1 /* 2 * Copyright (c) 1997, 2017, 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 "gc/shared/barrierSet.hpp" 27 #include "gc/shared/c2/barrierSetC2.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "memory/resourceArea.hpp" 30 #include "opto/ad.hpp" 31 #include "opto/addnode.hpp" 32 #include "opto/callnode.hpp" 33 #include "opto/idealGraphPrinter.hpp" 34 #include "opto/matcher.hpp" 35 #include "opto/memnode.hpp" 36 #include "opto/movenode.hpp" 37 #include "opto/opcodes.hpp" 38 #include "opto/regmask.hpp" 39 #include "opto/rootnode.hpp" 40 #include "opto/runtime.hpp" 41 #include "opto/type.hpp" 42 #include "opto/vectornode.hpp" 43 #include "runtime/os.hpp" 44 #include "runtime/sharedRuntime.hpp" 45 #include "utilities/align.hpp" 46 47 OptoReg::Name OptoReg::c_frame_pointer; 48 49 const RegMask *Matcher::idealreg2regmask[_last_machine_leaf]; 50 RegMask Matcher::mreg2regmask[_last_Mach_Reg]; 51 RegMask Matcher::STACK_ONLY_mask; 52 RegMask Matcher::c_frame_ptr_mask; 53 const uint Matcher::_begin_rematerialize = _BEGIN_REMATERIALIZE; 54 const uint Matcher::_end_rematerialize = _END_REMATERIALIZE; 55 56 //---------------------------Matcher------------------------------------------- 57 Matcher::Matcher() 58 : PhaseTransform( Phase::Ins_Select ), 59 _states_arena(Chunk::medium_size, mtCompiler), 60 _visited(&_states_arena), 61 _shared(&_states_arena), 62 _dontcare(&_states_arena), 63 _reduceOp(reduceOp), _leftOp(leftOp), _rightOp(rightOp), 64 _swallowed(swallowed), 65 _begin_inst_chain_rule(_BEGIN_INST_CHAIN_RULE), 66 _end_inst_chain_rule(_END_INST_CHAIN_RULE), 67 _must_clone(must_clone), 68 _shared_nodes(C->comp_arena()), 69 #ifdef ASSERT 70 _old2new_map(C->comp_arena()), 71 _new2old_map(C->comp_arena()), 72 #endif 73 _allocation_started(false), 74 _ruleName(ruleName), 75 _register_save_policy(register_save_policy), 76 _c_reg_save_policy(c_reg_save_policy), 77 _register_save_type(register_save_type) { 78 C->set_matcher(this); 79 80 idealreg2spillmask [Op_RegI] = NULL; 81 idealreg2spillmask [Op_RegN] = NULL; 82 idealreg2spillmask [Op_RegL] = NULL; 83 idealreg2spillmask [Op_RegF] = NULL; 84 idealreg2spillmask [Op_RegD] = NULL; 85 idealreg2spillmask [Op_RegP] = NULL; 86 idealreg2spillmask [Op_VecS] = NULL; 87 idealreg2spillmask [Op_VecD] = NULL; 88 idealreg2spillmask [Op_VecX] = NULL; 89 idealreg2spillmask [Op_VecY] = NULL; 90 idealreg2spillmask [Op_VecZ] = NULL; 91 idealreg2spillmask [Op_RegFlags] = NULL; 92 93 idealreg2debugmask [Op_RegI] = NULL; 94 idealreg2debugmask [Op_RegN] = NULL; 95 idealreg2debugmask [Op_RegL] = NULL; 96 idealreg2debugmask [Op_RegF] = NULL; 97 idealreg2debugmask [Op_RegD] = NULL; 98 idealreg2debugmask [Op_RegP] = NULL; 99 idealreg2debugmask [Op_VecS] = NULL; 100 idealreg2debugmask [Op_VecD] = NULL; 101 idealreg2debugmask [Op_VecX] = NULL; 102 idealreg2debugmask [Op_VecY] = NULL; 103 idealreg2debugmask [Op_VecZ] = NULL; 104 idealreg2debugmask [Op_RegFlags] = NULL; 105 106 idealreg2mhdebugmask[Op_RegI] = NULL; 107 idealreg2mhdebugmask[Op_RegN] = NULL; 108 idealreg2mhdebugmask[Op_RegL] = NULL; 109 idealreg2mhdebugmask[Op_RegF] = NULL; 110 idealreg2mhdebugmask[Op_RegD] = NULL; 111 idealreg2mhdebugmask[Op_RegP] = NULL; 112 idealreg2mhdebugmask[Op_VecS] = NULL; 113 idealreg2mhdebugmask[Op_VecD] = NULL; 114 idealreg2mhdebugmask[Op_VecX] = NULL; 115 idealreg2mhdebugmask[Op_VecY] = NULL; 116 idealreg2mhdebugmask[Op_VecZ] = NULL; 117 idealreg2mhdebugmask[Op_RegFlags] = NULL; 118 119 debug_only(_mem_node = NULL;) // Ideal memory node consumed by mach node 120 } 121 122 //------------------------------warp_incoming_stk_arg------------------------ 123 // This warps a VMReg into an OptoReg::Name 124 OptoReg::Name Matcher::warp_incoming_stk_arg( VMReg reg ) { 125 OptoReg::Name warped; 126 if( reg->is_stack() ) { // Stack slot argument? 127 warped = OptoReg::add(_old_SP, reg->reg2stack() ); 128 warped = OptoReg::add(warped, C->out_preserve_stack_slots()); 129 if( warped >= _in_arg_limit ) 130 _in_arg_limit = OptoReg::add(warped, 1); // Bump max stack slot seen 131 if (!RegMask::can_represent_arg(warped)) { 132 // the compiler cannot represent this method's calling sequence 133 C->record_method_not_compilable("unsupported incoming calling sequence"); 134 return OptoReg::Bad; 135 } 136 return warped; 137 } 138 return OptoReg::as_OptoReg(reg); 139 } 140 141 //---------------------------compute_old_SP------------------------------------ 142 OptoReg::Name Compile::compute_old_SP() { 143 int fixed = fixed_slots(); 144 int preserve = in_preserve_stack_slots(); 145 return OptoReg::stack2reg(align_up(fixed + preserve, (int)Matcher::stack_alignment_in_slots())); 146 } 147 148 149 150 #ifdef ASSERT 151 void Matcher::verify_new_nodes_only(Node* xroot) { 152 // Make sure that the new graph only references new nodes 153 ResourceMark rm; 154 Unique_Node_List worklist; 155 VectorSet visited(Thread::current()->resource_area()); 156 worklist.push(xroot); 157 while (worklist.size() > 0) { 158 Node* n = worklist.pop(); 159 visited <<= n->_idx; 160 assert(C->node_arena()->contains(n), "dead node"); 161 for (uint j = 0; j < n->req(); j++) { 162 Node* in = n->in(j); 163 if (in != NULL) { 164 assert(C->node_arena()->contains(in), "dead node"); 165 if (!visited.test(in->_idx)) { 166 worklist.push(in); 167 } 168 } 169 } 170 } 171 } 172 #endif 173 174 // Array of RegMask, one per returned values (value type instances can 175 // be returned as multiple return values, one per field) 176 RegMask* Matcher::return_values_mask(const TypeTuple *range) { 177 uint cnt = range->cnt() - TypeFunc::Parms; 178 if (cnt == 0) { 179 return NULL; 180 } 181 RegMask* mask = NEW_RESOURCE_ARRAY(RegMask, cnt); 182 183 if (!ValueTypeReturnedAsFields) { 184 // Get ideal-register return type 185 uint ireg = range->field_at(TypeFunc::Parms)->ideal_reg(); 186 // Get machine return register 187 OptoRegPair regs = return_value(ireg, false); 188 189 // And mask for same 190 mask[0].Clear(); 191 mask[0].Insert(regs.first()); 192 if (OptoReg::is_valid(regs.second())) { 193 mask[0].Insert(regs.second()); 194 } 195 } else { 196 BasicType* sig_bt = NEW_RESOURCE_ARRAY(BasicType, cnt); 197 VMRegPair* vm_parm_regs = NEW_RESOURCE_ARRAY(VMRegPair, cnt); 198 199 for (uint i = 0; i < cnt; i++) { 200 sig_bt[i] = range->field_at(i+TypeFunc::Parms)->basic_type(); 201 } 202 203 int regs = SharedRuntime::java_return_convention(sig_bt, vm_parm_regs, cnt); 204 assert(regs > 0, "should have been tested during graph construction"); 205 for (uint i = 0; i < cnt; i++) { 206 mask[i].Clear(); 207 208 OptoReg::Name reg1 = OptoReg::as_OptoReg(vm_parm_regs[i].first()); 209 if (OptoReg::is_valid(reg1)) { 210 mask[i].Insert(reg1); 211 } 212 OptoReg::Name reg2 = OptoReg::as_OptoReg(vm_parm_regs[i].second()); 213 if (OptoReg::is_valid(reg2)) { 214 mask[i].Insert(reg2); 215 } 216 } 217 } 218 return mask; 219 } 220 221 //---------------------------match--------------------------------------------- 222 void Matcher::match( ) { 223 if( MaxLabelRootDepth < 100 ) { // Too small? 224 assert(false, "invalid MaxLabelRootDepth, increase it to 100 minimum"); 225 MaxLabelRootDepth = 100; 226 } 227 // One-time initialization of some register masks. 228 init_spill_mask( C->root()->in(1) ); 229 _return_addr_mask = return_addr(); 230 #ifdef _LP64 231 // Pointers take 2 slots in 64-bit land 232 _return_addr_mask.Insert(OptoReg::add(return_addr(),1)); 233 #endif 234 235 // Map Java-signature return types into return register-value 236 // machine registers. 237 const TypeTuple *range = C->tf()->range_cc(); 238 _return_values_mask = return_values_mask(range); 239 240 // --------------- 241 // Frame Layout 242 243 // Need the method signature to determine the incoming argument types, 244 // because the types determine which registers the incoming arguments are 245 // in, and this affects the matched code. 246 const TypeTuple *domain = C->tf()->domain_cc(); 247 uint argcnt = domain->cnt() - TypeFunc::Parms; 248 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt ); 249 VMRegPair *vm_parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt ); 250 _parm_regs = NEW_RESOURCE_ARRAY( OptoRegPair, argcnt ); 251 _calling_convention_mask = NEW_RESOURCE_ARRAY( RegMask, argcnt ); 252 uint i; 253 for( i = 0; i<argcnt; i++ ) { 254 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type(); 255 } 256 257 // Pass array of ideal registers and length to USER code (from the AD file) 258 // that will convert this to an array of register numbers. 259 const StartNode *start = C->start(); 260 start->calling_convention( sig_bt, vm_parm_regs, argcnt ); 261 #ifdef ASSERT 262 // Sanity check users' calling convention. Real handy while trying to 263 // get the initial port correct. 264 { for (uint i = 0; i<argcnt; i++) { 265 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) { 266 assert(domain->field_at(i+TypeFunc::Parms)==Type::HALF, "only allowed on halve" ); 267 _parm_regs[i].set_bad(); 268 continue; 269 } 270 VMReg parm_reg = vm_parm_regs[i].first(); 271 assert(parm_reg->is_valid(), "invalid arg?"); 272 if (parm_reg->is_reg()) { 273 OptoReg::Name opto_parm_reg = OptoReg::as_OptoReg(parm_reg); 274 assert(can_be_java_arg(opto_parm_reg) || 275 C->stub_function() == CAST_FROM_FN_PTR(address, OptoRuntime::rethrow_C) || 276 opto_parm_reg == inline_cache_reg(), 277 "parameters in register must be preserved by runtime stubs"); 278 } 279 for (uint j = 0; j < i; j++) { 280 assert(parm_reg != vm_parm_regs[j].first(), 281 "calling conv. must produce distinct regs"); 282 } 283 } 284 } 285 #endif 286 287 // Do some initial frame layout. 288 289 // Compute the old incoming SP (may be called FP) as 290 // OptoReg::stack0() + locks + in_preserve_stack_slots + pad2. 291 _old_SP = C->compute_old_SP(); 292 assert( is_even(_old_SP), "must be even" ); 293 294 // Compute highest incoming stack argument as 295 // _old_SP + out_preserve_stack_slots + incoming argument size. 296 _in_arg_limit = OptoReg::add(_old_SP, C->out_preserve_stack_slots()); 297 assert( is_even(_in_arg_limit), "out_preserve must be even" ); 298 for( i = 0; i < argcnt; i++ ) { 299 // Permit args to have no register 300 _calling_convention_mask[i].Clear(); 301 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) { 302 continue; 303 } 304 // calling_convention returns stack arguments as a count of 305 // slots beyond OptoReg::stack0()/VMRegImpl::stack0. We need to convert this to 306 // the allocators point of view, taking into account all the 307 // preserve area, locks & pad2. 308 309 OptoReg::Name reg1 = warp_incoming_stk_arg(vm_parm_regs[i].first()); 310 if( OptoReg::is_valid(reg1)) 311 _calling_convention_mask[i].Insert(reg1); 312 313 OptoReg::Name reg2 = warp_incoming_stk_arg(vm_parm_regs[i].second()); 314 if( OptoReg::is_valid(reg2)) 315 _calling_convention_mask[i].Insert(reg2); 316 317 // Saved biased stack-slot register number 318 _parm_regs[i].set_pair(reg2, reg1); 319 } 320 321 // Finally, make sure the incoming arguments take up an even number of 322 // words, in case the arguments or locals need to contain doubleword stack 323 // slots. The rest of the system assumes that stack slot pairs (in 324 // particular, in the spill area) which look aligned will in fact be 325 // aligned relative to the stack pointer in the target machine. Double 326 // stack slots will always be allocated aligned. 327 _new_SP = OptoReg::Name(align_up(_in_arg_limit, (int)RegMask::SlotsPerLong)); 328 329 // Compute highest outgoing stack argument as 330 // _new_SP + out_preserve_stack_slots + max(outgoing argument size). 331 _out_arg_limit = OptoReg::add(_new_SP, C->out_preserve_stack_slots()); 332 assert( is_even(_out_arg_limit), "out_preserve must be even" ); 333 334 if (!RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1))) { 335 // the compiler cannot represent this method's calling sequence 336 C->record_method_not_compilable("must be able to represent all call arguments in reg mask"); 337 } 338 339 if (C->failing()) return; // bailed out on incoming arg failure 340 341 // --------------- 342 // Collect roots of matcher trees. Every node for which 343 // _shared[_idx] is cleared is guaranteed to not be shared, and thus 344 // can be a valid interior of some tree. 345 find_shared( C->root() ); 346 find_shared( C->top() ); 347 348 C->print_method(PHASE_BEFORE_MATCHING); 349 350 // Create new ideal node ConP #NULL even if it does exist in old space 351 // to avoid false sharing if the corresponding mach node is not used. 352 // The corresponding mach node is only used in rare cases for derived 353 // pointers. 354 Node* new_ideal_null = ConNode::make(TypePtr::NULL_PTR); 355 356 // Swap out to old-space; emptying new-space 357 Arena *old = C->node_arena()->move_contents(C->old_arena()); 358 359 // Save debug and profile information for nodes in old space: 360 _old_node_note_array = C->node_note_array(); 361 if (_old_node_note_array != NULL) { 362 C->set_node_note_array(new(C->comp_arena()) GrowableArray<Node_Notes*> 363 (C->comp_arena(), _old_node_note_array->length(), 364 0, NULL)); 365 } 366 367 // Pre-size the new_node table to avoid the need for range checks. 368 grow_new_node_array(C->unique()); 369 370 // Reset node counter so MachNodes start with _idx at 0 371 int live_nodes = C->live_nodes(); 372 C->set_unique(0); 373 C->reset_dead_node_list(); 374 375 // Recursively match trees from old space into new space. 376 // Correct leaves of new-space Nodes; they point to old-space. 377 _visited.Clear(); // Clear visit bits for xform call 378 C->set_cached_top_node(xform( C->top(), live_nodes )); 379 if (!C->failing()) { 380 Node* xroot = xform( C->root(), 1 ); 381 if (xroot == NULL) { 382 Matcher::soft_match_failure(); // recursive matching process failed 383 C->record_method_not_compilable("instruction match failed"); 384 } else { 385 // During matching shared constants were attached to C->root() 386 // because xroot wasn't available yet, so transfer the uses to 387 // the xroot. 388 for( DUIterator_Fast jmax, j = C->root()->fast_outs(jmax); j < jmax; j++ ) { 389 Node* n = C->root()->fast_out(j); 390 if (C->node_arena()->contains(n)) { 391 assert(n->in(0) == C->root(), "should be control user"); 392 n->set_req(0, xroot); 393 --j; 394 --jmax; 395 } 396 } 397 398 // Generate new mach node for ConP #NULL 399 assert(new_ideal_null != NULL, "sanity"); 400 _mach_null = match_tree(new_ideal_null); 401 // Don't set control, it will confuse GCM since there are no uses. 402 // The control will be set when this node is used first time 403 // in find_base_for_derived(). 404 assert(_mach_null != NULL, ""); 405 406 C->set_root(xroot->is_Root() ? xroot->as_Root() : NULL); 407 408 #ifdef ASSERT 409 verify_new_nodes_only(xroot); 410 #endif 411 } 412 } 413 if (C->top() == NULL || C->root() == NULL) { 414 C->record_method_not_compilable("graph lost"); // %%% cannot happen? 415 } 416 if (C->failing()) { 417 // delete old; 418 old->destruct_contents(); 419 return; 420 } 421 assert( C->top(), "" ); 422 assert( C->root(), "" ); 423 validate_null_checks(); 424 425 // Now smoke old-space 426 NOT_DEBUG( old->destruct_contents() ); 427 428 // ------------------------ 429 // Set up save-on-entry registers 430 Fixup_Save_On_Entry( ); 431 } 432 433 434 //------------------------------Fixup_Save_On_Entry---------------------------- 435 // The stated purpose of this routine is to take care of save-on-entry 436 // registers. However, the overall goal of the Match phase is to convert into 437 // machine-specific instructions which have RegMasks to guide allocation. 438 // So what this procedure really does is put a valid RegMask on each input 439 // to the machine-specific variations of all Return, TailCall and Halt 440 // instructions. It also adds edgs to define the save-on-entry values (and of 441 // course gives them a mask). 442 443 static RegMask *init_input_masks( uint size, RegMask &ret_adr, RegMask &fp ) { 444 RegMask *rms = NEW_RESOURCE_ARRAY( RegMask, size ); 445 // Do all the pre-defined register masks 446 rms[TypeFunc::Control ] = RegMask::Empty; 447 rms[TypeFunc::I_O ] = RegMask::Empty; 448 rms[TypeFunc::Memory ] = RegMask::Empty; 449 rms[TypeFunc::ReturnAdr] = ret_adr; 450 rms[TypeFunc::FramePtr ] = fp; 451 return rms; 452 } 453 454 //---------------------------init_first_stack_mask----------------------------- 455 // Create the initial stack mask used by values spilling to the stack. 456 // Disallow any debug info in outgoing argument areas by setting the 457 // initial mask accordingly. 458 void Matcher::init_first_stack_mask() { 459 460 // Allocate storage for spill masks as masks for the appropriate load type. 461 RegMask *rms = (RegMask*)C->comp_arena()->Amalloc_D(sizeof(RegMask) * (3*6+5)); 462 463 idealreg2spillmask [Op_RegN] = &rms[0]; 464 idealreg2spillmask [Op_RegI] = &rms[1]; 465 idealreg2spillmask [Op_RegL] = &rms[2]; 466 idealreg2spillmask [Op_RegF] = &rms[3]; 467 idealreg2spillmask [Op_RegD] = &rms[4]; 468 idealreg2spillmask [Op_RegP] = &rms[5]; 469 470 idealreg2debugmask [Op_RegN] = &rms[6]; 471 idealreg2debugmask [Op_RegI] = &rms[7]; 472 idealreg2debugmask [Op_RegL] = &rms[8]; 473 idealreg2debugmask [Op_RegF] = &rms[9]; 474 idealreg2debugmask [Op_RegD] = &rms[10]; 475 idealreg2debugmask [Op_RegP] = &rms[11]; 476 477 idealreg2mhdebugmask[Op_RegN] = &rms[12]; 478 idealreg2mhdebugmask[Op_RegI] = &rms[13]; 479 idealreg2mhdebugmask[Op_RegL] = &rms[14]; 480 idealreg2mhdebugmask[Op_RegF] = &rms[15]; 481 idealreg2mhdebugmask[Op_RegD] = &rms[16]; 482 idealreg2mhdebugmask[Op_RegP] = &rms[17]; 483 484 idealreg2spillmask [Op_VecS] = &rms[18]; 485 idealreg2spillmask [Op_VecD] = &rms[19]; 486 idealreg2spillmask [Op_VecX] = &rms[20]; 487 idealreg2spillmask [Op_VecY] = &rms[21]; 488 idealreg2spillmask [Op_VecZ] = &rms[22]; 489 490 OptoReg::Name i; 491 492 // At first, start with the empty mask 493 C->FIRST_STACK_mask().Clear(); 494 495 // Add in the incoming argument area 496 OptoReg::Name init_in = OptoReg::add(_old_SP, C->out_preserve_stack_slots()); 497 for (i = init_in; i < _in_arg_limit; i = OptoReg::add(i,1)) { 498 C->FIRST_STACK_mask().Insert(i); 499 } 500 501 // Check if the method has a reserved entry in the argument stack area that 502 // should not be used for spilling because it may hold the return address. 503 if (C->method() != NULL && C->method()->has_scalarized_args()) { 504 ExtendedSignature sig_cc = ExtendedSignature(C->method()->get_sig_cc(), SigEntryFilter()); 505 for (int off = 0; !sig_cc.at_end(); ) { 506 BasicType bt = (*sig_cc)._bt; 507 off += type2size[bt]; 508 while (SigEntry::next_is_reserved(sig_cc, bt)) { 509 // Remove reserved stack slot from mask to avoid spilling 510 OptoRegPair reg = _parm_regs[off]; 511 assert(OptoReg::is_valid(reg.first()), "invalid reserved register"); 512 C->FIRST_STACK_mask().Remove(reg.first()); 513 C->FIRST_STACK_mask().Remove(reg.first()+1); // Always occupies two stack slots 514 off += type2size[bt]; 515 } 516 } 517 } 518 519 // Add in all bits past the outgoing argument area 520 guarantee(RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1)), 521 "must be able to represent all call arguments in reg mask"); 522 OptoReg::Name init = _out_arg_limit; 523 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) { 524 C->FIRST_STACK_mask().Insert(i); 525 } 526 // Finally, set the "infinite stack" bit. 527 C->FIRST_STACK_mask().set_AllStack(); 528 529 // Make spill masks. Registers for their class, plus FIRST_STACK_mask. 530 RegMask aligned_stack_mask = C->FIRST_STACK_mask(); 531 // Keep spill masks aligned. 532 aligned_stack_mask.clear_to_pairs(); 533 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack"); 534 535 *idealreg2spillmask[Op_RegP] = *idealreg2regmask[Op_RegP]; 536 #ifdef _LP64 537 *idealreg2spillmask[Op_RegN] = *idealreg2regmask[Op_RegN]; 538 idealreg2spillmask[Op_RegN]->OR(C->FIRST_STACK_mask()); 539 idealreg2spillmask[Op_RegP]->OR(aligned_stack_mask); 540 #else 541 idealreg2spillmask[Op_RegP]->OR(C->FIRST_STACK_mask()); 542 #endif 543 *idealreg2spillmask[Op_RegI] = *idealreg2regmask[Op_RegI]; 544 idealreg2spillmask[Op_RegI]->OR(C->FIRST_STACK_mask()); 545 *idealreg2spillmask[Op_RegL] = *idealreg2regmask[Op_RegL]; 546 idealreg2spillmask[Op_RegL]->OR(aligned_stack_mask); 547 *idealreg2spillmask[Op_RegF] = *idealreg2regmask[Op_RegF]; 548 idealreg2spillmask[Op_RegF]->OR(C->FIRST_STACK_mask()); 549 *idealreg2spillmask[Op_RegD] = *idealreg2regmask[Op_RegD]; 550 idealreg2spillmask[Op_RegD]->OR(aligned_stack_mask); 551 552 if (Matcher::vector_size_supported(T_BYTE,4)) { 553 *idealreg2spillmask[Op_VecS] = *idealreg2regmask[Op_VecS]; 554 idealreg2spillmask[Op_VecS]->OR(C->FIRST_STACK_mask()); 555 } 556 if (Matcher::vector_size_supported(T_FLOAT,2)) { 557 // For VecD we need dual alignment and 8 bytes (2 slots) for spills. 558 // RA guarantees such alignment since it is needed for Double and Long values. 559 *idealreg2spillmask[Op_VecD] = *idealreg2regmask[Op_VecD]; 560 idealreg2spillmask[Op_VecD]->OR(aligned_stack_mask); 561 } 562 if (Matcher::vector_size_supported(T_FLOAT,4)) { 563 // For VecX we need quadro alignment and 16 bytes (4 slots) for spills. 564 // 565 // RA can use input arguments stack slots for spills but until RA 566 // we don't know frame size and offset of input arg stack slots. 567 // 568 // Exclude last input arg stack slots to avoid spilling vectors there 569 // otherwise vector spills could stomp over stack slots in caller frame. 570 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1); 571 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecX); k++) { 572 aligned_stack_mask.Remove(in); 573 in = OptoReg::add(in, -1); 574 } 575 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecX); 576 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack"); 577 *idealreg2spillmask[Op_VecX] = *idealreg2regmask[Op_VecX]; 578 idealreg2spillmask[Op_VecX]->OR(aligned_stack_mask); 579 } 580 if (Matcher::vector_size_supported(T_FLOAT,8)) { 581 // For VecY we need octo alignment and 32 bytes (8 slots) for spills. 582 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1); 583 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecY); k++) { 584 aligned_stack_mask.Remove(in); 585 in = OptoReg::add(in, -1); 586 } 587 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecY); 588 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack"); 589 *idealreg2spillmask[Op_VecY] = *idealreg2regmask[Op_VecY]; 590 idealreg2spillmask[Op_VecY]->OR(aligned_stack_mask); 591 } 592 if (Matcher::vector_size_supported(T_FLOAT,16)) { 593 // For VecZ we need enough alignment and 64 bytes (16 slots) for spills. 594 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1); 595 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecZ); k++) { 596 aligned_stack_mask.Remove(in); 597 in = OptoReg::add(in, -1); 598 } 599 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecZ); 600 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack"); 601 *idealreg2spillmask[Op_VecZ] = *idealreg2regmask[Op_VecZ]; 602 idealreg2spillmask[Op_VecZ]->OR(aligned_stack_mask); 603 } 604 if (UseFPUForSpilling) { 605 // This mask logic assumes that the spill operations are 606 // symmetric and that the registers involved are the same size. 607 // On sparc for instance we may have to use 64 bit moves will 608 // kill 2 registers when used with F0-F31. 609 idealreg2spillmask[Op_RegI]->OR(*idealreg2regmask[Op_RegF]); 610 idealreg2spillmask[Op_RegF]->OR(*idealreg2regmask[Op_RegI]); 611 #ifdef _LP64 612 idealreg2spillmask[Op_RegN]->OR(*idealreg2regmask[Op_RegF]); 613 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]); 614 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]); 615 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegD]); 616 #else 617 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegF]); 618 #ifdef ARM 619 // ARM has support for moving 64bit values between a pair of 620 // integer registers and a double register 621 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]); 622 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]); 623 #endif 624 #endif 625 } 626 627 // Make up debug masks. Any spill slot plus callee-save registers. 628 // Caller-save registers are assumed to be trashable by the various 629 // inline-cache fixup routines. 630 *idealreg2debugmask [Op_RegN]= *idealreg2spillmask[Op_RegN]; 631 *idealreg2debugmask [Op_RegI]= *idealreg2spillmask[Op_RegI]; 632 *idealreg2debugmask [Op_RegL]= *idealreg2spillmask[Op_RegL]; 633 *idealreg2debugmask [Op_RegF]= *idealreg2spillmask[Op_RegF]; 634 *idealreg2debugmask [Op_RegD]= *idealreg2spillmask[Op_RegD]; 635 *idealreg2debugmask [Op_RegP]= *idealreg2spillmask[Op_RegP]; 636 637 *idealreg2mhdebugmask[Op_RegN]= *idealreg2spillmask[Op_RegN]; 638 *idealreg2mhdebugmask[Op_RegI]= *idealreg2spillmask[Op_RegI]; 639 *idealreg2mhdebugmask[Op_RegL]= *idealreg2spillmask[Op_RegL]; 640 *idealreg2mhdebugmask[Op_RegF]= *idealreg2spillmask[Op_RegF]; 641 *idealreg2mhdebugmask[Op_RegD]= *idealreg2spillmask[Op_RegD]; 642 *idealreg2mhdebugmask[Op_RegP]= *idealreg2spillmask[Op_RegP]; 643 644 // Prevent stub compilations from attempting to reference 645 // callee-saved registers from debug info 646 bool exclude_soe = !Compile::current()->is_method_compilation(); 647 648 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) { 649 // registers the caller has to save do not work 650 if( _register_save_policy[i] == 'C' || 651 _register_save_policy[i] == 'A' || 652 (_register_save_policy[i] == 'E' && exclude_soe) ) { 653 idealreg2debugmask [Op_RegN]->Remove(i); 654 idealreg2debugmask [Op_RegI]->Remove(i); // Exclude save-on-call 655 idealreg2debugmask [Op_RegL]->Remove(i); // registers from debug 656 idealreg2debugmask [Op_RegF]->Remove(i); // masks 657 idealreg2debugmask [Op_RegD]->Remove(i); 658 idealreg2debugmask [Op_RegP]->Remove(i); 659 660 idealreg2mhdebugmask[Op_RegN]->Remove(i); 661 idealreg2mhdebugmask[Op_RegI]->Remove(i); 662 idealreg2mhdebugmask[Op_RegL]->Remove(i); 663 idealreg2mhdebugmask[Op_RegF]->Remove(i); 664 idealreg2mhdebugmask[Op_RegD]->Remove(i); 665 idealreg2mhdebugmask[Op_RegP]->Remove(i); 666 } 667 } 668 669 // Subtract the register we use to save the SP for MethodHandle 670 // invokes to from the debug mask. 671 const RegMask save_mask = method_handle_invoke_SP_save_mask(); 672 idealreg2mhdebugmask[Op_RegN]->SUBTRACT(save_mask); 673 idealreg2mhdebugmask[Op_RegI]->SUBTRACT(save_mask); 674 idealreg2mhdebugmask[Op_RegL]->SUBTRACT(save_mask); 675 idealreg2mhdebugmask[Op_RegF]->SUBTRACT(save_mask); 676 idealreg2mhdebugmask[Op_RegD]->SUBTRACT(save_mask); 677 idealreg2mhdebugmask[Op_RegP]->SUBTRACT(save_mask); 678 } 679 680 //---------------------------is_save_on_entry---------------------------------- 681 bool Matcher::is_save_on_entry( int reg ) { 682 return 683 _register_save_policy[reg] == 'E' || 684 _register_save_policy[reg] == 'A' || // Save-on-entry register? 685 // Also save argument registers in the trampolining stubs 686 (C->save_argument_registers() && is_spillable_arg(reg)); 687 } 688 689 //---------------------------Fixup_Save_On_Entry------------------------------- 690 void Matcher::Fixup_Save_On_Entry( ) { 691 init_first_stack_mask(); 692 693 Node *root = C->root(); // Short name for root 694 // Count number of save-on-entry registers. 695 uint soe_cnt = number_of_saved_registers(); 696 uint i; 697 698 // Find the procedure Start Node 699 StartNode *start = C->start(); 700 assert( start, "Expect a start node" ); 701 702 // Save argument registers in the trampolining stubs 703 if( C->save_argument_registers() ) 704 for( i = 0; i < _last_Mach_Reg; i++ ) 705 if( is_spillable_arg(i) ) 706 soe_cnt++; 707 708 // Input RegMask array shared by all Returns. 709 // The type for doubles and longs has a count of 2, but 710 // there is only 1 returned value 711 uint ret_edge_cnt = C->tf()->range_cc()->cnt(); 712 RegMask *ret_rms = init_input_masks( ret_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 713 for (i = TypeFunc::Parms; i < ret_edge_cnt; i++) { 714 ret_rms[i] = _return_values_mask[i-TypeFunc::Parms]; 715 } 716 717 // Input RegMask array shared by all Rethrows. 718 uint reth_edge_cnt = TypeFunc::Parms+1; 719 RegMask *reth_rms = init_input_masks( reth_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 720 // Rethrow takes exception oop only, but in the argument 0 slot. 721 OptoReg::Name reg = find_receiver(false); 722 if (reg >= 0) { 723 reth_rms[TypeFunc::Parms] = mreg2regmask[reg]; 724 #ifdef _LP64 725 // Need two slots for ptrs in 64-bit land 726 reth_rms[TypeFunc::Parms].Insert(OptoReg::add(OptoReg::Name(reg), 1)); 727 #endif 728 } 729 730 // Input RegMask array shared by all TailCalls 731 uint tail_call_edge_cnt = TypeFunc::Parms+2; 732 RegMask *tail_call_rms = init_input_masks( tail_call_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 733 734 // Input RegMask array shared by all TailJumps 735 uint tail_jump_edge_cnt = TypeFunc::Parms+2; 736 RegMask *tail_jump_rms = init_input_masks( tail_jump_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 737 738 // TailCalls have 2 returned values (target & moop), whose masks come 739 // from the usual MachNode/MachOper mechanism. Find a sample 740 // TailCall to extract these masks and put the correct masks into 741 // the tail_call_rms array. 742 for( i=1; i < root->req(); i++ ) { 743 MachReturnNode *m = root->in(i)->as_MachReturn(); 744 if( m->ideal_Opcode() == Op_TailCall ) { 745 tail_call_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0); 746 tail_call_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1); 747 break; 748 } 749 } 750 751 // TailJumps have 2 returned values (target & ex_oop), whose masks come 752 // from the usual MachNode/MachOper mechanism. Find a sample 753 // TailJump to extract these masks and put the correct masks into 754 // the tail_jump_rms array. 755 for( i=1; i < root->req(); i++ ) { 756 MachReturnNode *m = root->in(i)->as_MachReturn(); 757 if( m->ideal_Opcode() == Op_TailJump ) { 758 tail_jump_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0); 759 tail_jump_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1); 760 break; 761 } 762 } 763 764 // Input RegMask array shared by all Halts 765 uint halt_edge_cnt = TypeFunc::Parms; 766 RegMask *halt_rms = init_input_masks( halt_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 767 768 // Capture the return input masks into each exit flavor 769 for( i=1; i < root->req(); i++ ) { 770 MachReturnNode *exit = root->in(i)->as_MachReturn(); 771 switch( exit->ideal_Opcode() ) { 772 case Op_Return : exit->_in_rms = ret_rms; break; 773 case Op_Rethrow : exit->_in_rms = reth_rms; break; 774 case Op_TailCall : exit->_in_rms = tail_call_rms; break; 775 case Op_TailJump : exit->_in_rms = tail_jump_rms; break; 776 case Op_Halt : exit->_in_rms = halt_rms; break; 777 default : ShouldNotReachHere(); 778 } 779 } 780 781 // Next unused projection number from Start. 782 int proj_cnt = C->tf()->domain_cc()->cnt(); 783 784 // Do all the save-on-entry registers. Make projections from Start for 785 // them, and give them a use at the exit points. To the allocator, they 786 // look like incoming register arguments. 787 for( i = 0; i < _last_Mach_Reg; i++ ) { 788 if( is_save_on_entry(i) ) { 789 790 // Add the save-on-entry to the mask array 791 ret_rms [ ret_edge_cnt] = mreg2regmask[i]; 792 reth_rms [ reth_edge_cnt] = mreg2regmask[i]; 793 tail_call_rms[tail_call_edge_cnt] = mreg2regmask[i]; 794 tail_jump_rms[tail_jump_edge_cnt] = mreg2regmask[i]; 795 // Halts need the SOE registers, but only in the stack as debug info. 796 // A just-prior uncommon-trap or deoptimization will use the SOE regs. 797 halt_rms [ halt_edge_cnt] = *idealreg2spillmask[_register_save_type[i]]; 798 799 Node *mproj; 800 801 // Is this a RegF low half of a RegD? Double up 2 adjacent RegF's 802 // into a single RegD. 803 if( (i&1) == 0 && 804 _register_save_type[i ] == Op_RegF && 805 _register_save_type[i+1] == Op_RegF && 806 is_save_on_entry(i+1) ) { 807 // Add other bit for double 808 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1)); 809 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1)); 810 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1)); 811 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1)); 812 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1)); 813 mproj = new MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegD ); 814 proj_cnt += 2; // Skip 2 for doubles 815 } 816 else if( (i&1) == 1 && // Else check for high half of double 817 _register_save_type[i-1] == Op_RegF && 818 _register_save_type[i ] == Op_RegF && 819 is_save_on_entry(i-1) ) { 820 ret_rms [ ret_edge_cnt] = RegMask::Empty; 821 reth_rms [ reth_edge_cnt] = RegMask::Empty; 822 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty; 823 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty; 824 halt_rms [ halt_edge_cnt] = RegMask::Empty; 825 mproj = C->top(); 826 } 827 // Is this a RegI low half of a RegL? Double up 2 adjacent RegI's 828 // into a single RegL. 829 else if( (i&1) == 0 && 830 _register_save_type[i ] == Op_RegI && 831 _register_save_type[i+1] == Op_RegI && 832 is_save_on_entry(i+1) ) { 833 // Add other bit for long 834 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1)); 835 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1)); 836 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1)); 837 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1)); 838 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1)); 839 mproj = new MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegL ); 840 proj_cnt += 2; // Skip 2 for longs 841 } 842 else if( (i&1) == 1 && // Else check for high half of long 843 _register_save_type[i-1] == Op_RegI && 844 _register_save_type[i ] == Op_RegI && 845 is_save_on_entry(i-1) ) { 846 ret_rms [ ret_edge_cnt] = RegMask::Empty; 847 reth_rms [ reth_edge_cnt] = RegMask::Empty; 848 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty; 849 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty; 850 halt_rms [ halt_edge_cnt] = RegMask::Empty; 851 mproj = C->top(); 852 } else { 853 // Make a projection for it off the Start 854 mproj = new MachProjNode( start, proj_cnt++, ret_rms[ret_edge_cnt], _register_save_type[i] ); 855 } 856 857 ret_edge_cnt ++; 858 reth_edge_cnt ++; 859 tail_call_edge_cnt ++; 860 tail_jump_edge_cnt ++; 861 halt_edge_cnt ++; 862 863 // Add a use of the SOE register to all exit paths 864 for( uint j=1; j < root->req(); j++ ) 865 root->in(j)->add_req(mproj); 866 } // End of if a save-on-entry register 867 } // End of for all machine registers 868 } 869 870 //------------------------------init_spill_mask-------------------------------- 871 void Matcher::init_spill_mask( Node *ret ) { 872 if( idealreg2regmask[Op_RegI] ) return; // One time only init 873 874 OptoReg::c_frame_pointer = c_frame_pointer(); 875 c_frame_ptr_mask = c_frame_pointer(); 876 #ifdef _LP64 877 // pointers are twice as big 878 c_frame_ptr_mask.Insert(OptoReg::add(c_frame_pointer(),1)); 879 #endif 880 881 // Start at OptoReg::stack0() 882 STACK_ONLY_mask.Clear(); 883 OptoReg::Name init = OptoReg::stack2reg(0); 884 // STACK_ONLY_mask is all stack bits 885 OptoReg::Name i; 886 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) 887 STACK_ONLY_mask.Insert(i); 888 // Also set the "infinite stack" bit. 889 STACK_ONLY_mask.set_AllStack(); 890 891 // Copy the register names over into the shared world 892 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) { 893 // SharedInfo::regName[i] = regName[i]; 894 // Handy RegMasks per machine register 895 mreg2regmask[i].Insert(i); 896 } 897 898 // Grab the Frame Pointer 899 Node *fp = ret->in(TypeFunc::FramePtr); 900 Node *mem = ret->in(TypeFunc::Memory); 901 const TypePtr* atp = TypePtr::BOTTOM; 902 // Share frame pointer while making spill ops 903 set_shared(fp); 904 905 // Compute generic short-offset Loads 906 #ifdef _LP64 907 MachNode *spillCP = match_tree(new LoadNNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM,MemNode::unordered)); 908 #endif 909 MachNode *spillI = match_tree(new LoadINode(NULL,mem,fp,atp,TypeInt::INT,MemNode::unordered)); 910 MachNode *spillL = match_tree(new LoadLNode(NULL,mem,fp,atp,TypeLong::LONG,MemNode::unordered, LoadNode::DependsOnlyOnTest, false)); 911 MachNode *spillF = match_tree(new LoadFNode(NULL,mem,fp,atp,Type::FLOAT,MemNode::unordered)); 912 MachNode *spillD = match_tree(new LoadDNode(NULL,mem,fp,atp,Type::DOUBLE,MemNode::unordered)); 913 MachNode *spillP = match_tree(new LoadPNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM,MemNode::unordered)); 914 assert(spillI != NULL && spillL != NULL && spillF != NULL && 915 spillD != NULL && spillP != NULL, ""); 916 // Get the ADLC notion of the right regmask, for each basic type. 917 #ifdef _LP64 918 idealreg2regmask[Op_RegN] = &spillCP->out_RegMask(); 919 #endif 920 idealreg2regmask[Op_RegI] = &spillI->out_RegMask(); 921 idealreg2regmask[Op_RegL] = &spillL->out_RegMask(); 922 idealreg2regmask[Op_RegF] = &spillF->out_RegMask(); 923 idealreg2regmask[Op_RegD] = &spillD->out_RegMask(); 924 idealreg2regmask[Op_RegP] = &spillP->out_RegMask(); 925 926 // Vector regmasks. 927 if (Matcher::vector_size_supported(T_BYTE,4)) { 928 TypeVect::VECTS = TypeVect::make(T_BYTE, 4); 929 MachNode *spillVectS = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTS)); 930 idealreg2regmask[Op_VecS] = &spillVectS->out_RegMask(); 931 } 932 if (Matcher::vector_size_supported(T_FLOAT,2)) { 933 MachNode *spillVectD = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTD)); 934 idealreg2regmask[Op_VecD] = &spillVectD->out_RegMask(); 935 } 936 if (Matcher::vector_size_supported(T_FLOAT,4)) { 937 MachNode *spillVectX = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTX)); 938 idealreg2regmask[Op_VecX] = &spillVectX->out_RegMask(); 939 } 940 if (Matcher::vector_size_supported(T_FLOAT,8)) { 941 MachNode *spillVectY = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTY)); 942 idealreg2regmask[Op_VecY] = &spillVectY->out_RegMask(); 943 } 944 if (Matcher::vector_size_supported(T_FLOAT,16)) { 945 MachNode *spillVectZ = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTZ)); 946 idealreg2regmask[Op_VecZ] = &spillVectZ->out_RegMask(); 947 } 948 } 949 950 #ifdef ASSERT 951 static void match_alias_type(Compile* C, Node* n, Node* m) { 952 if (!VerifyAliases) return; // do not go looking for trouble by default 953 const TypePtr* nat = n->adr_type(); 954 const TypePtr* mat = m->adr_type(); 955 int nidx = C->get_alias_index(nat); 956 int midx = C->get_alias_index(mat); 957 // Detune the assert for cases like (AndI 0xFF (LoadB p)). 958 if (nidx == Compile::AliasIdxTop && midx >= Compile::AliasIdxRaw) { 959 for (uint i = 1; i < n->req(); i++) { 960 Node* n1 = n->in(i); 961 const TypePtr* n1at = n1->adr_type(); 962 if (n1at != NULL) { 963 nat = n1at; 964 nidx = C->get_alias_index(n1at); 965 } 966 } 967 } 968 // %%% Kludgery. Instead, fix ideal adr_type methods for all these cases: 969 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxRaw) { 970 switch (n->Opcode()) { 971 case Op_PrefetchAllocation: 972 nidx = Compile::AliasIdxRaw; 973 nat = TypeRawPtr::BOTTOM; 974 break; 975 } 976 } 977 if (nidx == Compile::AliasIdxRaw && midx == Compile::AliasIdxTop) { 978 switch (n->Opcode()) { 979 case Op_ClearArray: 980 midx = Compile::AliasIdxRaw; 981 mat = TypeRawPtr::BOTTOM; 982 break; 983 } 984 } 985 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxBot) { 986 switch (n->Opcode()) { 987 case Op_Return: 988 case Op_Rethrow: 989 case Op_Halt: 990 case Op_TailCall: 991 case Op_TailJump: 992 nidx = Compile::AliasIdxBot; 993 nat = TypePtr::BOTTOM; 994 break; 995 } 996 } 997 if (nidx == Compile::AliasIdxBot && midx == Compile::AliasIdxTop) { 998 switch (n->Opcode()) { 999 case Op_StrComp: 1000 case Op_StrEquals: 1001 case Op_StrIndexOf: 1002 case Op_StrIndexOfChar: 1003 case Op_AryEq: 1004 case Op_HasNegatives: 1005 case Op_MemBarVolatile: 1006 case Op_MemBarCPUOrder: // %%% these ideals should have narrower adr_type? 1007 case Op_StrInflatedCopy: 1008 case Op_StrCompressedCopy: 1009 case Op_OnSpinWait: 1010 case Op_EncodeISOArray: 1011 nidx = Compile::AliasIdxTop; 1012 nat = NULL; 1013 break; 1014 } 1015 } 1016 if (nidx != midx) { 1017 if (PrintOpto || (PrintMiscellaneous && (WizardMode || Verbose))) { 1018 tty->print_cr("==== Matcher alias shift %d => %d", nidx, midx); 1019 n->dump(); 1020 m->dump(); 1021 } 1022 assert(C->subsume_loads() && C->must_alias(nat, midx), 1023 "must not lose alias info when matching"); 1024 } 1025 } 1026 #endif 1027 1028 //------------------------------xform------------------------------------------ 1029 // Given a Node in old-space, Match him (Label/Reduce) to produce a machine 1030 // Node in new-space. Given a new-space Node, recursively walk his children. 1031 Node *Matcher::transform( Node *n ) { ShouldNotCallThis(); return n; } 1032 Node *Matcher::xform( Node *n, int max_stack ) { 1033 // Use one stack to keep both: child's node/state and parent's node/index 1034 MStack mstack(max_stack * 2 * 2); // usually: C->live_nodes() * 2 * 2 1035 mstack.push(n, Visit, NULL, -1); // set NULL as parent to indicate root 1036 while (mstack.is_nonempty()) { 1037 C->check_node_count(NodeLimitFudgeFactor, "too many nodes matching instructions"); 1038 if (C->failing()) return NULL; 1039 n = mstack.node(); // Leave node on stack 1040 Node_State nstate = mstack.state(); 1041 if (nstate == Visit) { 1042 mstack.set_state(Post_Visit); 1043 Node *oldn = n; 1044 // Old-space or new-space check 1045 if (!C->node_arena()->contains(n)) { 1046 // Old space! 1047 Node* m; 1048 if (has_new_node(n)) { // Not yet Label/Reduced 1049 m = new_node(n); 1050 } else { 1051 if (!is_dontcare(n)) { // Matcher can match this guy 1052 // Calls match special. They match alone with no children. 1053 // Their children, the incoming arguments, match normally. 1054 m = n->is_SafePoint() ? match_sfpt(n->as_SafePoint()):match_tree(n); 1055 if (C->failing()) return NULL; 1056 if (m == NULL) { Matcher::soft_match_failure(); return NULL; } 1057 if (n->is_MemBar()) { 1058 m->as_MachMemBar()->set_adr_type(n->adr_type()); 1059 } 1060 } else { // Nothing the matcher cares about 1061 if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Multi()) { // Projections? 1062 // Convert to machine-dependent projection 1063 RegMask* mask = NULL; 1064 if (n->in(0)->is_Call()) { 1065 mask = return_values_mask(n->in(0)->as_Call()->tf()->range_cc()); 1066 } 1067 m = n->in(0)->as_Multi()->match(n->as_Proj(), this, mask); 1068 #ifdef ASSERT 1069 _new2old_map.map(m->_idx, n); 1070 #endif 1071 if (m->in(0) != NULL) // m might be top 1072 collect_null_checks(m, n); 1073 } else { // Else just a regular 'ol guy 1074 m = n->clone(); // So just clone into new-space 1075 #ifdef ASSERT 1076 _new2old_map.map(m->_idx, n); 1077 #endif 1078 // Def-Use edges will be added incrementally as Uses 1079 // of this node are matched. 1080 assert(m->outcnt() == 0, "no Uses of this clone yet"); 1081 } 1082 } 1083 1084 set_new_node(n, m); // Map old to new 1085 if (_old_node_note_array != NULL) { 1086 Node_Notes* nn = C->locate_node_notes(_old_node_note_array, 1087 n->_idx); 1088 C->set_node_notes_at(m->_idx, nn); 1089 } 1090 debug_only(match_alias_type(C, n, m)); 1091 } 1092 n = m; // n is now a new-space node 1093 mstack.set_node(n); 1094 } 1095 1096 // New space! 1097 if (_visited.test_set(n->_idx)) continue; // while(mstack.is_nonempty()) 1098 1099 int i; 1100 // Put precedence edges on stack first (match them last). 1101 for (i = oldn->req(); (uint)i < oldn->len(); i++) { 1102 Node *m = oldn->in(i); 1103 if (m == NULL) break; 1104 // set -1 to call add_prec() instead of set_req() during Step1 1105 mstack.push(m, Visit, n, -1); 1106 } 1107 1108 // Handle precedence edges for interior nodes 1109 for (i = n->len()-1; (uint)i >= n->req(); i--) { 1110 Node *m = n->in(i); 1111 if (m == NULL || C->node_arena()->contains(m)) continue; 1112 n->rm_prec(i); 1113 // set -1 to call add_prec() instead of set_req() during Step1 1114 mstack.push(m, Visit, n, -1); 1115 } 1116 1117 // For constant debug info, I'd rather have unmatched constants. 1118 int cnt = n->req(); 1119 JVMState* jvms = n->jvms(); 1120 int debug_cnt = jvms ? jvms->debug_start() : cnt; 1121 1122 // Now do only debug info. Clone constants rather than matching. 1123 // Constants are represented directly in the debug info without 1124 // the need for executable machine instructions. 1125 // Monitor boxes are also represented directly. 1126 for (i = cnt - 1; i >= debug_cnt; --i) { // For all debug inputs do 1127 Node *m = n->in(i); // Get input 1128 int op = m->Opcode(); 1129 assert((op == Op_BoxLock) == jvms->is_monitor_use(i), "boxes only at monitor sites"); 1130 if( op == Op_ConI || op == Op_ConP || op == Op_ConN || op == Op_ConNKlass || 1131 op == Op_ConF || op == Op_ConD || op == Op_ConL 1132 // || op == Op_BoxLock // %%%% enable this and remove (+++) in chaitin.cpp 1133 ) { 1134 m = m->clone(); 1135 #ifdef ASSERT 1136 _new2old_map.map(m->_idx, n); 1137 #endif 1138 mstack.push(m, Post_Visit, n, i); // Don't need to visit 1139 mstack.push(m->in(0), Visit, m, 0); 1140 } else { 1141 mstack.push(m, Visit, n, i); 1142 } 1143 } 1144 1145 // And now walk his children, and convert his inputs to new-space. 1146 for( ; i >= 0; --i ) { // For all normal inputs do 1147 Node *m = n->in(i); // Get input 1148 if(m != NULL) 1149 mstack.push(m, Visit, n, i); 1150 } 1151 1152 } 1153 else if (nstate == Post_Visit) { 1154 // Set xformed input 1155 Node *p = mstack.parent(); 1156 if (p != NULL) { // root doesn't have parent 1157 int i = (int)mstack.index(); 1158 if (i >= 0) 1159 p->set_req(i, n); // required input 1160 else if (i == -1) 1161 p->add_prec(n); // precedence input 1162 else 1163 ShouldNotReachHere(); 1164 } 1165 mstack.pop(); // remove processed node from stack 1166 } 1167 else { 1168 ShouldNotReachHere(); 1169 } 1170 } // while (mstack.is_nonempty()) 1171 return n; // Return new-space Node 1172 } 1173 1174 //------------------------------warp_outgoing_stk_arg------------------------ 1175 OptoReg::Name Matcher::warp_outgoing_stk_arg( VMReg reg, OptoReg::Name begin_out_arg_area, OptoReg::Name &out_arg_limit_per_call ) { 1176 // Convert outgoing argument location to a pre-biased stack offset 1177 if (reg->is_stack()) { 1178 OptoReg::Name warped = reg->reg2stack(); 1179 // Adjust the stack slot offset to be the register number used 1180 // by the allocator. 1181 warped = OptoReg::add(begin_out_arg_area, warped); 1182 // Keep track of the largest numbered stack slot used for an arg. 1183 // Largest used slot per call-site indicates the amount of stack 1184 // that is killed by the call. 1185 if( warped >= out_arg_limit_per_call ) 1186 out_arg_limit_per_call = OptoReg::add(warped,1); 1187 if (!RegMask::can_represent_arg(warped)) { 1188 C->record_method_not_compilable("unsupported calling sequence"); 1189 return OptoReg::Bad; 1190 } 1191 return warped; 1192 } 1193 return OptoReg::as_OptoReg(reg); 1194 } 1195 1196 1197 //------------------------------match_sfpt------------------------------------- 1198 // Helper function to match call instructions. Calls match special. 1199 // They match alone with no children. Their children, the incoming 1200 // arguments, match normally. 1201 MachNode *Matcher::match_sfpt( SafePointNode *sfpt ) { 1202 MachSafePointNode *msfpt = NULL; 1203 MachCallNode *mcall = NULL; 1204 uint cnt; 1205 // Split out case for SafePoint vs Call 1206 CallNode *call; 1207 const TypeTuple *domain; 1208 ciMethod* method = NULL; 1209 bool is_method_handle_invoke = false; // for special kill effects 1210 if( sfpt->is_Call() ) { 1211 call = sfpt->as_Call(); 1212 domain = call->tf()->domain_cc(); 1213 cnt = domain->cnt(); 1214 1215 // Match just the call, nothing else 1216 MachNode *m = match_tree(call); 1217 if (C->failing()) return NULL; 1218 if( m == NULL ) { Matcher::soft_match_failure(); return NULL; } 1219 1220 // Copy data from the Ideal SafePoint to the machine version 1221 mcall = m->as_MachCall(); 1222 1223 mcall->set_tf( call->tf()); 1224 mcall->set_entry_point(call->entry_point()); 1225 mcall->set_cnt( call->cnt()); 1226 1227 if( mcall->is_MachCallJava() ) { 1228 MachCallJavaNode *mcall_java = mcall->as_MachCallJava(); 1229 const CallJavaNode *call_java = call->as_CallJava(); 1230 assert(call_java->validate_symbolic_info(), "inconsistent info"); 1231 method = call_java->method(); 1232 mcall_java->_method = method; 1233 mcall_java->_bci = call_java->_bci; 1234 mcall_java->_optimized_virtual = call_java->is_optimized_virtual(); 1235 is_method_handle_invoke = call_java->is_method_handle_invoke(); 1236 mcall_java->_method_handle_invoke = is_method_handle_invoke; 1237 mcall_java->_override_symbolic_info = call_java->override_symbolic_info(); 1238 if (is_method_handle_invoke) { 1239 C->set_has_method_handle_invokes(true); 1240 } 1241 if( mcall_java->is_MachCallStaticJava() ) 1242 mcall_java->as_MachCallStaticJava()->_name = 1243 call_java->as_CallStaticJava()->_name; 1244 if( mcall_java->is_MachCallDynamicJava() ) 1245 mcall_java->as_MachCallDynamicJava()->_vtable_index = 1246 call_java->as_CallDynamicJava()->_vtable_index; 1247 } 1248 else if( mcall->is_MachCallRuntime() ) { 1249 mcall->as_MachCallRuntime()->_name = call->as_CallRuntime()->_name; 1250 } 1251 msfpt = mcall; 1252 } 1253 // This is a non-call safepoint 1254 else { 1255 call = NULL; 1256 domain = NULL; 1257 MachNode *mn = match_tree(sfpt); 1258 if (C->failing()) return NULL; 1259 msfpt = mn->as_MachSafePoint(); 1260 cnt = TypeFunc::Parms; 1261 } 1262 1263 // Advertise the correct memory effects (for anti-dependence computation). 1264 msfpt->set_adr_type(sfpt->adr_type()); 1265 1266 // Allocate a private array of RegMasks. These RegMasks are not shared. 1267 msfpt->_in_rms = NEW_RESOURCE_ARRAY( RegMask, cnt ); 1268 // Empty them all. 1269 for (uint i = 0; i < cnt; i++) ::new (&(msfpt->_in_rms[i])) RegMask(); 1270 1271 // Do all the pre-defined non-Empty register masks 1272 msfpt->_in_rms[TypeFunc::ReturnAdr] = _return_addr_mask; 1273 msfpt->_in_rms[TypeFunc::FramePtr ] = c_frame_ptr_mask; 1274 1275 // Place first outgoing argument can possibly be put. 1276 OptoReg::Name begin_out_arg_area = OptoReg::add(_new_SP, C->out_preserve_stack_slots()); 1277 assert( is_even(begin_out_arg_area), "" ); 1278 // Compute max outgoing register number per call site. 1279 OptoReg::Name out_arg_limit_per_call = begin_out_arg_area; 1280 // Calls to C may hammer extra stack slots above and beyond any arguments. 1281 // These are usually backing store for register arguments for varargs. 1282 if( call != NULL && call->is_CallRuntime() ) 1283 out_arg_limit_per_call = OptoReg::add(out_arg_limit_per_call,C->varargs_C_out_slots_killed()); 1284 1285 1286 // Do the normal argument list (parameters) register masks 1287 // Null entry point is a special cast where the target of the call 1288 // is in a register. 1289 int adj = (call != NULL && call->entry_point() == NULL) ? 1 : 0; 1290 int argcnt = cnt - TypeFunc::Parms - adj; 1291 if( argcnt > 0 ) { // Skip it all if we have no args 1292 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt ); 1293 VMRegPair *parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt ); 1294 int i; 1295 for( i = 0; i < argcnt; i++ ) { 1296 sig_bt[i] = domain->field_at(i+TypeFunc::Parms+adj)->basic_type(); 1297 } 1298 // V-call to pick proper calling convention 1299 call->calling_convention( sig_bt, parm_regs, argcnt ); 1300 1301 #ifdef ASSERT 1302 // Sanity check users' calling convention. Really handy during 1303 // the initial porting effort. Fairly expensive otherwise. 1304 { for (int i = 0; i<argcnt; i++) { 1305 if( !parm_regs[i].first()->is_valid() && 1306 !parm_regs[i].second()->is_valid() ) continue; 1307 VMReg reg1 = parm_regs[i].first(); 1308 VMReg reg2 = parm_regs[i].second(); 1309 for (int j = 0; j < i; j++) { 1310 if( !parm_regs[j].first()->is_valid() && 1311 !parm_regs[j].second()->is_valid() ) continue; 1312 VMReg reg3 = parm_regs[j].first(); 1313 VMReg reg4 = parm_regs[j].second(); 1314 if( !reg1->is_valid() ) { 1315 assert( !reg2->is_valid(), "valid halvsies" ); 1316 } else if( !reg3->is_valid() ) { 1317 assert( !reg4->is_valid(), "valid halvsies" ); 1318 } else { 1319 assert( reg1 != reg2, "calling conv. must produce distinct regs"); 1320 assert( reg1 != reg3, "calling conv. must produce distinct regs"); 1321 assert( reg1 != reg4, "calling conv. must produce distinct regs"); 1322 assert( reg2 != reg3, "calling conv. must produce distinct regs"); 1323 assert( reg2 != reg4 || !reg2->is_valid(), "calling conv. must produce distinct regs"); 1324 assert( reg3 != reg4, "calling conv. must produce distinct regs"); 1325 } 1326 } 1327 } 1328 } 1329 #endif 1330 1331 // Visit each argument. Compute its outgoing register mask. 1332 // Return results now can have 2 bits returned. 1333 // Compute max over all outgoing arguments both per call-site 1334 // and over the entire method. 1335 for( i = 0; i < argcnt; i++ ) { 1336 // Address of incoming argument mask to fill in 1337 RegMask *rm = &mcall->_in_rms[i+TypeFunc::Parms+adj]; 1338 if( !parm_regs[i].first()->is_valid() && 1339 !parm_regs[i].second()->is_valid() ) { 1340 continue; // Avoid Halves 1341 } 1342 // Grab first register, adjust stack slots and insert in mask. 1343 OptoReg::Name reg1 = warp_outgoing_stk_arg(parm_regs[i].first(), begin_out_arg_area, out_arg_limit_per_call ); 1344 if (OptoReg::is_valid(reg1)) { 1345 rm->Insert( reg1 ); 1346 } 1347 // Grab second register (if any), adjust stack slots and insert in mask. 1348 OptoReg::Name reg2 = warp_outgoing_stk_arg(parm_regs[i].second(), begin_out_arg_area, out_arg_limit_per_call ); 1349 if (OptoReg::is_valid(reg2)) { 1350 rm->Insert( reg2 ); 1351 } 1352 } // End of for all arguments 1353 1354 // Compute number of stack slots needed to restore stack in case of 1355 // Pascal-style argument popping. 1356 mcall->_argsize = out_arg_limit_per_call - begin_out_arg_area; 1357 } 1358 1359 // Compute the max stack slot killed by any call. These will not be 1360 // available for debug info, and will be used to adjust FIRST_STACK_mask 1361 // after all call sites have been visited. 1362 if( _out_arg_limit < out_arg_limit_per_call) 1363 _out_arg_limit = out_arg_limit_per_call; 1364 1365 if (mcall) { 1366 // Kill the outgoing argument area, including any non-argument holes and 1367 // any legacy C-killed slots. Use Fat-Projections to do the killing. 1368 // Since the max-per-method covers the max-per-call-site and debug info 1369 // is excluded on the max-per-method basis, debug info cannot land in 1370 // this killed area. 1371 uint r_cnt = mcall->tf()->range_sig()->cnt(); 1372 MachProjNode *proj = new MachProjNode( mcall, r_cnt+10000, RegMask::Empty, MachProjNode::fat_proj ); 1373 if (!RegMask::can_represent_arg(OptoReg::Name(out_arg_limit_per_call-1))) { 1374 C->record_method_not_compilable("unsupported outgoing calling sequence"); 1375 } else { 1376 for (int i = begin_out_arg_area; i < out_arg_limit_per_call; i++) 1377 proj->_rout.Insert(OptoReg::Name(i)); 1378 } 1379 if (proj->_rout.is_NotEmpty()) { 1380 push_projection(proj); 1381 } 1382 } 1383 // Transfer the safepoint information from the call to the mcall 1384 // Move the JVMState list 1385 msfpt->set_jvms(sfpt->jvms()); 1386 for (JVMState* jvms = msfpt->jvms(); jvms; jvms = jvms->caller()) { 1387 jvms->set_map(sfpt); 1388 } 1389 1390 // Debug inputs begin just after the last incoming parameter 1391 assert((mcall == NULL) || (mcall->jvms() == NULL) || 1392 (mcall->jvms()->debug_start() + mcall->_jvmadj == mcall->tf()->domain_cc()->cnt()), ""); 1393 1394 // Move the OopMap 1395 msfpt->_oop_map = sfpt->_oop_map; 1396 1397 // Add additional edges. 1398 if (msfpt->mach_constant_base_node_input() != (uint)-1 && !msfpt->is_MachCallLeaf()) { 1399 // For these calls we can not add MachConstantBase in expand(), as the 1400 // ins are not complete then. 1401 msfpt->ins_req(msfpt->mach_constant_base_node_input(), C->mach_constant_base_node()); 1402 if (msfpt->jvms() && 1403 msfpt->mach_constant_base_node_input() <= msfpt->jvms()->debug_start() + msfpt->_jvmadj) { 1404 // We added an edge before jvms, so we must adapt the position of the ins. 1405 msfpt->jvms()->adapt_position(+1); 1406 } 1407 } 1408 1409 // Registers killed by the call are set in the local scheduling pass 1410 // of Global Code Motion. 1411 return msfpt; 1412 } 1413 1414 //---------------------------match_tree---------------------------------------- 1415 // Match a Ideal Node DAG - turn it into a tree; Label & Reduce. Used as part 1416 // of the whole-sale conversion from Ideal to Mach Nodes. Also used for 1417 // making GotoNodes while building the CFG and in init_spill_mask() to identify 1418 // a Load's result RegMask for memoization in idealreg2regmask[] 1419 MachNode *Matcher::match_tree( const Node *n ) { 1420 assert( n->Opcode() != Op_Phi, "cannot match" ); 1421 assert( !n->is_block_start(), "cannot match" ); 1422 // Set the mark for all locally allocated State objects. 1423 // When this call returns, the _states_arena arena will be reset 1424 // freeing all State objects. 1425 ResourceMark rm( &_states_arena ); 1426 1427 LabelRootDepth = 0; 1428 1429 // StoreNodes require their Memory input to match any LoadNodes 1430 Node *mem = n->is_Store() ? n->in(MemNode::Memory) : (Node*)1 ; 1431 #ifdef ASSERT 1432 Node* save_mem_node = _mem_node; 1433 _mem_node = n->is_Store() ? (Node*)n : NULL; 1434 #endif 1435 // State object for root node of match tree 1436 // Allocate it on _states_arena - stack allocation can cause stack overflow. 1437 State *s = new (&_states_arena) State; 1438 s->_kids[0] = NULL; 1439 s->_kids[1] = NULL; 1440 s->_leaf = (Node*)n; 1441 // Label the input tree, allocating labels from top-level arena 1442 Label_Root( n, s, n->in(0), mem ); 1443 if (C->failing()) return NULL; 1444 1445 // The minimum cost match for the whole tree is found at the root State 1446 uint mincost = max_juint; 1447 uint cost = max_juint; 1448 uint i; 1449 for( i = 0; i < NUM_OPERANDS; i++ ) { 1450 if( s->valid(i) && // valid entry and 1451 s->_cost[i] < cost && // low cost and 1452 s->_rule[i] >= NUM_OPERANDS ) // not an operand 1453 cost = s->_cost[mincost=i]; 1454 } 1455 if (mincost == max_juint) { 1456 #ifndef PRODUCT 1457 tty->print("No matching rule for:"); 1458 s->dump(); 1459 #endif 1460 Matcher::soft_match_failure(); 1461 return NULL; 1462 } 1463 // Reduce input tree based upon the state labels to machine Nodes 1464 MachNode *m = ReduceInst( s, s->_rule[mincost], mem ); 1465 #ifdef ASSERT 1466 _old2new_map.map(n->_idx, m); 1467 _new2old_map.map(m->_idx, (Node*)n); 1468 #endif 1469 1470 // Add any Matcher-ignored edges 1471 uint cnt = n->req(); 1472 uint start = 1; 1473 if( mem != (Node*)1 ) start = MemNode::Memory+1; 1474 if( n->is_AddP() ) { 1475 assert( mem == (Node*)1, "" ); 1476 start = AddPNode::Base+1; 1477 } 1478 for( i = start; i < cnt; i++ ) { 1479 if( !n->match_edge(i) ) { 1480 if( i < m->req() ) 1481 m->ins_req( i, n->in(i) ); 1482 else 1483 m->add_req( n->in(i) ); 1484 } 1485 } 1486 1487 debug_only( _mem_node = save_mem_node; ) 1488 return m; 1489 } 1490 1491 1492 //------------------------------match_into_reg--------------------------------- 1493 // Choose to either match this Node in a register or part of the current 1494 // match tree. Return true for requiring a register and false for matching 1495 // as part of the current match tree. 1496 static bool match_into_reg( const Node *n, Node *m, Node *control, int i, bool shared ) { 1497 1498 const Type *t = m->bottom_type(); 1499 1500 if (t->singleton()) { 1501 // Never force constants into registers. Allow them to match as 1502 // constants or registers. Copies of the same value will share 1503 // the same register. See find_shared_node. 1504 return false; 1505 } else { // Not a constant 1506 // Stop recursion if they have different Controls. 1507 Node* m_control = m->in(0); 1508 // Control of load's memory can post-dominates load's control. 1509 // So use it since load can't float above its memory. 1510 Node* mem_control = (m->is_Load()) ? m->in(MemNode::Memory)->in(0) : NULL; 1511 if (control && m_control && control != m_control && control != mem_control) { 1512 1513 // Actually, we can live with the most conservative control we 1514 // find, if it post-dominates the others. This allows us to 1515 // pick up load/op/store trees where the load can float a little 1516 // above the store. 1517 Node *x = control; 1518 const uint max_scan = 6; // Arbitrary scan cutoff 1519 uint j; 1520 for (j=0; j<max_scan; j++) { 1521 if (x->is_Region()) // Bail out at merge points 1522 return true; 1523 x = x->in(0); 1524 if (x == m_control) // Does 'control' post-dominate 1525 break; // m->in(0)? If so, we can use it 1526 if (x == mem_control) // Does 'control' post-dominate 1527 break; // mem_control? If so, we can use it 1528 } 1529 if (j == max_scan) // No post-domination before scan end? 1530 return true; // Then break the match tree up 1531 } 1532 if ((m->is_DecodeN() && Matcher::narrow_oop_use_complex_address()) || 1533 (m->is_DecodeNKlass() && Matcher::narrow_klass_use_complex_address())) { 1534 // These are commonly used in address expressions and can 1535 // efficiently fold into them on X64 in some cases. 1536 return false; 1537 } 1538 } 1539 1540 // Not forceable cloning. If shared, put it into a register. 1541 return shared; 1542 } 1543 1544 1545 //------------------------------Instruction Selection-------------------------- 1546 // Label method walks a "tree" of nodes, using the ADLC generated DFA to match 1547 // ideal nodes to machine instructions. Trees are delimited by shared Nodes, 1548 // things the Matcher does not match (e.g., Memory), and things with different 1549 // Controls (hence forced into different blocks). We pass in the Control 1550 // selected for this entire State tree. 1551 1552 // The Matcher works on Trees, but an Intel add-to-memory requires a DAG: the 1553 // Store and the Load must have identical Memories (as well as identical 1554 // pointers). Since the Matcher does not have anything for Memory (and 1555 // does not handle DAGs), I have to match the Memory input myself. If the 1556 // Tree root is a Store, I require all Loads to have the identical memory. 1557 Node *Matcher::Label_Root( const Node *n, State *svec, Node *control, const Node *mem){ 1558 // Since Label_Root is a recursive function, its possible that we might run 1559 // out of stack space. See bugs 6272980 & 6227033 for more info. 1560 LabelRootDepth++; 1561 if (LabelRootDepth > MaxLabelRootDepth) { 1562 C->record_method_not_compilable("Out of stack space, increase MaxLabelRootDepth"); 1563 return NULL; 1564 } 1565 uint care = 0; // Edges matcher cares about 1566 uint cnt = n->req(); 1567 uint i = 0; 1568 1569 // Examine children for memory state 1570 // Can only subsume a child into your match-tree if that child's memory state 1571 // is not modified along the path to another input. 1572 // It is unsafe even if the other inputs are separate roots. 1573 Node *input_mem = NULL; 1574 for( i = 1; i < cnt; i++ ) { 1575 if( !n->match_edge(i) ) continue; 1576 Node *m = n->in(i); // Get ith input 1577 assert( m, "expect non-null children" ); 1578 if( m->is_Load() ) { 1579 if( input_mem == NULL ) { 1580 input_mem = m->in(MemNode::Memory); 1581 } else if( input_mem != m->in(MemNode::Memory) ) { 1582 input_mem = NodeSentinel; 1583 } 1584 } 1585 } 1586 1587 for( i = 1; i < cnt; i++ ){// For my children 1588 if( !n->match_edge(i) ) continue; 1589 Node *m = n->in(i); // Get ith input 1590 // Allocate states out of a private arena 1591 State *s = new (&_states_arena) State; 1592 svec->_kids[care++] = s; 1593 assert( care <= 2, "binary only for now" ); 1594 1595 // Recursively label the State tree. 1596 s->_kids[0] = NULL; 1597 s->_kids[1] = NULL; 1598 s->_leaf = m; 1599 1600 // Check for leaves of the State Tree; things that cannot be a part of 1601 // the current tree. If it finds any, that value is matched as a 1602 // register operand. If not, then the normal matching is used. 1603 if( match_into_reg(n, m, control, i, is_shared(m)) || 1604 // 1605 // Stop recursion if this is LoadNode and the root of this tree is a 1606 // StoreNode and the load & store have different memories. 1607 ((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) || 1608 // Can NOT include the match of a subtree when its memory state 1609 // is used by any of the other subtrees 1610 (input_mem == NodeSentinel) ) { 1611 // Print when we exclude matching due to different memory states at input-loads 1612 if (PrintOpto && (Verbose && WizardMode) && (input_mem == NodeSentinel) 1613 && !((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem)) { 1614 tty->print_cr("invalid input_mem"); 1615 } 1616 // Switch to a register-only opcode; this value must be in a register 1617 // and cannot be subsumed as part of a larger instruction. 1618 s->DFA( m->ideal_reg(), m ); 1619 1620 } else { 1621 // If match tree has no control and we do, adopt it for entire tree 1622 if( control == NULL && m->in(0) != NULL && m->req() > 1 ) 1623 control = m->in(0); // Pick up control 1624 // Else match as a normal part of the match tree. 1625 control = Label_Root(m,s,control,mem); 1626 if (C->failing()) return NULL; 1627 } 1628 } 1629 1630 1631 // Call DFA to match this node, and return 1632 svec->DFA( n->Opcode(), n ); 1633 1634 #ifdef ASSERT 1635 uint x; 1636 for( x = 0; x < _LAST_MACH_OPER; x++ ) 1637 if( svec->valid(x) ) 1638 break; 1639 1640 if (x >= _LAST_MACH_OPER) { 1641 n->dump(); 1642 svec->dump(); 1643 assert( false, "bad AD file" ); 1644 } 1645 #endif 1646 return control; 1647 } 1648 1649 1650 // Con nodes reduced using the same rule can share their MachNode 1651 // which reduces the number of copies of a constant in the final 1652 // program. The register allocator is free to split uses later to 1653 // split live ranges. 1654 MachNode* Matcher::find_shared_node(Node* leaf, uint rule) { 1655 if (!leaf->is_Con() && !leaf->is_DecodeNarrowPtr()) return NULL; 1656 1657 // See if this Con has already been reduced using this rule. 1658 if (_shared_nodes.Size() <= leaf->_idx) return NULL; 1659 MachNode* last = (MachNode*)_shared_nodes.at(leaf->_idx); 1660 if (last != NULL && rule == last->rule()) { 1661 // Don't expect control change for DecodeN 1662 if (leaf->is_DecodeNarrowPtr()) 1663 return last; 1664 // Get the new space root. 1665 Node* xroot = new_node(C->root()); 1666 if (xroot == NULL) { 1667 // This shouldn't happen give the order of matching. 1668 return NULL; 1669 } 1670 1671 // Shared constants need to have their control be root so they 1672 // can be scheduled properly. 1673 Node* control = last->in(0); 1674 if (control != xroot) { 1675 if (control == NULL || control == C->root()) { 1676 last->set_req(0, xroot); 1677 } else { 1678 assert(false, "unexpected control"); 1679 return NULL; 1680 } 1681 } 1682 return last; 1683 } 1684 return NULL; 1685 } 1686 1687 1688 //------------------------------ReduceInst------------------------------------- 1689 // Reduce a State tree (with given Control) into a tree of MachNodes. 1690 // This routine (and it's cohort ReduceOper) convert Ideal Nodes into 1691 // complicated machine Nodes. Each MachNode covers some tree of Ideal Nodes. 1692 // Each MachNode has a number of complicated MachOper operands; each 1693 // MachOper also covers a further tree of Ideal Nodes. 1694 1695 // The root of the Ideal match tree is always an instruction, so we enter 1696 // the recursion here. After building the MachNode, we need to recurse 1697 // the tree checking for these cases: 1698 // (1) Child is an instruction - 1699 // Build the instruction (recursively), add it as an edge. 1700 // Build a simple operand (register) to hold the result of the instruction. 1701 // (2) Child is an interior part of an instruction - 1702 // Skip over it (do nothing) 1703 // (3) Child is the start of a operand - 1704 // Build the operand, place it inside the instruction 1705 // Call ReduceOper. 1706 MachNode *Matcher::ReduceInst( State *s, int rule, Node *&mem ) { 1707 assert( rule >= NUM_OPERANDS, "called with operand rule" ); 1708 1709 MachNode* shared_node = find_shared_node(s->_leaf, rule); 1710 if (shared_node != NULL) { 1711 return shared_node; 1712 } 1713 1714 // Build the object to represent this state & prepare for recursive calls 1715 MachNode *mach = s->MachNodeGenerator(rule); 1716 guarantee(mach != NULL, "Missing MachNode"); 1717 mach->_opnds[0] = s->MachOperGenerator(_reduceOp[rule]); 1718 assert( mach->_opnds[0] != NULL, "Missing result operand" ); 1719 Node *leaf = s->_leaf; 1720 // Check for instruction or instruction chain rule 1721 if( rule >= _END_INST_CHAIN_RULE || rule < _BEGIN_INST_CHAIN_RULE ) { 1722 assert(C->node_arena()->contains(s->_leaf) || !has_new_node(s->_leaf), 1723 "duplicating node that's already been matched"); 1724 // Instruction 1725 mach->add_req( leaf->in(0) ); // Set initial control 1726 // Reduce interior of complex instruction 1727 ReduceInst_Interior( s, rule, mem, mach, 1 ); 1728 } else { 1729 // Instruction chain rules are data-dependent on their inputs 1730 mach->add_req(0); // Set initial control to none 1731 ReduceInst_Chain_Rule( s, rule, mem, mach ); 1732 } 1733 1734 // If a Memory was used, insert a Memory edge 1735 if( mem != (Node*)1 ) { 1736 mach->ins_req(MemNode::Memory,mem); 1737 #ifdef ASSERT 1738 // Verify adr type after matching memory operation 1739 const MachOper* oper = mach->memory_operand(); 1740 if (oper != NULL && oper != (MachOper*)-1) { 1741 // It has a unique memory operand. Find corresponding ideal mem node. 1742 Node* m = NULL; 1743 if (leaf->is_Mem()) { 1744 m = leaf; 1745 } else { 1746 m = _mem_node; 1747 assert(m != NULL && m->is_Mem(), "expecting memory node"); 1748 } 1749 const Type* mach_at = mach->adr_type(); 1750 // DecodeN node consumed by an address may have different type 1751 // than its input. Don't compare types for such case. 1752 if (m->adr_type() != mach_at && 1753 (m->in(MemNode::Address)->is_DecodeNarrowPtr() || 1754 (m->in(MemNode::Address)->is_AddP() && 1755 m->in(MemNode::Address)->in(AddPNode::Address)->is_DecodeNarrowPtr()) || 1756 (m->in(MemNode::Address)->is_AddP() && 1757 m->in(MemNode::Address)->in(AddPNode::Address)->is_AddP() && 1758 m->in(MemNode::Address)->in(AddPNode::Address)->in(AddPNode::Address)->is_DecodeNarrowPtr()))) { 1759 mach_at = m->adr_type(); 1760 } 1761 if (m->adr_type() != mach_at) { 1762 m->dump(); 1763 tty->print_cr("mach:"); 1764 mach->dump(1); 1765 } 1766 assert(m->adr_type() == mach_at, "matcher should not change adr type"); 1767 } 1768 #endif 1769 } 1770 1771 // If the _leaf is an AddP, insert the base edge 1772 if (leaf->is_AddP()) { 1773 mach->ins_req(AddPNode::Base,leaf->in(AddPNode::Base)); 1774 } 1775 1776 uint number_of_projections_prior = number_of_projections(); 1777 1778 // Perform any 1-to-many expansions required 1779 MachNode *ex = mach->Expand(s, _projection_list, mem); 1780 if (ex != mach) { 1781 assert(ex->ideal_reg() == mach->ideal_reg(), "ideal types should match"); 1782 if( ex->in(1)->is_Con() ) 1783 ex->in(1)->set_req(0, C->root()); 1784 // Remove old node from the graph 1785 for( uint i=0; i<mach->req(); i++ ) { 1786 mach->set_req(i,NULL); 1787 } 1788 #ifdef ASSERT 1789 _new2old_map.map(ex->_idx, s->_leaf); 1790 #endif 1791 } 1792 1793 // PhaseChaitin::fixup_spills will sometimes generate spill code 1794 // via the matcher. By the time, nodes have been wired into the CFG, 1795 // and any further nodes generated by expand rules will be left hanging 1796 // in space, and will not get emitted as output code. Catch this. 1797 // Also, catch any new register allocation constraints ("projections") 1798 // generated belatedly during spill code generation. 1799 if (_allocation_started) { 1800 guarantee(ex == mach, "no expand rules during spill generation"); 1801 guarantee(number_of_projections_prior == number_of_projections(), "no allocation during spill generation"); 1802 } 1803 1804 if (leaf->is_Con() || leaf->is_DecodeNarrowPtr()) { 1805 // Record the con for sharing 1806 _shared_nodes.map(leaf->_idx, ex); 1807 } 1808 1809 return ex; 1810 } 1811 1812 void Matcher::handle_precedence_edges(Node* n, MachNode *mach) { 1813 for (uint i = n->req(); i < n->len(); i++) { 1814 if (n->in(i) != NULL) { 1815 mach->add_prec(n->in(i)); 1816 } 1817 } 1818 } 1819 1820 void Matcher::ReduceInst_Chain_Rule( State *s, int rule, Node *&mem, MachNode *mach ) { 1821 // 'op' is what I am expecting to receive 1822 int op = _leftOp[rule]; 1823 // Operand type to catch childs result 1824 // This is what my child will give me. 1825 int opnd_class_instance = s->_rule[op]; 1826 // Choose between operand class or not. 1827 // This is what I will receive. 1828 int catch_op = (FIRST_OPERAND_CLASS <= op && op < NUM_OPERANDS) ? opnd_class_instance : op; 1829 // New rule for child. Chase operand classes to get the actual rule. 1830 int newrule = s->_rule[catch_op]; 1831 1832 if( newrule < NUM_OPERANDS ) { 1833 // Chain from operand or operand class, may be output of shared node 1834 assert( 0 <= opnd_class_instance && opnd_class_instance < NUM_OPERANDS, 1835 "Bad AD file: Instruction chain rule must chain from operand"); 1836 // Insert operand into array of operands for this instruction 1837 mach->_opnds[1] = s->MachOperGenerator(opnd_class_instance); 1838 1839 ReduceOper( s, newrule, mem, mach ); 1840 } else { 1841 // Chain from the result of an instruction 1842 assert( newrule >= _LAST_MACH_OPER, "Do NOT chain from internal operand"); 1843 mach->_opnds[1] = s->MachOperGenerator(_reduceOp[catch_op]); 1844 Node *mem1 = (Node*)1; 1845 debug_only(Node *save_mem_node = _mem_node;) 1846 mach->add_req( ReduceInst(s, newrule, mem1) ); 1847 debug_only(_mem_node = save_mem_node;) 1848 } 1849 return; 1850 } 1851 1852 1853 uint Matcher::ReduceInst_Interior( State *s, int rule, Node *&mem, MachNode *mach, uint num_opnds ) { 1854 handle_precedence_edges(s->_leaf, mach); 1855 1856 if( s->_leaf->is_Load() ) { 1857 Node *mem2 = s->_leaf->in(MemNode::Memory); 1858 assert( mem == (Node*)1 || mem == mem2, "multiple Memories being matched at once?" ); 1859 debug_only( if( mem == (Node*)1 ) _mem_node = s->_leaf;) 1860 mem = mem2; 1861 } 1862 if( s->_leaf->in(0) != NULL && s->_leaf->req() > 1) { 1863 if( mach->in(0) == NULL ) 1864 mach->set_req(0, s->_leaf->in(0)); 1865 } 1866 1867 // Now recursively walk the state tree & add operand list. 1868 for( uint i=0; i<2; i++ ) { // binary tree 1869 State *newstate = s->_kids[i]; 1870 if( newstate == NULL ) break; // Might only have 1 child 1871 // 'op' is what I am expecting to receive 1872 int op; 1873 if( i == 0 ) { 1874 op = _leftOp[rule]; 1875 } else { 1876 op = _rightOp[rule]; 1877 } 1878 // Operand type to catch childs result 1879 // This is what my child will give me. 1880 int opnd_class_instance = newstate->_rule[op]; 1881 // Choose between operand class or not. 1882 // This is what I will receive. 1883 int catch_op = (op >= FIRST_OPERAND_CLASS && op < NUM_OPERANDS) ? opnd_class_instance : op; 1884 // New rule for child. Chase operand classes to get the actual rule. 1885 int newrule = newstate->_rule[catch_op]; 1886 1887 if( newrule < NUM_OPERANDS ) { // Operand/operandClass or internalOp/instruction? 1888 // Operand/operandClass 1889 // Insert operand into array of operands for this instruction 1890 mach->_opnds[num_opnds++] = newstate->MachOperGenerator(opnd_class_instance); 1891 ReduceOper( newstate, newrule, mem, mach ); 1892 1893 } else { // Child is internal operand or new instruction 1894 if( newrule < _LAST_MACH_OPER ) { // internal operand or instruction? 1895 // internal operand --> call ReduceInst_Interior 1896 // Interior of complex instruction. Do nothing but recurse. 1897 num_opnds = ReduceInst_Interior( newstate, newrule, mem, mach, num_opnds ); 1898 } else { 1899 // instruction --> call build operand( ) to catch result 1900 // --> ReduceInst( newrule ) 1901 mach->_opnds[num_opnds++] = s->MachOperGenerator(_reduceOp[catch_op]); 1902 Node *mem1 = (Node*)1; 1903 debug_only(Node *save_mem_node = _mem_node;) 1904 mach->add_req( ReduceInst( newstate, newrule, mem1 ) ); 1905 debug_only(_mem_node = save_mem_node;) 1906 } 1907 } 1908 assert( mach->_opnds[num_opnds-1], "" ); 1909 } 1910 return num_opnds; 1911 } 1912 1913 // This routine walks the interior of possible complex operands. 1914 // At each point we check our children in the match tree: 1915 // (1) No children - 1916 // We are a leaf; add _leaf field as an input to the MachNode 1917 // (2) Child is an internal operand - 1918 // Skip over it ( do nothing ) 1919 // (3) Child is an instruction - 1920 // Call ReduceInst recursively and 1921 // and instruction as an input to the MachNode 1922 void Matcher::ReduceOper( State *s, int rule, Node *&mem, MachNode *mach ) { 1923 assert( rule < _LAST_MACH_OPER, "called with operand rule" ); 1924 State *kid = s->_kids[0]; 1925 assert( kid == NULL || s->_leaf->in(0) == NULL, "internal operands have no control" ); 1926 1927 // Leaf? And not subsumed? 1928 if( kid == NULL && !_swallowed[rule] ) { 1929 mach->add_req( s->_leaf ); // Add leaf pointer 1930 return; // Bail out 1931 } 1932 1933 if( s->_leaf->is_Load() ) { 1934 assert( mem == (Node*)1, "multiple Memories being matched at once?" ); 1935 mem = s->_leaf->in(MemNode::Memory); 1936 debug_only(_mem_node = s->_leaf;) 1937 } 1938 1939 handle_precedence_edges(s->_leaf, mach); 1940 1941 if( s->_leaf->in(0) && s->_leaf->req() > 1) { 1942 if( !mach->in(0) ) 1943 mach->set_req(0,s->_leaf->in(0)); 1944 else { 1945 assert( s->_leaf->in(0) == mach->in(0), "same instruction, differing controls?" ); 1946 } 1947 } 1948 1949 for( uint i=0; kid != NULL && i<2; kid = s->_kids[1], i++ ) { // binary tree 1950 int newrule; 1951 if( i == 0) 1952 newrule = kid->_rule[_leftOp[rule]]; 1953 else 1954 newrule = kid->_rule[_rightOp[rule]]; 1955 1956 if( newrule < _LAST_MACH_OPER ) { // Operand or instruction? 1957 // Internal operand; recurse but do nothing else 1958 ReduceOper( kid, newrule, mem, mach ); 1959 1960 } else { // Child is a new instruction 1961 // Reduce the instruction, and add a direct pointer from this 1962 // machine instruction to the newly reduced one. 1963 Node *mem1 = (Node*)1; 1964 debug_only(Node *save_mem_node = _mem_node;) 1965 mach->add_req( ReduceInst( kid, newrule, mem1 ) ); 1966 debug_only(_mem_node = save_mem_node;) 1967 } 1968 } 1969 } 1970 1971 1972 // ------------------------------------------------------------------------- 1973 // Java-Java calling convention 1974 // (what you use when Java calls Java) 1975 1976 //------------------------------find_receiver---------------------------------- 1977 // For a given signature, return the OptoReg for parameter 0. 1978 OptoReg::Name Matcher::find_receiver( bool is_outgoing ) { 1979 VMRegPair regs; 1980 BasicType sig_bt = T_OBJECT; 1981 calling_convention(&sig_bt, ®s, 1, is_outgoing); 1982 // Return argument 0 register. In the LP64 build pointers 1983 // take 2 registers, but the VM wants only the 'main' name. 1984 return OptoReg::as_OptoReg(regs.first()); 1985 } 1986 1987 // This function identifies sub-graphs in which a 'load' node is 1988 // input to two different nodes, and such that it can be matched 1989 // with BMI instructions like blsi, blsr, etc. 1990 // Example : for b = -a[i] & a[i] can be matched to blsi r32, m32. 1991 // The graph is (AndL (SubL Con0 LoadL*) LoadL*), where LoadL* 1992 // refers to the same node. 1993 #ifdef X86 1994 // Match the generic fused operations pattern (op1 (op2 Con{ConType} mop) mop) 1995 // This is a temporary solution until we make DAGs expressible in ADL. 1996 template<typename ConType> 1997 class FusedPatternMatcher { 1998 Node* _op1_node; 1999 Node* _mop_node; 2000 int _con_op; 2001 2002 static int match_next(Node* n, int next_op, int next_op_idx) { 2003 if (n->in(1) == NULL || n->in(2) == NULL) { 2004 return -1; 2005 } 2006 2007 if (next_op_idx == -1) { // n is commutative, try rotations 2008 if (n->in(1)->Opcode() == next_op) { 2009 return 1; 2010 } else if (n->in(2)->Opcode() == next_op) { 2011 return 2; 2012 } 2013 } else { 2014 assert(next_op_idx > 0 && next_op_idx <= 2, "Bad argument index"); 2015 if (n->in(next_op_idx)->Opcode() == next_op) { 2016 return next_op_idx; 2017 } 2018 } 2019 return -1; 2020 } 2021 public: 2022 FusedPatternMatcher(Node* op1_node, Node *mop_node, int con_op) : 2023 _op1_node(op1_node), _mop_node(mop_node), _con_op(con_op) { } 2024 2025 bool match(int op1, int op1_op2_idx, // op1 and the index of the op1->op2 edge, -1 if op1 is commutative 2026 int op2, int op2_con_idx, // op2 and the index of the op2->con edge, -1 if op2 is commutative 2027 typename ConType::NativeType con_value) { 2028 if (_op1_node->Opcode() != op1) { 2029 return false; 2030 } 2031 if (_mop_node->outcnt() > 2) { 2032 return false; 2033 } 2034 op1_op2_idx = match_next(_op1_node, op2, op1_op2_idx); 2035 if (op1_op2_idx == -1) { 2036 return false; 2037 } 2038 // Memory operation must be the other edge 2039 int op1_mop_idx = (op1_op2_idx & 1) + 1; 2040 2041 // Check that the mop node is really what we want 2042 if (_op1_node->in(op1_mop_idx) == _mop_node) { 2043 Node *op2_node = _op1_node->in(op1_op2_idx); 2044 if (op2_node->outcnt() > 1) { 2045 return false; 2046 } 2047 assert(op2_node->Opcode() == op2, "Should be"); 2048 op2_con_idx = match_next(op2_node, _con_op, op2_con_idx); 2049 if (op2_con_idx == -1) { 2050 return false; 2051 } 2052 // Memory operation must be the other edge 2053 int op2_mop_idx = (op2_con_idx & 1) + 1; 2054 // Check that the memory operation is the same node 2055 if (op2_node->in(op2_mop_idx) == _mop_node) { 2056 // Now check the constant 2057 const Type* con_type = op2_node->in(op2_con_idx)->bottom_type(); 2058 if (con_type != Type::TOP && ConType::as_self(con_type)->get_con() == con_value) { 2059 return true; 2060 } 2061 } 2062 } 2063 return false; 2064 } 2065 }; 2066 2067 2068 bool Matcher::is_bmi_pattern(Node *n, Node *m) { 2069 if (n != NULL && m != NULL) { 2070 if (m->Opcode() == Op_LoadI) { 2071 FusedPatternMatcher<TypeInt> bmii(n, m, Op_ConI); 2072 return bmii.match(Op_AndI, -1, Op_SubI, 1, 0) || 2073 bmii.match(Op_AndI, -1, Op_AddI, -1, -1) || 2074 bmii.match(Op_XorI, -1, Op_AddI, -1, -1); 2075 } else if (m->Opcode() == Op_LoadL) { 2076 FusedPatternMatcher<TypeLong> bmil(n, m, Op_ConL); 2077 return bmil.match(Op_AndL, -1, Op_SubL, 1, 0) || 2078 bmil.match(Op_AndL, -1, Op_AddL, -1, -1) || 2079 bmil.match(Op_XorL, -1, Op_AddL, -1, -1); 2080 } 2081 } 2082 return false; 2083 } 2084 #endif // X86 2085 2086 bool Matcher::clone_base_plus_offset_address(AddPNode* m, Matcher::MStack& mstack, VectorSet& address_visited) { 2087 Node *off = m->in(AddPNode::Offset); 2088 if (off->is_Con()) { 2089 address_visited.test_set(m->_idx); // Flag as address_visited 2090 mstack.push(m->in(AddPNode::Address), Pre_Visit); 2091 // Clone X+offset as it also folds into most addressing expressions 2092 mstack.push(off, Visit); 2093 mstack.push(m->in(AddPNode::Base), Pre_Visit); 2094 return true; 2095 } 2096 return false; 2097 } 2098 2099 // A method-klass-holder may be passed in the inline_cache_reg 2100 // and then expanded into the inline_cache_reg and a method_oop register 2101 // defined in ad_<arch>.cpp 2102 2103 //------------------------------find_shared------------------------------------ 2104 // Set bits if Node is shared or otherwise a root 2105 void Matcher::find_shared( Node *n ) { 2106 // Allocate stack of size C->live_nodes() * 2 to avoid frequent realloc 2107 MStack mstack(C->live_nodes() * 2); 2108 // Mark nodes as address_visited if they are inputs to an address expression 2109 VectorSet address_visited(Thread::current()->resource_area()); 2110 mstack.push(n, Visit); // Don't need to pre-visit root node 2111 while (mstack.is_nonempty()) { 2112 n = mstack.node(); // Leave node on stack 2113 Node_State nstate = mstack.state(); 2114 uint nop = n->Opcode(); 2115 if (nstate == Pre_Visit) { 2116 if (address_visited.test(n->_idx)) { // Visited in address already? 2117 // Flag as visited and shared now. 2118 set_visited(n); 2119 } 2120 if (is_visited(n)) { // Visited already? 2121 // Node is shared and has no reason to clone. Flag it as shared. 2122 // This causes it to match into a register for the sharing. 2123 set_shared(n); // Flag as shared and 2124 mstack.pop(); // remove node from stack 2125 continue; 2126 } 2127 nstate = Visit; // Not already visited; so visit now 2128 } 2129 if (nstate == Visit) { 2130 mstack.set_state(Post_Visit); 2131 set_visited(n); // Flag as visited now 2132 bool mem_op = false; 2133 int mem_addr_idx = MemNode::Address; 2134 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->matcher_find_shared_visit(this, mstack, n, nop, mem_op, mem_addr_idx); 2135 if (!gc_handled) { 2136 if (find_shared_visit(mstack, n, nop, mem_op, mem_addr_idx)) { 2137 continue; 2138 } 2139 } 2140 for(int i = n->req() - 1; i >= 0; --i) { // For my children 2141 Node *m = n->in(i); // Get ith input 2142 if (m == NULL) continue; // Ignore NULLs 2143 uint mop = m->Opcode(); 2144 2145 // Must clone all producers of flags, or we will not match correctly. 2146 // Suppose a compare setting int-flags is shared (e.g., a switch-tree) 2147 // then it will match into an ideal Op_RegFlags. Alas, the fp-flags 2148 // are also there, so we may match a float-branch to int-flags and 2149 // expect the allocator to haul the flags from the int-side to the 2150 // fp-side. No can do. 2151 if( _must_clone[mop] ) { 2152 mstack.push(m, Visit); 2153 continue; // for(int i = ...) 2154 } 2155 2156 if( mop == Op_AddP && m->in(AddPNode::Base)->is_DecodeNarrowPtr()) { 2157 // Bases used in addresses must be shared but since 2158 // they are shared through a DecodeN they may appear 2159 // to have a single use so force sharing here. 2160 set_shared(m->in(AddPNode::Base)->in(1)); 2161 } 2162 2163 // if 'n' and 'm' are part of a graph for BMI instruction, clone this node. 2164 #ifdef X86 2165 if (UseBMI1Instructions && is_bmi_pattern(n, m)) { 2166 mstack.push(m, Visit); 2167 continue; 2168 } 2169 #endif 2170 2171 // Clone addressing expressions as they are "free" in memory access instructions 2172 if (mem_op && i == mem_addr_idx && mop == Op_AddP && 2173 // When there are other uses besides address expressions 2174 // put it on stack and mark as shared. 2175 !is_visited(m)) { 2176 // Some inputs for address expression are not put on stack 2177 // to avoid marking them as shared and forcing them into register 2178 // if they are used only in address expressions. 2179 // But they should be marked as shared if there are other uses 2180 // besides address expressions. 2181 2182 if (clone_address_expressions(m->as_AddP(), mstack, address_visited)) { 2183 continue; 2184 } 2185 } // if( mem_op && 2186 mstack.push(m, Pre_Visit); 2187 } // for(int i = ...) 2188 } 2189 else if (nstate == Alt_Post_Visit) { 2190 mstack.pop(); // Remove node from stack 2191 // We cannot remove the Cmp input from the Bool here, as the Bool may be 2192 // shared and all users of the Bool need to move the Cmp in parallel. 2193 // This leaves both the Bool and the If pointing at the Cmp. To 2194 // prevent the Matcher from trying to Match the Cmp along both paths 2195 // BoolNode::match_edge always returns a zero. 2196 2197 // We reorder the Op_If in a pre-order manner, so we can visit without 2198 // accidentally sharing the Cmp (the Bool and the If make 2 users). 2199 n->add_req( n->in(1)->in(1) ); // Add the Cmp next to the Bool 2200 } 2201 else if (nstate == Post_Visit) { 2202 mstack.pop(); // Remove node from stack 2203 2204 // Now hack a few special opcodes 2205 uint opcode = n->Opcode(); 2206 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->matcher_find_shared_post_visit(this, n, opcode); 2207 if (!gc_handled) { 2208 find_shared_post_visit(n, opcode); 2209 } 2210 } 2211 else { 2212 ShouldNotReachHere(); 2213 } 2214 } // end of while (mstack.is_nonempty()) 2215 } 2216 2217 bool Matcher::find_shared_visit(MStack& mstack, Node* n, uint opcode, bool& mem_op, int& mem_addr_idx) { 2218 switch(opcode) { // Handle some opcodes special 2219 case Op_Phi: // Treat Phis as shared roots 2220 case Op_Parm: 2221 case Op_Proj: // All handled specially during matching 2222 case Op_SafePointScalarObject: 2223 set_shared(n); 2224 set_dontcare(n); 2225 break; 2226 case Op_If: 2227 case Op_CountedLoopEnd: 2228 mstack.set_state(Alt_Post_Visit); // Alternative way 2229 // Convert (If (Bool (CmpX A B))) into (If (Bool) (CmpX A B)). Helps 2230 // with matching cmp/branch in 1 instruction. The Matcher needs the 2231 // Bool and CmpX side-by-side, because it can only get at constants 2232 // that are at the leaves of Match trees, and the Bool's condition acts 2233 // as a constant here. 2234 mstack.push(n->in(1), Visit); // Clone the Bool 2235 mstack.push(n->in(0), Pre_Visit); // Visit control input 2236 return true; // while (mstack.is_nonempty()) 2237 case Op_ConvI2D: // These forms efficiently match with a prior 2238 case Op_ConvI2F: // Load but not a following Store 2239 if( n->in(1)->is_Load() && // Prior load 2240 n->outcnt() == 1 && // Not already shared 2241 n->unique_out()->is_Store() ) // Following store 2242 set_shared(n); // Force it to be a root 2243 break; 2244 case Op_ReverseBytesI: 2245 case Op_ReverseBytesL: 2246 if( n->in(1)->is_Load() && // Prior load 2247 n->outcnt() == 1 ) // Not already shared 2248 set_shared(n); // Force it to be a root 2249 break; 2250 case Op_BoxLock: // Cant match until we get stack-regs in ADLC 2251 case Op_IfFalse: 2252 case Op_IfTrue: 2253 case Op_MachProj: 2254 case Op_MergeMem: 2255 case Op_Catch: 2256 case Op_CatchProj: 2257 case Op_CProj: 2258 case Op_JumpProj: 2259 case Op_JProj: 2260 case Op_NeverBranch: 2261 set_dontcare(n); 2262 break; 2263 case Op_Jump: 2264 mstack.push(n->in(1), Pre_Visit); // Switch Value (could be shared) 2265 mstack.push(n->in(0), Pre_Visit); // Visit Control input 2266 return true; // while (mstack.is_nonempty()) 2267 case Op_StrComp: 2268 case Op_StrEquals: 2269 case Op_StrIndexOf: 2270 case Op_StrIndexOfChar: 2271 case Op_AryEq: 2272 case Op_HasNegatives: 2273 case Op_StrInflatedCopy: 2274 case Op_StrCompressedCopy: 2275 case Op_EncodeISOArray: 2276 case Op_FmaD: 2277 case Op_FmaF: 2278 case Op_FmaVD: 2279 case Op_FmaVF: 2280 set_shared(n); // Force result into register (it will be anyways) 2281 break; 2282 case Op_ConP: { // Convert pointers above the centerline to NUL 2283 TypeNode *tn = n->as_Type(); // Constants derive from type nodes 2284 const TypePtr* tp = tn->type()->is_ptr(); 2285 if (tp->_ptr == TypePtr::AnyNull) { 2286 tn->set_type(TypePtr::NULL_PTR); 2287 } 2288 break; 2289 } 2290 case Op_ConN: { // Convert narrow pointers above the centerline to NUL 2291 TypeNode *tn = n->as_Type(); // Constants derive from type nodes 2292 const TypePtr* tp = tn->type()->make_ptr(); 2293 if (tp && tp->_ptr == TypePtr::AnyNull) { 2294 tn->set_type(TypeNarrowOop::NULL_PTR); 2295 } 2296 break; 2297 } 2298 case Op_Binary: // These are introduced in the Post_Visit state. 2299 ShouldNotReachHere(); 2300 break; 2301 case Op_ClearArray: 2302 case Op_SafePoint: 2303 mem_op = true; 2304 break; 2305 default: 2306 if( n->is_Store() ) { 2307 // Do match stores, despite no ideal reg 2308 mem_op = true; 2309 break; 2310 } 2311 if( n->is_Mem() ) { // Loads and LoadStores 2312 mem_op = true; 2313 // Loads must be root of match tree due to prior load conflict 2314 if( C->subsume_loads() == false ) 2315 set_shared(n); 2316 } 2317 // Fall into default case 2318 if( !n->ideal_reg() ) 2319 set_dontcare(n); // Unmatchable Nodes 2320 } // end_switch 2321 return false; 2322 } 2323 2324 void Matcher::find_shared_post_visit(Node* n, uint opcode) { 2325 switch(opcode) { // Handle some opcodes special 2326 case Op_StorePConditional: 2327 case Op_StoreIConditional: 2328 case Op_StoreLConditional: 2329 case Op_CompareAndExchangeB: 2330 case Op_CompareAndExchangeS: 2331 case Op_CompareAndExchangeI: 2332 case Op_CompareAndExchangeL: 2333 case Op_CompareAndExchangeP: 2334 case Op_CompareAndExchangeN: 2335 case Op_WeakCompareAndSwapB: 2336 case Op_WeakCompareAndSwapS: 2337 case Op_WeakCompareAndSwapI: 2338 case Op_WeakCompareAndSwapL: 2339 case Op_WeakCompareAndSwapP: 2340 case Op_WeakCompareAndSwapN: 2341 case Op_CompareAndSwapB: 2342 case Op_CompareAndSwapS: 2343 case Op_CompareAndSwapI: 2344 case Op_CompareAndSwapL: 2345 case Op_CompareAndSwapP: 2346 case Op_CompareAndSwapN: { // Convert trinary to binary-tree 2347 Node* newval = n->in(MemNode::ValueIn); 2348 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn); 2349 Node* pair = new BinaryNode(oldval, newval); 2350 n->set_req(MemNode::ValueIn, pair); 2351 n->del_req(LoadStoreConditionalNode::ExpectedIn); 2352 break; 2353 } 2354 case Op_CMoveD: // Convert trinary to binary-tree 2355 case Op_CMoveF: 2356 case Op_CMoveI: 2357 case Op_CMoveL: 2358 case Op_CMoveN: 2359 case Op_CMoveP: 2360 case Op_CMoveVF: 2361 case Op_CMoveVD: { 2362 // Restructure into a binary tree for Matching. It's possible that 2363 // we could move this code up next to the graph reshaping for IfNodes 2364 // or vice-versa, but I do not want to debug this for Ladybird. 2365 // 10/2/2000 CNC. 2366 Node* pair1 = new BinaryNode(n->in(1), n->in(1)->in(1)); 2367 n->set_req(1, pair1); 2368 Node* pair2 = new BinaryNode(n->in(2), n->in(3)); 2369 n->set_req(2, pair2); 2370 n->del_req(3); 2371 break; 2372 } 2373 case Op_LoopLimit: { 2374 Node* pair1 = new BinaryNode(n->in(1), n->in(2)); 2375 n->set_req(1, pair1); 2376 n->set_req(2, n->in(3)); 2377 n->del_req(3); 2378 break; 2379 } 2380 case Op_StrEquals: 2381 case Op_StrIndexOfChar: { 2382 Node* pair1 = new BinaryNode(n->in(2), n->in(3)); 2383 n->set_req(2, pair1); 2384 n->set_req(3, n->in(4)); 2385 n->del_req(4); 2386 break; 2387 } 2388 case Op_StrComp: 2389 case Op_StrIndexOf: { 2390 Node* pair1 = new BinaryNode(n->in(2), n->in(3)); 2391 n->set_req(2, pair1); 2392 Node* pair2 = new BinaryNode(n->in(4),n->in(5)); 2393 n->set_req(3, pair2); 2394 n->del_req(5); 2395 n->del_req(4); 2396 break; 2397 } 2398 case Op_StrCompressedCopy: 2399 case Op_StrInflatedCopy: 2400 case Op_EncodeISOArray: { 2401 // Restructure into a binary tree for Matching. 2402 Node* pair = new BinaryNode(n->in(3), n->in(4)); 2403 n->set_req(3, pair); 2404 n->del_req(4); 2405 break; 2406 } 2407 case Op_FmaD: 2408 case Op_FmaF: 2409 case Op_FmaVD: 2410 case Op_FmaVF: { 2411 // Restructure into a binary tree for Matching. 2412 Node* pair = new BinaryNode(n->in(1), n->in(2)); 2413 n->set_req(2, pair); 2414 n->set_req(1, n->in(3)); 2415 n->del_req(3); 2416 break; 2417 } 2418 case Op_MulAddS2I: { 2419 Node* pair1 = new BinaryNode(n->in(1), n->in(2)); 2420 Node* pair2 = new BinaryNode(n->in(3), n->in(4)); 2421 n->set_req(1, pair1); 2422 n->set_req(2, pair2); 2423 n->del_req(4); 2424 n->del_req(3); 2425 break; 2426 } 2427 case Op_ClearArray: { 2428 Node* pair = new BinaryNode(n->in(2), n->in(3)); 2429 n->set_req(2, pair); 2430 n->set_req(3, n->in(4)); 2431 n->del_req(4); 2432 break; 2433 } 2434 default: 2435 break; 2436 } 2437 } 2438 2439 #ifdef ASSERT 2440 // machine-independent root to machine-dependent root 2441 void Matcher::dump_old2new_map() { 2442 _old2new_map.dump(); 2443 } 2444 #endif 2445 2446 //---------------------------collect_null_checks------------------------------- 2447 // Find null checks in the ideal graph; write a machine-specific node for 2448 // it. Used by later implicit-null-check handling. Actually collects 2449 // either an IfTrue or IfFalse for the common NOT-null path, AND the ideal 2450 // value being tested. 2451 void Matcher::collect_null_checks( Node *proj, Node *orig_proj ) { 2452 Node *iff = proj->in(0); 2453 if( iff->Opcode() == Op_If ) { 2454 // During matching If's have Bool & Cmp side-by-side 2455 BoolNode *b = iff->in(1)->as_Bool(); 2456 Node *cmp = iff->in(2); 2457 int opc = cmp->Opcode(); 2458 if (opc != Op_CmpP && opc != Op_CmpN) return; 2459 2460 const Type* ct = cmp->in(2)->bottom_type(); 2461 if (ct == TypePtr::NULL_PTR || 2462 (opc == Op_CmpN && ct == TypeNarrowOop::NULL_PTR)) { 2463 2464 bool push_it = false; 2465 if( proj->Opcode() == Op_IfTrue ) { 2466 #ifndef PRODUCT 2467 extern int all_null_checks_found; 2468 all_null_checks_found++; 2469 #endif 2470 if( b->_test._test == BoolTest::ne ) { 2471 push_it = true; 2472 } 2473 } else { 2474 assert( proj->Opcode() == Op_IfFalse, "" ); 2475 if( b->_test._test == BoolTest::eq ) { 2476 push_it = true; 2477 } 2478 } 2479 if( push_it ) { 2480 _null_check_tests.push(proj); 2481 Node* val = cmp->in(1); 2482 #ifdef _LP64 2483 if (val->bottom_type()->isa_narrowoop() && 2484 !Matcher::narrow_oop_use_complex_address()) { 2485 // 2486 // Look for DecodeN node which should be pinned to orig_proj. 2487 // On platforms (Sparc) which can not handle 2 adds 2488 // in addressing mode we have to keep a DecodeN node and 2489 // use it to do implicit NULL check in address. 2490 // 2491 // DecodeN node was pinned to non-null path (orig_proj) during 2492 // CastPP transformation in final_graph_reshaping_impl(). 2493 // 2494 uint cnt = orig_proj->outcnt(); 2495 for (uint i = 0; i < orig_proj->outcnt(); i++) { 2496 Node* d = orig_proj->raw_out(i); 2497 if (d->is_DecodeN() && d->in(1) == val) { 2498 val = d; 2499 val->set_req(0, NULL); // Unpin now. 2500 // Mark this as special case to distinguish from 2501 // a regular case: CmpP(DecodeN, NULL). 2502 val = (Node*)(((intptr_t)val) | 1); 2503 break; 2504 } 2505 } 2506 } 2507 #endif 2508 _null_check_tests.push(val); 2509 } 2510 } 2511 } 2512 } 2513 2514 //---------------------------validate_null_checks------------------------------ 2515 // Its possible that the value being NULL checked is not the root of a match 2516 // tree. If so, I cannot use the value in an implicit null check. 2517 void Matcher::validate_null_checks( ) { 2518 uint cnt = _null_check_tests.size(); 2519 for( uint i=0; i < cnt; i+=2 ) { 2520 Node *test = _null_check_tests[i]; 2521 Node *val = _null_check_tests[i+1]; 2522 bool is_decoden = ((intptr_t)val) & 1; 2523 val = (Node*)(((intptr_t)val) & ~1); 2524 if (has_new_node(val)) { 2525 Node* new_val = new_node(val); 2526 if (is_decoden) { 2527 assert(val->is_DecodeNarrowPtr() && val->in(0) == NULL, "sanity"); 2528 // Note: new_val may have a control edge if 2529 // the original ideal node DecodeN was matched before 2530 // it was unpinned in Matcher::collect_null_checks(). 2531 // Unpin the mach node and mark it. 2532 new_val->set_req(0, NULL); 2533 new_val = (Node*)(((intptr_t)new_val) | 1); 2534 } 2535 // Is a match-tree root, so replace with the matched value 2536 _null_check_tests.map(i+1, new_val); 2537 } else { 2538 // Yank from candidate list 2539 _null_check_tests.map(i+1,_null_check_tests[--cnt]); 2540 _null_check_tests.map(i,_null_check_tests[--cnt]); 2541 _null_check_tests.pop(); 2542 _null_check_tests.pop(); 2543 i-=2; 2544 } 2545 } 2546 } 2547 2548 // Used by the DFA in dfa_xxx.cpp. Check for a following barrier or 2549 // atomic instruction acting as a store_load barrier without any 2550 // intervening volatile load, and thus we don't need a barrier here. 2551 // We retain the Node to act as a compiler ordering barrier. 2552 bool Matcher::post_store_load_barrier(const Node* vmb) { 2553 Compile* C = Compile::current(); 2554 assert(vmb->is_MemBar(), ""); 2555 assert(vmb->Opcode() != Op_MemBarAcquire && vmb->Opcode() != Op_LoadFence, ""); 2556 const MemBarNode* membar = vmb->as_MemBar(); 2557 2558 // Get the Ideal Proj node, ctrl, that can be used to iterate forward 2559 Node* ctrl = NULL; 2560 for (DUIterator_Fast imax, i = membar->fast_outs(imax); i < imax; i++) { 2561 Node* p = membar->fast_out(i); 2562 assert(p->is_Proj(), "only projections here"); 2563 if ((p->as_Proj()->_con == TypeFunc::Control) && 2564 !C->node_arena()->contains(p)) { // Unmatched old-space only 2565 ctrl = p; 2566 break; 2567 } 2568 } 2569 assert((ctrl != NULL), "missing control projection"); 2570 2571 for (DUIterator_Fast jmax, j = ctrl->fast_outs(jmax); j < jmax; j++) { 2572 Node *x = ctrl->fast_out(j); 2573 int xop = x->Opcode(); 2574 2575 // We don't need current barrier if we see another or a lock 2576 // before seeing volatile load. 2577 // 2578 // Op_Fastunlock previously appeared in the Op_* list below. 2579 // With the advent of 1-0 lock operations we're no longer guaranteed 2580 // that a monitor exit operation contains a serializing instruction. 2581 2582 if (xop == Op_MemBarVolatile || 2583 xop == Op_CompareAndExchangeB || 2584 xop == Op_CompareAndExchangeS || 2585 xop == Op_CompareAndExchangeI || 2586 xop == Op_CompareAndExchangeL || 2587 xop == Op_CompareAndExchangeP || 2588 xop == Op_CompareAndExchangeN || 2589 xop == Op_WeakCompareAndSwapB || 2590 xop == Op_WeakCompareAndSwapS || 2591 xop == Op_WeakCompareAndSwapL || 2592 xop == Op_WeakCompareAndSwapP || 2593 xop == Op_WeakCompareAndSwapN || 2594 xop == Op_WeakCompareAndSwapI || 2595 xop == Op_CompareAndSwapB || 2596 xop == Op_CompareAndSwapS || 2597 xop == Op_CompareAndSwapL || 2598 xop == Op_CompareAndSwapP || 2599 xop == Op_CompareAndSwapN || 2600 xop == Op_CompareAndSwapI || 2601 BarrierSet::barrier_set()->barrier_set_c2()->matcher_is_store_load_barrier(x, xop)) { 2602 return true; 2603 } 2604 2605 // Op_FastLock previously appeared in the Op_* list above. 2606 // With biased locking we're no longer guaranteed that a monitor 2607 // enter operation contains a serializing instruction. 2608 if ((xop == Op_FastLock) && !UseBiasedLocking) { 2609 return true; 2610 } 2611 2612 if (x->is_MemBar()) { 2613 // We must retain this membar if there is an upcoming volatile 2614 // load, which will be followed by acquire membar. 2615 if (xop == Op_MemBarAcquire || xop == Op_LoadFence) { 2616 return false; 2617 } else { 2618 // For other kinds of barriers, check by pretending we 2619 // are them, and seeing if we can be removed. 2620 return post_store_load_barrier(x->as_MemBar()); 2621 } 2622 } 2623 2624 // probably not necessary to check for these 2625 if (x->is_Call() || x->is_SafePoint() || x->is_block_proj()) { 2626 return false; 2627 } 2628 } 2629 return false; 2630 } 2631 2632 // Check whether node n is a branch to an uncommon trap that we could 2633 // optimize as test with very high branch costs in case of going to 2634 // the uncommon trap. The code must be able to be recompiled to use 2635 // a cheaper test. 2636 bool Matcher::branches_to_uncommon_trap(const Node *n) { 2637 // Don't do it for natives, adapters, or runtime stubs 2638 Compile *C = Compile::current(); 2639 if (!C->is_method_compilation()) return false; 2640 2641 assert(n->is_If(), "You should only call this on if nodes."); 2642 IfNode *ifn = n->as_If(); 2643 2644 Node *ifFalse = NULL; 2645 for (DUIterator_Fast imax, i = ifn->fast_outs(imax); i < imax; i++) { 2646 if (ifn->fast_out(i)->is_IfFalse()) { 2647 ifFalse = ifn->fast_out(i); 2648 break; 2649 } 2650 } 2651 assert(ifFalse, "An If should have an ifFalse. Graph is broken."); 2652 2653 Node *reg = ifFalse; 2654 int cnt = 4; // We must protect against cycles. Limit to 4 iterations. 2655 // Alternatively use visited set? Seems too expensive. 2656 while (reg != NULL && cnt > 0) { 2657 CallNode *call = NULL; 2658 RegionNode *nxt_reg = NULL; 2659 for (DUIterator_Fast imax, i = reg->fast_outs(imax); i < imax; i++) { 2660 Node *o = reg->fast_out(i); 2661 if (o->is_Call()) { 2662 call = o->as_Call(); 2663 } 2664 if (o->is_Region()) { 2665 nxt_reg = o->as_Region(); 2666 } 2667 } 2668 2669 if (call && 2670 call->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) { 2671 const Type* trtype = call->in(TypeFunc::Parms)->bottom_type(); 2672 if (trtype->isa_int() && trtype->is_int()->is_con()) { 2673 jint tr_con = trtype->is_int()->get_con(); 2674 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con); 2675 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con); 2676 assert((int)reason < (int)BitsPerInt, "recode bit map"); 2677 2678 if (is_set_nth_bit(C->allowed_deopt_reasons(), (int)reason) 2679 && action != Deoptimization::Action_none) { 2680 // This uncommon trap is sure to recompile, eventually. 2681 // When that happens, C->too_many_traps will prevent 2682 // this transformation from happening again. 2683 return true; 2684 } 2685 } 2686 } 2687 2688 reg = nxt_reg; 2689 cnt--; 2690 } 2691 2692 return false; 2693 } 2694 2695 //============================================================================= 2696 //---------------------------State--------------------------------------------- 2697 State::State(void) { 2698 #ifdef ASSERT 2699 _id = 0; 2700 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe); 2701 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d); 2702 //memset(_cost, -1, sizeof(_cost)); 2703 //memset(_rule, -1, sizeof(_rule)); 2704 #endif 2705 memset(_valid, 0, sizeof(_valid)); 2706 } 2707 2708 #ifdef ASSERT 2709 State::~State() { 2710 _id = 99; 2711 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe); 2712 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d); 2713 memset(_cost, -3, sizeof(_cost)); 2714 memset(_rule, -3, sizeof(_rule)); 2715 } 2716 #endif 2717 2718 #ifndef PRODUCT 2719 //---------------------------dump---------------------------------------------- 2720 void State::dump() { 2721 tty->print("\n"); 2722 dump(0); 2723 } 2724 2725 void State::dump(int depth) { 2726 for( int j = 0; j < depth; j++ ) 2727 tty->print(" "); 2728 tty->print("--N: "); 2729 _leaf->dump(); 2730 uint i; 2731 for( i = 0; i < _LAST_MACH_OPER; i++ ) 2732 // Check for valid entry 2733 if( valid(i) ) { 2734 for( int j = 0; j < depth; j++ ) 2735 tty->print(" "); 2736 assert(_cost[i] != max_juint, "cost must be a valid value"); 2737 assert(_rule[i] < _last_Mach_Node, "rule[i] must be valid rule"); 2738 tty->print_cr("%s %d %s", 2739 ruleName[i], _cost[i], ruleName[_rule[i]] ); 2740 } 2741 tty->cr(); 2742 2743 for( i=0; i<2; i++ ) 2744 if( _kids[i] ) 2745 _kids[i]->dump(depth+1); 2746 } 2747 #endif