Mercurial > hg > truffle
annotate src/share/vm/opto/postaloc.cpp @ 70:b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
Summary: copy elimination of a constant value results in incorrect execution
Reviewed-by: kvn, sgoldman, rasbold
| author | never |
|---|---|
| date | Wed, 19 Mar 2008 15:14:36 -0700 |
| parents | a61af66fc99e |
| children | d1605aabd0a1 |
| rev | line source |
|---|---|
| 0 | 1 /* |
| 2 * Copyright 1998-2007 Sun Microsystems, Inc. All Rights Reserved. | |
| 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. | |
| 4 * | |
| 5 * This code is free software; you can redistribute it and/or modify it | |
| 6 * under the terms of the GNU General Public License version 2 only, as | |
| 7 * published by the Free Software Foundation. | |
| 8 * | |
| 9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
| 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
| 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
| 12 * version 2 for more details (a copy is included in the LICENSE file that | |
| 13 * accompanied this code). | |
| 14 * | |
| 15 * You should have received a copy of the GNU General Public License version | |
| 16 * 2 along with this work; if not, write to the Free Software Foundation, | |
| 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
| 18 * | |
| 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
| 20 * CA 95054 USA or visit www.sun.com if you need additional information or | |
| 21 * have any questions. | |
| 22 * | |
| 23 */ | |
| 24 | |
| 25 #include "incls/_precompiled.incl" | |
| 26 #include "incls/_postaloc.cpp.incl" | |
| 27 | |
| 28 // see if this register kind does not requires two registers | |
| 29 static bool is_single_register(uint x) { | |
| 30 #ifdef _LP64 | |
| 31 return (x != Op_RegD && x != Op_RegL && x != Op_RegP); | |
| 32 #else | |
| 33 return (x != Op_RegD && x != Op_RegL); | |
| 34 #endif | |
| 35 } | |
| 36 | |
| 37 //------------------------------may_be_copy_of_callee----------------------------- | |
| 38 // Check to see if we can possibly be a copy of a callee-save value. | |
| 39 bool PhaseChaitin::may_be_copy_of_callee( Node *def ) const { | |
| 40 // Short circuit if there are no callee save registers | |
| 41 if (_matcher.number_of_saved_registers() == 0) return false; | |
| 42 | |
| 43 // Expect only a spill-down and reload on exit for callee-save spills. | |
| 44 // Chains of copies cannot be deep. | |
| 45 // 5008997 - This is wishful thinking. Register allocator seems to | |
| 46 // be splitting live ranges for callee save registers to such | |
| 47 // an extent that in large methods the chains can be very long | |
| 48 // (50+). The conservative answer is to return true if we don't | |
| 49 // know as this prevents optimizations from occuring. | |
| 50 | |
| 51 const int limit = 60; | |
| 52 int i; | |
| 53 for( i=0; i < limit; i++ ) { | |
| 54 if( def->is_Proj() && def->in(0)->is_Start() && | |
| 55 _matcher.is_save_on_entry(lrgs(n2lidx(def)).reg()) ) | |
| 56 return true; // Direct use of callee-save proj | |
| 57 if( def->is_Copy() ) // Copies carry value through | |
| 58 def = def->in(def->is_Copy()); | |
| 59 else if( def->is_Phi() ) // Phis can merge it from any direction | |
| 60 def = def->in(1); | |
| 61 else | |
| 62 break; | |
| 63 guarantee(def != NULL, "must not resurrect dead copy"); | |
| 64 } | |
| 65 // If we reached the end and didn't find a callee save proj | |
| 66 // then this may be a callee save proj so we return true | |
| 67 // as the conservative answer. If we didn't reach then end | |
| 68 // we must have discovered that it was not a callee save | |
| 69 // else we would have returned. | |
| 70 return i == limit; | |
| 71 } | |
| 72 | |
| 73 | |
| 74 | |
| 75 //------------------------------yank_if_dead----------------------------------- | |
| 76 // Removed an edge from 'old'. Yank if dead. Return adjustment counts to | |
| 77 // iterators in the current block. | |
| 78 int PhaseChaitin::yank_if_dead( Node *old, Block *current_block, Node_List *value, Node_List *regnd ) { | |
| 79 int blk_adjust=0; | |
| 80 while (old->outcnt() == 0 && old != C->top()) { | |
| 81 Block *oldb = _cfg._bbs[old->_idx]; | |
| 82 oldb->find_remove(old); | |
| 83 // Count 1 if deleting an instruction from the current block | |
| 84 if( oldb == current_block ) blk_adjust++; | |
| 85 _cfg._bbs.map(old->_idx,NULL); | |
| 86 OptoReg::Name old_reg = lrgs(n2lidx(old)).reg(); | |
| 87 if( regnd && (*regnd)[old_reg]==old ) { // Instruction is currently available? | |
| 88 value->map(old_reg,NULL); // Yank from value/regnd maps | |
| 89 regnd->map(old_reg,NULL); // This register's value is now unknown | |
| 90 } | |
| 91 Node *tmp = old->req() > 1 ? old->in(1) : NULL; | |
| 92 old->disconnect_inputs(NULL); | |
| 93 if( !tmp ) break; | |
| 94 old = tmp; | |
| 95 } | |
| 96 return blk_adjust; | |
| 97 } | |
| 98 | |
| 99 //------------------------------use_prior_register----------------------------- | |
| 100 // Use the prior value instead of the current value, in an effort to make | |
| 101 // the current value go dead. Return block iterator adjustment, in case | |
| 102 // we yank some instructions from this block. | |
| 103 int PhaseChaitin::use_prior_register( Node *n, uint idx, Node *def, Block *current_block, Node_List &value, Node_List ®nd ) { | |
| 104 // No effect? | |
| 105 if( def == n->in(idx) ) return 0; | |
| 106 // Def is currently dead and can be removed? Do not resurrect | |
| 107 if( def->outcnt() == 0 ) return 0; | |
| 108 | |
| 109 // Not every pair of physical registers are assignment compatible, | |
| 110 // e.g. on sparc floating point registers are not assignable to integer | |
| 111 // registers. | |
| 112 const LRG &def_lrg = lrgs(n2lidx(def)); | |
| 113 OptoReg::Name def_reg = def_lrg.reg(); | |
| 114 const RegMask &use_mask = n->in_RegMask(idx); | |
| 115 bool can_use = ( RegMask::can_represent(def_reg) ? (use_mask.Member(def_reg) != 0) | |
| 116 : (use_mask.is_AllStack() != 0)); | |
| 117 // Check for a copy to or from a misaligned pair. | |
| 118 can_use = can_use && !use_mask.is_misaligned_Pair() && !def_lrg.mask().is_misaligned_Pair(); | |
| 119 | |
| 120 if (!can_use) | |
| 121 return 0; | |
| 122 | |
| 123 // Capture the old def in case it goes dead... | |
| 124 Node *old = n->in(idx); | |
| 125 | |
| 126 // Save-on-call copies can only be elided if the entire copy chain can go | |
| 127 // away, lest we get the same callee-save value alive in 2 locations at | |
| 128 // once. We check for the obvious trivial case here. Although it can | |
| 129 // sometimes be elided with cooperation outside our scope, here we will just | |
| 130 // miss the opportunity. :-( | |
| 131 if( may_be_copy_of_callee(def) ) { | |
| 132 if( old->outcnt() > 1 ) return 0; // We're the not last user | |
| 133 int idx = old->is_Copy(); | |
| 134 assert( idx, "chain of copies being removed" ); | |
| 135 Node *old2 = old->in(idx); // Chain of copies | |
| 136 if( old2->outcnt() > 1 ) return 0; // old is not the last user | |
| 137 int idx2 = old2->is_Copy(); | |
| 138 if( !idx2 ) return 0; // Not a chain of 2 copies | |
| 139 if( def != old2->in(idx2) ) return 0; // Chain of exactly 2 copies | |
| 140 } | |
| 141 | |
| 142 // Use the new def | |
| 143 n->set_req(idx,def); | |
| 144 _post_alloc++; | |
| 145 | |
| 146 // Is old def now dead? We successfully yanked a copy? | |
| 147 return yank_if_dead(old,current_block,&value,®nd); | |
| 148 } | |
| 149 | |
| 150 | |
| 151 //------------------------------skip_copies------------------------------------ | |
| 152 // Skip through any number of copies (that don't mod oop-i-ness) | |
| 153 Node *PhaseChaitin::skip_copies( Node *c ) { | |
| 154 int idx = c->is_Copy(); | |
| 155 uint is_oop = lrgs(n2lidx(c))._is_oop; | |
| 156 while (idx != 0) { | |
| 157 guarantee(c->in(idx) != NULL, "must not resurrect dead copy"); | |
| 158 if (lrgs(n2lidx(c->in(idx)))._is_oop != is_oop) | |
| 159 break; // casting copy, not the same value | |
| 160 c = c->in(idx); | |
| 161 idx = c->is_Copy(); | |
| 162 } | |
| 163 return c; | |
| 164 } | |
| 165 | |
| 166 //------------------------------elide_copy------------------------------------- | |
| 167 // Remove (bypass) copies along Node n, edge k. | |
| 168 int PhaseChaitin::elide_copy( Node *n, int k, Block *current_block, Node_List &value, Node_List ®nd, bool can_change_regs ) { | |
| 169 int blk_adjust = 0; | |
| 170 | |
| 171 uint nk_idx = n2lidx(n->in(k)); | |
| 172 OptoReg::Name nk_reg = lrgs(nk_idx ).reg(); | |
| 173 | |
| 174 // Remove obvious same-register copies | |
| 175 Node *x = n->in(k); | |
| 176 int idx; | |
| 177 while( (idx=x->is_Copy()) != 0 ) { | |
| 178 Node *copy = x->in(idx); | |
| 179 guarantee(copy != NULL, "must not resurrect dead copy"); | |
| 180 if( lrgs(n2lidx(copy)).reg() != nk_reg ) break; | |
| 181 blk_adjust += use_prior_register(n,k,copy,current_block,value,regnd); | |
| 182 if( n->in(k) != copy ) break; // Failed for some cutout? | |
| 183 x = copy; // Progress, try again | |
| 184 } | |
| 185 | |
| 186 // Phis and 2-address instructions cannot change registers so easily - their | |
| 187 // outputs must match their input. | |
| 188 if( !can_change_regs ) | |
| 189 return blk_adjust; // Only check stupid copies! | |
| 190 | |
| 191 // Loop backedges won't have a value-mapping yet | |
| 192 if( &value == NULL ) return blk_adjust; | |
| 193 | |
| 194 // Skip through all copies to the _value_ being used. Do not change from | |
| 195 // int to pointer. This attempts to jump through a chain of copies, where | |
| 196 // intermediate copies might be illegal, i.e., value is stored down to stack | |
| 197 // then reloaded BUT survives in a register the whole way. | |
| 198 Node *val = skip_copies(n->in(k)); | |
| 199 | |
| 200 if( val == x ) return blk_adjust; // No progress? | |
| 201 | |
| 202 bool single = is_single_register(val->ideal_reg()); | |
| 203 uint val_idx = n2lidx(val); | |
| 204 OptoReg::Name val_reg = lrgs(val_idx).reg(); | |
| 205 | |
| 206 // See if it happens to already be in the correct register! | |
| 207 // (either Phi's direct register, or the common case of the name | |
| 208 // never-clobbered original-def register) | |
| 209 if( value[val_reg] == val && | |
| 210 // Doubles check both halves | |
| 211 ( single || value[val_reg-1] == val ) ) { | |
| 212 blk_adjust += use_prior_register(n,k,regnd[val_reg],current_block,value,regnd); | |
| 213 if( n->in(k) == regnd[val_reg] ) // Success! Quit trying | |
| 214 return blk_adjust; | |
| 215 } | |
| 216 | |
| 217 // See if we can skip the copy by changing registers. Don't change from | |
| 218 // using a register to using the stack unless we know we can remove a | |
| 219 // copy-load. Otherwise we might end up making a pile of Intel cisc-spill | |
| 220 // ops reading from memory instead of just loading once and using the | |
| 221 // register. | |
| 222 | |
| 223 // Also handle duplicate copies here. | |
| 224 const Type *t = val->is_Con() ? val->bottom_type() : NULL; | |
| 225 | |
| 226 // Scan all registers to see if this value is around already | |
| 227 for( uint reg = 0; reg < (uint)_max_reg; reg++ ) { | |
| 228 Node *vv = value[reg]; | |
| 229 if( !single ) { // Doubles check for aligned-adjacent pair | |
| 230 if( (reg&1)==0 ) continue; // Wrong half of a pair | |
| 231 if( vv != value[reg-1] ) continue; // Not a complete pair | |
| 232 } | |
| 233 if( vv == val || // Got a direct hit? | |
| 234 (t && vv && vv->bottom_type() == t && vv->is_Mach() && | |
| 235 vv->as_Mach()->rule() == val->as_Mach()->rule()) ) { // Or same constant? | |
| 236 assert( !n->is_Phi(), "cannot change registers at a Phi so easily" ); | |
| 237 if( OptoReg::is_stack(nk_reg) || // CISC-loading from stack OR | |
| 238 OptoReg::is_reg(reg) || // turning into a register use OR | |
| 239 regnd[reg]->outcnt()==1 ) { // last use of a spill-load turns into a CISC use | |
| 240 blk_adjust += use_prior_register(n,k,regnd[reg],current_block,value,regnd); | |
| 241 if( n->in(k) == regnd[reg] ) // Success! Quit trying | |
| 242 return blk_adjust; | |
| 243 } // End of if not degrading to a stack | |
| 244 } // End of if found value in another register | |
| 245 } // End of scan all machine registers | |
| 246 return blk_adjust; | |
| 247 } | |
| 248 | |
| 249 | |
| 250 // | |
| 251 // Check if nreg already contains the constant value val. Normal copy | |
| 252 // elimination doesn't doesn't work on constants because multiple | |
| 253 // nodes can represent the same constant so the type and rule of the | |
| 254 // MachNode must be checked to ensure equivalence. | |
| 255 // | |
|
70
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
256 bool PhaseChaitin::eliminate_copy_of_constant(Node* val, Node* n, |
|
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
257 Block *current_block, |
| 0 | 258 Node_List& value, Node_List& regnd, |
| 259 OptoReg::Name nreg, OptoReg::Name nreg2) { | |
| 260 if (value[nreg] != val && val->is_Con() && | |
| 261 value[nreg] != NULL && value[nreg]->is_Con() && | |
| 262 (nreg2 == OptoReg::Bad || value[nreg] == value[nreg2]) && | |
| 263 value[nreg]->bottom_type() == val->bottom_type() && | |
| 264 value[nreg]->as_Mach()->rule() == val->as_Mach()->rule()) { | |
| 265 // This code assumes that two MachNodes representing constants | |
| 266 // which have the same rule and the same bottom type will produce | |
| 267 // identical effects into a register. This seems like it must be | |
| 268 // objectively true unless there are hidden inputs to the nodes | |
| 269 // but if that were to change this code would need to updated. | |
| 270 // Since they are equivalent the second one if redundant and can | |
| 271 // be removed. | |
| 272 // | |
|
70
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
273 // n will be replaced with the old value but n might have |
| 0 | 274 // kills projections associated with it so remove them now so that |
| 275 // yank_if_dead will be able to elminate the copy once the uses | |
| 276 // have been transferred to the old[value]. | |
|
70
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
277 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { |
|
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
278 Node* use = n->fast_out(i); |
| 0 | 279 if (use->is_Proj() && use->outcnt() == 0) { |
| 280 // Kill projections have no users and one input | |
| 281 use->set_req(0, C->top()); | |
| 282 yank_if_dead(use, current_block, &value, ®nd); | |
| 283 --i; --imax; | |
| 284 } | |
| 285 } | |
| 286 _post_alloc++; | |
| 287 return true; | |
| 288 } | |
| 289 return false; | |
| 290 } | |
| 291 | |
| 292 | |
| 293 //------------------------------post_allocate_copy_removal--------------------- | |
| 294 // Post-Allocation peephole copy removal. We do this in 1 pass over the | |
| 295 // basic blocks. We maintain a mapping of registers to Nodes (an array of | |
| 296 // Nodes indexed by machine register or stack slot number). NULL means that a | |
| 297 // register is not mapped to any Node. We can (want to have!) have several | |
| 298 // registers map to the same Node. We walk forward over the instructions | |
| 299 // updating the mapping as we go. At merge points we force a NULL if we have | |
| 300 // to merge 2 different Nodes into the same register. Phi functions will give | |
| 301 // us a new Node if there is a proper value merging. Since the blocks are | |
| 302 // arranged in some RPO, we will visit all parent blocks before visiting any | |
| 303 // successor blocks (except at loops). | |
| 304 // | |
| 305 // If we find a Copy we look to see if the Copy's source register is a stack | |
| 306 // slot and that value has already been loaded into some machine register; if | |
| 307 // so we use machine register directly. This turns a Load into a reg-reg | |
| 308 // Move. We also look for reloads of identical constants. | |
| 309 // | |
| 310 // When we see a use from a reg-reg Copy, we will attempt to use the copy's | |
| 311 // source directly and make the copy go dead. | |
| 312 void PhaseChaitin::post_allocate_copy_removal() { | |
| 313 NOT_PRODUCT( Compile::TracePhase t3("postAllocCopyRemoval", &_t_postAllocCopyRemoval, TimeCompiler); ) | |
| 314 ResourceMark rm; | |
| 315 | |
| 316 // Need a mapping from basic block Node_Lists. We need a Node_List to | |
| 317 // map from register number to value-producing Node. | |
| 318 Node_List **blk2value = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1); | |
| 319 memset( blk2value, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) ); | |
| 320 // Need a mapping from basic block Node_Lists. We need a Node_List to | |
| 321 // map from register number to register-defining Node. | |
| 322 Node_List **blk2regnd = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1); | |
| 323 memset( blk2regnd, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) ); | |
| 324 | |
| 325 // We keep unused Node_Lists on a free_list to avoid wasting | |
| 326 // memory. | |
| 327 GrowableArray<Node_List*> free_list = GrowableArray<Node_List*>(16); | |
| 328 | |
| 329 // For all blocks | |
| 330 for( uint i = 0; i < _cfg._num_blocks; i++ ) { | |
| 331 uint j; | |
| 332 Block *b = _cfg._blocks[i]; | |
| 333 | |
| 334 // Count of Phis in block | |
| 335 uint phi_dex; | |
| 336 for( phi_dex = 1; phi_dex < b->_nodes.size(); phi_dex++ ) { | |
| 337 Node *phi = b->_nodes[phi_dex]; | |
| 338 if( !phi->is_Phi() ) | |
| 339 break; | |
| 340 } | |
| 341 | |
| 342 // If any predecessor has not been visited, we do not know the state | |
| 343 // of registers at the start. Check for this, while updating copies | |
| 344 // along Phi input edges | |
| 345 bool missing_some_inputs = false; | |
| 346 Block *freed = NULL; | |
| 347 for( j = 1; j < b->num_preds(); j++ ) { | |
| 348 Block *pb = _cfg._bbs[b->pred(j)->_idx]; | |
| 349 // Remove copies along phi edges | |
| 350 for( uint k=1; k<phi_dex; k++ ) | |
| 351 elide_copy( b->_nodes[k], j, b, *blk2value[pb->_pre_order], *blk2regnd[pb->_pre_order], false ); | |
| 352 if( blk2value[pb->_pre_order] ) { // Have a mapping on this edge? | |
| 353 // See if this predecessor's mappings have been used by everybody | |
| 354 // who wants them. If so, free 'em. | |
| 355 uint k; | |
| 356 for( k=0; k<pb->_num_succs; k++ ) { | |
| 357 Block *pbsucc = pb->_succs[k]; | |
| 358 if( !blk2value[pbsucc->_pre_order] && pbsucc != b ) | |
| 359 break; // Found a future user | |
| 360 } | |
| 361 if( k >= pb->_num_succs ) { // No more uses, free! | |
| 362 freed = pb; // Record last block freed | |
| 363 free_list.push(blk2value[pb->_pre_order]); | |
| 364 free_list.push(blk2regnd[pb->_pre_order]); | |
| 365 } | |
| 366 } else { // This block has unvisited (loopback) inputs | |
| 367 missing_some_inputs = true; | |
| 368 } | |
| 369 } | |
| 370 | |
| 371 | |
| 372 // Extract Node_List mappings. If 'freed' is non-zero, we just popped | |
| 373 // 'freed's blocks off the list | |
| 374 Node_List ®nd = *(free_list.is_empty() ? new Node_List() : free_list.pop()); | |
| 375 Node_List &value = *(free_list.is_empty() ? new Node_List() : free_list.pop()); | |
| 376 assert( !freed || blk2value[freed->_pre_order] == &value, "" ); | |
| 377 value.map(_max_reg,NULL); | |
| 378 regnd.map(_max_reg,NULL); | |
| 379 // Set mappings as OUR mappings | |
| 380 blk2value[b->_pre_order] = &value; | |
| 381 blk2regnd[b->_pre_order] = ®nd; | |
| 382 | |
| 383 // Initialize value & regnd for this block | |
| 384 if( missing_some_inputs ) { | |
| 385 // Some predecessor has not yet been visited; zap map to empty | |
| 386 for( uint k = 0; k < (uint)_max_reg; k++ ) { | |
| 387 value.map(k,NULL); | |
| 388 regnd.map(k,NULL); | |
| 389 } | |
| 390 } else { | |
| 391 if( !freed ) { // Didn't get a freebie prior block | |
| 392 // Must clone some data | |
| 393 freed = _cfg._bbs[b->pred(1)->_idx]; | |
| 394 Node_List &f_value = *blk2value[freed->_pre_order]; | |
| 395 Node_List &f_regnd = *blk2regnd[freed->_pre_order]; | |
| 396 for( uint k = 0; k < (uint)_max_reg; k++ ) { | |
| 397 value.map(k,f_value[k]); | |
| 398 regnd.map(k,f_regnd[k]); | |
| 399 } | |
| 400 } | |
| 401 // Merge all inputs together, setting to NULL any conflicts. | |
| 402 for( j = 1; j < b->num_preds(); j++ ) { | |
| 403 Block *pb = _cfg._bbs[b->pred(j)->_idx]; | |
| 404 if( pb == freed ) continue; // Did self already via freelist | |
| 405 Node_List &p_regnd = *blk2regnd[pb->_pre_order]; | |
| 406 for( uint k = 0; k < (uint)_max_reg; k++ ) { | |
| 407 if( regnd[k] != p_regnd[k] ) { // Conflict on reaching defs? | |
| 408 value.map(k,NULL); // Then no value handy | |
| 409 regnd.map(k,NULL); | |
| 410 } | |
| 411 } | |
| 412 } | |
| 413 } | |
| 414 | |
| 415 // For all Phi's | |
| 416 for( j = 1; j < phi_dex; j++ ) { | |
| 417 uint k; | |
| 418 Node *phi = b->_nodes[j]; | |
| 419 uint pidx = n2lidx(phi); | |
| 420 OptoReg::Name preg = lrgs(n2lidx(phi)).reg(); | |
| 421 | |
| 422 // Remove copies remaining on edges. Check for junk phi. | |
| 423 Node *u = NULL; | |
| 424 for( k=1; k<phi->req(); k++ ) { | |
| 425 Node *x = phi->in(k); | |
| 426 if( phi != x && u != x ) // Found a different input | |
| 427 u = u ? NodeSentinel : x; // Capture unique input, or NodeSentinel for 2nd input | |
| 428 } | |
| 429 if( u != NodeSentinel ) { // Junk Phi. Remove | |
| 430 b->_nodes.remove(j--); phi_dex--; | |
| 431 _cfg._bbs.map(phi->_idx,NULL); | |
| 432 phi->replace_by(u); | |
| 433 phi->disconnect_inputs(NULL); | |
| 434 continue; | |
| 435 } | |
| 436 // Note that if value[pidx] exists, then we merged no new values here | |
| 437 // and the phi is useless. This can happen even with the above phi | |
| 438 // removal for complex flows. I cannot keep the better known value here | |
| 439 // because locally the phi appears to define a new merged value. If I | |
| 440 // keep the better value then a copy of the phi, being unable to use the | |
| 441 // global flow analysis, can't "peek through" the phi to the original | |
| 442 // reaching value and so will act like it's defining a new value. This | |
| 443 // can lead to situations where some uses are from the old and some from | |
| 444 // the new values. Not illegal by itself but throws the over-strong | |
| 445 // assert in scheduling. | |
| 446 if( pidx ) { | |
| 447 value.map(preg,phi); | |
| 448 regnd.map(preg,phi); | |
| 449 OptoReg::Name preg_lo = OptoReg::add(preg,-1); | |
| 450 if( !is_single_register(phi->ideal_reg()) ) { | |
| 451 value.map(preg_lo,phi); | |
| 452 regnd.map(preg_lo,phi); | |
| 453 } | |
| 454 } | |
| 455 } | |
| 456 | |
| 457 // For all remaining instructions | |
| 458 for( j = phi_dex; j < b->_nodes.size(); j++ ) { | |
| 459 Node *n = b->_nodes[j]; | |
| 460 | |
| 461 if( n->outcnt() == 0 && // Dead? | |
| 462 n != C->top() && // (ignore TOP, it has no du info) | |
| 463 !n->is_Proj() ) { // fat-proj kills | |
| 464 j -= yank_if_dead(n,b,&value,®nd); | |
| 465 continue; | |
| 466 } | |
| 467 | |
| 468 // Improve reaching-def info. Occasionally post-alloc's liveness gives | |
| 469 // up (at loop backedges, because we aren't doing a full flow pass). | |
| 470 // The presence of a live use essentially asserts that the use's def is | |
| 471 // alive and well at the use (or else the allocator fubar'd). Take | |
| 472 // advantage of this info to set a reaching def for the use-reg. | |
| 473 uint k; | |
| 474 for( k = 1; k < n->req(); k++ ) { | |
| 475 Node *def = n->in(k); // n->in(k) is a USE; def is the DEF for this USE | |
| 476 guarantee(def != NULL, "no disconnected nodes at this point"); | |
| 477 uint useidx = n2lidx(def); // useidx is the live range index for this USE | |
| 478 | |
| 479 if( useidx ) { | |
| 480 OptoReg::Name ureg = lrgs(useidx).reg(); | |
| 481 if( !value[ureg] ) { | |
| 482 int idx; // Skip occasional useless copy | |
| 483 while( (idx=def->is_Copy()) != 0 && | |
| 484 def->in(idx) != NULL && // NULL should not happen | |
| 485 ureg == lrgs(n2lidx(def->in(idx))).reg() ) | |
| 486 def = def->in(idx); | |
| 487 Node *valdef = skip_copies(def); // tighten up val through non-useless copies | |
| 488 value.map(ureg,valdef); // record improved reaching-def info | |
| 489 regnd.map(ureg, def); | |
| 490 // Record other half of doubles | |
| 491 OptoReg::Name ureg_lo = OptoReg::add(ureg,-1); | |
| 492 if( !is_single_register(def->ideal_reg()) && | |
| 493 ( !RegMask::can_represent(ureg_lo) || | |
| 494 lrgs(useidx).mask().Member(ureg_lo) ) && // Nearly always adjacent | |
| 495 !value[ureg_lo] ) { | |
| 496 value.map(ureg_lo,valdef); // record improved reaching-def info | |
| 497 regnd.map(ureg_lo, def); | |
| 498 } | |
| 499 } | |
| 500 } | |
| 501 } | |
| 502 | |
| 503 const uint two_adr = n->is_Mach() ? n->as_Mach()->two_adr() : 0; | |
| 504 | |
| 505 // Remove copies along input edges | |
| 506 for( k = 1; k < n->req(); k++ ) | |
| 507 j -= elide_copy( n, k, b, value, regnd, two_adr!=k ); | |
| 508 | |
| 509 // Unallocated Nodes define no registers | |
| 510 uint lidx = n2lidx(n); | |
| 511 if( !lidx ) continue; | |
| 512 | |
| 513 // Update the register defined by this instruction | |
| 514 OptoReg::Name nreg = lrgs(lidx).reg(); | |
| 515 // Skip through all copies to the _value_ being defined. | |
| 516 // Do not change from int to pointer | |
| 517 Node *val = skip_copies(n); | |
| 518 | |
| 519 uint n_ideal_reg = n->ideal_reg(); | |
| 520 if( is_single_register(n_ideal_reg) ) { | |
| 521 // If Node 'n' does not change the value mapped by the register, | |
| 522 // then 'n' is a useless copy. Do not update the register->node | |
| 523 // mapping so 'n' will go dead. | |
| 524 if( value[nreg] != val ) { | |
|
70
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
525 if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, OptoReg::Bad)) { |
| 0 | 526 n->replace_by(regnd[nreg]); |
| 527 j -= yank_if_dead(n,b,&value,®nd); | |
| 528 } else { | |
| 529 // Update the mapping: record new Node defined by the register | |
| 530 regnd.map(nreg,n); | |
| 531 // Update mapping for defined *value*, which is the defined | |
| 532 // Node after skipping all copies. | |
| 533 value.map(nreg,val); | |
| 534 } | |
| 535 } else if( !may_be_copy_of_callee(n) && regnd[nreg]->outcnt() != 0 ) { | |
| 536 assert( n->is_Copy(), "" ); | |
| 537 n->replace_by(regnd[nreg]); | |
| 538 j -= yank_if_dead(n,b,&value,®nd); | |
| 539 } | |
| 540 } else { | |
| 541 // If the value occupies a register pair, record same info | |
| 542 // in both registers. | |
| 543 OptoReg::Name nreg_lo = OptoReg::add(nreg,-1); | |
| 544 if( RegMask::can_represent(nreg_lo) && // Either a spill slot, or | |
| 545 !lrgs(lidx).mask().Member(nreg_lo) ) { // Nearly always adjacent | |
| 546 // Sparc occasionally has non-adjacent pairs. | |
| 547 // Find the actual other value | |
| 548 RegMask tmp = lrgs(lidx).mask(); | |
| 549 tmp.Remove(nreg); | |
| 550 nreg_lo = tmp.find_first_elem(); | |
| 551 } | |
| 552 if( value[nreg] != val || value[nreg_lo] != val ) { | |
|
70
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
553 if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, nreg_lo)) { |
| 0 | 554 n->replace_by(regnd[nreg]); |
| 555 j -= yank_if_dead(n,b,&value,®nd); | |
| 556 } else { | |
| 557 regnd.map(nreg , n ); | |
| 558 regnd.map(nreg_lo, n ); | |
| 559 value.map(nreg ,val); | |
| 560 value.map(nreg_lo,val); | |
| 561 } | |
| 562 } else if( !may_be_copy_of_callee(n) && regnd[nreg]->outcnt() != 0 ) { | |
| 563 assert( n->is_Copy(), "" ); | |
| 564 n->replace_by(regnd[nreg]); | |
| 565 j -= yank_if_dead(n,b,&value,®nd); | |
| 566 } | |
| 567 } | |
| 568 | |
| 569 // Fat projections kill many registers | |
| 570 if( n_ideal_reg == MachProjNode::fat_proj ) { | |
| 571 RegMask rm = n->out_RegMask(); | |
| 572 // wow, what an expensive iterator... | |
| 573 nreg = rm.find_first_elem(); | |
| 574 while( OptoReg::is_valid(nreg)) { | |
| 575 rm.Remove(nreg); | |
| 576 value.map(nreg,n); | |
| 577 regnd.map(nreg,n); | |
| 578 nreg = rm.find_first_elem(); | |
| 579 } | |
| 580 } | |
| 581 | |
| 582 } // End of for all instructions in the block | |
| 583 | |
| 584 } // End for all blocks | |
| 585 } |