Fri, 12 Feb 2010 15:27:36 -0800
Merge
duke@435 | 1 | /* |
xdono@1383 | 2 | * Copyright 1998-2009 Sun Microsystems, Inc. All Rights Reserved. |
duke@435 | 3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
duke@435 | 4 | * |
duke@435 | 5 | * This code is free software; you can redistribute it and/or modify it |
duke@435 | 6 | * under the terms of the GNU General Public License version 2 only, as |
duke@435 | 7 | * published by the Free Software Foundation. |
duke@435 | 8 | * |
duke@435 | 9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
duke@435 | 10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
duke@435 | 11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
duke@435 | 12 | * version 2 for more details (a copy is included in the LICENSE file that |
duke@435 | 13 | * accompanied this code). |
duke@435 | 14 | * |
duke@435 | 15 | * You should have received a copy of the GNU General Public License version |
duke@435 | 16 | * 2 along with this work; if not, write to the Free Software Foundation, |
duke@435 | 17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
duke@435 | 18 | * |
duke@435 | 19 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, |
duke@435 | 20 | * CA 95054 USA or visit www.sun.com if you need additional information or |
duke@435 | 21 | * have any questions. |
duke@435 | 22 | * |
duke@435 | 23 | */ |
duke@435 | 24 | |
duke@435 | 25 | #include "incls/_precompiled.incl" |
duke@435 | 26 | #include "incls/_postaloc.cpp.incl" |
duke@435 | 27 | |
duke@435 | 28 | // see if this register kind does not requires two registers |
duke@435 | 29 | static bool is_single_register(uint x) { |
duke@435 | 30 | #ifdef _LP64 |
duke@435 | 31 | return (x != Op_RegD && x != Op_RegL && x != Op_RegP); |
duke@435 | 32 | #else |
duke@435 | 33 | return (x != Op_RegD && x != Op_RegL); |
duke@435 | 34 | #endif |
duke@435 | 35 | } |
duke@435 | 36 | |
kvn@835 | 37 | //---------------------------may_be_copy_of_callee----------------------------- |
duke@435 | 38 | // Check to see if we can possibly be a copy of a callee-save value. |
duke@435 | 39 | bool PhaseChaitin::may_be_copy_of_callee( Node *def ) const { |
duke@435 | 40 | // Short circuit if there are no callee save registers |
duke@435 | 41 | if (_matcher.number_of_saved_registers() == 0) return false; |
duke@435 | 42 | |
duke@435 | 43 | // Expect only a spill-down and reload on exit for callee-save spills. |
duke@435 | 44 | // Chains of copies cannot be deep. |
duke@435 | 45 | // 5008997 - This is wishful thinking. Register allocator seems to |
duke@435 | 46 | // be splitting live ranges for callee save registers to such |
duke@435 | 47 | // an extent that in large methods the chains can be very long |
duke@435 | 48 | // (50+). The conservative answer is to return true if we don't |
twisti@1040 | 49 | // know as this prevents optimizations from occurring. |
duke@435 | 50 | |
duke@435 | 51 | const int limit = 60; |
duke@435 | 52 | int i; |
duke@435 | 53 | for( i=0; i < limit; i++ ) { |
duke@435 | 54 | if( def->is_Proj() && def->in(0)->is_Start() && |
duke@435 | 55 | _matcher.is_save_on_entry(lrgs(n2lidx(def)).reg()) ) |
duke@435 | 56 | return true; // Direct use of callee-save proj |
duke@435 | 57 | if( def->is_Copy() ) // Copies carry value through |
duke@435 | 58 | def = def->in(def->is_Copy()); |
duke@435 | 59 | else if( def->is_Phi() ) // Phis can merge it from any direction |
duke@435 | 60 | def = def->in(1); |
duke@435 | 61 | else |
duke@435 | 62 | break; |
duke@435 | 63 | guarantee(def != NULL, "must not resurrect dead copy"); |
duke@435 | 64 | } |
duke@435 | 65 | // If we reached the end and didn't find a callee save proj |
duke@435 | 66 | // then this may be a callee save proj so we return true |
duke@435 | 67 | // as the conservative answer. If we didn't reach then end |
duke@435 | 68 | // we must have discovered that it was not a callee save |
duke@435 | 69 | // else we would have returned. |
duke@435 | 70 | return i == limit; |
duke@435 | 71 | } |
duke@435 | 72 | |
duke@435 | 73 | |
duke@435 | 74 | |
duke@435 | 75 | //------------------------------yank_if_dead----------------------------------- |
duke@435 | 76 | // Removed an edge from 'old'. Yank if dead. Return adjustment counts to |
duke@435 | 77 | // iterators in the current block. |
duke@435 | 78 | int PhaseChaitin::yank_if_dead( Node *old, Block *current_block, Node_List *value, Node_List *regnd ) { |
duke@435 | 79 | int blk_adjust=0; |
duke@435 | 80 | while (old->outcnt() == 0 && old != C->top()) { |
duke@435 | 81 | Block *oldb = _cfg._bbs[old->_idx]; |
duke@435 | 82 | oldb->find_remove(old); |
duke@435 | 83 | // Count 1 if deleting an instruction from the current block |
duke@435 | 84 | if( oldb == current_block ) blk_adjust++; |
duke@435 | 85 | _cfg._bbs.map(old->_idx,NULL); |
duke@435 | 86 | OptoReg::Name old_reg = lrgs(n2lidx(old)).reg(); |
duke@435 | 87 | if( regnd && (*regnd)[old_reg]==old ) { // Instruction is currently available? |
duke@435 | 88 | value->map(old_reg,NULL); // Yank from value/regnd maps |
duke@435 | 89 | regnd->map(old_reg,NULL); // This register's value is now unknown |
duke@435 | 90 | } |
never@1358 | 91 | assert(old->req() <= 2, "can't handle more inputs"); |
duke@435 | 92 | Node *tmp = old->req() > 1 ? old->in(1) : NULL; |
duke@435 | 93 | old->disconnect_inputs(NULL); |
duke@435 | 94 | if( !tmp ) break; |
duke@435 | 95 | old = tmp; |
duke@435 | 96 | } |
duke@435 | 97 | return blk_adjust; |
duke@435 | 98 | } |
duke@435 | 99 | |
duke@435 | 100 | //------------------------------use_prior_register----------------------------- |
duke@435 | 101 | // Use the prior value instead of the current value, in an effort to make |
duke@435 | 102 | // the current value go dead. Return block iterator adjustment, in case |
duke@435 | 103 | // we yank some instructions from this block. |
duke@435 | 104 | int PhaseChaitin::use_prior_register( Node *n, uint idx, Node *def, Block *current_block, Node_List &value, Node_List ®nd ) { |
duke@435 | 105 | // No effect? |
duke@435 | 106 | if( def == n->in(idx) ) return 0; |
duke@435 | 107 | // Def is currently dead and can be removed? Do not resurrect |
duke@435 | 108 | if( def->outcnt() == 0 ) return 0; |
duke@435 | 109 | |
duke@435 | 110 | // Not every pair of physical registers are assignment compatible, |
duke@435 | 111 | // e.g. on sparc floating point registers are not assignable to integer |
duke@435 | 112 | // registers. |
duke@435 | 113 | const LRG &def_lrg = lrgs(n2lidx(def)); |
duke@435 | 114 | OptoReg::Name def_reg = def_lrg.reg(); |
duke@435 | 115 | const RegMask &use_mask = n->in_RegMask(idx); |
duke@435 | 116 | bool can_use = ( RegMask::can_represent(def_reg) ? (use_mask.Member(def_reg) != 0) |
duke@435 | 117 | : (use_mask.is_AllStack() != 0)); |
duke@435 | 118 | // Check for a copy to or from a misaligned pair. |
duke@435 | 119 | can_use = can_use && !use_mask.is_misaligned_Pair() && !def_lrg.mask().is_misaligned_Pair(); |
duke@435 | 120 | |
duke@435 | 121 | if (!can_use) |
duke@435 | 122 | return 0; |
duke@435 | 123 | |
duke@435 | 124 | // Capture the old def in case it goes dead... |
duke@435 | 125 | Node *old = n->in(idx); |
duke@435 | 126 | |
duke@435 | 127 | // Save-on-call copies can only be elided if the entire copy chain can go |
duke@435 | 128 | // away, lest we get the same callee-save value alive in 2 locations at |
duke@435 | 129 | // once. We check for the obvious trivial case here. Although it can |
duke@435 | 130 | // sometimes be elided with cooperation outside our scope, here we will just |
duke@435 | 131 | // miss the opportunity. :-( |
duke@435 | 132 | if( may_be_copy_of_callee(def) ) { |
duke@435 | 133 | if( old->outcnt() > 1 ) return 0; // We're the not last user |
duke@435 | 134 | int idx = old->is_Copy(); |
duke@435 | 135 | assert( idx, "chain of copies being removed" ); |
duke@435 | 136 | Node *old2 = old->in(idx); // Chain of copies |
duke@435 | 137 | if( old2->outcnt() > 1 ) return 0; // old is not the last user |
duke@435 | 138 | int idx2 = old2->is_Copy(); |
duke@435 | 139 | if( !idx2 ) return 0; // Not a chain of 2 copies |
duke@435 | 140 | if( def != old2->in(idx2) ) return 0; // Chain of exactly 2 copies |
duke@435 | 141 | } |
duke@435 | 142 | |
duke@435 | 143 | // Use the new def |
duke@435 | 144 | n->set_req(idx,def); |
duke@435 | 145 | _post_alloc++; |
duke@435 | 146 | |
duke@435 | 147 | // Is old def now dead? We successfully yanked a copy? |
duke@435 | 148 | return yank_if_dead(old,current_block,&value,®nd); |
duke@435 | 149 | } |
duke@435 | 150 | |
duke@435 | 151 | |
duke@435 | 152 | //------------------------------skip_copies------------------------------------ |
duke@435 | 153 | // Skip through any number of copies (that don't mod oop-i-ness) |
duke@435 | 154 | Node *PhaseChaitin::skip_copies( Node *c ) { |
duke@435 | 155 | int idx = c->is_Copy(); |
duke@435 | 156 | uint is_oop = lrgs(n2lidx(c))._is_oop; |
duke@435 | 157 | while (idx != 0) { |
duke@435 | 158 | guarantee(c->in(idx) != NULL, "must not resurrect dead copy"); |
duke@435 | 159 | if (lrgs(n2lidx(c->in(idx)))._is_oop != is_oop) |
duke@435 | 160 | break; // casting copy, not the same value |
duke@435 | 161 | c = c->in(idx); |
duke@435 | 162 | idx = c->is_Copy(); |
duke@435 | 163 | } |
duke@435 | 164 | return c; |
duke@435 | 165 | } |
duke@435 | 166 | |
duke@435 | 167 | //------------------------------elide_copy------------------------------------- |
duke@435 | 168 | // Remove (bypass) copies along Node n, edge k. |
duke@435 | 169 | int PhaseChaitin::elide_copy( Node *n, int k, Block *current_block, Node_List &value, Node_List ®nd, bool can_change_regs ) { |
duke@435 | 170 | int blk_adjust = 0; |
duke@435 | 171 | |
duke@435 | 172 | uint nk_idx = n2lidx(n->in(k)); |
duke@435 | 173 | OptoReg::Name nk_reg = lrgs(nk_idx ).reg(); |
duke@435 | 174 | |
duke@435 | 175 | // Remove obvious same-register copies |
duke@435 | 176 | Node *x = n->in(k); |
duke@435 | 177 | int idx; |
duke@435 | 178 | while( (idx=x->is_Copy()) != 0 ) { |
duke@435 | 179 | Node *copy = x->in(idx); |
duke@435 | 180 | guarantee(copy != NULL, "must not resurrect dead copy"); |
duke@435 | 181 | if( lrgs(n2lidx(copy)).reg() != nk_reg ) break; |
duke@435 | 182 | blk_adjust += use_prior_register(n,k,copy,current_block,value,regnd); |
duke@435 | 183 | if( n->in(k) != copy ) break; // Failed for some cutout? |
duke@435 | 184 | x = copy; // Progress, try again |
duke@435 | 185 | } |
duke@435 | 186 | |
duke@435 | 187 | // Phis and 2-address instructions cannot change registers so easily - their |
duke@435 | 188 | // outputs must match their input. |
duke@435 | 189 | if( !can_change_regs ) |
duke@435 | 190 | return blk_adjust; // Only check stupid copies! |
duke@435 | 191 | |
duke@435 | 192 | // Loop backedges won't have a value-mapping yet |
duke@435 | 193 | if( &value == NULL ) return blk_adjust; |
duke@435 | 194 | |
duke@435 | 195 | // Skip through all copies to the _value_ being used. Do not change from |
duke@435 | 196 | // int to pointer. This attempts to jump through a chain of copies, where |
duke@435 | 197 | // intermediate copies might be illegal, i.e., value is stored down to stack |
duke@435 | 198 | // then reloaded BUT survives in a register the whole way. |
duke@435 | 199 | Node *val = skip_copies(n->in(k)); |
duke@435 | 200 | |
duke@435 | 201 | if( val == x ) return blk_adjust; // No progress? |
duke@435 | 202 | |
duke@435 | 203 | bool single = is_single_register(val->ideal_reg()); |
duke@435 | 204 | uint val_idx = n2lidx(val); |
duke@435 | 205 | OptoReg::Name val_reg = lrgs(val_idx).reg(); |
duke@435 | 206 | |
duke@435 | 207 | // See if it happens to already be in the correct register! |
duke@435 | 208 | // (either Phi's direct register, or the common case of the name |
duke@435 | 209 | // never-clobbered original-def register) |
duke@435 | 210 | if( value[val_reg] == val && |
duke@435 | 211 | // Doubles check both halves |
duke@435 | 212 | ( single || value[val_reg-1] == val ) ) { |
duke@435 | 213 | blk_adjust += use_prior_register(n,k,regnd[val_reg],current_block,value,regnd); |
duke@435 | 214 | if( n->in(k) == regnd[val_reg] ) // Success! Quit trying |
duke@435 | 215 | return blk_adjust; |
duke@435 | 216 | } |
duke@435 | 217 | |
duke@435 | 218 | // See if we can skip the copy by changing registers. Don't change from |
duke@435 | 219 | // using a register to using the stack unless we know we can remove a |
duke@435 | 220 | // copy-load. Otherwise we might end up making a pile of Intel cisc-spill |
duke@435 | 221 | // ops reading from memory instead of just loading once and using the |
duke@435 | 222 | // register. |
duke@435 | 223 | |
duke@435 | 224 | // Also handle duplicate copies here. |
duke@435 | 225 | const Type *t = val->is_Con() ? val->bottom_type() : NULL; |
duke@435 | 226 | |
duke@435 | 227 | // Scan all registers to see if this value is around already |
duke@435 | 228 | for( uint reg = 0; reg < (uint)_max_reg; reg++ ) { |
kvn@835 | 229 | if (reg == (uint)nk_reg) { |
kvn@835 | 230 | // Found ourselves so check if there is only one user of this |
kvn@835 | 231 | // copy and keep on searching for a better copy if so. |
kvn@835 | 232 | bool ignore_self = true; |
kvn@835 | 233 | x = n->in(k); |
kvn@835 | 234 | DUIterator_Fast imax, i = x->fast_outs(imax); |
kvn@835 | 235 | Node* first = x->fast_out(i); i++; |
kvn@835 | 236 | while (i < imax && ignore_self) { |
kvn@835 | 237 | Node* use = x->fast_out(i); i++; |
kvn@835 | 238 | if (use != first) ignore_self = false; |
kvn@835 | 239 | } |
kvn@835 | 240 | if (ignore_self) continue; |
kvn@835 | 241 | } |
kvn@835 | 242 | |
duke@435 | 243 | Node *vv = value[reg]; |
duke@435 | 244 | if( !single ) { // Doubles check for aligned-adjacent pair |
duke@435 | 245 | if( (reg&1)==0 ) continue; // Wrong half of a pair |
duke@435 | 246 | if( vv != value[reg-1] ) continue; // Not a complete pair |
duke@435 | 247 | } |
duke@435 | 248 | if( vv == val || // Got a direct hit? |
duke@435 | 249 | (t && vv && vv->bottom_type() == t && vv->is_Mach() && |
duke@435 | 250 | vv->as_Mach()->rule() == val->as_Mach()->rule()) ) { // Or same constant? |
duke@435 | 251 | assert( !n->is_Phi(), "cannot change registers at a Phi so easily" ); |
duke@435 | 252 | if( OptoReg::is_stack(nk_reg) || // CISC-loading from stack OR |
duke@435 | 253 | OptoReg::is_reg(reg) || // turning into a register use OR |
duke@435 | 254 | regnd[reg]->outcnt()==1 ) { // last use of a spill-load turns into a CISC use |
duke@435 | 255 | blk_adjust += use_prior_register(n,k,regnd[reg],current_block,value,regnd); |
duke@435 | 256 | if( n->in(k) == regnd[reg] ) // Success! Quit trying |
duke@435 | 257 | return blk_adjust; |
duke@435 | 258 | } // End of if not degrading to a stack |
duke@435 | 259 | } // End of if found value in another register |
duke@435 | 260 | } // End of scan all machine registers |
duke@435 | 261 | return blk_adjust; |
duke@435 | 262 | } |
duke@435 | 263 | |
duke@435 | 264 | |
duke@435 | 265 | // |
duke@435 | 266 | // Check if nreg already contains the constant value val. Normal copy |
duke@435 | 267 | // elimination doesn't doesn't work on constants because multiple |
duke@435 | 268 | // nodes can represent the same constant so the type and rule of the |
duke@435 | 269 | // MachNode must be checked to ensure equivalence. |
duke@435 | 270 | // |
never@505 | 271 | bool PhaseChaitin::eliminate_copy_of_constant(Node* val, Node* n, |
never@505 | 272 | Block *current_block, |
duke@435 | 273 | Node_List& value, Node_List& regnd, |
duke@435 | 274 | OptoReg::Name nreg, OptoReg::Name nreg2) { |
duke@435 | 275 | if (value[nreg] != val && val->is_Con() && |
duke@435 | 276 | value[nreg] != NULL && value[nreg]->is_Con() && |
duke@435 | 277 | (nreg2 == OptoReg::Bad || value[nreg] == value[nreg2]) && |
duke@435 | 278 | value[nreg]->bottom_type() == val->bottom_type() && |
duke@435 | 279 | value[nreg]->as_Mach()->rule() == val->as_Mach()->rule()) { |
duke@435 | 280 | // This code assumes that two MachNodes representing constants |
duke@435 | 281 | // which have the same rule and the same bottom type will produce |
duke@435 | 282 | // identical effects into a register. This seems like it must be |
duke@435 | 283 | // objectively true unless there are hidden inputs to the nodes |
duke@435 | 284 | // but if that were to change this code would need to updated. |
duke@435 | 285 | // Since they are equivalent the second one if redundant and can |
duke@435 | 286 | // be removed. |
duke@435 | 287 | // |
never@505 | 288 | // n will be replaced with the old value but n might have |
duke@435 | 289 | // kills projections associated with it so remove them now so that |
twisti@1040 | 290 | // yank_if_dead will be able to eliminate the copy once the uses |
duke@435 | 291 | // have been transferred to the old[value]. |
never@505 | 292 | for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { |
never@505 | 293 | Node* use = n->fast_out(i); |
duke@435 | 294 | if (use->is_Proj() && use->outcnt() == 0) { |
duke@435 | 295 | // Kill projections have no users and one input |
duke@435 | 296 | use->set_req(0, C->top()); |
duke@435 | 297 | yank_if_dead(use, current_block, &value, ®nd); |
duke@435 | 298 | --i; --imax; |
duke@435 | 299 | } |
duke@435 | 300 | } |
duke@435 | 301 | _post_alloc++; |
duke@435 | 302 | return true; |
duke@435 | 303 | } |
duke@435 | 304 | return false; |
duke@435 | 305 | } |
duke@435 | 306 | |
duke@435 | 307 | |
duke@435 | 308 | //------------------------------post_allocate_copy_removal--------------------- |
duke@435 | 309 | // Post-Allocation peephole copy removal. We do this in 1 pass over the |
duke@435 | 310 | // basic blocks. We maintain a mapping of registers to Nodes (an array of |
duke@435 | 311 | // Nodes indexed by machine register or stack slot number). NULL means that a |
duke@435 | 312 | // register is not mapped to any Node. We can (want to have!) have several |
duke@435 | 313 | // registers map to the same Node. We walk forward over the instructions |
duke@435 | 314 | // updating the mapping as we go. At merge points we force a NULL if we have |
duke@435 | 315 | // to merge 2 different Nodes into the same register. Phi functions will give |
duke@435 | 316 | // us a new Node if there is a proper value merging. Since the blocks are |
duke@435 | 317 | // arranged in some RPO, we will visit all parent blocks before visiting any |
duke@435 | 318 | // successor blocks (except at loops). |
duke@435 | 319 | // |
duke@435 | 320 | // If we find a Copy we look to see if the Copy's source register is a stack |
duke@435 | 321 | // slot and that value has already been loaded into some machine register; if |
duke@435 | 322 | // so we use machine register directly. This turns a Load into a reg-reg |
duke@435 | 323 | // Move. We also look for reloads of identical constants. |
duke@435 | 324 | // |
duke@435 | 325 | // When we see a use from a reg-reg Copy, we will attempt to use the copy's |
duke@435 | 326 | // source directly and make the copy go dead. |
duke@435 | 327 | void PhaseChaitin::post_allocate_copy_removal() { |
duke@435 | 328 | NOT_PRODUCT( Compile::TracePhase t3("postAllocCopyRemoval", &_t_postAllocCopyRemoval, TimeCompiler); ) |
duke@435 | 329 | ResourceMark rm; |
duke@435 | 330 | |
duke@435 | 331 | // Need a mapping from basic block Node_Lists. We need a Node_List to |
duke@435 | 332 | // map from register number to value-producing Node. |
duke@435 | 333 | Node_List **blk2value = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1); |
duke@435 | 334 | memset( blk2value, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) ); |
duke@435 | 335 | // Need a mapping from basic block Node_Lists. We need a Node_List to |
duke@435 | 336 | // map from register number to register-defining Node. |
duke@435 | 337 | Node_List **blk2regnd = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1); |
duke@435 | 338 | memset( blk2regnd, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) ); |
duke@435 | 339 | |
duke@435 | 340 | // We keep unused Node_Lists on a free_list to avoid wasting |
duke@435 | 341 | // memory. |
duke@435 | 342 | GrowableArray<Node_List*> free_list = GrowableArray<Node_List*>(16); |
duke@435 | 343 | |
duke@435 | 344 | // For all blocks |
duke@435 | 345 | for( uint i = 0; i < _cfg._num_blocks; i++ ) { |
duke@435 | 346 | uint j; |
duke@435 | 347 | Block *b = _cfg._blocks[i]; |
duke@435 | 348 | |
duke@435 | 349 | // Count of Phis in block |
duke@435 | 350 | uint phi_dex; |
duke@435 | 351 | for( phi_dex = 1; phi_dex < b->_nodes.size(); phi_dex++ ) { |
duke@435 | 352 | Node *phi = b->_nodes[phi_dex]; |
duke@435 | 353 | if( !phi->is_Phi() ) |
duke@435 | 354 | break; |
duke@435 | 355 | } |
duke@435 | 356 | |
duke@435 | 357 | // If any predecessor has not been visited, we do not know the state |
duke@435 | 358 | // of registers at the start. Check for this, while updating copies |
duke@435 | 359 | // along Phi input edges |
duke@435 | 360 | bool missing_some_inputs = false; |
duke@435 | 361 | Block *freed = NULL; |
duke@435 | 362 | for( j = 1; j < b->num_preds(); j++ ) { |
duke@435 | 363 | Block *pb = _cfg._bbs[b->pred(j)->_idx]; |
duke@435 | 364 | // Remove copies along phi edges |
duke@435 | 365 | for( uint k=1; k<phi_dex; k++ ) |
duke@435 | 366 | elide_copy( b->_nodes[k], j, b, *blk2value[pb->_pre_order], *blk2regnd[pb->_pre_order], false ); |
duke@435 | 367 | if( blk2value[pb->_pre_order] ) { // Have a mapping on this edge? |
duke@435 | 368 | // See if this predecessor's mappings have been used by everybody |
duke@435 | 369 | // who wants them. If so, free 'em. |
duke@435 | 370 | uint k; |
duke@435 | 371 | for( k=0; k<pb->_num_succs; k++ ) { |
duke@435 | 372 | Block *pbsucc = pb->_succs[k]; |
duke@435 | 373 | if( !blk2value[pbsucc->_pre_order] && pbsucc != b ) |
duke@435 | 374 | break; // Found a future user |
duke@435 | 375 | } |
duke@435 | 376 | if( k >= pb->_num_succs ) { // No more uses, free! |
duke@435 | 377 | freed = pb; // Record last block freed |
duke@435 | 378 | free_list.push(blk2value[pb->_pre_order]); |
duke@435 | 379 | free_list.push(blk2regnd[pb->_pre_order]); |
duke@435 | 380 | } |
duke@435 | 381 | } else { // This block has unvisited (loopback) inputs |
duke@435 | 382 | missing_some_inputs = true; |
duke@435 | 383 | } |
duke@435 | 384 | } |
duke@435 | 385 | |
duke@435 | 386 | |
duke@435 | 387 | // Extract Node_List mappings. If 'freed' is non-zero, we just popped |
duke@435 | 388 | // 'freed's blocks off the list |
duke@435 | 389 | Node_List ®nd = *(free_list.is_empty() ? new Node_List() : free_list.pop()); |
duke@435 | 390 | Node_List &value = *(free_list.is_empty() ? new Node_List() : free_list.pop()); |
duke@435 | 391 | assert( !freed || blk2value[freed->_pre_order] == &value, "" ); |
duke@435 | 392 | value.map(_max_reg,NULL); |
duke@435 | 393 | regnd.map(_max_reg,NULL); |
duke@435 | 394 | // Set mappings as OUR mappings |
duke@435 | 395 | blk2value[b->_pre_order] = &value; |
duke@435 | 396 | blk2regnd[b->_pre_order] = ®nd; |
duke@435 | 397 | |
duke@435 | 398 | // Initialize value & regnd for this block |
duke@435 | 399 | if( missing_some_inputs ) { |
duke@435 | 400 | // Some predecessor has not yet been visited; zap map to empty |
duke@435 | 401 | for( uint k = 0; k < (uint)_max_reg; k++ ) { |
duke@435 | 402 | value.map(k,NULL); |
duke@435 | 403 | regnd.map(k,NULL); |
duke@435 | 404 | } |
duke@435 | 405 | } else { |
duke@435 | 406 | if( !freed ) { // Didn't get a freebie prior block |
duke@435 | 407 | // Must clone some data |
duke@435 | 408 | freed = _cfg._bbs[b->pred(1)->_idx]; |
duke@435 | 409 | Node_List &f_value = *blk2value[freed->_pre_order]; |
duke@435 | 410 | Node_List &f_regnd = *blk2regnd[freed->_pre_order]; |
duke@435 | 411 | for( uint k = 0; k < (uint)_max_reg; k++ ) { |
duke@435 | 412 | value.map(k,f_value[k]); |
duke@435 | 413 | regnd.map(k,f_regnd[k]); |
duke@435 | 414 | } |
duke@435 | 415 | } |
duke@435 | 416 | // Merge all inputs together, setting to NULL any conflicts. |
duke@435 | 417 | for( j = 1; j < b->num_preds(); j++ ) { |
duke@435 | 418 | Block *pb = _cfg._bbs[b->pred(j)->_idx]; |
duke@435 | 419 | if( pb == freed ) continue; // Did self already via freelist |
duke@435 | 420 | Node_List &p_regnd = *blk2regnd[pb->_pre_order]; |
duke@435 | 421 | for( uint k = 0; k < (uint)_max_reg; k++ ) { |
duke@435 | 422 | if( regnd[k] != p_regnd[k] ) { // Conflict on reaching defs? |
duke@435 | 423 | value.map(k,NULL); // Then no value handy |
duke@435 | 424 | regnd.map(k,NULL); |
duke@435 | 425 | } |
duke@435 | 426 | } |
duke@435 | 427 | } |
duke@435 | 428 | } |
duke@435 | 429 | |
duke@435 | 430 | // For all Phi's |
duke@435 | 431 | for( j = 1; j < phi_dex; j++ ) { |
duke@435 | 432 | uint k; |
duke@435 | 433 | Node *phi = b->_nodes[j]; |
duke@435 | 434 | uint pidx = n2lidx(phi); |
duke@435 | 435 | OptoReg::Name preg = lrgs(n2lidx(phi)).reg(); |
duke@435 | 436 | |
duke@435 | 437 | // Remove copies remaining on edges. Check for junk phi. |
duke@435 | 438 | Node *u = NULL; |
duke@435 | 439 | for( k=1; k<phi->req(); k++ ) { |
duke@435 | 440 | Node *x = phi->in(k); |
duke@435 | 441 | if( phi != x && u != x ) // Found a different input |
duke@435 | 442 | u = u ? NodeSentinel : x; // Capture unique input, or NodeSentinel for 2nd input |
duke@435 | 443 | } |
duke@435 | 444 | if( u != NodeSentinel ) { // Junk Phi. Remove |
duke@435 | 445 | b->_nodes.remove(j--); phi_dex--; |
duke@435 | 446 | _cfg._bbs.map(phi->_idx,NULL); |
duke@435 | 447 | phi->replace_by(u); |
duke@435 | 448 | phi->disconnect_inputs(NULL); |
duke@435 | 449 | continue; |
duke@435 | 450 | } |
duke@435 | 451 | // Note that if value[pidx] exists, then we merged no new values here |
duke@435 | 452 | // and the phi is useless. This can happen even with the above phi |
duke@435 | 453 | // removal for complex flows. I cannot keep the better known value here |
duke@435 | 454 | // because locally the phi appears to define a new merged value. If I |
duke@435 | 455 | // keep the better value then a copy of the phi, being unable to use the |
duke@435 | 456 | // global flow analysis, can't "peek through" the phi to the original |
duke@435 | 457 | // reaching value and so will act like it's defining a new value. This |
duke@435 | 458 | // can lead to situations where some uses are from the old and some from |
duke@435 | 459 | // the new values. Not illegal by itself but throws the over-strong |
duke@435 | 460 | // assert in scheduling. |
duke@435 | 461 | if( pidx ) { |
duke@435 | 462 | value.map(preg,phi); |
duke@435 | 463 | regnd.map(preg,phi); |
duke@435 | 464 | OptoReg::Name preg_lo = OptoReg::add(preg,-1); |
duke@435 | 465 | if( !is_single_register(phi->ideal_reg()) ) { |
duke@435 | 466 | value.map(preg_lo,phi); |
duke@435 | 467 | regnd.map(preg_lo,phi); |
duke@435 | 468 | } |
duke@435 | 469 | } |
duke@435 | 470 | } |
duke@435 | 471 | |
duke@435 | 472 | // For all remaining instructions |
duke@435 | 473 | for( j = phi_dex; j < b->_nodes.size(); j++ ) { |
duke@435 | 474 | Node *n = b->_nodes[j]; |
duke@435 | 475 | |
duke@435 | 476 | if( n->outcnt() == 0 && // Dead? |
duke@435 | 477 | n != C->top() && // (ignore TOP, it has no du info) |
duke@435 | 478 | !n->is_Proj() ) { // fat-proj kills |
duke@435 | 479 | j -= yank_if_dead(n,b,&value,®nd); |
duke@435 | 480 | continue; |
duke@435 | 481 | } |
duke@435 | 482 | |
duke@435 | 483 | // Improve reaching-def info. Occasionally post-alloc's liveness gives |
duke@435 | 484 | // up (at loop backedges, because we aren't doing a full flow pass). |
duke@435 | 485 | // The presence of a live use essentially asserts that the use's def is |
duke@435 | 486 | // alive and well at the use (or else the allocator fubar'd). Take |
duke@435 | 487 | // advantage of this info to set a reaching def for the use-reg. |
duke@435 | 488 | uint k; |
duke@435 | 489 | for( k = 1; k < n->req(); k++ ) { |
duke@435 | 490 | Node *def = n->in(k); // n->in(k) is a USE; def is the DEF for this USE |
duke@435 | 491 | guarantee(def != NULL, "no disconnected nodes at this point"); |
duke@435 | 492 | uint useidx = n2lidx(def); // useidx is the live range index for this USE |
duke@435 | 493 | |
duke@435 | 494 | if( useidx ) { |
duke@435 | 495 | OptoReg::Name ureg = lrgs(useidx).reg(); |
duke@435 | 496 | if( !value[ureg] ) { |
duke@435 | 497 | int idx; // Skip occasional useless copy |
duke@435 | 498 | while( (idx=def->is_Copy()) != 0 && |
duke@435 | 499 | def->in(idx) != NULL && // NULL should not happen |
duke@435 | 500 | ureg == lrgs(n2lidx(def->in(idx))).reg() ) |
duke@435 | 501 | def = def->in(idx); |
duke@435 | 502 | Node *valdef = skip_copies(def); // tighten up val through non-useless copies |
duke@435 | 503 | value.map(ureg,valdef); // record improved reaching-def info |
duke@435 | 504 | regnd.map(ureg, def); |
duke@435 | 505 | // Record other half of doubles |
duke@435 | 506 | OptoReg::Name ureg_lo = OptoReg::add(ureg,-1); |
duke@435 | 507 | if( !is_single_register(def->ideal_reg()) && |
duke@435 | 508 | ( !RegMask::can_represent(ureg_lo) || |
duke@435 | 509 | lrgs(useidx).mask().Member(ureg_lo) ) && // Nearly always adjacent |
duke@435 | 510 | !value[ureg_lo] ) { |
duke@435 | 511 | value.map(ureg_lo,valdef); // record improved reaching-def info |
duke@435 | 512 | regnd.map(ureg_lo, def); |
duke@435 | 513 | } |
duke@435 | 514 | } |
duke@435 | 515 | } |
duke@435 | 516 | } |
duke@435 | 517 | |
duke@435 | 518 | const uint two_adr = n->is_Mach() ? n->as_Mach()->two_adr() : 0; |
duke@435 | 519 | |
duke@435 | 520 | // Remove copies along input edges |
duke@435 | 521 | for( k = 1; k < n->req(); k++ ) |
duke@435 | 522 | j -= elide_copy( n, k, b, value, regnd, two_adr!=k ); |
duke@435 | 523 | |
duke@435 | 524 | // Unallocated Nodes define no registers |
duke@435 | 525 | uint lidx = n2lidx(n); |
duke@435 | 526 | if( !lidx ) continue; |
duke@435 | 527 | |
duke@435 | 528 | // Update the register defined by this instruction |
duke@435 | 529 | OptoReg::Name nreg = lrgs(lidx).reg(); |
duke@435 | 530 | // Skip through all copies to the _value_ being defined. |
duke@435 | 531 | // Do not change from int to pointer |
duke@435 | 532 | Node *val = skip_copies(n); |
duke@435 | 533 | |
never@1358 | 534 | // Clear out a dead definition before starting so that the |
never@1358 | 535 | // elimination code doesn't have to guard against it. The |
never@1358 | 536 | // definition could in fact be a kill projection with a count of |
never@1358 | 537 | // 0 which is safe but since those are uninteresting for copy |
never@1358 | 538 | // elimination just delete them as well. |
never@1358 | 539 | if (regnd[nreg] != NULL && regnd[nreg]->outcnt() == 0) { |
never@1358 | 540 | regnd.map(nreg, NULL); |
never@1358 | 541 | value.map(nreg, NULL); |
never@1358 | 542 | } |
never@1358 | 543 | |
duke@435 | 544 | uint n_ideal_reg = n->ideal_reg(); |
duke@435 | 545 | if( is_single_register(n_ideal_reg) ) { |
duke@435 | 546 | // If Node 'n' does not change the value mapped by the register, |
duke@435 | 547 | // then 'n' is a useless copy. Do not update the register->node |
duke@435 | 548 | // mapping so 'n' will go dead. |
duke@435 | 549 | if( value[nreg] != val ) { |
never@505 | 550 | if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, OptoReg::Bad)) { |
never@1358 | 551 | j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); |
duke@435 | 552 | } else { |
duke@435 | 553 | // Update the mapping: record new Node defined by the register |
duke@435 | 554 | regnd.map(nreg,n); |
duke@435 | 555 | // Update mapping for defined *value*, which is the defined |
duke@435 | 556 | // Node after skipping all copies. |
duke@435 | 557 | value.map(nreg,val); |
duke@435 | 558 | } |
never@1358 | 559 | } else if( !may_be_copy_of_callee(n) ) { |
duke@435 | 560 | assert( n->is_Copy(), "" ); |
never@1358 | 561 | j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); |
duke@435 | 562 | } |
duke@435 | 563 | } else { |
duke@435 | 564 | // If the value occupies a register pair, record same info |
duke@435 | 565 | // in both registers. |
duke@435 | 566 | OptoReg::Name nreg_lo = OptoReg::add(nreg,-1); |
duke@435 | 567 | if( RegMask::can_represent(nreg_lo) && // Either a spill slot, or |
duke@435 | 568 | !lrgs(lidx).mask().Member(nreg_lo) ) { // Nearly always adjacent |
duke@435 | 569 | // Sparc occasionally has non-adjacent pairs. |
duke@435 | 570 | // Find the actual other value |
duke@435 | 571 | RegMask tmp = lrgs(lidx).mask(); |
duke@435 | 572 | tmp.Remove(nreg); |
duke@435 | 573 | nreg_lo = tmp.find_first_elem(); |
duke@435 | 574 | } |
duke@435 | 575 | if( value[nreg] != val || value[nreg_lo] != val ) { |
never@505 | 576 | if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, nreg_lo)) { |
never@1358 | 577 | j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); |
duke@435 | 578 | } else { |
duke@435 | 579 | regnd.map(nreg , n ); |
duke@435 | 580 | regnd.map(nreg_lo, n ); |
duke@435 | 581 | value.map(nreg ,val); |
duke@435 | 582 | value.map(nreg_lo,val); |
duke@435 | 583 | } |
never@1358 | 584 | } else if( !may_be_copy_of_callee(n) ) { |
duke@435 | 585 | assert( n->is_Copy(), "" ); |
never@1358 | 586 | j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); |
duke@435 | 587 | } |
duke@435 | 588 | } |
duke@435 | 589 | |
duke@435 | 590 | // Fat projections kill many registers |
duke@435 | 591 | if( n_ideal_reg == MachProjNode::fat_proj ) { |
duke@435 | 592 | RegMask rm = n->out_RegMask(); |
duke@435 | 593 | // wow, what an expensive iterator... |
duke@435 | 594 | nreg = rm.find_first_elem(); |
duke@435 | 595 | while( OptoReg::is_valid(nreg)) { |
duke@435 | 596 | rm.Remove(nreg); |
duke@435 | 597 | value.map(nreg,n); |
duke@435 | 598 | regnd.map(nreg,n); |
duke@435 | 599 | nreg = rm.find_first_elem(); |
duke@435 | 600 | } |
duke@435 | 601 | } |
duke@435 | 602 | |
duke@435 | 603 | } // End of for all instructions in the block |
duke@435 | 604 | |
duke@435 | 605 | } // End for all blocks |
duke@435 | 606 | } |