Mon, 24 Sep 2018 17:18:38 -0400
8131048: ppc implement CRC32 intrinsic
Reviewed-by: goetz
duke@435 | 1 | /* |
mikael@6198 | 2 | * Copyright (c) 1998, 2013, Oracle and/or its affiliates. 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 | * |
trims@1907 | 19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
trims@1907 | 20 | * or visit www.oracle.com if you need additional information or have any |
trims@1907 | 21 | * questions. |
duke@435 | 22 | * |
duke@435 | 23 | */ |
duke@435 | 24 | |
stefank@2314 | 25 | #include "precompiled.hpp" |
stefank@2314 | 26 | #include "memory/allocation.inline.hpp" |
stefank@2314 | 27 | #include "opto/chaitin.hpp" |
stefank@2314 | 28 | #include "opto/machnode.hpp" |
duke@435 | 29 | |
kvn@3882 | 30 | // See if this register (or pairs, or vector) already contains the value. |
kvn@3882 | 31 | static bool register_contains_value(Node* val, OptoReg::Name reg, int n_regs, |
kvn@3882 | 32 | Node_List& value) { |
kvn@3882 | 33 | for (int i = 0; i < n_regs; i++) { |
kvn@3882 | 34 | OptoReg::Name nreg = OptoReg::add(reg,-i); |
kvn@3882 | 35 | if (value[nreg] != val) |
kvn@3882 | 36 | return false; |
kvn@3882 | 37 | } |
kvn@3882 | 38 | return true; |
duke@435 | 39 | } |
duke@435 | 40 | |
kvn@835 | 41 | //---------------------------may_be_copy_of_callee----------------------------- |
duke@435 | 42 | // Check to see if we can possibly be a copy of a callee-save value. |
duke@435 | 43 | bool PhaseChaitin::may_be_copy_of_callee( Node *def ) const { |
duke@435 | 44 | // Short circuit if there are no callee save registers |
duke@435 | 45 | if (_matcher.number_of_saved_registers() == 0) return false; |
duke@435 | 46 | |
duke@435 | 47 | // Expect only a spill-down and reload on exit for callee-save spills. |
duke@435 | 48 | // Chains of copies cannot be deep. |
duke@435 | 49 | // 5008997 - This is wishful thinking. Register allocator seems to |
duke@435 | 50 | // be splitting live ranges for callee save registers to such |
duke@435 | 51 | // an extent that in large methods the chains can be very long |
duke@435 | 52 | // (50+). The conservative answer is to return true if we don't |
twisti@1040 | 53 | // know as this prevents optimizations from occurring. |
duke@435 | 54 | |
duke@435 | 55 | const int limit = 60; |
duke@435 | 56 | int i; |
duke@435 | 57 | for( i=0; i < limit; i++ ) { |
duke@435 | 58 | if( def->is_Proj() && def->in(0)->is_Start() && |
neliasso@4949 | 59 | _matcher.is_save_on_entry(lrgs(_lrg_map.live_range_id(def)).reg())) |
duke@435 | 60 | return true; // Direct use of callee-save proj |
duke@435 | 61 | if( def->is_Copy() ) // Copies carry value through |
duke@435 | 62 | def = def->in(def->is_Copy()); |
duke@435 | 63 | else if( def->is_Phi() ) // Phis can merge it from any direction |
duke@435 | 64 | def = def->in(1); |
duke@435 | 65 | else |
duke@435 | 66 | break; |
duke@435 | 67 | guarantee(def != NULL, "must not resurrect dead copy"); |
duke@435 | 68 | } |
duke@435 | 69 | // If we reached the end and didn't find a callee save proj |
duke@435 | 70 | // then this may be a callee save proj so we return true |
duke@435 | 71 | // as the conservative answer. If we didn't reach then end |
duke@435 | 72 | // we must have discovered that it was not a callee save |
duke@435 | 73 | // else we would have returned. |
duke@435 | 74 | return i == limit; |
duke@435 | 75 | } |
duke@435 | 76 | |
roland@3133 | 77 | //------------------------------yank----------------------------------- |
roland@3133 | 78 | // Helper function for yank_if_dead |
roland@3133 | 79 | int PhaseChaitin::yank( Node *old, Block *current_block, Node_List *value, Node_List *regnd ) { |
roland@3133 | 80 | int blk_adjust=0; |
adlertz@5509 | 81 | Block *oldb = _cfg.get_block_for_node(old); |
roland@3133 | 82 | oldb->find_remove(old); |
roland@3133 | 83 | // Count 1 if deleting an instruction from the current block |
adlertz@5509 | 84 | if (oldb == current_block) { |
adlertz@5509 | 85 | blk_adjust++; |
adlertz@5509 | 86 | } |
adlertz@5509 | 87 | _cfg.unmap_node_from_block(old); |
neliasso@4949 | 88 | OptoReg::Name old_reg = lrgs(_lrg_map.live_range_id(old)).reg(); |
roland@3133 | 89 | if( regnd && (*regnd)[old_reg]==old ) { // Instruction is currently available? |
roland@3133 | 90 | value->map(old_reg,NULL); // Yank from value/regnd maps |
roland@3133 | 91 | regnd->map(old_reg,NULL); // This register's value is now unknown |
roland@3133 | 92 | } |
roland@3133 | 93 | return blk_adjust; |
roland@3133 | 94 | } |
duke@435 | 95 | |
kvn@3405 | 96 | #ifdef ASSERT |
kvn@3405 | 97 | static bool expected_yanked_node(Node *old, Node *orig_old) { |
kvn@3405 | 98 | // This code is expected only next original nodes: |
kvn@3405 | 99 | // - load from constant table node which may have next data input nodes: |
adlertz@6041 | 100 | // MachConstantBase, MachTemp, MachSpillCopy |
adlertz@6041 | 101 | // - Phi nodes that are considered Junk |
kvn@3405 | 102 | // - load constant node which may have next data input nodes: |
kvn@3405 | 103 | // MachTemp, MachSpillCopy |
kvn@3405 | 104 | // - MachSpillCopy |
kvn@3405 | 105 | // - MachProj and Copy dead nodes |
kvn@3405 | 106 | if (old->is_MachSpillCopy()) { |
kvn@3405 | 107 | return true; |
kvn@3405 | 108 | } else if (old->is_Con()) { |
kvn@3405 | 109 | return true; |
kvn@3405 | 110 | } else if (old->is_MachProj()) { // Dead kills projection of Con node |
kvn@3405 | 111 | return (old == orig_old); |
kvn@3405 | 112 | } else if (old->is_Copy()) { // Dead copy of a callee-save value |
kvn@3405 | 113 | return (old == orig_old); |
kvn@3405 | 114 | } else if (old->is_MachTemp()) { |
kvn@3405 | 115 | return orig_old->is_Con(); |
adlertz@6041 | 116 | } else if (old->is_Phi()) { // Junk phi's |
adlertz@6041 | 117 | return true; |
adlertz@6041 | 118 | } else if (old->is_MachConstantBase()) { |
kvn@3405 | 119 | return (orig_old->is_Con() && orig_old->is_MachConstant()); |
kvn@3405 | 120 | } |
kvn@3405 | 121 | return false; |
kvn@3405 | 122 | } |
kvn@3405 | 123 | #endif |
kvn@3405 | 124 | |
duke@435 | 125 | //------------------------------yank_if_dead----------------------------------- |
kvn@3405 | 126 | // Removed edges from 'old'. Yank if dead. Return adjustment counts to |
duke@435 | 127 | // iterators in the current block. |
kvn@3405 | 128 | int PhaseChaitin::yank_if_dead_recurse(Node *old, Node *orig_old, Block *current_block, |
kvn@3405 | 129 | Node_List *value, Node_List *regnd) { |
duke@435 | 130 | int blk_adjust=0; |
kvn@3405 | 131 | if (old->outcnt() == 0 && old != C->top()) { |
kvn@3405 | 132 | #ifdef ASSERT |
kvn@3405 | 133 | if (!expected_yanked_node(old, orig_old)) { |
kvn@3405 | 134 | tty->print_cr("=============================================="); |
kvn@3405 | 135 | tty->print_cr("orig_old:"); |
kvn@3405 | 136 | orig_old->dump(); |
kvn@3405 | 137 | tty->print_cr("old:"); |
kvn@3405 | 138 | old->dump(); |
kvn@3405 | 139 | assert(false, "unexpected yanked node"); |
kvn@3405 | 140 | } |
kvn@3405 | 141 | if (old->is_Con()) |
kvn@3405 | 142 | orig_old = old; // Reset to satisfy expected nodes checks. |
kvn@3405 | 143 | #endif |
roland@3133 | 144 | blk_adjust += yank(old, current_block, value, regnd); |
roland@3133 | 145 | |
roland@3133 | 146 | for (uint i = 1; i < old->req(); i++) { |
kvn@3405 | 147 | Node* n = old->in(i); |
kvn@3405 | 148 | if (n != NULL) { |
kvn@3405 | 149 | old->set_req(i, NULL); |
kvn@3405 | 150 | blk_adjust += yank_if_dead_recurse(n, orig_old, current_block, value, regnd); |
roland@3133 | 151 | } |
duke@435 | 152 | } |
kvn@3405 | 153 | // Disconnect control and remove precedence edges if any exist |
bharadwaj@4315 | 154 | old->disconnect_inputs(NULL, C); |
duke@435 | 155 | } |
duke@435 | 156 | return blk_adjust; |
duke@435 | 157 | } |
duke@435 | 158 | |
duke@435 | 159 | //------------------------------use_prior_register----------------------------- |
duke@435 | 160 | // Use the prior value instead of the current value, in an effort to make |
duke@435 | 161 | // the current value go dead. Return block iterator adjustment, in case |
duke@435 | 162 | // we yank some instructions from this block. |
duke@435 | 163 | int PhaseChaitin::use_prior_register( Node *n, uint idx, Node *def, Block *current_block, Node_List &value, Node_List ®nd ) { |
duke@435 | 164 | // No effect? |
duke@435 | 165 | if( def == n->in(idx) ) return 0; |
duke@435 | 166 | // Def is currently dead and can be removed? Do not resurrect |
duke@435 | 167 | if( def->outcnt() == 0 ) return 0; |
duke@435 | 168 | |
duke@435 | 169 | // Not every pair of physical registers are assignment compatible, |
duke@435 | 170 | // e.g. on sparc floating point registers are not assignable to integer |
duke@435 | 171 | // registers. |
neliasso@4949 | 172 | const LRG &def_lrg = lrgs(_lrg_map.live_range_id(def)); |
duke@435 | 173 | OptoReg::Name def_reg = def_lrg.reg(); |
duke@435 | 174 | const RegMask &use_mask = n->in_RegMask(idx); |
duke@435 | 175 | bool can_use = ( RegMask::can_represent(def_reg) ? (use_mask.Member(def_reg) != 0) |
duke@435 | 176 | : (use_mask.is_AllStack() != 0)); |
kvn@3882 | 177 | if (!RegMask::is_vector(def->ideal_reg())) { |
kvn@3882 | 178 | // Check for a copy to or from a misaligned pair. |
kvn@3882 | 179 | // It is workaround for a sparc with misaligned pairs. |
kvn@3882 | 180 | can_use = can_use && !use_mask.is_misaligned_pair() && !def_lrg.mask().is_misaligned_pair(); |
kvn@3882 | 181 | } |
duke@435 | 182 | if (!can_use) |
duke@435 | 183 | return 0; |
duke@435 | 184 | |
duke@435 | 185 | // Capture the old def in case it goes dead... |
duke@435 | 186 | Node *old = n->in(idx); |
duke@435 | 187 | |
duke@435 | 188 | // Save-on-call copies can only be elided if the entire copy chain can go |
duke@435 | 189 | // away, lest we get the same callee-save value alive in 2 locations at |
duke@435 | 190 | // once. We check for the obvious trivial case here. Although it can |
duke@435 | 191 | // sometimes be elided with cooperation outside our scope, here we will just |
duke@435 | 192 | // miss the opportunity. :-( |
duke@435 | 193 | if( may_be_copy_of_callee(def) ) { |
duke@435 | 194 | if( old->outcnt() > 1 ) return 0; // We're the not last user |
duke@435 | 195 | int idx = old->is_Copy(); |
duke@435 | 196 | assert( idx, "chain of copies being removed" ); |
duke@435 | 197 | Node *old2 = old->in(idx); // Chain of copies |
duke@435 | 198 | if( old2->outcnt() > 1 ) return 0; // old is not the last user |
duke@435 | 199 | int idx2 = old2->is_Copy(); |
duke@435 | 200 | if( !idx2 ) return 0; // Not a chain of 2 copies |
duke@435 | 201 | if( def != old2->in(idx2) ) return 0; // Chain of exactly 2 copies |
duke@435 | 202 | } |
duke@435 | 203 | |
duke@435 | 204 | // Use the new def |
duke@435 | 205 | n->set_req(idx,def); |
duke@435 | 206 | _post_alloc++; |
duke@435 | 207 | |
duke@435 | 208 | // Is old def now dead? We successfully yanked a copy? |
duke@435 | 209 | return yank_if_dead(old,current_block,&value,®nd); |
duke@435 | 210 | } |
duke@435 | 211 | |
duke@435 | 212 | |
duke@435 | 213 | //------------------------------skip_copies------------------------------------ |
duke@435 | 214 | // Skip through any number of copies (that don't mod oop-i-ness) |
duke@435 | 215 | Node *PhaseChaitin::skip_copies( Node *c ) { |
duke@435 | 216 | int idx = c->is_Copy(); |
neliasso@4949 | 217 | uint is_oop = lrgs(_lrg_map.live_range_id(c))._is_oop; |
duke@435 | 218 | while (idx != 0) { |
duke@435 | 219 | guarantee(c->in(idx) != NULL, "must not resurrect dead copy"); |
neliasso@4949 | 220 | if (lrgs(_lrg_map.live_range_id(c->in(idx)))._is_oop != is_oop) { |
duke@435 | 221 | break; // casting copy, not the same value |
neliasso@4949 | 222 | } |
duke@435 | 223 | c = c->in(idx); |
duke@435 | 224 | idx = c->is_Copy(); |
duke@435 | 225 | } |
duke@435 | 226 | return c; |
duke@435 | 227 | } |
duke@435 | 228 | |
duke@435 | 229 | //------------------------------elide_copy------------------------------------- |
duke@435 | 230 | // Remove (bypass) copies along Node n, edge k. |
duke@435 | 231 | int PhaseChaitin::elide_copy( Node *n, int k, Block *current_block, Node_List &value, Node_List ®nd, bool can_change_regs ) { |
duke@435 | 232 | int blk_adjust = 0; |
duke@435 | 233 | |
neliasso@4949 | 234 | uint nk_idx = _lrg_map.live_range_id(n->in(k)); |
neliasso@4949 | 235 | OptoReg::Name nk_reg = lrgs(nk_idx).reg(); |
duke@435 | 236 | |
duke@435 | 237 | // Remove obvious same-register copies |
duke@435 | 238 | Node *x = n->in(k); |
duke@435 | 239 | int idx; |
duke@435 | 240 | while( (idx=x->is_Copy()) != 0 ) { |
duke@435 | 241 | Node *copy = x->in(idx); |
duke@435 | 242 | guarantee(copy != NULL, "must not resurrect dead copy"); |
neliasso@4949 | 243 | if(lrgs(_lrg_map.live_range_id(copy)).reg() != nk_reg) { |
neliasso@4949 | 244 | break; |
neliasso@4949 | 245 | } |
duke@435 | 246 | blk_adjust += use_prior_register(n,k,copy,current_block,value,regnd); |
neliasso@4949 | 247 | if (n->in(k) != copy) { |
neliasso@4949 | 248 | break; // Failed for some cutout? |
neliasso@4949 | 249 | } |
duke@435 | 250 | x = copy; // Progress, try again |
duke@435 | 251 | } |
duke@435 | 252 | |
duke@435 | 253 | // Phis and 2-address instructions cannot change registers so easily - their |
duke@435 | 254 | // outputs must match their input. |
duke@435 | 255 | if( !can_change_regs ) |
duke@435 | 256 | return blk_adjust; // Only check stupid copies! |
duke@435 | 257 | |
duke@435 | 258 | // Loop backedges won't have a value-mapping yet |
duke@435 | 259 | if( &value == NULL ) return blk_adjust; |
duke@435 | 260 | |
duke@435 | 261 | // Skip through all copies to the _value_ being used. Do not change from |
duke@435 | 262 | // int to pointer. This attempts to jump through a chain of copies, where |
duke@435 | 263 | // intermediate copies might be illegal, i.e., value is stored down to stack |
duke@435 | 264 | // then reloaded BUT survives in a register the whole way. |
duke@435 | 265 | Node *val = skip_copies(n->in(k)); |
kvn@3882 | 266 | if (val == x) return blk_adjust; // No progress? |
duke@435 | 267 | |
kvn@3882 | 268 | int n_regs = RegMask::num_registers(val->ideal_reg()); |
neliasso@4949 | 269 | uint val_idx = _lrg_map.live_range_id(val); |
duke@435 | 270 | OptoReg::Name val_reg = lrgs(val_idx).reg(); |
duke@435 | 271 | |
duke@435 | 272 | // See if it happens to already be in the correct register! |
duke@435 | 273 | // (either Phi's direct register, or the common case of the name |
duke@435 | 274 | // never-clobbered original-def register) |
kvn@3882 | 275 | if (register_contains_value(val, val_reg, n_regs, value)) { |
duke@435 | 276 | blk_adjust += use_prior_register(n,k,regnd[val_reg],current_block,value,regnd); |
duke@435 | 277 | if( n->in(k) == regnd[val_reg] ) // Success! Quit trying |
duke@435 | 278 | return blk_adjust; |
duke@435 | 279 | } |
duke@435 | 280 | |
duke@435 | 281 | // See if we can skip the copy by changing registers. Don't change from |
duke@435 | 282 | // using a register to using the stack unless we know we can remove a |
duke@435 | 283 | // copy-load. Otherwise we might end up making a pile of Intel cisc-spill |
duke@435 | 284 | // ops reading from memory instead of just loading once and using the |
duke@435 | 285 | // register. |
duke@435 | 286 | |
duke@435 | 287 | // Also handle duplicate copies here. |
duke@435 | 288 | const Type *t = val->is_Con() ? val->bottom_type() : NULL; |
duke@435 | 289 | |
duke@435 | 290 | // Scan all registers to see if this value is around already |
duke@435 | 291 | for( uint reg = 0; reg < (uint)_max_reg; reg++ ) { |
kvn@835 | 292 | if (reg == (uint)nk_reg) { |
kvn@835 | 293 | // Found ourselves so check if there is only one user of this |
kvn@835 | 294 | // copy and keep on searching for a better copy if so. |
kvn@835 | 295 | bool ignore_self = true; |
kvn@835 | 296 | x = n->in(k); |
kvn@835 | 297 | DUIterator_Fast imax, i = x->fast_outs(imax); |
kvn@835 | 298 | Node* first = x->fast_out(i); i++; |
kvn@835 | 299 | while (i < imax && ignore_self) { |
kvn@835 | 300 | Node* use = x->fast_out(i); i++; |
kvn@835 | 301 | if (use != first) ignore_self = false; |
kvn@835 | 302 | } |
kvn@835 | 303 | if (ignore_self) continue; |
kvn@835 | 304 | } |
kvn@835 | 305 | |
duke@435 | 306 | Node *vv = value[reg]; |
kvn@3888 | 307 | if (n_regs > 1) { // Doubles and vectors check for aligned-adjacent set |
kvn@3888 | 308 | uint last = (n_regs-1); // Looking for the last part of a set |
kvn@3888 | 309 | if ((reg&last) != last) continue; // Wrong part of a set |
kvn@3888 | 310 | if (!register_contains_value(vv, reg, n_regs, value)) continue; // Different value |
duke@435 | 311 | } |
duke@435 | 312 | if( vv == val || // Got a direct hit? |
duke@435 | 313 | (t && vv && vv->bottom_type() == t && vv->is_Mach() && |
duke@435 | 314 | vv->as_Mach()->rule() == val->as_Mach()->rule()) ) { // Or same constant? |
duke@435 | 315 | assert( !n->is_Phi(), "cannot change registers at a Phi so easily" ); |
duke@435 | 316 | if( OptoReg::is_stack(nk_reg) || // CISC-loading from stack OR |
duke@435 | 317 | OptoReg::is_reg(reg) || // turning into a register use OR |
duke@435 | 318 | regnd[reg]->outcnt()==1 ) { // last use of a spill-load turns into a CISC use |
duke@435 | 319 | blk_adjust += use_prior_register(n,k,regnd[reg],current_block,value,regnd); |
duke@435 | 320 | if( n->in(k) == regnd[reg] ) // Success! Quit trying |
duke@435 | 321 | return blk_adjust; |
duke@435 | 322 | } // End of if not degrading to a stack |
duke@435 | 323 | } // End of if found value in another register |
duke@435 | 324 | } // End of scan all machine registers |
duke@435 | 325 | return blk_adjust; |
duke@435 | 326 | } |
duke@435 | 327 | |
duke@435 | 328 | |
duke@435 | 329 | // |
duke@435 | 330 | // Check if nreg already contains the constant value val. Normal copy |
duke@435 | 331 | // elimination doesn't doesn't work on constants because multiple |
duke@435 | 332 | // nodes can represent the same constant so the type and rule of the |
duke@435 | 333 | // MachNode must be checked to ensure equivalence. |
duke@435 | 334 | // |
never@505 | 335 | bool PhaseChaitin::eliminate_copy_of_constant(Node* val, Node* n, |
never@505 | 336 | Block *current_block, |
duke@435 | 337 | Node_List& value, Node_List& regnd, |
duke@435 | 338 | OptoReg::Name nreg, OptoReg::Name nreg2) { |
duke@435 | 339 | if (value[nreg] != val && val->is_Con() && |
duke@435 | 340 | value[nreg] != NULL && value[nreg]->is_Con() && |
duke@435 | 341 | (nreg2 == OptoReg::Bad || value[nreg] == value[nreg2]) && |
duke@435 | 342 | value[nreg]->bottom_type() == val->bottom_type() && |
duke@435 | 343 | value[nreg]->as_Mach()->rule() == val->as_Mach()->rule()) { |
duke@435 | 344 | // This code assumes that two MachNodes representing constants |
duke@435 | 345 | // which have the same rule and the same bottom type will produce |
duke@435 | 346 | // identical effects into a register. This seems like it must be |
duke@435 | 347 | // objectively true unless there are hidden inputs to the nodes |
duke@435 | 348 | // but if that were to change this code would need to updated. |
duke@435 | 349 | // Since they are equivalent the second one if redundant and can |
duke@435 | 350 | // be removed. |
duke@435 | 351 | // |
never@505 | 352 | // n will be replaced with the old value but n might have |
duke@435 | 353 | // kills projections associated with it so remove them now so that |
twisti@1040 | 354 | // yank_if_dead will be able to eliminate the copy once the uses |
duke@435 | 355 | // have been transferred to the old[value]. |
never@505 | 356 | for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { |
never@505 | 357 | Node* use = n->fast_out(i); |
duke@435 | 358 | if (use->is_Proj() && use->outcnt() == 0) { |
duke@435 | 359 | // Kill projections have no users and one input |
duke@435 | 360 | use->set_req(0, C->top()); |
duke@435 | 361 | yank_if_dead(use, current_block, &value, ®nd); |
duke@435 | 362 | --i; --imax; |
duke@435 | 363 | } |
duke@435 | 364 | } |
duke@435 | 365 | _post_alloc++; |
duke@435 | 366 | return true; |
duke@435 | 367 | } |
duke@435 | 368 | return false; |
duke@435 | 369 | } |
duke@435 | 370 | |
iveresov@7564 | 371 | // The algorithms works as follows: |
iveresov@7564 | 372 | // We traverse the block top to bottom. possibly_merge_multidef() is invoked for every input edge k |
iveresov@7564 | 373 | // of the instruction n. We check to see if the input is a multidef lrg. If it is, we record the fact that we've |
iveresov@7564 | 374 | // seen a definition (coming as an input) and add that fact to the reg2defuse array. The array maps registers to their |
iveresov@7564 | 375 | // current reaching definitions (we track only multidefs though). With each definition we also associate the first |
iveresov@7564 | 376 | // instruction we saw use it. If we encounter the situation when we observe an def (an input) that is a part of the |
iveresov@7564 | 377 | // same lrg but is different from the previous seen def we merge the two with a MachMerge node and substitute |
iveresov@7564 | 378 | // all the uses that we've seen so far to use the merge. After that we keep replacing the new defs in the same lrg |
iveresov@7564 | 379 | // as they get encountered with the merge node and keep adding these defs to the merge inputs. |
iveresov@7564 | 380 | void PhaseChaitin::merge_multidefs() { |
iveresov@7564 | 381 | NOT_PRODUCT( Compile::TracePhase t3("mergeMultidefs", &_t_mergeMultidefs, TimeCompiler); ) |
iveresov@7564 | 382 | ResourceMark rm; |
iveresov@7564 | 383 | // Keep track of the defs seen in registers and collect their uses in the block. |
iveresov@7564 | 384 | RegToDefUseMap reg2defuse(_max_reg, _max_reg, RegDefUse()); |
iveresov@7564 | 385 | for (uint i = 0; i < _cfg.number_of_blocks(); i++) { |
iveresov@7564 | 386 | Block* block = _cfg.get_block(i); |
iveresov@7564 | 387 | for (uint j = 1; j < block->number_of_nodes(); j++) { |
iveresov@7564 | 388 | Node* n = block->get_node(j); |
iveresov@7564 | 389 | if (n->is_Phi()) continue; |
iveresov@7564 | 390 | for (uint k = 1; k < n->req(); k++) { |
iveresov@7564 | 391 | j += possibly_merge_multidef(n, k, block, reg2defuse); |
iveresov@7564 | 392 | } |
iveresov@7564 | 393 | // Null out the value produced by the instruction itself, since we're only interested in defs |
iveresov@7564 | 394 | // implicitly defined by the uses. We are actually interested in tracking only redefinitions |
iveresov@7564 | 395 | // of the multidef lrgs in the same register. For that matter it's enough to track changes in |
iveresov@7564 | 396 | // the base register only and ignore other effects of multi-register lrgs and fat projections. |
iveresov@7564 | 397 | // It is also ok to ignore defs coming from singledefs. After an implicit overwrite by one of |
iveresov@7564 | 398 | // those our register is guaranteed to be used by another lrg and we won't attempt to merge it. |
iveresov@7564 | 399 | uint lrg = _lrg_map.live_range_id(n); |
iveresov@7564 | 400 | if (lrg > 0 && lrgs(lrg).is_multidef()) { |
iveresov@7564 | 401 | OptoReg::Name reg = lrgs(lrg).reg(); |
iveresov@7564 | 402 | reg2defuse.at(reg).clear(); |
iveresov@7564 | 403 | } |
iveresov@7564 | 404 | } |
iveresov@7564 | 405 | // Clear reg->def->use tracking for the next block |
iveresov@7564 | 406 | for (int j = 0; j < reg2defuse.length(); j++) { |
iveresov@7564 | 407 | reg2defuse.at(j).clear(); |
iveresov@7564 | 408 | } |
iveresov@7564 | 409 | } |
iveresov@7564 | 410 | } |
iveresov@7564 | 411 | |
iveresov@7564 | 412 | int PhaseChaitin::possibly_merge_multidef(Node *n, uint k, Block *block, RegToDefUseMap& reg2defuse) { |
iveresov@7564 | 413 | int blk_adjust = 0; |
iveresov@7564 | 414 | |
iveresov@7564 | 415 | uint lrg = _lrg_map.live_range_id(n->in(k)); |
iveresov@7564 | 416 | if (lrg > 0 && lrgs(lrg).is_multidef()) { |
iveresov@7564 | 417 | OptoReg::Name reg = lrgs(lrg).reg(); |
iveresov@7564 | 418 | |
iveresov@7564 | 419 | Node* def = reg2defuse.at(reg).def(); |
iveresov@7564 | 420 | if (def != NULL && lrg == _lrg_map.live_range_id(def) && def != n->in(k)) { |
iveresov@7564 | 421 | // Same lrg but different node, we have to merge. |
iveresov@7564 | 422 | MachMergeNode* merge; |
iveresov@7564 | 423 | if (def->is_MachMerge()) { // is it already a merge? |
iveresov@7564 | 424 | merge = def->as_MachMerge(); |
iveresov@7564 | 425 | } else { |
iveresov@7564 | 426 | merge = new (C) MachMergeNode(def); |
iveresov@7564 | 427 | |
iveresov@7564 | 428 | // Insert the merge node into the block before the first use. |
iveresov@7564 | 429 | uint use_index = block->find_node(reg2defuse.at(reg).first_use()); |
iveresov@7564 | 430 | block->insert_node(merge, use_index++); |
iveresov@7570 | 431 | _cfg.map_node_to_block(merge, block); |
iveresov@7564 | 432 | |
iveresov@7564 | 433 | // Let the allocator know about the new node, use the same lrg |
iveresov@7564 | 434 | _lrg_map.extend(merge->_idx, lrg); |
iveresov@7564 | 435 | blk_adjust++; |
iveresov@7564 | 436 | |
iveresov@7564 | 437 | // Fixup all the uses (there is at least one) that happened between the first |
iveresov@7564 | 438 | // use and before the current one. |
iveresov@7564 | 439 | for (; use_index < block->number_of_nodes(); use_index++) { |
iveresov@7564 | 440 | Node* use = block->get_node(use_index); |
iveresov@7564 | 441 | if (use == n) { |
iveresov@7564 | 442 | break; |
iveresov@7564 | 443 | } |
iveresov@7564 | 444 | use->replace_edge(def, merge); |
iveresov@7564 | 445 | } |
iveresov@7564 | 446 | } |
iveresov@7564 | 447 | if (merge->find_edge(n->in(k)) == -1) { |
iveresov@7564 | 448 | merge->add_req(n->in(k)); |
iveresov@7564 | 449 | } |
iveresov@7564 | 450 | n->set_req(k, merge); |
iveresov@7564 | 451 | } |
iveresov@7564 | 452 | |
iveresov@7564 | 453 | // update the uses |
iveresov@7564 | 454 | reg2defuse.at(reg).update(n->in(k), n); |
iveresov@7564 | 455 | } |
iveresov@7564 | 456 | |
iveresov@7564 | 457 | return blk_adjust; |
iveresov@7564 | 458 | } |
iveresov@7564 | 459 | |
duke@435 | 460 | |
duke@435 | 461 | //------------------------------post_allocate_copy_removal--------------------- |
duke@435 | 462 | // Post-Allocation peephole copy removal. We do this in 1 pass over the |
duke@435 | 463 | // basic blocks. We maintain a mapping of registers to Nodes (an array of |
duke@435 | 464 | // Nodes indexed by machine register or stack slot number). NULL means that a |
duke@435 | 465 | // register is not mapped to any Node. We can (want to have!) have several |
duke@435 | 466 | // registers map to the same Node. We walk forward over the instructions |
duke@435 | 467 | // updating the mapping as we go. At merge points we force a NULL if we have |
duke@435 | 468 | // to merge 2 different Nodes into the same register. Phi functions will give |
duke@435 | 469 | // us a new Node if there is a proper value merging. Since the blocks are |
duke@435 | 470 | // arranged in some RPO, we will visit all parent blocks before visiting any |
duke@435 | 471 | // successor blocks (except at loops). |
duke@435 | 472 | // |
duke@435 | 473 | // If we find a Copy we look to see if the Copy's source register is a stack |
duke@435 | 474 | // slot and that value has already been loaded into some machine register; if |
duke@435 | 475 | // so we use machine register directly. This turns a Load into a reg-reg |
duke@435 | 476 | // Move. We also look for reloads of identical constants. |
duke@435 | 477 | // |
duke@435 | 478 | // When we see a use from a reg-reg Copy, we will attempt to use the copy's |
duke@435 | 479 | // source directly and make the copy go dead. |
duke@435 | 480 | void PhaseChaitin::post_allocate_copy_removal() { |
duke@435 | 481 | NOT_PRODUCT( Compile::TracePhase t3("postAllocCopyRemoval", &_t_postAllocCopyRemoval, TimeCompiler); ) |
duke@435 | 482 | ResourceMark rm; |
duke@435 | 483 | |
duke@435 | 484 | // Need a mapping from basic block Node_Lists. We need a Node_List to |
duke@435 | 485 | // map from register number to value-producing Node. |
adlertz@5539 | 486 | Node_List **blk2value = NEW_RESOURCE_ARRAY( Node_List *, _cfg.number_of_blocks() + 1); |
adlertz@5539 | 487 | memset(blk2value, 0, sizeof(Node_List*) * (_cfg.number_of_blocks() + 1)); |
duke@435 | 488 | // Need a mapping from basic block Node_Lists. We need a Node_List to |
duke@435 | 489 | // map from register number to register-defining Node. |
adlertz@5539 | 490 | Node_List **blk2regnd = NEW_RESOURCE_ARRAY( Node_List *, _cfg.number_of_blocks() + 1); |
adlertz@5539 | 491 | memset(blk2regnd, 0, sizeof(Node_List*) * (_cfg.number_of_blocks() + 1)); |
duke@435 | 492 | |
duke@435 | 493 | // We keep unused Node_Lists on a free_list to avoid wasting |
duke@435 | 494 | // memory. |
duke@435 | 495 | GrowableArray<Node_List*> free_list = GrowableArray<Node_List*>(16); |
duke@435 | 496 | |
duke@435 | 497 | // For all blocks |
adlertz@5539 | 498 | for (uint i = 0; i < _cfg.number_of_blocks(); i++) { |
duke@435 | 499 | uint j; |
adlertz@5539 | 500 | Block* block = _cfg.get_block(i); |
duke@435 | 501 | |
duke@435 | 502 | // Count of Phis in block |
duke@435 | 503 | uint phi_dex; |
adlertz@5635 | 504 | for (phi_dex = 1; phi_dex < block->number_of_nodes(); phi_dex++) { |
adlertz@5635 | 505 | Node* phi = block->get_node(phi_dex); |
adlertz@5539 | 506 | if (!phi->is_Phi()) { |
duke@435 | 507 | break; |
adlertz@5539 | 508 | } |
duke@435 | 509 | } |
duke@435 | 510 | |
duke@435 | 511 | // If any predecessor has not been visited, we do not know the state |
duke@435 | 512 | // of registers at the start. Check for this, while updating copies |
duke@435 | 513 | // along Phi input edges |
duke@435 | 514 | bool missing_some_inputs = false; |
duke@435 | 515 | Block *freed = NULL; |
adlertz@5539 | 516 | for (j = 1; j < block->num_preds(); j++) { |
adlertz@5539 | 517 | Block* pb = _cfg.get_block_for_node(block->pred(j)); |
duke@435 | 518 | // Remove copies along phi edges |
adlertz@5539 | 519 | for (uint k = 1; k < phi_dex; k++) { |
adlertz@5635 | 520 | elide_copy(block->get_node(k), j, block, *blk2value[pb->_pre_order], *blk2regnd[pb->_pre_order], false); |
adlertz@5539 | 521 | } |
adlertz@5539 | 522 | if (blk2value[pb->_pre_order]) { // Have a mapping on this edge? |
duke@435 | 523 | // See if this predecessor's mappings have been used by everybody |
duke@435 | 524 | // who wants them. If so, free 'em. |
duke@435 | 525 | uint k; |
adlertz@5539 | 526 | for (k = 0; k < pb->_num_succs; k++) { |
adlertz@5539 | 527 | Block* pbsucc = pb->_succs[k]; |
adlertz@5539 | 528 | if (!blk2value[pbsucc->_pre_order] && pbsucc != block) { |
duke@435 | 529 | break; // Found a future user |
adlertz@5539 | 530 | } |
duke@435 | 531 | } |
adlertz@5539 | 532 | if (k >= pb->_num_succs) { // No more uses, free! |
duke@435 | 533 | freed = pb; // Record last block freed |
duke@435 | 534 | free_list.push(blk2value[pb->_pre_order]); |
duke@435 | 535 | free_list.push(blk2regnd[pb->_pre_order]); |
duke@435 | 536 | } |
duke@435 | 537 | } else { // This block has unvisited (loopback) inputs |
duke@435 | 538 | missing_some_inputs = true; |
duke@435 | 539 | } |
duke@435 | 540 | } |
duke@435 | 541 | |
duke@435 | 542 | |
duke@435 | 543 | // Extract Node_List mappings. If 'freed' is non-zero, we just popped |
duke@435 | 544 | // 'freed's blocks off the list |
duke@435 | 545 | Node_List ®nd = *(free_list.is_empty() ? new Node_List() : free_list.pop()); |
duke@435 | 546 | Node_List &value = *(free_list.is_empty() ? new Node_List() : free_list.pop()); |
duke@435 | 547 | assert( !freed || blk2value[freed->_pre_order] == &value, "" ); |
duke@435 | 548 | value.map(_max_reg,NULL); |
duke@435 | 549 | regnd.map(_max_reg,NULL); |
duke@435 | 550 | // Set mappings as OUR mappings |
adlertz@5539 | 551 | blk2value[block->_pre_order] = &value; |
adlertz@5539 | 552 | blk2regnd[block->_pre_order] = ®nd; |
duke@435 | 553 | |
duke@435 | 554 | // Initialize value & regnd for this block |
adlertz@5539 | 555 | if (missing_some_inputs) { |
duke@435 | 556 | // Some predecessor has not yet been visited; zap map to empty |
adlertz@5539 | 557 | for (uint k = 0; k < (uint)_max_reg; k++) { |
duke@435 | 558 | value.map(k,NULL); |
duke@435 | 559 | regnd.map(k,NULL); |
duke@435 | 560 | } |
duke@435 | 561 | } else { |
duke@435 | 562 | if( !freed ) { // Didn't get a freebie prior block |
duke@435 | 563 | // Must clone some data |
adlertz@5539 | 564 | freed = _cfg.get_block_for_node(block->pred(1)); |
duke@435 | 565 | Node_List &f_value = *blk2value[freed->_pre_order]; |
duke@435 | 566 | Node_List &f_regnd = *blk2regnd[freed->_pre_order]; |
duke@435 | 567 | for( uint k = 0; k < (uint)_max_reg; k++ ) { |
duke@435 | 568 | value.map(k,f_value[k]); |
duke@435 | 569 | regnd.map(k,f_regnd[k]); |
duke@435 | 570 | } |
duke@435 | 571 | } |
duke@435 | 572 | // Merge all inputs together, setting to NULL any conflicts. |
adlertz@5539 | 573 | for (j = 1; j < block->num_preds(); j++) { |
adlertz@5539 | 574 | Block* pb = _cfg.get_block_for_node(block->pred(j)); |
adlertz@5539 | 575 | if (pb == freed) { |
adlertz@5539 | 576 | continue; // Did self already via freelist |
adlertz@5539 | 577 | } |
duke@435 | 578 | Node_List &p_regnd = *blk2regnd[pb->_pre_order]; |
duke@435 | 579 | for( uint k = 0; k < (uint)_max_reg; k++ ) { |
duke@435 | 580 | if( regnd[k] != p_regnd[k] ) { // Conflict on reaching defs? |
duke@435 | 581 | value.map(k,NULL); // Then no value handy |
duke@435 | 582 | regnd.map(k,NULL); |
duke@435 | 583 | } |
duke@435 | 584 | } |
duke@435 | 585 | } |
duke@435 | 586 | } |
duke@435 | 587 | |
duke@435 | 588 | // For all Phi's |
adlertz@5539 | 589 | for (j = 1; j < phi_dex; j++) { |
duke@435 | 590 | uint k; |
adlertz@5635 | 591 | Node *phi = block->get_node(j); |
neliasso@4949 | 592 | uint pidx = _lrg_map.live_range_id(phi); |
neliasso@4949 | 593 | OptoReg::Name preg = lrgs(_lrg_map.live_range_id(phi)).reg(); |
duke@435 | 594 | |
duke@435 | 595 | // Remove copies remaining on edges. Check for junk phi. |
duke@435 | 596 | Node *u = NULL; |
neliasso@4949 | 597 | for (k = 1; k < phi->req(); k++) { |
duke@435 | 598 | Node *x = phi->in(k); |
duke@435 | 599 | if( phi != x && u != x ) // Found a different input |
duke@435 | 600 | u = u ? NodeSentinel : x; // Capture unique input, or NodeSentinel for 2nd input |
duke@435 | 601 | } |
adlertz@5539 | 602 | if (u != NodeSentinel) { // Junk Phi. Remove |
adlertz@6041 | 603 | phi->replace_by(u); |
adlertz@6041 | 604 | j -= yank_if_dead(phi, block, &value, ®nd); |
adlertz@5509 | 605 | phi_dex--; |
duke@435 | 606 | continue; |
duke@435 | 607 | } |
duke@435 | 608 | // Note that if value[pidx] exists, then we merged no new values here |
duke@435 | 609 | // and the phi is useless. This can happen even with the above phi |
duke@435 | 610 | // removal for complex flows. I cannot keep the better known value here |
duke@435 | 611 | // because locally the phi appears to define a new merged value. If I |
duke@435 | 612 | // keep the better value then a copy of the phi, being unable to use the |
duke@435 | 613 | // global flow analysis, can't "peek through" the phi to the original |
duke@435 | 614 | // reaching value and so will act like it's defining a new value. This |
duke@435 | 615 | // can lead to situations where some uses are from the old and some from |
duke@435 | 616 | // the new values. Not illegal by itself but throws the over-strong |
duke@435 | 617 | // assert in scheduling. |
duke@435 | 618 | if( pidx ) { |
duke@435 | 619 | value.map(preg,phi); |
duke@435 | 620 | regnd.map(preg,phi); |
kvn@3882 | 621 | int n_regs = RegMask::num_registers(phi->ideal_reg()); |
kvn@3882 | 622 | for (int l = 1; l < n_regs; l++) { |
kvn@3882 | 623 | OptoReg::Name preg_lo = OptoReg::add(preg,-l); |
duke@435 | 624 | value.map(preg_lo,phi); |
duke@435 | 625 | regnd.map(preg_lo,phi); |
duke@435 | 626 | } |
duke@435 | 627 | } |
duke@435 | 628 | } |
duke@435 | 629 | |
duke@435 | 630 | // For all remaining instructions |
adlertz@5635 | 631 | for (j = phi_dex; j < block->number_of_nodes(); j++) { |
adlertz@5635 | 632 | Node* n = block->get_node(j); |
duke@435 | 633 | |
adlertz@5539 | 634 | if(n->outcnt() == 0 && // Dead? |
adlertz@5539 | 635 | n != C->top() && // (ignore TOP, it has no du info) |
adlertz@5539 | 636 | !n->is_Proj() ) { // fat-proj kills |
adlertz@5539 | 637 | j -= yank_if_dead(n, block, &value, ®nd); |
duke@435 | 638 | continue; |
duke@435 | 639 | } |
duke@435 | 640 | |
duke@435 | 641 | // Improve reaching-def info. Occasionally post-alloc's liveness gives |
duke@435 | 642 | // up (at loop backedges, because we aren't doing a full flow pass). |
duke@435 | 643 | // The presence of a live use essentially asserts that the use's def is |
duke@435 | 644 | // alive and well at the use (or else the allocator fubar'd). Take |
duke@435 | 645 | // advantage of this info to set a reaching def for the use-reg. |
duke@435 | 646 | uint k; |
neliasso@4949 | 647 | for (k = 1; k < n->req(); k++) { |
duke@435 | 648 | Node *def = n->in(k); // n->in(k) is a USE; def is the DEF for this USE |
duke@435 | 649 | guarantee(def != NULL, "no disconnected nodes at this point"); |
neliasso@4949 | 650 | uint useidx = _lrg_map.live_range_id(def); // useidx is the live range index for this USE |
duke@435 | 651 | |
duke@435 | 652 | if( useidx ) { |
duke@435 | 653 | OptoReg::Name ureg = lrgs(useidx).reg(); |
duke@435 | 654 | if( !value[ureg] ) { |
duke@435 | 655 | int idx; // Skip occasional useless copy |
duke@435 | 656 | while( (idx=def->is_Copy()) != 0 && |
duke@435 | 657 | def->in(idx) != NULL && // NULL should not happen |
neliasso@4949 | 658 | ureg == lrgs(_lrg_map.live_range_id(def->in(idx))).reg()) |
duke@435 | 659 | def = def->in(idx); |
duke@435 | 660 | Node *valdef = skip_copies(def); // tighten up val through non-useless copies |
duke@435 | 661 | value.map(ureg,valdef); // record improved reaching-def info |
duke@435 | 662 | regnd.map(ureg, def); |
duke@435 | 663 | // Record other half of doubles |
kvn@3882 | 664 | uint def_ideal_reg = def->ideal_reg(); |
kvn@3882 | 665 | int n_regs = RegMask::num_registers(def_ideal_reg); |
kvn@3882 | 666 | for (int l = 1; l < n_regs; l++) { |
kvn@3882 | 667 | OptoReg::Name ureg_lo = OptoReg::add(ureg,-l); |
kvn@3882 | 668 | if (!value[ureg_lo] && |
kvn@3882 | 669 | (!RegMask::can_represent(ureg_lo) || |
kvn@3882 | 670 | lrgs(useidx).mask().Member(ureg_lo))) { // Nearly always adjacent |
kvn@3882 | 671 | value.map(ureg_lo,valdef); // record improved reaching-def info |
kvn@3882 | 672 | regnd.map(ureg_lo, def); |
kvn@3882 | 673 | } |
duke@435 | 674 | } |
duke@435 | 675 | } |
duke@435 | 676 | } |
duke@435 | 677 | } |
duke@435 | 678 | |
duke@435 | 679 | const uint two_adr = n->is_Mach() ? n->as_Mach()->two_adr() : 0; |
duke@435 | 680 | |
duke@435 | 681 | // Remove copies along input edges |
adlertz@5539 | 682 | for (k = 1; k < n->req(); k++) { |
adlertz@5539 | 683 | j -= elide_copy(n, k, block, value, regnd, two_adr != k); |
adlertz@5539 | 684 | } |
duke@435 | 685 | |
duke@435 | 686 | // Unallocated Nodes define no registers |
neliasso@4949 | 687 | uint lidx = _lrg_map.live_range_id(n); |
neliasso@4949 | 688 | if (!lidx) { |
neliasso@4949 | 689 | continue; |
neliasso@4949 | 690 | } |
duke@435 | 691 | |
duke@435 | 692 | // Update the register defined by this instruction |
duke@435 | 693 | OptoReg::Name nreg = lrgs(lidx).reg(); |
duke@435 | 694 | // Skip through all copies to the _value_ being defined. |
duke@435 | 695 | // Do not change from int to pointer |
duke@435 | 696 | Node *val = skip_copies(n); |
duke@435 | 697 | |
never@1358 | 698 | // Clear out a dead definition before starting so that the |
never@1358 | 699 | // elimination code doesn't have to guard against it. The |
never@1358 | 700 | // definition could in fact be a kill projection with a count of |
never@1358 | 701 | // 0 which is safe but since those are uninteresting for copy |
never@1358 | 702 | // elimination just delete them as well. |
never@1358 | 703 | if (regnd[nreg] != NULL && regnd[nreg]->outcnt() == 0) { |
never@1358 | 704 | regnd.map(nreg, NULL); |
never@1358 | 705 | value.map(nreg, NULL); |
never@1358 | 706 | } |
never@1358 | 707 | |
duke@435 | 708 | uint n_ideal_reg = n->ideal_reg(); |
kvn@3882 | 709 | int n_regs = RegMask::num_registers(n_ideal_reg); |
kvn@3882 | 710 | if (n_regs == 1) { |
duke@435 | 711 | // If Node 'n' does not change the value mapped by the register, |
duke@435 | 712 | // then 'n' is a useless copy. Do not update the register->node |
duke@435 | 713 | // mapping so 'n' will go dead. |
duke@435 | 714 | if( value[nreg] != val ) { |
adlertz@5539 | 715 | if (eliminate_copy_of_constant(val, n, block, value, regnd, nreg, OptoReg::Bad)) { |
adlertz@5539 | 716 | j -= replace_and_yank_if_dead(n, nreg, block, value, regnd); |
duke@435 | 717 | } else { |
duke@435 | 718 | // Update the mapping: record new Node defined by the register |
duke@435 | 719 | regnd.map(nreg,n); |
duke@435 | 720 | // Update mapping for defined *value*, which is the defined |
duke@435 | 721 | // Node after skipping all copies. |
duke@435 | 722 | value.map(nreg,val); |
duke@435 | 723 | } |
never@1358 | 724 | } else if( !may_be_copy_of_callee(n) ) { |
adlertz@5539 | 725 | assert(n->is_Copy(), ""); |
adlertz@5539 | 726 | j -= replace_and_yank_if_dead(n, nreg, block, value, regnd); |
duke@435 | 727 | } |
kvn@3882 | 728 | } else if (RegMask::is_vector(n_ideal_reg)) { |
kvn@3882 | 729 | // If Node 'n' does not change the value mapped by the register, |
kvn@3882 | 730 | // then 'n' is a useless copy. Do not update the register->node |
kvn@3882 | 731 | // mapping so 'n' will go dead. |
kvn@3882 | 732 | if (!register_contains_value(val, nreg, n_regs, value)) { |
kvn@3882 | 733 | // Update the mapping: record new Node defined by the register |
kvn@3882 | 734 | regnd.map(nreg,n); |
kvn@3882 | 735 | // Update mapping for defined *value*, which is the defined |
kvn@3882 | 736 | // Node after skipping all copies. |
kvn@3882 | 737 | value.map(nreg,val); |
kvn@3882 | 738 | for (int l = 1; l < n_regs; l++) { |
kvn@3882 | 739 | OptoReg::Name nreg_lo = OptoReg::add(nreg,-l); |
kvn@3882 | 740 | regnd.map(nreg_lo, n ); |
kvn@3882 | 741 | value.map(nreg_lo,val); |
kvn@3882 | 742 | } |
kvn@3882 | 743 | } else if (n->is_Copy()) { |
kvn@3882 | 744 | // Note: vector can't be constant and can't be copy of calee. |
adlertz@5539 | 745 | j -= replace_and_yank_if_dead(n, nreg, block, value, regnd); |
kvn@3882 | 746 | } |
duke@435 | 747 | } else { |
duke@435 | 748 | // If the value occupies a register pair, record same info |
duke@435 | 749 | // in both registers. |
duke@435 | 750 | OptoReg::Name nreg_lo = OptoReg::add(nreg,-1); |
duke@435 | 751 | if( RegMask::can_represent(nreg_lo) && // Either a spill slot, or |
duke@435 | 752 | !lrgs(lidx).mask().Member(nreg_lo) ) { // Nearly always adjacent |
duke@435 | 753 | // Sparc occasionally has non-adjacent pairs. |
duke@435 | 754 | // Find the actual other value |
duke@435 | 755 | RegMask tmp = lrgs(lidx).mask(); |
duke@435 | 756 | tmp.Remove(nreg); |
duke@435 | 757 | nreg_lo = tmp.find_first_elem(); |
duke@435 | 758 | } |
adlertz@5539 | 759 | if (value[nreg] != val || value[nreg_lo] != val) { |
adlertz@5539 | 760 | if (eliminate_copy_of_constant(val, n, block, value, regnd, nreg, nreg_lo)) { |
adlertz@5539 | 761 | j -= replace_and_yank_if_dead(n, nreg, block, value, regnd); |
duke@435 | 762 | } else { |
duke@435 | 763 | regnd.map(nreg , n ); |
duke@435 | 764 | regnd.map(nreg_lo, n ); |
duke@435 | 765 | value.map(nreg ,val); |
duke@435 | 766 | value.map(nreg_lo,val); |
duke@435 | 767 | } |
adlertz@5539 | 768 | } else if (!may_be_copy_of_callee(n)) { |
adlertz@5539 | 769 | assert(n->is_Copy(), ""); |
adlertz@5539 | 770 | j -= replace_and_yank_if_dead(n, nreg, block, value, regnd); |
duke@435 | 771 | } |
duke@435 | 772 | } |
duke@435 | 773 | |
duke@435 | 774 | // Fat projections kill many registers |
duke@435 | 775 | if( n_ideal_reg == MachProjNode::fat_proj ) { |
duke@435 | 776 | RegMask rm = n->out_RegMask(); |
duke@435 | 777 | // wow, what an expensive iterator... |
duke@435 | 778 | nreg = rm.find_first_elem(); |
duke@435 | 779 | while( OptoReg::is_valid(nreg)) { |
duke@435 | 780 | rm.Remove(nreg); |
duke@435 | 781 | value.map(nreg,n); |
duke@435 | 782 | regnd.map(nreg,n); |
duke@435 | 783 | nreg = rm.find_first_elem(); |
duke@435 | 784 | } |
duke@435 | 785 | } |
duke@435 | 786 | |
duke@435 | 787 | } // End of for all instructions in the block |
duke@435 | 788 | |
duke@435 | 789 | } // End for all blocks |
duke@435 | 790 | } |