Tue, 09 Mar 2010 20:16:19 +0100
6919934: JSR 292 needs to support x86 C1
Summary: This implements JSR 292 support for C1 x86.
Reviewed-by: never, jrose, kvn
duke@435 | 1 | /* |
never@1708 | 2 | * Copyright 1999-2010 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/_split_if.cpp.incl" |
duke@435 | 27 | |
duke@435 | 28 | |
duke@435 | 29 | //------------------------------split_thru_region------------------------------ |
duke@435 | 30 | // Split Node 'n' through merge point. |
duke@435 | 31 | Node *PhaseIdealLoop::split_thru_region( Node *n, Node *region ) { |
duke@435 | 32 | uint wins = 0; |
duke@435 | 33 | assert( n->is_CFG(), "" ); |
duke@435 | 34 | assert( region->is_Region(), "" ); |
duke@435 | 35 | Node *r = new (C, region->req()) RegionNode( region->req() ); |
duke@435 | 36 | IdealLoopTree *loop = get_loop( n ); |
duke@435 | 37 | for( uint i = 1; i < region->req(); i++ ) { |
duke@435 | 38 | Node *x = n->clone(); |
duke@435 | 39 | Node *in0 = n->in(0); |
duke@435 | 40 | if( in0->in(0) == region ) x->set_req( 0, in0->in(i) ); |
duke@435 | 41 | for( uint j = 1; j < n->req(); j++ ) { |
duke@435 | 42 | Node *in = n->in(j); |
duke@435 | 43 | if( get_ctrl(in) == region ) |
duke@435 | 44 | x->set_req( j, in->in(i) ); |
duke@435 | 45 | } |
duke@435 | 46 | _igvn.register_new_node_with_optimizer(x); |
duke@435 | 47 | set_loop(x, loop); |
duke@435 | 48 | set_idom(x, x->in(0), dom_depth(x->in(0))+1); |
duke@435 | 49 | r->init_req(i, x); |
duke@435 | 50 | } |
duke@435 | 51 | |
duke@435 | 52 | // Record region |
duke@435 | 53 | r->set_req(0,region); // Not a TRUE RegionNode |
duke@435 | 54 | _igvn.register_new_node_with_optimizer(r); |
duke@435 | 55 | set_loop(r, loop); |
duke@435 | 56 | if( !loop->_child ) |
duke@435 | 57 | loop->_body.push(r); |
duke@435 | 58 | return r; |
duke@435 | 59 | } |
duke@435 | 60 | |
duke@435 | 61 | //------------------------------split_up--------------------------------------- |
duke@435 | 62 | // Split block-local op up through the phis to empty the current block |
duke@435 | 63 | bool PhaseIdealLoop::split_up( Node *n, Node *blk1, Node *blk2 ) { |
duke@435 | 64 | if( n->is_CFG() ) { |
duke@435 | 65 | assert( n->in(0) != blk1, "Lousy candidate for split-if" ); |
duke@435 | 66 | return false; |
duke@435 | 67 | } |
duke@435 | 68 | if( get_ctrl(n) != blk1 && get_ctrl(n) != blk2 ) |
duke@435 | 69 | return false; // Not block local |
duke@435 | 70 | if( n->is_Phi() ) return false; // Local PHIs are expected |
duke@435 | 71 | |
duke@435 | 72 | // Recursively split-up inputs |
duke@435 | 73 | for (uint i = 1; i < n->req(); i++) { |
duke@435 | 74 | if( split_up( n->in(i), blk1, blk2 ) ) { |
duke@435 | 75 | // Got split recursively and self went dead? |
duke@435 | 76 | if (n->outcnt() == 0) |
duke@435 | 77 | _igvn.remove_dead_node(n); |
duke@435 | 78 | return true; |
duke@435 | 79 | } |
duke@435 | 80 | } |
duke@435 | 81 | |
duke@435 | 82 | // Check for needing to clone-up a compare. Can't do that, it forces |
duke@435 | 83 | // another (nested) split-if transform. Instead, clone it "down". |
duke@435 | 84 | if( n->is_Cmp() ) { |
duke@435 | 85 | assert(get_ctrl(n) == blk2 || get_ctrl(n) == blk1, "must be in block with IF"); |
duke@435 | 86 | // Check for simple Cmp/Bool/CMove which we can clone-up. Cmp/Bool/CMove |
duke@435 | 87 | // sequence can have no other users and it must all reside in the split-if |
duke@435 | 88 | // block. Non-simple Cmp/Bool/CMove sequences are 'cloned-down' below - |
duke@435 | 89 | // private, per-use versions of the Cmp and Bool are made. These sink to |
duke@435 | 90 | // the CMove block. If the CMove is in the split-if block, then in the |
duke@435 | 91 | // next iteration this will become a simple Cmp/Bool/CMove set to clone-up. |
duke@435 | 92 | Node *bol, *cmov; |
duke@435 | 93 | if( !(n->outcnt() == 1 && n->unique_out()->is_Bool() && |
duke@435 | 94 | (bol = n->unique_out()->as_Bool()) && |
duke@435 | 95 | (get_ctrl(bol) == blk1 || |
duke@435 | 96 | get_ctrl(bol) == blk2) && |
duke@435 | 97 | bol->outcnt() == 1 && |
duke@435 | 98 | bol->unique_out()->is_CMove() && |
duke@435 | 99 | (cmov = bol->unique_out()->as_CMove()) && |
duke@435 | 100 | (get_ctrl(cmov) == blk1 || |
duke@435 | 101 | get_ctrl(cmov) == blk2) ) ) { |
duke@435 | 102 | |
duke@435 | 103 | // Must clone down |
duke@435 | 104 | #ifndef PRODUCT |
duke@435 | 105 | if( PrintOpto && VerifyLoopOptimizations ) { |
duke@435 | 106 | tty->print("Cloning down: "); |
duke@435 | 107 | n->dump(); |
duke@435 | 108 | } |
duke@435 | 109 | #endif |
duke@435 | 110 | // Clone down any block-local BoolNode uses of this CmpNode |
duke@435 | 111 | for (DUIterator i = n->outs(); n->has_out(i); i++) { |
duke@435 | 112 | Node* bol = n->out(i); |
duke@435 | 113 | assert( bol->is_Bool(), "" ); |
duke@435 | 114 | if (bol->outcnt() == 1) { |
duke@435 | 115 | Node* use = bol->unique_out(); |
duke@435 | 116 | Node *use_c = use->is_If() ? use->in(0) : get_ctrl(use); |
duke@435 | 117 | if (use_c == blk1 || use_c == blk2) { |
duke@435 | 118 | continue; |
duke@435 | 119 | } |
duke@435 | 120 | } |
duke@435 | 121 | if (get_ctrl(bol) == blk1 || get_ctrl(bol) == blk2) { |
duke@435 | 122 | // Recursively sink any BoolNode |
duke@435 | 123 | #ifndef PRODUCT |
duke@435 | 124 | if( PrintOpto && VerifyLoopOptimizations ) { |
duke@435 | 125 | tty->print("Cloning down: "); |
duke@435 | 126 | bol->dump(); |
duke@435 | 127 | } |
duke@435 | 128 | #endif |
duke@435 | 129 | for (DUIterator_Last jmin, j = bol->last_outs(jmin); j >= jmin; --j) { |
duke@435 | 130 | // Uses are either IfNodes or CMoves |
duke@435 | 131 | Node* iff = bol->last_out(j); |
duke@435 | 132 | assert( iff->in(1) == bol, "" ); |
duke@435 | 133 | // Get control block of either the CMove or the If input |
duke@435 | 134 | Node *iff_ctrl = iff->is_If() ? iff->in(0) : get_ctrl(iff); |
duke@435 | 135 | Node *x = bol->clone(); |
duke@435 | 136 | register_new_node(x, iff_ctrl); |
duke@435 | 137 | _igvn.hash_delete(iff); |
duke@435 | 138 | iff->set_req(1, x); |
duke@435 | 139 | _igvn._worklist.push(iff); |
duke@435 | 140 | } |
duke@435 | 141 | _igvn.remove_dead_node( bol ); |
duke@435 | 142 | --i; |
duke@435 | 143 | } |
duke@435 | 144 | } |
duke@435 | 145 | // Clone down this CmpNode |
duke@435 | 146 | for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; --j) { |
duke@435 | 147 | Node* bol = n->last_out(j); |
duke@435 | 148 | assert( bol->in(1) == n, "" ); |
duke@435 | 149 | Node *x = n->clone(); |
duke@435 | 150 | register_new_node(x, get_ctrl(bol)); |
duke@435 | 151 | _igvn.hash_delete(bol); |
duke@435 | 152 | bol->set_req(1, x); |
duke@435 | 153 | _igvn._worklist.push(bol); |
duke@435 | 154 | } |
duke@435 | 155 | _igvn.remove_dead_node( n ); |
duke@435 | 156 | |
duke@435 | 157 | return true; |
duke@435 | 158 | } |
duke@435 | 159 | } |
duke@435 | 160 | |
duke@435 | 161 | // See if splitting-up a Store. Any anti-dep loads must go up as |
duke@435 | 162 | // well. An anti-dep load might be in the wrong block, because in |
duke@435 | 163 | // this particular layout/schedule we ignored anti-deps and allow |
duke@435 | 164 | // memory to be alive twice. This only works if we do the same |
duke@435 | 165 | // operations on anti-dep loads as we do their killing stores. |
duke@435 | 166 | if( n->is_Store() && n->in(MemNode::Memory)->in(0) == n->in(0) ) { |
duke@435 | 167 | // Get store's memory slice |
duke@435 | 168 | int alias_idx = C->get_alias_index(_igvn.type(n->in(MemNode::Address))->is_ptr()); |
duke@435 | 169 | |
duke@435 | 170 | // Get memory-phi anti-dep loads will be using |
duke@435 | 171 | Node *memphi = n->in(MemNode::Memory); |
duke@435 | 172 | assert( memphi->is_Phi(), "" ); |
duke@435 | 173 | // Hoist any anti-dep load to the splitting block; |
duke@435 | 174 | // it will then "split-up". |
duke@435 | 175 | for (DUIterator_Fast imax,i = memphi->fast_outs(imax); i < imax; i++) { |
duke@435 | 176 | Node *load = memphi->fast_out(i); |
duke@435 | 177 | if( load->is_Load() && alias_idx == C->get_alias_index(_igvn.type(load->in(MemNode::Address))->is_ptr()) ) |
duke@435 | 178 | set_ctrl(load,blk1); |
duke@435 | 179 | } |
duke@435 | 180 | } |
duke@435 | 181 | |
duke@435 | 182 | // Found some other Node; must clone it up |
duke@435 | 183 | #ifndef PRODUCT |
duke@435 | 184 | if( PrintOpto && VerifyLoopOptimizations ) { |
duke@435 | 185 | tty->print("Cloning up: "); |
duke@435 | 186 | n->dump(); |
duke@435 | 187 | } |
duke@435 | 188 | #endif |
duke@435 | 189 | |
never@1708 | 190 | // ConvI2L may have type information on it which becomes invalid if |
never@1708 | 191 | // it moves up in the graph so change any clones so widen the type |
never@1708 | 192 | // to TypeLong::INT when pushing it up. |
never@1708 | 193 | const Type* rtype = NULL; |
never@1708 | 194 | if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::INT) { |
never@1708 | 195 | rtype = TypeLong::INT; |
never@1708 | 196 | } |
never@1708 | 197 | |
duke@435 | 198 | // Now actually split-up this guy. One copy per control path merging. |
duke@435 | 199 | Node *phi = PhiNode::make_blank(blk1, n); |
duke@435 | 200 | for( uint j = 1; j < blk1->req(); j++ ) { |
duke@435 | 201 | Node *x = n->clone(); |
never@1708 | 202 | // Widen the type of the ConvI2L when pushing up. |
never@1708 | 203 | if (rtype != NULL) x->as_Type()->set_type(rtype); |
duke@435 | 204 | if( n->in(0) && n->in(0) == blk1 ) |
duke@435 | 205 | x->set_req( 0, blk1->in(j) ); |
duke@435 | 206 | for( uint i = 1; i < n->req(); i++ ) { |
duke@435 | 207 | Node *m = n->in(i); |
duke@435 | 208 | if( get_ctrl(m) == blk1 ) { |
duke@435 | 209 | assert( m->in(0) == blk1, "" ); |
duke@435 | 210 | x->set_req( i, m->in(j) ); |
duke@435 | 211 | } |
duke@435 | 212 | } |
duke@435 | 213 | register_new_node( x, blk1->in(j) ); |
duke@435 | 214 | phi->init_req( j, x ); |
duke@435 | 215 | } |
duke@435 | 216 | // Announce phi to optimizer |
duke@435 | 217 | register_new_node(phi, blk1); |
duke@435 | 218 | |
duke@435 | 219 | // Remove cloned-up value from optimizer; use phi instead |
duke@435 | 220 | _igvn.hash_delete(n); |
duke@435 | 221 | _igvn.subsume_node( n, phi ); |
duke@435 | 222 | |
duke@435 | 223 | // (There used to be a self-recursive call to split_up() here, |
duke@435 | 224 | // but it is not needed. All necessary forward walking is done |
duke@435 | 225 | // by do_split_if() below.) |
duke@435 | 226 | |
duke@435 | 227 | return true; |
duke@435 | 228 | } |
duke@435 | 229 | |
duke@435 | 230 | //------------------------------register_new_node------------------------------ |
duke@435 | 231 | void PhaseIdealLoop::register_new_node( Node *n, Node *blk ) { |
cfang@1607 | 232 | assert(!n->is_CFG(), "must be data node"); |
duke@435 | 233 | _igvn.register_new_node_with_optimizer(n); |
duke@435 | 234 | set_ctrl(n, blk); |
duke@435 | 235 | IdealLoopTree *loop = get_loop(blk); |
duke@435 | 236 | if( !loop->_child ) |
duke@435 | 237 | loop->_body.push(n); |
duke@435 | 238 | } |
duke@435 | 239 | |
duke@435 | 240 | //------------------------------small_cache------------------------------------ |
duke@435 | 241 | struct small_cache : public Dict { |
duke@435 | 242 | |
duke@435 | 243 | small_cache() : Dict( cmpkey, hashptr ) {} |
duke@435 | 244 | Node *probe( Node *use_blk ) { return (Node*)((*this)[use_blk]); } |
duke@435 | 245 | void lru_insert( Node *use_blk, Node *new_def ) { Insert(use_blk,new_def); } |
duke@435 | 246 | }; |
duke@435 | 247 | |
duke@435 | 248 | //------------------------------spinup----------------------------------------- |
duke@435 | 249 | // "Spin up" the dominator tree, starting at the use site and stopping when we |
duke@435 | 250 | // find the post-dominating point. |
duke@435 | 251 | |
duke@435 | 252 | // We must be at the merge point which post-dominates 'new_false' and |
duke@435 | 253 | // 'new_true'. Figure out which edges into the RegionNode eventually lead up |
duke@435 | 254 | // to false and which to true. Put in a PhiNode to merge values; plug in |
duke@435 | 255 | // the appropriate false-arm or true-arm values. If some path leads to the |
duke@435 | 256 | // original IF, then insert a Phi recursively. |
duke@435 | 257 | Node *PhaseIdealLoop::spinup( Node *iff_dom, Node *new_false, Node *new_true, Node *use_blk, Node *def, small_cache *cache ) { |
duke@435 | 258 | if (use_blk->is_top()) // Handle dead uses |
duke@435 | 259 | return use_blk; |
duke@435 | 260 | Node *prior_n = (Node*)0xdeadbeef; |
duke@435 | 261 | Node *n = use_blk; // Get path input |
duke@435 | 262 | assert( use_blk != iff_dom, "" ); |
duke@435 | 263 | // Here's the "spinup" the dominator tree loop. Do a cache-check |
duke@435 | 264 | // along the way, in case we've come this way before. |
duke@435 | 265 | while( n != iff_dom ) { // Found post-dominating point? |
duke@435 | 266 | prior_n = n; |
duke@435 | 267 | n = idom(n); // Search higher |
duke@435 | 268 | Node *s = cache->probe( prior_n ); // Check cache |
duke@435 | 269 | if( s ) return s; // Cache hit! |
duke@435 | 270 | } |
duke@435 | 271 | |
duke@435 | 272 | Node *phi_post; |
duke@435 | 273 | if( prior_n == new_false || prior_n == new_true ) { |
duke@435 | 274 | phi_post = def->clone(); |
duke@435 | 275 | phi_post->set_req(0, prior_n ); |
duke@435 | 276 | register_new_node(phi_post, prior_n); |
duke@435 | 277 | } else { |
duke@435 | 278 | // This method handles both control uses (looking for Regions) or data |
duke@435 | 279 | // uses (looking for Phis). If looking for a control use, then we need |
duke@435 | 280 | // to insert a Region instead of a Phi; however Regions always exist |
duke@435 | 281 | // previously (the hash_find_insert below would always hit) so we can |
duke@435 | 282 | // return the existing Region. |
duke@435 | 283 | if( def->is_CFG() ) { |
duke@435 | 284 | phi_post = prior_n; // If looking for CFG, return prior |
duke@435 | 285 | } else { |
duke@435 | 286 | assert( def->is_Phi(), "" ); |
duke@435 | 287 | assert( prior_n->is_Region(), "must be a post-dominating merge point" ); |
duke@435 | 288 | |
duke@435 | 289 | // Need a Phi here |
duke@435 | 290 | phi_post = PhiNode::make_blank(prior_n, def); |
duke@435 | 291 | // Search for both true and false on all paths till find one. |
duke@435 | 292 | for( uint i = 1; i < phi_post->req(); i++ ) // For all paths |
duke@435 | 293 | phi_post->init_req( i, spinup( iff_dom, new_false, new_true, prior_n->in(i), def, cache ) ); |
duke@435 | 294 | Node *t = _igvn.hash_find_insert(phi_post); |
duke@435 | 295 | if( t ) { // See if we already have this one |
duke@435 | 296 | // phi_post will not be used, so kill it |
duke@435 | 297 | _igvn.remove_dead_node(phi_post); |
duke@435 | 298 | phi_post->destruct(); |
duke@435 | 299 | phi_post = t; |
duke@435 | 300 | } else { |
duke@435 | 301 | register_new_node( phi_post, prior_n ); |
duke@435 | 302 | } |
duke@435 | 303 | } |
duke@435 | 304 | } |
duke@435 | 305 | |
duke@435 | 306 | // Update cache everywhere |
duke@435 | 307 | prior_n = (Node*)0xdeadbeef; // Reset IDOM walk |
duke@435 | 308 | n = use_blk; // Get path input |
duke@435 | 309 | // Spin-up the idom tree again, basically doing path-compression. |
duke@435 | 310 | // Insert cache entries along the way, so that if we ever hit this |
duke@435 | 311 | // point in the IDOM tree again we'll stop immediately on a cache hit. |
duke@435 | 312 | while( n != iff_dom ) { // Found post-dominating point? |
duke@435 | 313 | prior_n = n; |
duke@435 | 314 | n = idom(n); // Search higher |
duke@435 | 315 | cache->lru_insert( prior_n, phi_post ); // Fill cache |
duke@435 | 316 | } // End of while not gone high enough |
duke@435 | 317 | |
duke@435 | 318 | return phi_post; |
duke@435 | 319 | } |
duke@435 | 320 | |
duke@435 | 321 | //------------------------------find_use_block--------------------------------- |
duke@435 | 322 | // Find the block a USE is in. Normally USE's are in the same block as the |
duke@435 | 323 | // using instruction. For Phi-USE's, the USE is in the predecessor block |
duke@435 | 324 | // along the corresponding path. |
duke@435 | 325 | Node *PhaseIdealLoop::find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ) { |
duke@435 | 326 | // CFG uses are their own block |
duke@435 | 327 | if( use->is_CFG() ) |
duke@435 | 328 | return use; |
duke@435 | 329 | |
duke@435 | 330 | if( use->is_Phi() ) { // Phi uses in prior block |
duke@435 | 331 | // Grab the first Phi use; there may be many. |
twisti@1040 | 332 | // Each will be handled as a separate iteration of |
duke@435 | 333 | // the "while( phi->outcnt() )" loop. |
duke@435 | 334 | uint j; |
duke@435 | 335 | for( j = 1; j < use->req(); j++ ) |
duke@435 | 336 | if( use->in(j) == def ) |
duke@435 | 337 | break; |
duke@435 | 338 | assert( j < use->req(), "def should be among use's inputs" ); |
duke@435 | 339 | return use->in(0)->in(j); |
duke@435 | 340 | } |
duke@435 | 341 | // Normal (non-phi) use |
duke@435 | 342 | Node *use_blk = get_ctrl(use); |
duke@435 | 343 | // Some uses are directly attached to the old (and going away) |
duke@435 | 344 | // false and true branches. |
duke@435 | 345 | if( use_blk == old_false ) { |
duke@435 | 346 | use_blk = new_false; |
duke@435 | 347 | set_ctrl(use, new_false); |
duke@435 | 348 | } |
duke@435 | 349 | if( use_blk == old_true ) { |
duke@435 | 350 | use_blk = new_true; |
duke@435 | 351 | set_ctrl(use, new_true); |
duke@435 | 352 | } |
duke@435 | 353 | |
duke@435 | 354 | if (use_blk == NULL) { // He's dead, Jim |
duke@435 | 355 | _igvn.hash_delete(use); |
duke@435 | 356 | _igvn.subsume_node(use, C->top()); |
duke@435 | 357 | } |
duke@435 | 358 | |
duke@435 | 359 | return use_blk; |
duke@435 | 360 | } |
duke@435 | 361 | |
duke@435 | 362 | //------------------------------handle_use------------------------------------- |
duke@435 | 363 | // Handle uses of the merge point. Basically, split-if makes the merge point |
duke@435 | 364 | // go away so all uses of the merge point must go away as well. Most block |
duke@435 | 365 | // local uses have already been split-up, through the merge point. Uses from |
duke@435 | 366 | // far below the merge point can't always be split up (e.g., phi-uses are |
duke@435 | 367 | // pinned) and it makes too much stuff live. Instead we use a path-based |
duke@435 | 368 | // solution to move uses down. |
duke@435 | 369 | // |
duke@435 | 370 | // If the use is along the pre-split-CFG true branch, then the new use will |
duke@435 | 371 | // be from the post-split-CFG true merge point. Vice-versa for the false |
duke@435 | 372 | // path. Some uses will be along both paths; then we sink the use to the |
duke@435 | 373 | // post-dominating location; we may need to insert a Phi there. |
duke@435 | 374 | void PhaseIdealLoop::handle_use( Node *use, Node *def, small_cache *cache, Node *region_dom, Node *new_false, Node *new_true, Node *old_false, Node *old_true ) { |
duke@435 | 375 | |
duke@435 | 376 | Node *use_blk = find_use_block(use,def,old_false,new_false,old_true,new_true); |
duke@435 | 377 | if( !use_blk ) return; // He's dead, Jim |
duke@435 | 378 | |
duke@435 | 379 | // Walk up the dominator tree until I hit either the old IfFalse, the old |
duke@435 | 380 | // IfTrue or the old If. Insert Phis where needed. |
duke@435 | 381 | Node *new_def = spinup( region_dom, new_false, new_true, use_blk, def, cache ); |
duke@435 | 382 | |
duke@435 | 383 | // Found where this USE goes. Re-point him. |
duke@435 | 384 | uint i; |
duke@435 | 385 | for( i = 0; i < use->req(); i++ ) |
duke@435 | 386 | if( use->in(i) == def ) |
duke@435 | 387 | break; |
duke@435 | 388 | assert( i < use->req(), "def should be among use's inputs" ); |
duke@435 | 389 | _igvn.hash_delete(use); |
duke@435 | 390 | use->set_req(i, new_def); |
duke@435 | 391 | _igvn._worklist.push(use); |
duke@435 | 392 | } |
duke@435 | 393 | |
duke@435 | 394 | //------------------------------do_split_if------------------------------------ |
duke@435 | 395 | // Found an If getting its condition-code input from a Phi in the same block. |
duke@435 | 396 | // Split thru the Region. |
duke@435 | 397 | void PhaseIdealLoop::do_split_if( Node *iff ) { |
duke@435 | 398 | #ifndef PRODUCT |
duke@435 | 399 | if( PrintOpto && VerifyLoopOptimizations ) |
duke@435 | 400 | tty->print_cr("Split-if"); |
duke@435 | 401 | #endif |
duke@435 | 402 | C->set_major_progress(); |
duke@435 | 403 | Node *region = iff->in(0); |
duke@435 | 404 | Node *region_dom = idom(region); |
duke@435 | 405 | |
duke@435 | 406 | // We are going to clone this test (and the control flow with it) up through |
duke@435 | 407 | // the incoming merge point. We need to empty the current basic block. |
duke@435 | 408 | // Clone any instructions which must be in this block up through the merge |
duke@435 | 409 | // point. |
duke@435 | 410 | DUIterator i, j; |
duke@435 | 411 | bool progress = true; |
duke@435 | 412 | while (progress) { |
duke@435 | 413 | progress = false; |
duke@435 | 414 | for (i = region->outs(); region->has_out(i); i++) { |
duke@435 | 415 | Node* n = region->out(i); |
duke@435 | 416 | if( n == region ) continue; |
duke@435 | 417 | // The IF to be split is OK. |
duke@435 | 418 | if( n == iff ) continue; |
duke@435 | 419 | if( !n->is_Phi() ) { // Found pinned memory op or such |
duke@435 | 420 | if (split_up(n, region, iff)) { |
duke@435 | 421 | i = region->refresh_out_pos(i); |
duke@435 | 422 | progress = true; |
duke@435 | 423 | } |
duke@435 | 424 | continue; |
duke@435 | 425 | } |
duke@435 | 426 | assert( n->in(0) == region, "" ); |
duke@435 | 427 | |
duke@435 | 428 | // Recursively split up all users of a Phi |
duke@435 | 429 | for (j = n->outs(); n->has_out(j); j++) { |
duke@435 | 430 | Node* m = n->out(j); |
duke@435 | 431 | // If m is dead, throw it away, and declare progress |
duke@435 | 432 | if (_nodes[m->_idx] == NULL) { |
duke@435 | 433 | _igvn.remove_dead_node(m); |
duke@435 | 434 | // fall through |
duke@435 | 435 | } |
duke@435 | 436 | else if (m != iff && split_up(m, region, iff)) { |
duke@435 | 437 | // fall through |
duke@435 | 438 | } else { |
duke@435 | 439 | continue; |
duke@435 | 440 | } |
duke@435 | 441 | // Something unpredictable changed. |
duke@435 | 442 | // Tell the iterators to refresh themselves, and rerun the loop. |
duke@435 | 443 | i = region->refresh_out_pos(i); |
duke@435 | 444 | j = region->refresh_out_pos(j); |
duke@435 | 445 | progress = true; |
duke@435 | 446 | } |
duke@435 | 447 | } |
duke@435 | 448 | } |
duke@435 | 449 | |
duke@435 | 450 | // Now we have no instructions in the block containing the IF. |
duke@435 | 451 | // Split the IF. |
duke@435 | 452 | Node *new_iff = split_thru_region( iff, region ); |
duke@435 | 453 | |
duke@435 | 454 | // Replace both uses of 'new_iff' with Regions merging True/False |
duke@435 | 455 | // paths. This makes 'new_iff' go dead. |
duke@435 | 456 | Node *old_false, *old_true; |
duke@435 | 457 | Node *new_false, *new_true; |
duke@435 | 458 | for (DUIterator_Last j2min, j2 = iff->last_outs(j2min); j2 >= j2min; --j2) { |
duke@435 | 459 | Node *ifp = iff->last_out(j2); |
duke@435 | 460 | assert( ifp->Opcode() == Op_IfFalse || ifp->Opcode() == Op_IfTrue, "" ); |
duke@435 | 461 | ifp->set_req(0, new_iff); |
duke@435 | 462 | Node *ifpx = split_thru_region( ifp, region ); |
duke@435 | 463 | |
duke@435 | 464 | // Replace 'If' projection of a Region with a Region of |
duke@435 | 465 | // 'If' projections. |
duke@435 | 466 | ifpx->set_req(0, ifpx); // A TRUE RegionNode |
duke@435 | 467 | |
duke@435 | 468 | // Setup dominator info |
duke@435 | 469 | set_idom(ifpx, region_dom, dom_depth(region_dom) + 1); |
duke@435 | 470 | |
duke@435 | 471 | // Check for splitting loop tails |
duke@435 | 472 | if( get_loop(iff)->tail() == ifp ) |
duke@435 | 473 | get_loop(iff)->_tail = ifpx; |
duke@435 | 474 | |
duke@435 | 475 | // Replace in the graph with lazy-update mechanism |
duke@435 | 476 | new_iff->set_req(0, new_iff); // hook self so it does not go dead |
duke@435 | 477 | lazy_replace_proj( ifp, ifpx ); |
duke@435 | 478 | new_iff->set_req(0, region); |
duke@435 | 479 | |
duke@435 | 480 | // Record bits for later xforms |
duke@435 | 481 | if( ifp->Opcode() == Op_IfFalse ) { |
duke@435 | 482 | old_false = ifp; |
duke@435 | 483 | new_false = ifpx; |
duke@435 | 484 | } else { |
duke@435 | 485 | old_true = ifp; |
duke@435 | 486 | new_true = ifpx; |
duke@435 | 487 | } |
duke@435 | 488 | } |
duke@435 | 489 | _igvn.remove_dead_node(new_iff); |
duke@435 | 490 | // Lazy replace IDOM info with the region's dominator |
duke@435 | 491 | lazy_replace( iff, region_dom ); |
duke@435 | 492 | |
duke@435 | 493 | // Now make the original merge point go dead, by handling all its uses. |
duke@435 | 494 | small_cache region_cache; |
duke@435 | 495 | // Preload some control flow in region-cache |
duke@435 | 496 | region_cache.lru_insert( new_false, new_false ); |
duke@435 | 497 | region_cache.lru_insert( new_true , new_true ); |
duke@435 | 498 | // Now handle all uses of the splitting block |
duke@435 | 499 | for (DUIterator_Last kmin, k = region->last_outs(kmin); k >= kmin; --k) { |
duke@435 | 500 | Node* phi = region->last_out(k); |
duke@435 | 501 | if( !phi->in(0) ) { // Dead phi? Remove it |
duke@435 | 502 | _igvn.remove_dead_node(phi); |
duke@435 | 503 | continue; |
duke@435 | 504 | } |
duke@435 | 505 | assert( phi->in(0) == region, "" ); |
duke@435 | 506 | if( phi == region ) { // Found the self-reference |
duke@435 | 507 | phi->set_req(0, NULL); |
duke@435 | 508 | continue; // Break the self-cycle |
duke@435 | 509 | } |
duke@435 | 510 | // Expected common case: Phi hanging off of Region |
duke@435 | 511 | if( phi->is_Phi() ) { |
duke@435 | 512 | // Need a per-def cache. Phi represents a def, so make a cache |
duke@435 | 513 | small_cache phi_cache; |
duke@435 | 514 | |
duke@435 | 515 | // Inspect all Phi uses to make the Phi go dead |
duke@435 | 516 | for (DUIterator_Last lmin, l = phi->last_outs(lmin); l >= lmin; --l) { |
duke@435 | 517 | Node* use = phi->last_out(l); |
duke@435 | 518 | // Compute the new DEF for this USE. New DEF depends on the path |
duke@435 | 519 | // taken from the original DEF to the USE. The new DEF may be some |
duke@435 | 520 | // collection of PHI's merging values from different paths. The Phis |
duke@435 | 521 | // inserted depend only on the location of the USE. We use a |
duke@435 | 522 | // 2-element cache to handle multiple uses from the same block. |
duke@435 | 523 | handle_use( use, phi, &phi_cache, region_dom, new_false, new_true, old_false, old_true ); |
duke@435 | 524 | } // End of while phi has uses |
duke@435 | 525 | |
duke@435 | 526 | // Because handle_use might relocate region->_out, |
duke@435 | 527 | // we must refresh the iterator. |
duke@435 | 528 | k = region->last_outs(kmin); |
duke@435 | 529 | |
duke@435 | 530 | // Remove the dead Phi |
duke@435 | 531 | _igvn.remove_dead_node( phi ); |
duke@435 | 532 | |
duke@435 | 533 | } else { |
duke@435 | 534 | // Random memory op guarded by Region. Compute new DEF for USE. |
duke@435 | 535 | handle_use( phi, region, ®ion_cache, region_dom, new_false, new_true, old_false, old_true ); |
duke@435 | 536 | } |
duke@435 | 537 | |
duke@435 | 538 | } // End of while merge point has phis |
duke@435 | 539 | |
duke@435 | 540 | // Any leftover bits in the splitting block must not have depended on local |
duke@435 | 541 | // Phi inputs (these have already been split-up). Hence it's safe to hoist |
duke@435 | 542 | // these guys to the dominating point. |
duke@435 | 543 | lazy_replace( region, region_dom ); |
duke@435 | 544 | #ifndef PRODUCT |
duke@435 | 545 | if( VerifyLoopOptimizations ) verify(); |
duke@435 | 546 | #endif |
duke@435 | 547 | } |