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