal-register return type 157 int ireg = base2reg[range->field_at(TypeFunc::Parms)->base()]; 158 // Get machine return register 159 uint sop = C->start()->Opcode(); 160 OptoRegPair regs = return_value(ireg, false); 161 162 // And mask for same 163 _return_value_mask = RegMask(regs.first()); 164 if( OptoReg::is_valid(regs.second()) ) 165 _return_value_mask.Insert(regs.second()); 166 } 167 168 // --------------- 169 // Frame Layout 170 171 // Need the method signature to determine the incoming argument types, 172 // because the types determine which registers the incoming arguments are 173 // in, and this affects the matched code. 174 const TypeTuple *domain = C->tf()->domain(); 175 uint argcnt = domain->cnt() - TypeFunc::Parms; 176 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt ); 177 VMRegPair *vm_parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt ); 178 _parm_regs = NEW_RESOURCE_ARRAY( OptoRegPair, argcnt ); 179 _calling_convention_mask = NEW_RESOURCE_ARRAY( RegMask, argcnt ); 180 uint i; 181 for( i = 0; i<argcnt; i++ ) { 182 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type(); 183 } 184 185 // Pass array of ideal registers and length to USER code (from the AD file) 186 // that will convert this to an array of register numbers. 187 const StartNode *start = C->start(); 188 start->calling_convention( sig_bt, vm_parm_regs, argcnt ); 189 #ifdef ASSERT 190 // Sanity check users' calling convention. Real handy while trying to 191 // get the initial port correct. 192 { for (uint i = 0; i<argcnt; i++) { 193 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) { 194 assert(domain->field_at(i+TypeFunc::Parms)==Type::HALF, "only allowed on halve" ); 195 _parm_regs[i].set_bad(); 196 continue; 197 } 198 VMReg parm_reg = vm_parm_regs[i].first(); 199 assert(parm_reg->is_valid(), "invalid arg?"); 200 if (parm_reg->is_reg()) { 201 OptoReg::Name opto_parm_reg = OptoReg::as_OptoReg(parm_reg); 202 assert(can_be_java_arg(opto_parm_reg) || 203 C->stub_function() == CAST_FROM_FN_PTR(address, OptoRuntime::rethrow_C) || 204 opto_parm_reg == inline_cache_reg(), 205 "parameters in register must be preserved by runtime stubs"); 206 } 207 for (uint j = 0; j < i; j++) { 208 assert(parm_reg != vm_parm_regs[j].first(), 209 "calling conv. must produce distinct regs"); 210 } 211 } 212 } 213 #endif 214 215 // Do some initial frame layout. 216 217 // Compute the old incoming SP (may be called FP) as 218 // OptoReg::stack0() + locks + in_preserve_stack_slots + pad2. 219 _old_SP = C->compute_old_SP(); 220 assert( is_even(_old_SP), "must be even" ); 221 222 // Compute highest incoming stack argument as 223 // _old_SP + out_preserve_stack_slots + incoming argument size. 224 _in_arg_limit = OptoReg::add(_old_SP, C->out_preserve_stack_slots()); 225 assert( is_even(_in_arg_limit), "out_preserve must be even" ); 226 for( i = 0; i < argcnt; i++ ) { 227 // Permit args to have no register 228 _calling_convention_mask[i].Clear(); 229 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) { 230 continue; 231 } 232 // calling_convention returns stack arguments as a count of 233 // slots beyond OptoReg::stack0()/VMRegImpl::stack0. We need to convert this to 234 // the allocators point of view, taking into account all the 235 // preserve area, locks & pad2. 236 237 OptoReg::Name reg1 = warp_incoming_stk_arg(vm_parm_regs[i].first()); 238 if( OptoReg::is_valid(reg1)) 239 _calling_convention_mask[i].Insert(reg1); 240 241 OptoReg::Name reg2 = warp_incoming_stk_arg(vm_parm_regs[i].second()); 242 if( OptoReg::is_valid(reg2)) 243 _calling_convention_mask[i].Insert(reg2); 244 245 // Saved biased stack-slot register number 246 _parm_regs[i].set_pair(reg2, reg1); 247 } 248 249 // Finally, make sure the incoming arguments take up an even number of 250 // words, in case the arguments or locals need to contain doubleword stack 251 // slots. The rest of the system assumes that stack slot pairs (in 252 // particular, in the spill area) which look aligned will in fact be 253 // aligned relative to the stack pointer in the target machine. Double 254 // stack slots will always be allocated aligned. 255 _new_SP = OptoReg::Name(round_to(_in_arg_limit, RegMask::SlotsPerLong)); 256 257 // Compute highest outgoing stack argument as 258 // _new_SP + out_preserve_stack_slots + max(outgoing argument size). 259 _out_arg_limit = OptoReg::add(_new_SP, C->out_preserve_stack_slots()); 260 assert( is_even(_out_arg_limit), "out_preserve must be even" ); 261 262 if (!RegMask::can_represent(OptoReg::add(_out_arg_limit,-1))) { 263 // the compiler cannot represent this method's calling sequence 264 C->record_method_not_compilable("must be able to represent all call arguments in reg mask"); 265 } 266 267 if (C->failing()) return; // bailed out on incoming arg failure 268 269 // --------------- 270 // Collect roots of matcher trees. Every node for which 271 // _shared[_idx] is cleared is guaranteed to not be shared, and thus 272 // can be a valid interior of some tree. 273 find_shared( C->root() ); 274 find_shared( C->top() ); 275 276 C->print_method("Before Matching"); 277 278 // Create new ideal node ConP #NULL even if it does exist in old space 279 // to avoid false sharing if the corresponding mach node is not used. 280 // The corresponding mach node is only used in rare cases for derived 281 // pointers. Reuse existing ConP node only if it is shared. 282 Node* new_idealConP0 = ConNode::make(C, TypePtr::NULL_PTR); 283 Node* old_idealConP0 = NULL; 284 for( DUIterator_Fast jmax, j = C->root()->fast_outs(jmax); j < jmax; j++ ) { 285 Node* n = C->root()->fast_out(j); 286 if (n->Opcode() == Op_ConP && n->bottom_type() == TypePtr::NULL_PTR) { 287 old_idealConP0 = n; 288 break; 289 } 290 } 291 292 // Swap out to old-space; emptying new-space 293 Arena *old = C->node_arena()->move_contents(C->old_arena()); 294 295 // Save debug and profile information for nodes in old space: 296 _old_node_note_array = C->node_note_array(); 297 if (_old_node_note_array != NULL) { 298 C->set_node_note_array(new(C->comp_arena()) GrowableArray<Node_Notes*> 299 (C->comp_arena(), _old_node_note_array->length(), 300 0, NULL)); 301 } 302 303 // Pre-size the new_node table to avoid the need for range checks. 304 grow_new_node_array(C->unique()); 305 306 // Reset node counter so MachNodes start with _idx at 0 307 int nodes = C->unique(); // save value 308 C->set_unique(0); 309 310 // Recursively match trees from old space into new space. 311 // Correct leaves of new-space Nodes; they point to old-space. 312 _visited.Clear(); // Clear visit bits for xform call 313 C->set_cached_top_node(xform( C->top(), nodes )); 314 if (!C->failing()) { 315 Node* xroot = xform( C->root(), 1 ); 316 if (xroot == NULL) { 317 Matcher::soft_match_failure(); // recursive matching process failed 318 C->record_method_not_compilable("instruction match failed"); 319 } else { 320 // During matching shared constants were attached to C->root() 321 // because xroot wasn't available yet, so transfer the uses to 322 // the xroot. 323 for( DUIterator_Fast jmax, j = C->root()->fast_outs(jmax); j < jmax; j++ ) { 324 Node* n = C->root()->fast_out(j); 325 if (C->node_arena()->contains(n)) { 326 assert(n->in(0) == C->root(), "should be control user"); 327 n->set_req(0, xroot); 328 --j; 329 --jmax; 330 } 331 } 332 333 // Generate new mach node for ConP #NULL or use already generated. 334 if (old_idealConP0 != NULL && has_new_node(old_idealConP0)) { 335 _machConP0 = new_node(old_idealConP0)->as_Mach(); 336 } else { 337 assert(new_idealConP0 != NULL, "sanity"); 338 _machConP0 = match_tree(new_idealConP0); 339 // Don't set control, it will confuse GCM since there are no uses. 340 // The control will be set when this node is used first time 341 // in find_base_for_derived(). 342 } 343 assert(_machConP0 != NULL, ""); 344 345 C->set_root(xroot->is_Root() ? xroot->as_Root() : NULL); 346 347 #ifdef ASSERT 348 verify_new_nodes_only(xroot); 349 #endif 350 } 351 } 352 if (C->top() == NULL || C->root() == NULL) { 353 C->record_method_not_compilable("graph lost"); // %%% cannot happen? 354 } 355 if (C->failing()) { 356 // delete old; 357 old->destruct_contents(); 358 return; 359 } 360 assert( C->top(), "" ); 361 assert( C->root(), "" ); 362 validate_null_checks(); 363 364 // Now smoke old-space 365 NOT_DEBUG( old->destruct_contents() ); 366 367 // ------------------------ 368 // Set up save-on-entry registers 369 Fixup_Save_On_Entry( ); 370 } 371 372 373 //------------------------------Fixup_Save_On_Entry---------------------------- 374 // The stated purpose of this routine is to take care of save-on-entry 375 // registers. However, the overall goal of the Match phase is to convert into 376 // machine-specific instructions which have RegMasks to guide allocation. 377 // So what this procedure really does is put a valid RegMask on each input 378 // to the machine-specific variations of all Return, TailCall and Halt 379 // instructions. It also adds edgs to define the save-on-entry values (and of 380 // course gives them a mask). 381 382 static RegMask *init_input_masks( uint size, RegMask &ret_adr, RegMask &fp ) { 383 RegMask *rms = NEW_RESOURCE_ARRAY( RegMask, size ); 384 // Do all the pre-defined register masks 385 rms[TypeFunc::Control ] = RegMask::Empty; 386 rms[TypeFunc::I_O ] = RegMask::Empty; 387 rms[TypeFunc::Memory ] = RegMask::Empty; 388 rms[TypeFunc::ReturnAdr] = ret_adr; 389 rms[TypeFunc::FramePtr ] = fp; 390 return rms; 391 } 392 393 //---------------------------init_first_stack_mask----------------------------- 394 // Create the initial stack mask used by values spilling to the stack. 395 // Disallow any debug info in outgoing argument areas by setting the 396 // initial mask accordingly. 397 void Matcher::init_first_stack_mask() { 398 399 // Allocate storage for spill masks as masks for the appropriate load type. 400 RegMask *rms = (RegMask*)C->comp_arena()->Amalloc_D(sizeof(RegMask)*12); 401 idealreg2spillmask[Op_RegN] = &rms[0]; 402 idealreg2spillmask[Op_RegI] = &rms[1]; 403 idealreg2spillmask[Op_RegL] = &rms[2]; 404 idealreg2spillmask[Op_RegF] = &rms[3]; 405 idealreg2spillmask[Op_RegD] = &rms[4]; 406 idealreg2spillmask[Op_RegP] = &rms[5]; 407 idealreg2debugmask[Op_RegN] = &rms[6]; 408 idealreg2debugmask[Op_RegI] = &rms[7]; 409 idealreg2debugmask[Op_RegL] = &rms[8]; 410 idealreg2debugmask[Op_RegF] = &rms[9]; 411 idealreg2debugmask[Op_RegD] = &rms[10]; 412 idealreg2debugmask[Op_RegP] = &rms[11]; 413 414 OptoReg::Name i; 415 416 // At first, start with the empty mask 417 C->FIRST_STACK_mask().Clear(); 418 419 // Add in the incoming argument area 420 OptoReg::Name init = OptoReg::add(_old_SP, C->out_preserve_stack_slots()); 421 for (i = init; i < _in_arg_limit; i = OptoReg::add(i,1)) 422 C->FIRST_STACK_mask().Insert(i); 423 424 // Add in all bits past the outgoing argument area 425 guarantee(RegMask::can_represent(OptoReg::add(_out_arg_limit,-1)), 426 "must be able to represent all call arguments in reg mask"); 427 init = _out_arg_limit; 428 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) 429 C->FIRST_STACK_mask().Insert(i); 430 431 // Finally, set the "infinite stack" bit. 432 C->FIRST_STACK_mask().set_AllStack(); 433 434 // Make spill masks. Registers for their class, plus FIRST_STACK_mask. 435 #ifdef _LP64 436 *idealreg2spillmask[Op_RegN] = *idealreg2regmask[Op_RegN]; 437 idealreg2spillmask[Op_RegN]->OR(C->FIRST_STACK_mask()); 438 #endif 439 *idealreg2spillmask[Op_RegI] = *idealreg2regmask[Op_RegI]; 440 idealreg2spillmask[Op_RegI]->OR(C->FIRST_STACK_mask()); 441 *idealreg2spillmask[Op_RegL] = *idealreg2regmask[Op_RegL]; 442 idealreg2spillmask[Op_RegL]->OR(C->FIRST_STACK_mask()); 443 *idealreg2spillmask[Op_RegF] = *idealreg2regmask[Op_RegF]; 444 idealreg2spillmask[Op_RegF]->OR(C->FIRST_STACK_mask()); 445 *idealreg2spillmask[Op_RegD] = *idealreg2regmask[Op_RegD]; 446 idealreg2spillmask[Op_RegD]->OR(C->FIRST_STACK_mask()); 447 *idealreg2spillmask[Op_RegP] = *idealreg2regmask[Op_RegP]; 448 idealreg2spillmask[Op_RegP]->OR(C->FIRST_STACK_mask()); 449 450 // Make up debug masks. Any spill slot plus callee-save registers. 451 // Caller-save registers are assumed to be trashable by the various 452 // inline-cache fixup routines. 453 *idealreg2debugmask[Op_RegN]= *idealreg2spillmask[Op_RegN]; 454 *idealreg2debugmask[Op_RegI]= *idealreg2spillmask[Op_RegI]; 455 *idealreg2debugmask[Op_RegL]= *idealreg2spillmask[Op_RegL]; 456 *idealreg2debugmask[Op_RegF]= *idealreg2spillmask[Op_RegF]; 457 *idealreg2debugmask[Op_RegD]= *idealreg2spillmask[Op_RegD]; 458 *idealreg2debugmask[Op_RegP]= *idealreg2spillmask[Op_RegP]; 459 460 // Prevent stub compilations from attempting to reference 461 // callee-saved registers from debug info 462 bool exclude_soe = !Compile::current()->is_method_compilation(); 463 464 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) { 465 // registers the caller has to save do not work 466 if( _register_save_policy[i] == 'C' || 467 _register_save_policy[i] == 'A' || 468 (_register_save_policy[i] == 'E' && exclude_soe) ) { 469 idealreg2debugmask[Op_RegN]->Remove(i); 470 idealreg2debugmask[Op_RegI]->Remove(i); // Exclude save-on-call 471 idealreg2debugmask[Op_RegL]->Remove(i); // registers from debug 472 idealreg2debugmask[Op_RegF]->Remove(i); // masks 473 idealreg2debugmask[Op_RegD]->Remove(i); 474 idealreg2debugmask[Op_RegP]->Remove(i); 475 } 476 } 477 } 478 479 //---------------------------is_save_on_entry---------------------------------- 480 bool Matcher::is_save_on_entry( int reg ) { 481 return 482 _register_save_policy[reg] == 'E' || 483 _register_save_policy[reg] == 'A' || // Save-on-entry register? 484 // Also save argument registers in the trampolining stubs 485 (C->save_argument_registers() && is_spillable_arg(reg)); 486 } 487 488 //---------------------------Fixup_Save_On_Entry------------------------------- 489 void Matcher::Fixup_Save_On_Entry( ) { 490 init_first_stack_mask(); 491 492 Node *root = C->root(); // Short name for root 493 // Count number of save-on-entry registers. 494 uint soe_cnt = number_of_saved_registers(); 495 uint i; 496 497 // Find the procedure Start Node 498 StartNode *start = C->start(); 499 assert( start, "Expect a start node" ); 500 501 // Save argument registers in the trampolining stubs 502 if( C->save_argument_registers() ) 503 for( i = 0; i < _last_Mach_Reg; i++ ) 504 if( is_spillable_arg(i) ) 505 soe_cnt++; 506 507 // Input RegMask array shared by all Returns. 508 // The type for doubles and longs has a count of 2, but 509 // there is only 1 returned value 510 uint ret_edge_cnt = TypeFunc::Parms + ((C->tf()->range()->cnt() == TypeFunc::Parms) ? 0 : 1); 511 RegMask *ret_rms = init_input_masks( ret_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 512 // Returns have 0 or 1 returned values depending on call signature. 513 // Return register is specified by return_value in the AD file. 514 if (ret_edge_cnt > TypeFunc::Parms) 515 ret_rms[TypeFunc::Parms+0] = _return_value_mask; 516 517 // Input RegMask array shared by all Rethrows. 518 uint reth_edge_cnt = TypeFunc::Parms+1; 519 RegMask *reth_rms = init_input_masks( reth_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 520 // Rethrow takes exception oop only, but in the argument 0 slot. 521 reth_rms[TypeFunc::Parms] = mreg2regmask[find_receiver(false)]; 522 #ifdef _LP64 523 // Need two slots for ptrs in 64-bit land 524 reth_rms[TypeFunc::Parms].Insert(OptoReg::add(OptoReg::Name(find_receiver(false)),1)); 525 #endif 526 527 // Input RegMask array shared by all TailCalls 528 uint tail_call_edge_cnt = TypeFunc::Parms+2; 529 RegMask *tail_call_rms = init_input_masks( tail_call_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 530 531 // Input RegMask array shared by all TailJumps 532 uint tail_jump_edge_cnt = TypeFunc::Parms+2; 533 RegMask *tail_jump_rms = init_input_masks( tail_jump_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 534 535 // TailCalls have 2 returned values (target & moop), whose masks come 536 // from the usual MachNode/MachOper mechanism. Find a sample 537 // TailCall to extract these masks and put the correct masks into 538 // the tail_call_rms array. 539 for( i=1; i < root->req(); i++ ) { 540 MachReturnNode *m = root->in(i)->as_MachReturn(); 541 if( m->ideal_Opcode() == Op_TailCall ) { 542 tail_call_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0); 543 tail_call_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1); 544 break; 545 } 546 } 547 548 // TailJumps have 2 returned values (target & ex_oop), whose masks come 549 // from the usual MachNode/MachOper mechanism. Find a sample 550 // TailJump to extract these masks and put the correct masks into 551 // the tail_jump_rms array. 552 for( i=1; i < root->req(); i++ ) { 553 MachReturnNode *m = root->in(i)->as_MachReturn(); 554 if( m->ideal_Opcode() == Op_TailJump ) { 555 tail_jump_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0); 556 tail_jump_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1); 557 break; 558 } 559 } 560 561 // Input RegMask array shared by all Halts 562 uint halt_edge_cnt = TypeFunc::Parms; 563 RegMask *halt_rms = init_input_masks( halt_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 564 565 // Capture the return input masks into each exit flavor 566 for( i=1; i < root->req(); i++ ) { 567 MachReturnNode *exit = root->in(i)->as_MachReturn(); 568 switch( exit->ideal_Opcode() ) { 569 case Op_Return : exit->_in_rms = ret_rms; break; 570 case Op_Rethrow : exit->_in_rms = reth_rms; break; 571 case Op_TailCall : exit->_in_rms = tail_call_rms; break; 572 case Op_TailJump : exit->_in_rms = tail_jump_rms; break; 573 case Op_Halt : exit->_in_rms = halt_rms; break; 574 default : ShouldNotReachHere(); 575 } 576 } 577 578 // Next unused projection number from Start. 579 int proj_cnt = C->tf()->domain()->cnt(); 580 581 // Do all the save-on-entry registers. Make projections from Start for 582 // them, and give them a use at the exit points. To the allocator, they 583 // look like incoming register arguments. 584 for( i = 0; i < _last_Mach_Reg; i++ ) { 585 if( is_save_on_entry(i) ) { 586 587 // Add the save-on-entry to the mask array 588 ret_rms [ ret_edge_cnt] = mreg2regmask[i]; 589 reth_rms [ reth_edge_cnt] = mreg2regmask[i]; 590 tail_call_rms[tail_call_edge_cnt] = mreg2regmask[i]; 591 tail_jump_rms[tail_jump_edge_cnt] = mreg2regmask[i]; 592 // Halts need the SOE registers, but only in the stack as debug info. 593 // A just-prior uncommon-trap or deoptimization will use the SOE regs. 594 halt_rms [ halt_edge_cnt] = *idealreg2spillmask[_register_save_type[i]]; 595 596 Node *mproj; 597 598 // Is this a RegF low half of a RegD? Double up 2 adjacent RegF's 599 // into a single RegD. 600 if( (i&1) == 0 && 601 _register_save_type[i ] == Op_RegF && 602 _register_save_type[i+1] == Op_RegF && 603 is_save_on_entry(i+1) ) { 604 // Add other bit for double 605 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1)); 606 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1)); 607 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1)); 608 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1)); 609 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1)); 610 mproj = new (C, 1) MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegD ); 611 proj_cnt += 2; // Skip 2 for doubles 612 } 613 else if( (i&1) == 1 && // Else check for high half of double 614 _register_save_type[i-1] == Op_RegF && 615 _register_save_type[i ] == Op_RegF && 616 is_save_on_entry(i-1) ) { 617 ret_rms [ ret_edge_cnt] = RegMask::Empty; 618 reth_rms [ reth_edge_cnt] = RegMask::Empty; 619 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty; 620 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty; 621 halt_rms [ halt_edge_cnt] = RegMask::Empty; 622 mproj = C->top(); 623 } 624 // Is this a RegI low half of a RegL? Double up 2 adjacent RegI's 625 // into a single RegL. 626 else if( (i&1) == 0 && 627 _register_save_type[i ] == Op_RegI && 628 _register_save_type[i+1] == Op_RegI && 629 is_save_on_entry(i+1) ) { 630 // Add other bit for long 631 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1)); 632 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1)); 633 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1)); 634 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1)); 635 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1)); 636 mproj = new (C, 1) MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegL ); 637 proj_cnt += 2; // Skip 2 for longs 638 } 639 else if( (i&1) == 1 && // Else check for high half of long 640 _register_save_type[i-1] == Op_RegI && 641 _register_save_type[i ] == Op_RegI && 642 is_save_on_entry(i-1) ) { 643 ret_rms [ ret_edge_cnt] = RegMask::Empty; 644 reth_rms [ reth_edge_cnt] = RegMask::Empty; 645 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty; 646 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty; 647 halt_rms [ halt_edge_cnt] = RegMask::Empty; 648 mproj = C->top(); 649 } else { 650 // Make a projection for it off the Start 651 mproj = new (C, 1) MachProjNode( start, proj_cnt++, ret_rms[ret_edge_cnt], _register_save_type[i] ); 652 } 653 654 ret_edge_cnt ++; 655 reth_edge_cnt ++; 656 tail_call_edge_cnt ++; 657 tail_jump_edge_cnt ++; 658 halt_edge_cnt ++; 659 660 // Add a use of the SOE register to all exit paths 661 for( uint j=1; j < root->req(); j++ ) 662 root->in(j)->add_req(mproj); 663 } // End of if a save-on-entry register 664 } // End of for all machine registers 665 } 666 667 //------------------------------init_spill_mask-------------------------------- 668 void Matcher::init_spill_mask( Node *ret ) { 669 if( idealreg2regmask[Op_RegI] ) return; // One time only init 670 671 OptoReg::c_frame_pointer = c_frame_pointer(); 672 c_frame_ptr_mask = c_frame_pointer(); 673 #ifdef _LP64 674 // pointers are twice as big 675 c_frame_ptr_mask.Insert(OptoReg::add(c_frame_pointer(),1)); 676 #endif 677 678 // Start at OptoReg::stack0() 679 STACK_ONLY_mask.Clear(); 680 OptoReg::Name init = OptoReg::stack2reg(0); 681 // STACK_ONLY_mask is all stack bits 682 OptoReg::Name i; 683 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) 684 STACK_ONLY_mask.Insert(i); 685 // Also set the "infinite stack" bit. 686 STACK_ONLY_mask.set_AllStack(); 687 688 // Copy the register names over into the shared world 689 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) { 690 // SharedInfo::regName[i] = regName[i]; 691 // Handy RegMasks per machine register 692 mreg2regmask[i].Insert(i); 693 } 694 695 // Grab the Frame Pointer 696 Node *fp = ret->in(TypeFunc::FramePtr); 697 Node *mem = ret->in(TypeFunc::Memory); 698 const TypePtr* atp = TypePtr::BOTTOM; 699 // Share frame pointer while making spill ops 700 set_shared(fp); 701 702 // Compute generic short-offset Loads 703 #ifdef _LP64 704 MachNode *spillCP = match_tree(new (C, 3) LoadNNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM)); 705 #endif 706 MachNode *spillI = match_tree(new (C, 3) LoadINode(NULL,mem,fp,atp)); 707 MachNode *spillL = match_tree(new (C, 3) LoadLNode(NULL,mem,fp,atp)); 708 MachNode *spillF = match_tree(new (C, 3) LoadFNode(NULL,mem,fp,atp)); 709 MachNode *spillD = match_tree(new (C, 3) LoadDNode(NULL,mem,fp,atp)); 710 MachNode *spillP = match_tree(new (C, 3) LoadPNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM)); 711 assert(spillI != NULL && spillL != NULL && spillF != NULL && 712 spillD != NULL && spillP != NULL, ""); 713 714 // Get the ADLC notion of the right regmask, for each basic type. 715 #ifdef _LP64 716 idealreg2regmask[Op_RegN] = &spillCP->out_RegMask(); 717 #endif 718 idealreg2regmask[Op_RegI] = &spillI->out_RegMask(); 719 idealreg2regmask[Op_RegL] = &spillL->out_RegMask(); 720 idealreg2regmask[Op_RegF] = &spillF->out_RegMask(); 721 idealreg2regmask[Op_RegD] = &spillD->out_RegMask(); 722 idealreg2regmask[Op_RegP] = &spillP->out_RegMask(); 723 } 724 725 #ifdef ASSERT 726 static void match_alias_type(Compile* C, Node* n, Node* m) { 727 if (!VerifyAliases) return; // do not go looking for trouble by default 728 const TypePtr* nat = n->adr_type(); 729 const TypePtr* mat = m->adr_type(); 730 int nidx = C->get_alias_index(nat); 731 int midx = C->get_alias_index(mat); 732 // Detune the assert for cases like (AndI 0xFF (LoadB p)). 733 if (nidx == Compile::AliasIdxTop && midx >= Compile::AliasIdxRaw) { 734 for (uint i = 1; i < n->req(); i++) { 735 Node* n1 = n->in(i); 736 const TypePtr* n1at = n1->adr_type(); 737 if (n1at != NULL) { 738 nat = n1at; 739 nidx = C->get_alias_index(n1at); 740 } 741 } 742 } 743 // %%% Kludgery. Instead, fix ideal adr_type methods for all these cases: 744 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxRaw) { 745 switch (n->Opcode()) { 746 case Op_PrefetchRead: 747 case Op_PrefetchWrite: 748 nidx = Compile::AliasIdxRaw; 749 nat = TypeRawPtr::BOTTOM; 750 break; 751 } 752 } 753 if (nidx == Compile::AliasIdxRaw && midx == Compile::AliasIdxTop) { 754 switch (n->Opcode()) { 755 case Op_ClearArray: 756 midx = Compile::AliasIdxRaw; 757 mat = TypeRawPtr::BOTTOM; 758 break; 759 } 760 } 761 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxBot) { 762 switch (n->Opcode()) { 763 case Op_Return: 764 case Op_Rethrow: 765 case Op_Halt: 766 case Op_TailCall: 767 case Op_TailJump: 768 nidx = Compile::AliasIdxBot; 769 nat = TypePtr::BOTTOM; 770 break; 771 } 772 } 773 if (nidx == Compile::AliasIdxBot && midx == Compile::AliasIdxTop) { 774 switch (n->Opcode()) { 775 case Op_StrComp: 776 case Op_StrEquals: 777 case Op_StrIndexOf: 778 case Op_AryEq: 779 case Op_MemBarVolatile: 780 case Op_MemBarCPUOrder: // %%% these ideals should have narrower adr_type? 781 nidx = Compile::AliasIdxTop; 782 nat = NULL; 783 break; 784 } 785 } 786 if (nidx != midx) { 787 if (PrintOpto || (PrintMiscellaneous && (WizardMode || Verbose))) { 788 tty->print_cr("==== Matcher alias shift %d => %d", nidx, midx); 789 n->dump(); 790 m->dump(); 791 } 792 assert(C->subsume_loads() && C->must_alias(nat, midx), 793 "must not lose alias info when matching"); 794 } 795 } 796 #endif 797 798 799 //------------------------------MStack----------------------------------------- 800 // State and MStack class used in xform() and find_shared() iterative methods. 801 enum Node_State { Pre_Visit, // node has to be pre-visited 802 Visit, // visit node 803 Post_Visit, // post-visit node 804 Alt_Post_Visit // alternative post-visit path 805 }; 806 807 class MStack: public Node_Stack { 808 public: 809 MStack(int size) : Node_Stack(size) { } 810 811 void push(Node *n, Node_State ns) { 812 Node_Stack::push(n, (uint)ns); 813 } 814 void push(Node *n, Node_State ns, Node *parent, int indx) { 815 ++_inode_top; 816 if ((_inode_top + 1) >= _inode_max) grow(); 817 _inode_top->node = parent; 818 _inode_top->indx = (uint)indx; 819 ++_inode_top; 820 _inode_top->node = n; 821 _inode_top->indx = (uint)ns; 822 } 823 Node *parent() { 824 pop(); 825 return node(); 826 } 827 Node_State state() const { 828 return (Node_State)index(); 829 } 830 void set_state(Node_State ns) { 831 set_index((uint)ns); 832 } 833 }; 834 835 836 //------------------------------xform------------------------------------------ 837 // Given a Node in old-space, Match him (Label/Reduce) to produce a machine 838 // Node in new-space. Given a new-space Node, recursively walk his children. 839 Node *Matcher::transform( Node *n ) { ShouldNotCallThis(); return n; } 840 Node *Matcher::xform( Node *n, int max_stack ) { 841 // Use one stack to keep both: child's node/state and parent's node/index 842 MStack mstack(max_stack * 2 * 2); // C->unique() * 2 * 2 843 mstack.push(n, Visit, NULL, -1); // set NULL as parent to indicate root 844 845 while (mstack.is_nonempty()) { 846 n = mstack.node(); // Leave node on stack 847 Node_State nstate = mstack.state(); 848 if (nstate == Visit) { 849 mstack.set_state(Post_Visit); 850 Node *oldn = n; 851 // Old-space or new-space check 852 if (!C->node_arena()->contains(n)) { 853 // Old space! 854 Node* m; 855 if (has_new_node(n)) { // Not yet Label/Reduced 856 m = new_node(n); 857 } else { 858 if (!is_dontcare(n)) { // Matcher can match this guy 859 // Calls match special. They match alone with no children. 860 // Their children, the incoming arguments, match normally. 861 m = n->is_SafePoint() ? match_sfpt(n->as_SafePoint()):match_tree(n); 862 if (C->failing()) return NULL; 863 if (m == NULL) { Matcher::soft_match_failure(); return NULL; } 864 } else { // Nothing the matcher cares about 865 if( n->is_Proj() && n->in(0)->is_Multi()) { // Projections? 866 // Convert to machine-dependent projection 867 m = n->in(0)->as_Multi()->match( n->as_Proj(), this ); 868 #ifdef ASSERT 869 _new2old_map.map(m->_idx, n); 870 #endif 871 if (m->in(0) != NULL) // m might be top 872 collect_null_checks(m, n); 873 } else { // Else just a regular 'ol guy 874 m = n->clone(); // So just clone into new-space 875 #ifdef ASSERT 876 _new2old_map.map(m->_idx, n); 877 #endif 878 // Def-Use edges will be added incrementally as Uses 879 // of this node are matched. 880 assert(m->outcnt() == 0, "no Uses of this clone yet"); 881 } 882 } 883 884 set_new_node(n, m); // Map old to new 885 if (_old_node_note_array != NULL) { 886 Node_Notes* nn = C->locate_node_notes(_old_node_note_array, 887 n->_idx); 888 C->set_node_notes_at(m->_idx, nn); 889 } 890 debug_only(match_alias_type(C, n, m)); 891 } 892 n = m; // n is now a new-space node 893 mstack.set_node(n); 894 } 895 896 // New space! 897 if (_visited.test_set(n->_idx)) continue; // while(mstack.is_nonempty()) 898 899 int i; 900 // Put precedence edges on stack first (match them last). 901 for (i = oldn->req(); (uint)i < oldn->len(); i++) { 902 Node *m = oldn->in(i); 903 if (m == NULL) break; 904 // set -1 to call add_prec() instead of set_req() during Step1 905 mstack.push(m, Visit, n, -1); 906 } 907 908 // For constant debug info, I'd rather have unmatched constants. 909 int cnt = n->req(); 910 JVMState* jvms = n->jvms(); 911 int debug_cnt = jvms ? jvms->debug_start() : cnt; 912 913 // Now do only debug info. Clone constants rather than matching. 914 // Constants are represented directly in the debug info without 915 // the need for executable machine instructions. 916 // Monitor boxes are also represented directly. 917 for (i = cnt - 1; i >= debug_cnt; --i) { // For all debug inputs do 918 Node *m = n->in(i); // Get input 919 int op = m->Opcode(); 920 assert((op == Op_BoxLock) == jvms->is_monitor_use(i), "boxes only at monitor sites"); 921 if( op == Op_ConI || op == Op_ConP || op == Op_ConN || 922 op == Op_ConF || op == Op_ConD || op == Op_ConL 923 // || op == Op_BoxLock // %%%% enable this and remove (+++) in chaitin.cpp 924 ) { 925 m = m->clone(); 926 #ifdef ASSERT 927 _new2old_map.map(m->_idx, n); 928 #endif 929 mstack.push(m, Post_Visit, n, i); // Don't need to visit 930 mstack.push(m->in(0), Visit, m, 0); 931 } else { 932 mstack.push(m, Visit, n, i); 933 } 934 } 935 936 // And now walk his children, and convert his inputs to new-space. 937 for( ; i >= 0; --i ) { // For all normal inputs do 938 Node *m = n->in(i); // Get input 939 if(m != NULL) 940 mstack.push(m, Visit, n, i); 941 } 942 943 } 944 else if (nstate == Post_Visit) { 945 // Set xformed input 946 Node *p = mstack.parent(); 947 if (p != NULL) { // root doesn't have parent 948 int i = (int)mstack.index(); 949 if (i >= 0) 950 p->set_req(i, n); // required input 951 else if (i == -1) 952 p->add_prec(n); // precedence input 953 else 954 ShouldNotReachHere(); 955 } 956 mstack.pop(); // remove processed node from stack 957 } 958 else { 959 ShouldNotReachHere(); 960 } 961 } // while (mstack.is_nonempty()) 962 return n; // Return new-space Node 963 } 964 965 //------------------------------warp_outgoing_stk_arg------------------------ 966 OptoReg::Name Matcher::warp_outgoing_stk_arg( VMReg reg, OptoReg::Name begin_out_arg_area, OptoReg::Name &out_arg_limit_per_call ) { 967 // Convert outgoing argument location to a pre-biased stack offset 968 if (reg->is_stack()) { 969 OptoReg::Name warped = reg->reg2stack(); 970 // Adjust the stack slot offset to be the register number used 971 // by the allocator. 972 warped = OptoReg::add(begin_out_arg_area, warped); 973 // Keep track of the largest numbered stack slot used for an arg. 974 // Largest used slot per call-site indicates the amount of stack 975 // that is killed by the call. 976 if( warped >= out_arg_limit_per_call ) 977 out_arg_limit_per_call = OptoReg::add(warped,1); 978 if (!RegMask::can_represent(warped)) { 979 C->record_method_not_compilable_all_tiers("unsupported calling sequence"); 980 return OptoReg::Bad; 981 } 982 return warped; 983 } 984 return OptoReg::as_OptoReg(reg); 985 } 986 987 988 //------------------------------match_sfpt------------------------------------- 989 // Helper function to match call instructions. Calls match special. 990 // They match alone with no children. Their children, the incoming 991 // arguments, match normally. 992 MachNode *Matcher::match_sfpt( SafePointNode *sfpt ) { 993 MachSafePointNode *msfpt = NULL; 994 MachCallNode *mcall = NULL; 995 uint cnt; 996 // Split out case for SafePoint vs Call 997 CallNode *call; 998 const TypeTuple *domain; 999 ciMethod* method = NULL; 1000 if( sfpt->is_Call() ) { 1001 call = sfpt->as_Call(); 1002 domain = call->tf()->domain(); 1003 cnt = domain->cnt(); 1004 1005 // Match just the call, nothing else 1006 MachNode *m = match_tree(call); 1007 if (C->failing()) return NULL; 1008 if( m == NULL ) { Matcher::soft_match_failure(); return NULL; } 1009 1010 // Copy data from the Ideal SafePoint to the machine version 1011 mcall = m->as_MachCall(); 1012 1013 mcall->set_tf( call->tf()); 1014 mcall->set_entry_point(call->entry_point()); 1015 mcall->set_cnt( call->cnt()); 1016 1017 if( mcall->is_MachCallJava() ) { 1018 MachCallJavaNode *mcall_java = mcall->as_MachCallJava(); 1019 const CallJavaNode *call_java = call->as_CallJava(); 1020 method = call_java->method(); 1021 mcall_java->_method = method; 1022 mcall_java->_bci = call_java->_bci; 1023 mcall_java->_optimized_virtual = call_java->is_optimized_virtual(); 1024 if( mcall_java->is_MachCallStaticJava() ) 1025 mcall_java->as_MachCallStaticJava()->_name = 1026 call_java->as_CallStaticJava()->_name; 1027 if( mcall_java->is_MachCallDynamicJava() ) 1028 mcall_java->as_MachCallDynamicJava()->_vtable_index = 1029 call_java->as_CallDynamicJava()->_vtable_index; 1030 } 1031 else if( mcall->is_MachCallRuntime() ) { 1032 mcall->as_MachCallRuntime()->_name = call->as_CallRuntime()->_name; 1033 } 1034 msfpt = mcall; 1035 } 1036 // This is a non-call safepoint 1037 else { 1038 call = NULL; 1039 domain = NULL; 1040 MachNode *mn = match_tree(sfpt); 1041 if (C->failing()) return NULL; 1042 msfpt = mn->as_MachSafePoint(); 1043 cnt = TypeFunc::Parms; 1044 } 1045 1046 // Advertise the correct memory effects (for anti-dependence computation). 1047 msfpt->set_adr_type(sfpt->adr_type()); 1048 1049 // Allocate a private array of RegMasks. These RegMasks are not shared. 1050 msfpt->_in_rms = NEW_RESOURCE_ARRAY( RegMask, cnt ); 1051 // Empty them all. 1052 memset( msfpt->_in_rms, 0, sizeof(RegMask)*cnt ); 1053 1054 // Do all the pre-defined non-Empty register masks 1055 msfpt->_in_rms[TypeFunc::ReturnAdr] = _return_addr_mask; 1056 msfpt->_in_rms[TypeFunc::FramePtr ] = c_frame_ptr_mask; 1057 1058 // Place first outgoing argument can possibly be put. 1059 OptoReg::Name begin_out_arg_area = OptoReg::add(_new_SP, C->out_preserve_stack_slots()); 1060 assert( is_even(begin_out_arg_area), "" ); 1061 // Compute max outgoing register number per call site. 1062 OptoReg::Name out_arg_limit_per_call = begin_out_arg_area; 1063 // Calls to C may hammer extra stack slots above and beyond any arguments. 1064 // These are usually backing store for register arguments for varargs. 1065 if( call != NULL && call->is_CallRuntime() ) 1066 out_arg_limit_per_call = OptoReg::add(out_arg_limit_per_call,C->varargs_C_out_slots_killed()); 1067 1068 1069 // Do the normal argument list (parameters) register masks 1070 int argcnt = cnt - TypeFunc::Parms; 1071 if( argcnt > 0 ) { // Skip it all if we have no args 1072 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt ); 1073 VMRegPair *parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt ); 1074 int i; 1075 for( i = 0; i < argcnt; i++ ) { 1076 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type(); 1077 } 1078 // V-call to pick proper calling convention 1079 call->calling_convention( sig_bt, parm_regs, argcnt ); 1080 1081 #ifdef ASSERT 1082 // Sanity check users' calling convention. Really handy during 1083 // the initial porting effort. Fairly expensive otherwise. 1084 { for (int i = 0; i<argcnt; i++) { 1085 if( !parm_regs[i].first()->is_valid() && 1086 !parm_regs[i].second()->is_valid() ) continue; 1087 VMReg reg1 = parm_regs[i].first(); 1088 VMReg reg2 = parm_regs[i].second(); 1089 for (int j = 0; j < i; j++) { 1090 if( !parm_regs[j].first()->is_valid() && 1091 !parm_regs[j].second()->is_valid() ) continue; 1092 VMReg reg3 = parm_regs[j].first(); 1093 VMReg reg4 = parm_regs[j].second(); 1094 if( !reg1->is_valid() ) { 1095 assert( !reg2->is_valid(), "valid halvsies" ); 1096 } else if( !reg3->is_valid() ) { 1097 assert( !reg4->is_valid(), "valid halvsies" ); 1098 } else { 1099 assert( reg1 != reg2, "calling conv. must produce distinct regs"); 1100 assert( reg1 != reg3, "calling conv. must produce distinct regs"); 1101 assert( reg1 != reg4, "calling conv. must produce distinct regs"); 1102 assert( reg2 != reg3, "calling conv. must produce distinct regs"); 1103 assert( reg2 != reg4 || !reg2->is_valid(), "calling conv. must produce distinct regs"); 1104 assert( reg3 != reg4, "calling conv. must produce distinct regs"); 1105 } 1106 } 1107 } 1108 } 1109 #endif 1110 1111 // Visit each argument. Compute its outgoing register mask. 1112 // Return results now can have 2 bits returned. 1113 // Compute max over all outgoing arguments both per call-site 1114 // and over the entire method. 1115 for( i = 0; i < argcnt; i++ ) { 1116 // Address of incoming argument mask to fill in 1117 RegMask *rm = &mcall->_in_rms[i+TypeFunc::Parms]; 1118 if( !parm_regs[i].first()->is_valid() && 1119 !parm_regs[i].second()->is_valid() ) { 1120 continue; // Avoid Halves 1121 } 1122 // Grab first register, adjust stack slots and insert in mask. 1123 OptoReg::Name reg1 = warp_outgoing_stk_arg(parm_regs[i].first(), begin_out_arg_area, out_arg_limit_per_call ); 1124 if (OptoReg::is_valid(reg1)) 1125 rm->Insert( reg1 ); 1126 // Grab second register (if any), adjust stack slots and insert in mask. 1127 OptoReg::Name reg2 = warp_outgoing_stk_arg(parm_regs[i].second(), begin_out_arg_area, out_arg_limit_per_call ); 1128 if (OptoReg::is_valid(reg2)) 1129 rm->Insert( reg2 ); 1130 } // End of for all arguments 1131 1132 // Compute number of stack slots needed to restore stack in case of 1133 // Pascal-style argument popping. 1134 mcall->_argsize = out_arg_limit_per_call - begin_out_arg_area; 1135 } 1136 1137 // Compute the max stack slot killed by any call. These will not be 1138 // available for debug info, and will be used to adjust FIRST_STACK_mask 1139 // after all call sites have been visited. 1140 if( _out_arg_limit < out_arg_limit_per_call) 1141 _out_arg_limit = out_arg_limit_per_call; 1142 1143 if (mcall) { 1144 // Kill the outgoing argument area, including any non-argument holes and 1145 // any legacy C-killed slots. Use Fat-Projections to do the killing. 1146 // Since the max-per-method covers the max-per-call-site and debug info 1147 // is excluded on the max-per-method basis, debug info cannot land in 1148 // this killed area. 1149 uint r_cnt = mcall->tf()->range()->cnt(); 1150 MachProjNode *proj = new (C, 1) MachProjNode( mcall, r_cnt+10000, RegMask::Empty, MachProjNode::fat_proj ); 1151 if (!RegMask::can_represent(OptoReg::Name(out_arg_limit_per_call-1))) { 1152 C->record_method_not_compilable_all_tiers("unsupported outgoing calling sequence"); 1153 } else { 1154 for (int i = begin_out_arg_area; i < out_arg_limit_per_call; i++) 1155 proj->_rout.Insert(OptoReg::Name(i)); 1156 } 1157 if( proj->_rout.is_NotEmpty() ) 1158 _proj_list.push(proj); 1159 } 1160 // Transfer the safepoint information from the call to the mcall 1161 // Move the JVMState list 1162 msfpt->set_jvms(sfpt->jvms()); 1163 for (JVMState* jvms = msfpt->jvms(); jvms; jvms = jvms->caller()) { 1164 jvms->set_map(sfpt); 1165 } 1166 1167 // Debug inputs begin just after the last incoming parameter 1168 assert( (mcall == NULL) || (mcall->jvms() == NULL) || 1169 (mcall->jvms()->debug_start() + mcall->_jvmadj == mcall->tf()->domain()->cnt()), "" ); 1170 1171 // Move the OopMap 1172 msfpt->_oop_map = sfpt->_oop_map; 1173 1174 // Registers killed by the call are set in the local scheduling pass 1175 // of Global Code Motion. 1176 return msfpt; 1177 } 1178 1179 //---------------------------match_tree---------------------------------------- 1180 // Match a Ideal Node DAG - turn it into a tree; Label & Reduce. Used as part 1181 // of the whole-sale conversion from Ideal to Mach Nodes. Also used for 1182 // making GotoNodes while building the CFG and in init_spill_mask() to identify 1183 // a Load's result RegMask for memoization in idealreg2regmask[] 1184 MachNode *Matcher::match_tree( const Node *n ) { 1185 assert( n->Opcode() != Op_Phi, "cannot match" ); 1186 assert( !n->is_block_start(), "cannot match" ); 1187 // Set the mark for all locally allocated State objects. 1188 // When this call returns, the _states_arena arena will be reset 1189 // freeing all State objects. 1190 ResourceMark rm( &_states_arena ); 1191 1192 LabelRootDepth = 0; 1193 1194 // StoreNodes require their Memory input to match any LoadNodes 1195 Node *mem = n->is_Store() ? n->in(MemNode::Memory) : (Node*)1 ; 1196 #ifdef ASSERT 1197 Node* save_mem_node = _mem_node; 1198 _mem_node = n->is_Store() ? (Node*)n : NULL; 1199 #endif 1200 // State object for root node of match tree 1201 // Allocate it on _states_arena - stack allocation can cause stack overflow. 1202 State *s = new (&_states_arena) State; 1203 s->_kids[0] = NULL; 1204 s->_kids[1] = NULL; 1205 s->_leaf = (Node*)n; 1206 // Label the input tree, allocating labels from top-level arena 1207 Label_Root( n, s, n->in(0), mem ); 1208 if (C->failing()) return NULL; 1209 1210 // The minimum cost match for the whole tree is found at the root State 1211 uint mincost = max_juint; 1212 uint cost = max_juint; 1213 uint i; 1214 for( i = 0; i < NUM_OPERANDS; i++ ) { 1215 if( s->valid(i) && // valid entry and 1216 s->_cost[i] < cost && // low cost and 1217 s->_rule[i] >= NUM_OPERANDS ) // not an operand 1218 cost = s->_cost[mincost=i]; 1219 } 1220 if (mincost == max_juint) { 1221 #ifndef PRODUCT 1222 tty->print("No matching rule for:"); 1223 s->dump(); 1224 #endif 1225 Matcher::soft_match_failure(); 1226 return NULL; 1227 } 1228 // Reduce input tree based upon the state labels to machine Nodes 1229 MachNode *m = ReduceInst( s, s->_rule[mincost], mem ); 1230 #ifdef ASSERT 1231 _old2new_map.map(n->_idx, m); 1232 _new2old_map.map(m->_idx, (Node*)n); 1233 #endif 1234 1235 // Add any Matcher-ignored edges 1236 uint cnt = n->req(); 1237 uint start = 1; 1238 if( mem != (Node*)1 ) start = MemNode::Memory+1; 1239 if( n->is_AddP() ) { 1240 assert( mem == (Node*)1, "" ); 1241 start = AddPNode::Base+1; 1242 } 1243 for( i = start; i < cnt; i++ ) { 1244 if( !n->match_edge(i) ) { 1245 if( i < m->req() ) 1246 m->ins_req( i, n->in(i) ); 1247 else 1248 m->add_req( n->in(i) ); 1249 } 1250 } 1251 1252 debug_only( _mem_node = save_mem_node; ) 1253 return m; 1254 } 1255 1256 1257 //------------------------------match_into_reg--------------------------------- 1258 // Choose to either match this Node in a register or part of the current 1259 // match tree. Return true for requiring a register and false for matching 1260 // as part of the current match tree. 1261 static bool match_into_reg( const Node *n, Node *m, Node *control, int i, bool shared ) { 1262 1263 const Type *t = m->bottom_type(); 1264 1265 if( t->singleton() ) { 1266 // Never force constants into registers. Allow them to match as 1267 // constants or registers. Copies of the same value will share 1268 // the same register. See find_shared_node. 1269 return false; 1270 } else { // Not a constant 1271 // Stop recursion if they have different Controls. 1272 // Slot 0 of constants is not really a Control. 1273 if( control && m->in(0) && control != m->in(0) ) { 1274 1275 // Actually, we can live with the most conservative control we 1276 // find, if it post-dominates the others. This allows us to 1277 // pick up load/op/store trees where the load can float a little 1278 // above the store. 1279 Node *x = control; 1280 const uint max_scan = 6; // Arbitrary scan cutoff 1281 uint j; 1282 for( j=0; j<max_scan; j++ ) { 1283 if( x->is_Region() ) // Bail out at merge points 1284 return true; 1285 x = x->in(0); 1286 if( x == m->in(0) ) // Does 'control' post-dominate 1287 break; // m->in(0)? If so, we can use it 1288 } 1289 if( j == max_scan ) // No post-domination before scan end? 1290 return true; // Then break the match tree up 1291 } 1292 if (m->is_DecodeN() && Matcher::clone_shift_expressions) { 1293 // These are commonly used in address expressions and can 1294 // efficiently fold into them on X64 in some cases. 1295 return false; 1296 } 1297 } 1298 1299 // Not forceable cloning. If shared, put it into a register. 1300 return shared; 1301 } 1302 1303 1304 //------------------------------Instruction Selection-------------------------- 1305 // Label method walks a "tree" of nodes, using the ADLC generated DFA to match 1306 // ideal nodes to machine instructions. Trees are delimited by shared Nodes, 1307 // things the Matcher does not match (e.g., Memory), and things with different 1308 // Controls (hence forced into different blocks). We pass in the Control 1309 // selected for this entire State tree. 1310 1311 // The Matcher works on Trees, but an Intel add-to-memory requires a DAG: the 1312 // Store and the Load must have identical Memories (as well as identical 1313 // pointers). Since the Matcher does not have anything for Memory (and 1314 // does not handle DAGs), I have to match the Memory input myself. If the 1315 // Tree root is a Store, I require all Loads to have the identical memory. 1316 Node *Matcher::Label_Root( const Node *n, State *svec, Node *control, const Node *mem){ 1317 // Since Label_Root is a recursive function, its possible that we might run 1318 // out of stack space. See bugs 6272980 & 6227033 for more info. 1319 LabelRootDepth++; 1320 if (LabelRootDepth > MaxLabelRootDepth) { 1321 C->record_method_not_compilable_all_tiers("Out of stack space, increase MaxLabelRootDepth"); 1322 return NULL; 1323 } 1324 uint care = 0; // Edges matcher cares about 1325 uint cnt = n->req(); 1326 uint i = 0; 1327 1328 // Examine children for memory state 1329 // Can only subsume a child into your match-tree if that child's memory state 1330 // is not modified along the path to another input. 1331 // It is unsafe even if the other inputs are separate roots. 1332 Node *input_mem = NULL; 1333 for( i = 1; i < cnt; i++ ) { 1334 if( !n->match_edge(i) ) continue; 1335 Node *m = n->in(i); // Get ith input 1336 assert( m, "expect non-null children" ); 1337 if( m->is_Load() ) { 1338 if( input_mem == NULL ) { 1339 input_mem = m->in(MemNode::Memory); 1340 } else if( input_mem != m->in(MemNode::Memory) ) { 1341 input_mem = NodeSentinel; 1342 } 1343 } 1344 } 1345 1346 for( i = 1; i < cnt; i++ ){// For my children 1347 if( !n->match_edge(i) ) continue; 1348 Node *m = n->in(i); // Get ith input 1349 // Allocate states out of a private arena 1350 State *s = new (&_states_arena) State; 1351 svec->_kids[care++] = s; 1352 assert( care <= 2, "binary only for now" ); 1353 1354 // Recursively label the State tree. 1355 s->_kids[0] = NULL; 1356 s->_kids[1] = NULL; 1357 s->_leaf = m; 1358 1359 // Check for leaves of the State Tree; things that cannot be a part of 1360 // the current tree. If it finds any, that value is matched as a 1361 // register operand. If not, then the normal matching is used. 1362 if( match_into_reg(n, m, control, i, is_shared(m)) || 1363 // 1364 // Stop recursion if this is LoadNode and the root of this tree is a 1365 // StoreNode and the load & store have different memories. 1366 ((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) || 1367 // Can NOT include the match of a subtree when its memory state 1368 // is used by any of the other subtrees 1369 (input_mem == NodeSentinel) ) { 1370 #ifndef PRODUCT 1371 // Print when we exclude matching due to different memory states at input-loads 1372 if( PrintOpto && (Verbose && WizardMode) && (input_mem == NodeSentinel) 1373 && !((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) ) { 1374 tty->print_cr("invalid input_mem"); 1375 } 1376 #endif 1377 // Switch to a register-only opcode; this value must be in a register 1378 // and cannot be subsumed as part of a larger instruction. 1379 s->DFA( m->ideal_reg(), m ); 1380 1381 } else { 1382 // If match tree has no control and we do, adopt it for entire tree 1383 if( control == NULL && m->in(0) != NULL && m->req() > 1 ) 1384 control = m->in(0); // Pick up control 1385 // Else match as a normal part of the match tree. 1386 control = Label_Root(m,s,control,mem); 1387 if (C->failing()) return NULL; 1388 } 1389 } 1390 1391 1392 // Call DFA to match this node, and return 1393 svec->DFA( n->Opcode(), n ); 1394 1395 #ifdef ASSERT 1396 uint x; 1397 for( x = 0; x < _LAST_MACH_OPER; x++ ) 1398 if( svec->valid(x) ) 1399 break; 1400 1401 if (x >= _LAST_MACH_OPER) { 1402 n->dump(); 1403 svec->dump(); 1404 assert( false, "bad AD file" ); 1405 } 1406 #endif 1407 return control; 1408 } 1409 1410 1411 // Con nodes reduced using the same rule can share their MachNode 1412 // which reduces the number of copies of a constant in the final 1413 // program. The register allocator is free to split uses later to 1414 // split live ranges. 1415 MachNode* Matcher::find_shared_node(Node* leaf, uint rule) { 1416 if (!leaf->is_Con() && !leaf->is_DecodeN()) return NULL; 1417 1418 // See if this Con has already been reduced using this rule. 1419 if (_shared_nodes.Size() <= leaf->_idx) return NULL; 1420 MachNode* last = (MachNode*)_shared_nodes.at(leaf->_idx); 1421 if (last != NULL && rule == last->rule()) { 1422 // Don't expect control change for DecodeN 1423 if (leaf->is_DecodeN()) 1424 return last; 1425 // Get the new space root. 1426 Node* xroot = new_node(C->root()); 1427 if (xroot == NULL) { 1428 // This shouldn't happen give the order of matching. 1429 return NULL; 1430 } 1431 1432 // Shared constants need to have their control be root so they 1433 // can be scheduled properly. 1434 Node* control = last->in(0); 1435 if (control != xroot) { 1436 if (control == NULL || control == C->root()) { 1437 last->set_req(0, xroot); 1438 } else { 1439 assert(false, "unexpected control"); 1440 return NULL; 1441 } 1442 } 1443 return last; 1444 } 1445 return NULL; 1446 } 1447 1448 1449 //------------------------------ReduceInst------------------------------------- 1450 // Reduce a State tree (with given Control) into a tree of MachNodes. 1451 // This routine (and it's cohort ReduceOper) convert Ideal Nodes into 1452 // complicated machine Nodes. Each MachNode covers some tree of Ideal Nodes. 1453 // Each MachNode has a number of complicated MachOper operands; each 1454 // MachOper also covers a further tree of Ideal Nodes. 1455 1456 // The root of the Ideal match tree is always an instruction, so we enter 1457 // the recursion here. After building the MachNode, we need to recurse 1458 // the tree checking for these cases: 1459 // (1) Child is an instruction - 1460 // Build the instruction (recursively), add it as an edge. 1461 // Build a simple operand (register) to hold the result of the instruction. 1462 // (2) Child is an interior part of an instruction - 1463 // Skip over it (do nothing) 1464 // (3) Child is the start of a operand - 1465 // Build the operand, place it inside the instruction 1466 // Call ReduceOper. 1467 MachNode *Matcher::ReduceInst( State *s, int rule, Node *&mem ) { 1468 assert( rule >= NUM_OPERANDS, "called with operand rule" ); 1469 1470 MachNode* shared_node = find_shared_node(s->_leaf, rule); 1471 if (shared_node != NULL) { 1472 return shared_node; 1473 } 1474 1475 // Build the object to represent this state & prepare for recursive calls 1476 MachNode *mach = s->MachNodeGenerator( rule, C ); 1477 mach->_opnds[0] = s->MachOperGenerator( _reduceOp[rule], C ); 1478 assert( mach->_opnds[0] != NULL, "Missing result operand" ); 1479 Node *leaf = s->_leaf; 1480 // Check for instruction or instruction chain rule 1481 if( rule >= _END_INST_CHAIN_RULE || rule < _BEGIN_INST_CHAIN_RULE ) { 1482 assert(C->node_arena()->contains(s->_leaf) || !has_new_node(s->_leaf), 1483 "duplicating node that's already been matched"); 1484 // Instruction 1485 mach->add_req( leaf->in(0) ); // Set initial control 1486 // Reduce interior of complex instruction 1487 ReduceInst_Interior( s, rule, mem, mach, 1 ); 1488 } else { 1489 // Instruction chain rules are data-dependent on their inputs 1490 mach->add_req(0); // Set initial control to none 1491 ReduceInst_Chain_Rule( s, rule, mem, mach ); 1492 } 1493 1494 // If a Memory was used, insert a Memory edge 1495 if( mem != (Node*)1 ) { 1496 mach->ins_req(MemNode::Memory,mem); 1497 #ifdef ASSERT 1498 // Verify adr type after matching memory operation 1499 const MachOper* oper = mach->memory_operand(); 1500 if (oper != NULL && oper != (MachOper*)-1 && 1501 mach->adr_type() != TypeRawPtr::BOTTOM) { // non-direct addressing mode 1502 // It has a unique memory operand. Find corresponding ideal mem node. 1503 Node* m = NULL; 1504 if (leaf->is_Mem()) { 1505 m = leaf; 1506 } else { 1507 m = _mem_node; 1508 assert(m != NULL && m->is_Mem(), "expecting memory node"); 1509 } 1510 const Type* mach_at = mach->adr_type(); 1511 // DecodeN node consumed by an address may have different type 1512 // then its input. Don't compare types for such case. 1513 if (m->adr_type() != mach_at && 1514 (m->in(MemNode::Address)->is_DecodeN() || 1515 m->in(MemNode::Address)->is_AddP() && 1516 m->in(MemNode::Address)->in(AddPNode::Address)->is_DecodeN() || 1517 m->in(MemNode::Address)->is_AddP() && 1518 m->in(MemNode::Address)->in(AddPNode::Address)->is_AddP() && 1519 m->in(MemNode::Address)->in(AddPNode::Address)->in(AddPNode::Address)->is_DecodeN())) { 1520 mach_at = m->adr_type(); 1521 } 1522 if (m->adr_type() != mach_at) { 1523 m->dump(); 1524 tty->print_cr("mach:"); 1525 mach->dump(1); 1526 } 1527 assert(m->adr_type() == mach_at, "matcher should not change adr type"); 1528 } 1529 #endif 1530 } 1531 1532 // If the _leaf is an AddP, insert the base edge 1533 if( leaf->is_AddP() ) 1534 mach->ins_req(AddPNode::Base,leaf->in(AddPNode::Base)); 1535 1536 uint num_proj = _proj_list.size(); 1537 1538 // Perform any 1-to-many expansions required 1539 MachNode *ex = mach->Expand(s,_proj_list); 1540 if( ex != mach ) { 1541 assert(ex->ideal_reg() == mach->ideal_reg(), "ideal types should match"); 1542 if( ex->in(1)->is_Con() ) 1543 ex->in(1)->set_req(0, C->root()); 1544 // Remove old node from the graph 1545 for( uint i=0; i<mach->req(); i++ ) { 1546 mach->set_req(i,NULL); 1547 } 1548 #ifdef ASSERT 1549 _new2old_map.map(ex->_idx, s->_leaf); 1550 #endif 1551 } 1552 1553 // PhaseChaitin::fixup_spills will sometimes generate spill code 1554 // via the matcher. By the time, nodes have been wired into the CFG, 1555 // and any further nodes generated by expand rules will be left hanging 1556 // in space, and will not get emitted as output code. Catch this. 1557 // Also, catch any new register allocation constraints ("projections") 1558 // generated belatedly during spill code generation. 1559 if (_allocation_started) { 1560 guarantee(ex == mach, "no expand rules during spill generation"); 1561 guarantee(_proj_list.size() == num_proj, "no allocation during spill generation"); 1562 } 1563 1564 if (leaf->is_Con() || leaf->is_DecodeN()) { 1565 // Record the con for sharing 1566 _shared_nodes.map(leaf->_idx, ex); 1567 } 1568 1569 return ex; 1570 } 1571 1572 void Matcher::ReduceInst_Chain_Rule( State *s, int rule, Node *&mem, MachNode *mach ) { 1573 // 'op' is what I am expecting to receive 1574 int op = _leftOp[rule]; 1575 // Operand type to catch childs result 1576 // This is what my child will give me. 1577 int opnd_class_instance = s->_rule[op]; 1578 // Choose between operand class or not. 1579 // This is what I will receive. 1580 int catch_op = (FIRST_OPERAND_CLASS <= op && op < NUM_OPERANDS) ? opnd_class_instance : op; 1581 // New rule for child. Chase operand classes to get the actual rule. 1582 int newrule = s->_rule[catch_op]; 1583 1584 if( newrule < NUM_OPERANDS ) { 1585 // Chain from operand or operand class, may be output of shared node 1586 assert( 0 <= opnd_class_instance && opnd_class_instance < NUM_OPERANDS, 1587 "Bad AD file: Instruction chain rule must chain from operand"); 1588 // Insert operand into array of operands for this instruction 1589 mach->_opnds[1] = s->MachOperGenerator( opnd_class_instance, C ); 1590 1591 ReduceOper( s, newrule, mem, mach ); 1592 } else { 1593 // Chain from the result of an instruction 1594 assert( newrule >= _LAST_MACH_OPER, "Do NOT chain from internal operand"); 1595 mach->_opnds[1] = s->MachOperGenerator( _reduceOp[catch_op], C ); 1596 Node *mem1 = (Node*)1; 1597 debug_only(Node *save_mem_node = _mem_node;) 1598 mach->add_req( ReduceInst(s, newrule, mem1) ); 1599 debug_only(_mem_node = save_mem_node;) 1600 } 1601 return; 1602 } 1603 1604 1605 uint Matcher::ReduceInst_Interior( State *s, int rule, Node *&mem, MachNode *mach, uint num_opnds ) { 1606 if( s->_leaf->is_Load() ) { 1607 Node *mem2 = s->_leaf->in(MemNode::Memory); 1608 assert( mem == (Node*)1 || mem == mem2, "multiple Memories being matched at once?" ); 1609 debug_only( if( mem == (Node*)1 ) _mem_node = s->_leaf;) 1610 mem = mem2; 1611 } 1612 if( s->_leaf->in(0) != NULL && s->_leaf->req() > 1) { 1613 if( mach->in(0) == NULL ) 1614 mach->set_req(0, s->_leaf->in(0)); 1615 } 1616 1617 // Now recursively walk the state tree & add operand list. 1618 for( uint i=0; i<2; i++ ) { // binary tree 1619 State *newstate = s->_kids[i]; 1620 if( newstate == NULL ) break; // Might only have 1 child 1621 // 'op' is what I am expecting to receive 1622 int op; 1623 if( i == 0 ) { 1624 op = _leftOp[rule]; 1625 } else { 1626 op = _rightOp[rule]; 1627 } 1628 // Operand type to catch childs result 1629 // This is what my child will give me. 1630 int opnd_class_instance = newstate->_rule[op]; 1631 // Choose between operand class or not. 1632 // This is what I will receive. 1633 int catch_op = (op >= FIRST_OPERAND_CLASS && op < NUM_OPERANDS) ? opnd_class_instance : op; 1634 // New rule for child. Chase operand classes to get the actual rule. 1635 int newrule = newstate->_rule[catch_op]; 1636 1637 if( newrule < NUM_OPERANDS ) { // Operand/operandClass or internalOp/instruction? 1638 // Operand/operandClass 1639 // Insert operand into array of operands for this instruction 1640 mach->_opnds[num_opnds++] = newstate->MachOperGenerator( opnd_class_instance, C ); 1641 ReduceOper( newstate, newrule, mem, mach ); 1642 1643 } else { // Child is internal operand or new instruction 1644 if( newrule < _LAST_MACH_OPER ) { // internal operand or instruction? 1645 // internal operand --> call ReduceInst_Interior 1646 // Interior of complex instruction. Do nothing but recurse. 1647 num_opnds = ReduceInst_Interior( newstate, newrule, mem, mach, num_opnds ); 1648 } else { 1649 // instruction --> call build operand( ) to catch result 1650 // --> ReduceInst( newrule ) 1651 mach->_opnds[num_opnds++] = s->MachOperGenerator( _reduceOp[catch_op], C ); 1652 Node *mem1 = (Node*)1; 1653 debug_only(Node *save_mem_node = _mem_node;) 1654 mach->add_req( ReduceInst( newstate, newrule, mem1 ) ); 1655 debug_only(_mem_node = save_mem_node;) 1656 } 1657 } 1658 assert( mach->_opnds[num_opnds-1], "" ); 1659 } 1660 return num_opnds; 1661 } 1662 1663 // This routine walks the interior of possible complex operands. 1664 // At each point we check our children in the match tree: 1665 // (1) No children - 1666 // We are a leaf; add _leaf field as an input to the MachNode 1667 // (2) Child is an internal operand - 1668 // Skip over it ( do nothing ) 1669 // (3) Child is an instruction - 1670 // Call ReduceInst recursively and 1671 // and instruction as an input to the MachNode 1672 void Matcher::ReduceOper( State *s, int rule, Node *&mem, MachNode *mach ) { 1673 assert( rule < _LAST_MACH_OPER, "called with operand rule" ); 1674 State *kid = s->_kids[0]; 1675 assert( kid == NULL || s->_leaf->in(0) == NULL, "internal operands have no control" ); 1676 1677 // Leaf? And not subsumed? 1678 if( kid == NULL && !_swallowed[rule] ) { 1679 mach->add_req( s->_leaf ); // Add leaf pointer 1680 return; // Bail out 1681 } 1682 1683 if( s->_leaf->is_Load() ) { 1684 assert( mem == (Node*)1, "multiple Memories being matched at once?" ); 1685 mem = s->_leaf->in(MemNode::Memory); 1686 debug_only(_mem_node = s->_leaf;) 1687 } 1688 if( s->_leaf->in(0) && s->_leaf->req() > 1) { 1689 if( !mach->in(0) ) 1690 mach->set_req(0,s->_leaf->in(0)); 1691 else { 1692 assert( s->_leaf->in(0) == mach->in(0), "same instruction, differing controls?" ); 1693 } 1694 } 1695 1696 for( uint i=0; kid != NULL && i<2; kid = s->_kids[1], i++ ) { // binary tree 1697 int newrule; 1698 if( i == 0 ) 1699 newrule = kid->_rule[_leftOp[rule]]; 1700 else 1701 newrule = kid->_rule[_rightOp[rule]]; 1702 1703 if( newrule < _LAST_MACH_OPER ) { // Operand or instruction? 1704 // Internal operand; recurse but do nothing else 1705 ReduceOper( kid, newrule, mem, mach ); 1706 1707 } else { // Child is a new instruction 1708 // Reduce the instruction, and add a direct pointer from this 1709 // machine instruction to the newly reduced one. 1710 Node *mem1 = (Node*)1; 1711 debug_only(Node *save_mem_node = _mem_node;) 1712 mach->add_req( ReduceInst( kid, newrule, mem1 ) ); 1713 debug_only(_mem_node = save_mem_node;) 1714 } 1715 } 1716 } 1717 1718 1719 // ------------------------------------------------------------------------- 1720 // Java-Java calling convention 1721 // (what you use when Java calls Java) 1722 1723 //------------------------------find_receiver---------------------------------- 1724 // For a given signature, return the OptoReg for parameter 0. 1725 OptoReg::Name Matcher::find_receiver( bool is_outgoing ) { 1726 VMRegPair regs; 1727 BasicType sig_bt = T_OBJECT; 1728 calling_convention(&sig_bt, ®s, 1, is_outgoing); 1729 // Return argument 0 register. In the LP64 build pointers 1730 // take 2 registers, but the VM wants only the 'main' name. 1731 return OptoReg::as_OptoReg(regs.first()); 1732 } 1733 1734 // A method-klass-holder may be passed in the inline_cache_reg 1735 // and then expanded into the inline_cache_reg and a method_oop register 1736 // defined in ad_<arch>.cpp 1737 1738 1739 //------------------------------find_shared------------------------------------ 1740 // Set bits if Node is shared or otherwise a root 1741 void Matcher::find_shared( Node *n ) { 1742 // Allocate stack of size C->unique() * 2 to avoid frequent realloc 1743 MStack mstack(C->unique() * 2); 1744 // Mark nodes as address_visited if they are inputs to an address expression 1745 VectorSet address_visited(Thread::current()->resource_area()); 1746 mstack.push(n, Visit); // Don't need to pre-visit root node 1747 while (mstack.is_nonempty()) { 1748 n = mstack.node(); // Leave node on stack 1749 Node_State nstate = mstack.state(); 1750 uint nop = n->Opcode(); 1751 if (nstate == Pre_Visit) { 1752 if (address_visited.test(n->_idx)) { // Visited in address already? 1753 // Flag as visited and shared now. 1754 set_visited(n); 1755 } 1756 if (is_visited(n)) { // Visited already? 1757 // Node is shared and has no reason to clone. Flag it as shared. 1758 // This causes it to match into a register for the sharing. 1759 set_shared(n); // Flag as shared and 1760 mstack.pop(); // remove node from stack 1761 continue; 1762 } 1763 nstate = Visit; // Not already visited; so visit now 1764 } 1765 if (nstate == Visit) { 1766 mstack.set_state(Post_Visit); 1767 set_visited(n); // Flag as visited now 1768 bool mem_op = false; 1769 1770 switch( nop ) { // Handle some opcodes special 1771 case Op_Phi: // Treat Phis as shared roots 1772 case Op_Parm: 1773 case Op_Proj: // All handled specially during matching 1774 case Op_SafePointScalarObject: 1775 set_shared(n); 1776 set_dontcare(n); 1777 break; 1778 case Op_If: 1779 case Op_CountedLoopEnd: 1780 mstack.set_state(Alt_Post_Visit); // Alternative way 1781 // Convert (If (Bool (CmpX A B))) into (If (Bool) (CmpX A B)). Helps 1782 // with matching cmp/branch in 1 instruction. The Matcher needs the 1783 // Bool and CmpX side-by-side, because it can only get at constants 1784 // that are at the leaves of Match trees, and the Bool's condition acts 1785 // as a constant here. 1786 mstack.push(n->in(1), Visit); // Clone the Bool 1787 mstack.push(n->in(0), Pre_Visit); // Visit control input 1788 continue; // while (mstack.is_nonempty()) 1789 case Op_ConvI2D: // These forms efficiently match with a prior 1790 case Op_ConvI2F: // Load but not a following Store 1791 if( n->in(1)->is_Load() && // Prior load 1792 n->outcnt() == 1 && // Not already shared 1793 n->unique_out()->is_Store() ) // Following store 1794 set_shared(n); // Force it to be a root 1795 break; 1796 case Op_ReverseBytesI: 1797 case Op_ReverseBytesL: 1798 if( n->in(1)->is_Load() && // Prior load 1799 n->outcnt() == 1 ) // Not already shared 1800 set_shared(n); // Force it to be a root 1801 break; 1802 case Op_BoxLock: // Cant match until we get stack-regs in ADLC 1803 case Op_IfFalse: 1804 case Op_IfTrue: 1805 case Op_MachProj: 1806 case Op_MergeMem: 1807 case Op_Catch: 1808 case Op_CatchProj: 1809 case Op_CProj: 1810 case Op_JumpProj: 1811 case Op_JProj: 1812 case Op_NeverBranch: 1813 set_dontcare(n); 1814 break; 1815 case Op_Jump: 1816 mstack.push(n->in(1), Visit); // Switch Value 1817 mstack.push(n->in(0), Pre_Visit); // Visit Control input 1818 continue; // while (mstack.is_nonempty()) 1819 case Op_StrComp: 1820 case Op_StrEquals: 1821 case Op_StrIndexOf: 1822 case Op_AryEq: 1823 set_shared(n); // Force result into register (it will be anyways) 1824 break; 1825 case Op_ConP: { // Convert pointers above the centerline to NUL 1826 TypeNode *tn = n->as_Type(); // Constants derive from type nodes 1827 const TypePtr* tp = tn->type()->is_ptr(); 1828 if (tp->_ptr == TypePtr::AnyNull) { 1829 tn->set_type(TypePtr::NULL_PTR); 1830 } 1831 break; 1832 } 1833 case Op_ConN: { // Convert narrow pointers above the centerline to NUL 1834 TypeNode *tn = n->as_Type(); // Constants derive from type nodes 1835 const TypePtr* tp = tn->type()->make_ptr(); 1836 if (tp && tp->_ptr == TypePtr::AnyNull) { 1837 tn->set_type(TypeNarrowOop::NULL_PTR); 1838 } 1839 break; 1840 } 1841 case Op_Binary: // These are introduced in the Post_Visit state. 1842 ShouldNotReachHere(); 1843 break; 1844 case Op_StoreB: // Do match these, despite no ideal reg 1845 case Op_StoreC: 1846 case Op_StoreCM: 1847 case Op_StoreD: 1848 case Op_StoreF: 1849 case Op_StoreI: 1850 case Op_StoreL: 1851 case Op_StoreP: 1852 case Op_StoreN: 1853 case Op_Store16B: 1854 case Op_Store8B: 1855 case Op_Store4B: 1856 case Op_Store8C: 1857 case Op_Store4C: 1858 case Op_Store2C: 1859 case Op_Store4I: 1860 case Op_Store2I: 1861 case Op_Store2L: 1862 case Op_Store4F: 1863 case Op_Store2F: 1864 case Op_Store2D: 1865 case Op_ClearArray: 1866 case Op_SafePoint: 1867 mem_op = true; 1868 break; 1869 case Op_LoadB: 1870 case Op_LoadUS: 1871 case Op_LoadD: 1872 case Op_LoadF: 1873 case Op_LoadI: 1874 case Op_LoadKlass: 1875 case Op_LoadNKlass: 1876 case Op_LoadL: 1877 case Op_LoadS: 1878 case Op_LoadP: 1879 case Op_LoadN: 1880 case Op_LoadRange: 1881 case Op_LoadD_unaligned: 1882 case Op_LoadL_unaligned: 1883 case Op_Load16B: 1884 case Op_Load8B: 1885 case Op_Load4B: 1886 case Op_Load4C: 1887 case Op_Load2C: 1888 case Op_Load8C: 1889 case Op_Load8S: 1890 case Op_Load4S: 1891 case Op_Load2S: 1892 case Op_Load4I: 1893 case Op_Load2I: 1894 case Op_Load2L: 1895 case Op_Load4F: 1896 case Op_Load2F: 1897 case Op_Load2D: 1898 mem_op = true; 1899 // Must be root of match tree due to prior load conflict 1900 if( C->subsume_loads() == false ) { 1901 set_shared(n); 1902 } 1903 // Fall into default case 1904 default: 1905 if( !n->ideal_reg() ) 1906 set_dontcare(n); // Unmatchable Nodes 1907 } // end_switch 1908 1909 for(int i = n->req() - 1; i >= 0; --i) { // For my children 1910 Node *m = n->in(i); // Get ith input 1911 if (m == NULL) continue; // Ignore NULLs 1912 uint mop = m->Opcode(); 1913 1914 // Must clone all producers of flags, or we will not match correctly. 1915 // Suppose a compare setting int-flags is shared (e.g., a switch-tree) 1916 // then it will match into an ideal Op_RegFlags. Alas, the fp-flags 1917 // are also there, so we may match a float-branch to int-flags and 1918 // expect the allocator to haul the flags from the int-side to the 1919 // fp-side. No can do. 1920 if( _must_clone[mop] ) { 1921 mstack.push(m, Visit); 1922 continue; // for(int i = ...) 1923 } 1924 1925 // Clone addressing expressions as they are "free" in most instructions 1926 if( mem_op && i == MemNode::Address && mop == Op_AddP ) { 1927 if (m->in(AddPNode::Base)->Opcode() == Op_DecodeN) { 1928 // Bases used in addresses must be shared but since 1929 // they are shared through a DecodeN they may appear 1930 // to have a single use so force sharing here. 1931 set_shared(m->in(AddPNode::Base)->in(1)); 1932 } 1933 1934 // Some inputs for address expression are not put on stack 1935 // to avoid marking them as shared and forcing them into register 1936 // if they are used only in address expressions. 1937 // But they should be marked as shared if there are other uses 1938 // besides address expressions. 1939 1940 Node *off = m->in(AddPNode::Offset); 1941 if( off->is_Con() && 1942 // When there are other uses besides address expressions 1943 // put it on stack and mark as shared. 1944 !is_visited(m) ) { 1945 address_visited.test_set(m->_idx); // Flag as address_visited 1946 Node *adr = m->in(AddPNode::Address); 1947 1948 // Intel, ARM and friends can handle 2 adds in addressing mode 1949 if( clone_shift_expressions && adr->is_AddP() && 1950 // AtomicAdd is not an addressing expression. 1951 // Cheap to find it by looking for screwy base. 1952 !adr->in(AddPNode::Base)->is_top() && 1953 // Are there other uses besides address expressions? 1954 !is_visited(adr) ) { 1955 address_visited.set(adr->_idx); // Flag as address_visited 1956 Node *shift = adr->in(AddPNode::Offset); 1957 // Check for shift by small constant as well 1958 if( shift->Opcode() == Op_LShiftX && shift->in(2)->is_Con() && 1959 shift->in(2)->get_int() <= 3 && 1960 // Are there other uses besides address expressions? 1961 !is_visited(shift) ) { 1962 address_visited.set(shift->_idx); // Flag as address_visited 1963 mstack.push(shift->in(2), Visit); 1964 Node *conv = shift->in(1); 1965 #ifdef _LP64 1966 // Allow Matcher to match the rule which bypass 1967 // ConvI2L operation for an array index on LP64 1968 // if the index value is positive. 1969 if( conv->Opcode() == Op_ConvI2L && 1970 conv->as_Type()->type()->is_long()->_lo >= 0 && 1971 // Are there other uses besides address expressions? 1972 !is_visited(conv) ) { 1973 address_visited.set(conv->_idx); // Flag as address_visited 1974 mstack.push(conv->in(1), Pre_Visit); 1975 } else 1976 #endif 1977 mstack.push(conv, Pre_Visit); 1978 } else { 1979 mstack.push(shift, Pre_Visit); 1980 } 1981 mstack.push(adr->in(AddPNode::Address), Pre_Visit); 1982 mstack.push(adr->in(AddPNode::Base), Pre_Visit); 1983 } else { // Sparc, Alpha, PPC and friends 1984 mstack.push(adr, Pre_Visit); 1985 } 1986 1987 // Clone X+offset as it also folds into most addressing expressions 1988 mstack.push(off, Visit); 1989 mstack.push(m->in(AddPNode::Base), Pre_Visit); 1990 continue; // for(int i = ...) 1991 } // if( off->is_Con() ) 1992 } // if( mem_op && 1993 mstack.push(m, Pre_Visit); 1994 } // for(int i = ...) 1995 } 1996 else if (nstate == Alt_Post_Visit) { 1997 mstack.pop(); // Remove node from stack 1998 // We cannot remove the Cmp input from the Bool here, as the Bool may be 1999 // shared and all users of the Bool need to move the Cmp in parallel. 2000 // This leaves both the Bool and the If pointing at the Cmp. To 2001 // prevent the Matcher from trying to Match the Cmp along both paths 2002 // BoolNode::match_edge always returns a zero. 2003 2004 // We reorder the Op_If in a pre-order manner, so we can visit without 2005 // accidentally sharing the Cmp (the Bool and the If make 2 users). 2006 n->add_req( n->in(1)->in(1) ); // Add the Cmp next to the Bool 2007 } 2008 else if (nstate == Post_Visit) { 2009 mstack.pop(); // Remove node from stack 2010 2011 // Now hack a few special opcodes 2012 switch( n->Opcode() ) { // Handle some opcodes special 2013 case Op_StorePConditional: 2014 case Op_StoreIConditional: 2015 case Op_StoreLConditional: 2016 case Op_CompareAndSwapI: 2017 case Op_CompareAndSwapL: 2018 case Op_CompareAndSwapP: 2019 case Op_CompareAndSwapN: { // Convert trinary to binary-tree 2020 Node *newval = n->in(MemNode::ValueIn ); 2021 Node *oldval = n->in(LoadStoreNode::ExpectedIn); 2022 Node *pair = new (C, 3) BinaryNode( oldval, newval ); 2023 n->set_req(MemNode::ValueIn,pair); 2024 n->del_req(LoadStoreNode::ExpectedIn); 2025 break; 2026 } 2027 case Op_CMoveD: // Convert trinary to binary-tree 2028 case Op_CMoveF: 2029 case Op_CMoveI: 2030 case Op_CMoveL: 2031 case Op_CMoveN: 2032 case Op_CMoveP: { 2033 // Restructure into a binary tree for Matching. It's possible that 2034 // we could move this code up next to the graph reshaping for IfNodes 2035 // or vice-versa, but I do not want to debug this for Ladybird. 2036 // 10/2/2000 CNC. 2037 Node *pair1 = new (C, 3) BinaryNode(n->in(1),n->in(1)->in(1)); 2038 n->set_req(1,pair1); 2039 Node *pair2 = new (C, 3) BinaryNode(n->in(2),n->in(3)); 2040 n->set_req(2,pair2); 2041 n->del_req(3); 2042 break; 2043 } 2044 default: 2045 break; 2046 } 2047 } 2048 else { 2049 ShouldNotReachHere(); 2050 } 2051 } // end of while (mstack.is_nonempty()) 2052 } 2053 2054 #ifdef ASSERT 2055 // machine-independent root to machine-dependent root 2056 void Matcher::dump_old2new_map() { 2057 _old2new_map.dump(); 2058 } 2059 #endif 2060 2061 //---------------------------collect_null_checks------------------------------- 2062 // Find null checks in the ideal graph; write a machine-specific node for 2063 // it. Used by later implicit-null-check handling. Actually collects 2064 // either an IfTrue or IfFalse for the common NOT-null path, AND the ideal 2065 // value being tested. 2066 void Matcher::collect_null_checks( Node *proj, Node *orig_proj ) { 2067 Node *iff = proj->in(0); 2068 if( iff->Opcode() == Op_If ) { 2069 // During matching If's have Bool & Cmp side-by-side 2070 BoolNode *b = iff->in(1)->as_Bool(); 2071 Node *cmp = iff->in(2); 2072 int opc = cmp->Opcode(); 2073 if (opc != Op_CmpP && opc != Op_CmpN) return; 2074 2075 const Type* ct = cmp->in(2)->bottom_type(); 2076 if (ct == TypePtr::NULL_PTR || 2077 (opc == Op_CmpN && ct == TypeNarrowOop::NULL_PTR)) { 2078 2079 bool push_it = false; 2080 if( proj->Opcode() == Op_IfTrue ) { 2081 extern int all_null_checks_found; 2082 all_null_checks_found++; 2083 if( b->_test._test == BoolTest::ne ) { 2084 push_it = true; 2085 } 2086 } else { 2087 assert( proj->Opcode() == Op_IfFalse, "" ); 2088 if( b->_test._test == BoolTest::eq ) { 2089 push_it = true; 2090 } 2091 } 2092 if( push_it ) { 2093 _null_check_tests.push(proj); 2094 Node* val = cmp->in(1); 2095 #ifdef _LP64 2096 if (UseCompressedOops && !Matcher::clone_shift_expressions && 2097 val->bottom_type()->isa_narrowoop()) { 2098 // 2099 // Look for DecodeN node which should be pinned to orig_proj. 2100 // On platforms (Sparc) which can not handle 2 adds 2101 // in addressing mode we have to keep a DecodeN node and 2102 // use it to do implicit NULL check in address. 2103 // 2104 // DecodeN node was pinned to non-null path (orig_proj) during 2105 // CastPP transformation in final_graph_reshaping_impl(). 2106 // 2107 uint cnt = orig_proj->outcnt(); 2108 for (uint i = 0; i < orig_proj->outcnt(); i++) { 2109 Node* d = orig_proj->raw_out(i); 2110 if (d->is_DecodeN() && d->in(1) == val) { 2111 val = d; 2112 val->set_req(0, NULL); // Unpin now. 2113 break; 2114 } 2115 } 2116 } 2117 #endif 2118 _null_check_tests.push(val); 2119 } 2120 } 2121 } 2122 } 2123 2124 //---------------------------validate_null_checks------------------------------ 2125 // Its possible that the value being NULL checked is not the root of a match 2126 // tree. If so, I cannot use the value in an implicit null check. 2127 void Matcher::validate_null_checks( ) { 2128 uint cnt = _null_check_tests.size(); 2129 for( uint i=0; i < cnt; i+=2 ) { 2130 Node *test = _null_check_tests[i]; 2131 Node *val = _null_check_tests[i+1]; 2132 if (has_new_node(val)) { 2133 // Is a match-tree root, so replace with the matched value 2134 _null_check_tests.map(i+1, new_node(val)); 2135 } else { 2136 // Yank from candidate list 2137 _null_check_tests.map(i+1,_null_check_tests[--cnt]); 2138 _null_check_tests.map(i,_null_check_tests[--cnt]); 2139 _null_check_tests.pop(); 2140 _null_check_tests.pop(); 2141 i-=2; 2142 } 2143 } 2144 } 2145 2146 2147 // Used by the DFA in dfa_sparc.cpp. Check for a prior FastLock 2148 // acting as an Acquire and thus we don't need an Acquire here. We 2149 // retain the Node to act as a compiler ordering barrier. 2150 bool Matcher::prior_fast_lock( const Node *acq ) { 2151 Node *r = acq->in(0); 2152 if( !r->is_Region() || r->req() <= 1 ) return false; 2153 Node *proj = r->in(1); 2154 if( !proj->is_Proj() ) return false; 2155 Node *call = proj->in(0); 2156 if( !call->is_Call() || call->as_Call()->entry_point() != OptoRuntime::complete_monitor_locking_Java() ) 2157 return false; 2158 2159 return true; 2160 } 2161 2162 // Used by the DFA in dfa_sparc.cpp. Check for a following FastUnLock 2163 // acting as a Release and thus we don't need a Release here. We 2164 // retain the Node to act as a compiler ordering barrier. 2165 bool Matcher::post_fast_unlock( const Node *rel ) { 2166 Compile *C = Compile::current(); 2167 assert( rel->Opcode() == Op_MemBarRelease, "" ); 2168 const MemBarReleaseNode *mem = (const MemBarReleaseNode*)rel; 2169 DUIterator_Fast imax, i = mem->fast_outs(imax); 2170 Node *ctrl = NULL; 2171 while( true ) { 2172 ctrl = mem->fast_out(i); // Throw out-of-bounds if proj not found 2173 assert( ctrl->is_Proj(), "only projections here" ); 2174 ProjNode *proj = (ProjNode*)ctrl; 2175 if( proj->_con == TypeFunc::Control && 2176 !C->node_arena()->contains(ctrl) ) // Unmatched old-space only 2177 break; 2178 i++; 2179 } 2180 Node *iff = NULL; 2181 for( DUIterator_Fast jmax, j = ctrl->fast_outs(jmax); j < jmax; j++ ) { 2182 Node *x = ctrl->fast_out(j); 2183 if( x->is_If() && x->req() > 1 && 2184 !C->node_arena()->contains(x) ) { // Unmatched old-space only 2185 iff = x; 2186 break; 2187 } 2188 } 2189 if( !iff ) return false; 2190 Node *bol = iff->in(1); 2191 // The iff might be some random subclass of If or bol might be Con-Top 2192 if (!bol->is_Bool()) return false; 2193 assert( bol->req() > 1, "" ); 2194 return (bol->in(1)->Opcode() == Op_FastUnlock); 2195 } 2196 2197 // Used by the DFA in dfa_xxx.cpp. Check for a following barrier or 2198 // atomic instruction acting as a store_load barrier without any 2199 // intervening volatile load, and thus we don't need a barrier here. 2200 // We retain the Node to act as a compiler ordering barrier. 2201 bool Matcher::post_store_load_barrier(const Node *vmb) { 2202 Compile *C = Compile::current(); 2203 assert( vmb->is_MemBar(), "" ); 2204 assert( vmb->Opcode() != Op_MemBarAcquire, "" ); 2205 const MemBarNode *mem = (const MemBarNode*)vmb; 2206 2207 // Get the Proj node, ctrl, that can be used to iterate forward 2208 Node *ctrl = NULL; 2209 DUIterator_Fast imax, i = mem->fast_outs(imax); 2210 while( true ) { 2211 ctrl = mem->fast_out(i); // Throw out-of-bounds if proj not found 2212 assert( ctrl->is_Proj(), "only projections here" ); 2213 ProjNode *proj = (ProjNode*)ctrl; 2214 if( proj->_con == TypeFunc::Control && 2215 !C->node_arena()->contains(ctrl) ) // Unmatched old-space only 2216 break; 2217 i++; 2218 } 2219 2220 for( DUIterator_Fast jmax, j = ctrl->fast_outs(jmax); j < jmax; j++ ) { 2221 Node *x = ctrl->fast_out(j); 2222 int xop = x->Opcode(); 2223 2224 // We don't need current barrier if we see another or a lock 2225 // before seeing volatile load. 2226 // 2227 // Op_Fastunlock previously appeared in the Op_* list below. 2228 // With the advent of 1-0 lock operations we're no longer guaranteed 2229 // that a monitor exit operation contains a serializing instruction. 2230 2231 if (xop == Op_MemBarVolatile || 2232 xop == Op_FastLock || 2233 xop == Op_CompareAndSwapL || 2234 xop == Op_CompareAndSwapP || 2235 xop == Op_CompareAndSwapN || 2236 xop == Op_CompareAndSwapI) 2237 return true; 2238 2239 if (x->is_MemBar()) { 2240 // We must retain this membar if there is an upcoming volatile 2241 // load, which will be preceded by acquire membar. 2242 if (xop == Op_MemBarAcquire) 2243 return false; 2244 // For other kinds of barriers, check by pretending we 2245 // are them, and seeing if we can be removed. 2246 else 2247 return post_store_load_barrier((const MemBarNode*)x); 2248 } 2249 2250 // Delicate code to detect case of an upcoming fastlock block 2251 if( x->is_If() && x->req() > 1 && 2252 !C->node_arena()->contains(x) ) { // Unmatched old-space only 2253 Node *iff = x; 2254 Node *bol = iff->in(1); 2255 // The iff might be some random subclass of If or bol might be Con-Top 2256 if (!bol->is_Bool()) return false; 2257 assert( bol->req() > 1, "" ); 2258 return (bol->in(1)->Opcode() == Op_FastUnlock); 2259 } 2260 // probably not necessary to check for these 2261 if (x->is_Call() || x->is_SafePoint() || x->is_block_proj()) 2262 return false; 2263 } 2264 return false; 2265 } 2266 2267 //============================================================================= 2268 //---------------------------State--------------------------------------------- 2269 State::State(void) { 2270 #ifdef ASSERT 2271 _id = 0; 2272 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe); 2273 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d); 2274 //memset(_cost, -1, sizeof(_cost)); 2275 //memset(_rule, -1, sizeof(_rule)); 2276 #endif 2277 memset(_valid, 0, sizeof(_valid)); 2278 } 2279 2280 #ifdef ASSERT 2281 State::~State() { 2282 _id = 99; 2283 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe); 2284 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d); 2285 memset(_cost, -3, sizeof(_cost)); 2286 memset(_rule, -3, sizeof(_rule)); 2287 } 2288 #endif 2289 2290 #ifndef PRODUCT 2291 //---------------------------dump---------------------------------------------- 2292 void State::dump() { 2293 tty->print("\n"); 2294 dump(0); 2295 } 2296 2297 void State::dump(int depth) { 2298 for( int j = 0; j < depth; j++ ) 2299 tty->print(" "); 2300 tty->print("--N: "); 2301 _leaf->dump(); 2302 uint i; 2303 for( i = 0; i < _LAST_MACH_OPER; i++ ) 2304 // Check for valid entry 2305 if( valid(i) ) { 2306 for( int j = 0; j < depth; j++ ) 2307 tty->print(" "); 2308 assert(_cost[i] != max_juint, "cost must be a valid value"); 2309 assert(_rule[i] < _last_Mach_Node, "rule[i] must be valid rule"); 2310 tty->print_cr("%s %d %s", 2311 ruleName[i], _cost[i], ruleName[_rule[i]] ); 2312 } 2313 tty->print_cr(""); 2314 2315 for( i=0; i<2; i++ ) 2316 if( _kids[i] ) 2317 _kids[i]->dump(depth+1); 2318 } 2319 #endif