1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000 1.2 +++ b/src/share/vm/opto/coalesce.cpp Wed Apr 27 01:25:04 2016 +0800 1.3 @@ -0,0 +1,789 @@ 1.4 +/* 1.5 + * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved. 1.6 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 1.7 + * 1.8 + * This code is free software; you can redistribute it and/or modify it 1.9 + * under the terms of the GNU General Public License version 2 only, as 1.10 + * published by the Free Software Foundation. 1.11 + * 1.12 + * This code is distributed in the hope that it will be useful, but WITHOUT 1.13 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 1.14 + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 1.15 + * version 2 for more details (a copy is included in the LICENSE file that 1.16 + * accompanied this code). 1.17 + * 1.18 + * You should have received a copy of the GNU General Public License version 1.19 + * 2 along with this work; if not, write to the Free Software Foundation, 1.20 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 1.21 + * 1.22 + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 1.23 + * or visit www.oracle.com if you need additional information or have any 1.24 + * questions. 1.25 + * 1.26 + */ 1.27 + 1.28 +#include "precompiled.hpp" 1.29 +#include "memory/allocation.inline.hpp" 1.30 +#include "opto/block.hpp" 1.31 +#include "opto/cfgnode.hpp" 1.32 +#include "opto/chaitin.hpp" 1.33 +#include "opto/coalesce.hpp" 1.34 +#include "opto/connode.hpp" 1.35 +#include "opto/indexSet.hpp" 1.36 +#include "opto/machnode.hpp" 1.37 +#include "opto/matcher.hpp" 1.38 +#include "opto/regmask.hpp" 1.39 + 1.40 +#ifndef PRODUCT 1.41 +void PhaseCoalesce::dump(Node *n) const { 1.42 + // Being a const function means I cannot use 'Find' 1.43 + uint r = _phc._lrg_map.find(n); 1.44 + tty->print("L%d/N%d ",r,n->_idx); 1.45 +} 1.46 + 1.47 +void PhaseCoalesce::dump() const { 1.48 + // I know I have a block layout now, so I can print blocks in a loop 1.49 + for( uint i=0; i<_phc._cfg.number_of_blocks(); i++ ) { 1.50 + uint j; 1.51 + Block* b = _phc._cfg.get_block(i); 1.52 + // Print a nice block header 1.53 + tty->print("B%d: ",b->_pre_order); 1.54 + for( j=1; j<b->num_preds(); j++ ) 1.55 + tty->print("B%d ", _phc._cfg.get_block_for_node(b->pred(j))->_pre_order); 1.56 + tty->print("-> "); 1.57 + for( j=0; j<b->_num_succs; j++ ) 1.58 + tty->print("B%d ",b->_succs[j]->_pre_order); 1.59 + tty->print(" IDom: B%d/#%d\n", b->_idom ? b->_idom->_pre_order : 0, b->_dom_depth); 1.60 + uint cnt = b->number_of_nodes(); 1.61 + for( j=0; j<cnt; j++ ) { 1.62 + Node *n = b->get_node(j); 1.63 + dump( n ); 1.64 + tty->print("\t%s\t",n->Name()); 1.65 + 1.66 + // Dump the inputs 1.67 + uint k; // Exit value of loop 1.68 + for( k=0; k<n->req(); k++ ) // For all required inputs 1.69 + if( n->in(k) ) dump( n->in(k) ); 1.70 + else tty->print("_ "); 1.71 + int any_prec = 0; 1.72 + for( ; k<n->len(); k++ ) // For all precedence inputs 1.73 + if( n->in(k) ) { 1.74 + if( !any_prec++ ) tty->print(" |"); 1.75 + dump( n->in(k) ); 1.76 + } 1.77 + 1.78 + // Dump node-specific info 1.79 + n->dump_spec(tty); 1.80 + tty->print("\n"); 1.81 + 1.82 + } 1.83 + tty->print("\n"); 1.84 + } 1.85 +} 1.86 +#endif 1.87 + 1.88 +// Combine the live ranges def'd by these 2 Nodes. N2 is an input to N1. 1.89 +void PhaseCoalesce::combine_these_two(Node *n1, Node *n2) { 1.90 + uint lr1 = _phc._lrg_map.find(n1); 1.91 + uint lr2 = _phc._lrg_map.find(n2); 1.92 + if( lr1 != lr2 && // Different live ranges already AND 1.93 + !_phc._ifg->test_edge_sq( lr1, lr2 ) ) { // Do not interfere 1.94 + LRG *lrg1 = &_phc.lrgs(lr1); 1.95 + LRG *lrg2 = &_phc.lrgs(lr2); 1.96 + // Not an oop->int cast; oop->oop, int->int, AND int->oop are OK. 1.97 + 1.98 + // Now, why is int->oop OK? We end up declaring a raw-pointer as an oop 1.99 + // and in general that's a bad thing. However, int->oop conversions only 1.100 + // happen at GC points, so the lifetime of the misclassified raw-pointer 1.101 + // is from the CheckCastPP (that converts it to an oop) backwards up 1.102 + // through a merge point and into the slow-path call, and around the 1.103 + // diamond up to the heap-top check and back down into the slow-path call. 1.104 + // The misclassified raw pointer is NOT live across the slow-path call, 1.105 + // and so does not appear in any GC info, so the fact that it is 1.106 + // misclassified is OK. 1.107 + 1.108 + if( (lrg1->_is_oop || !lrg2->_is_oop) && // not an oop->int cast AND 1.109 + // Compatible final mask 1.110 + lrg1->mask().overlap( lrg2->mask() ) ) { 1.111 + // Merge larger into smaller. 1.112 + if( lr1 > lr2 ) { 1.113 + uint tmp = lr1; lr1 = lr2; lr2 = tmp; 1.114 + Node *n = n1; n1 = n2; n2 = n; 1.115 + LRG *ltmp = lrg1; lrg1 = lrg2; lrg2 = ltmp; 1.116 + } 1.117 + // Union lr2 into lr1 1.118 + _phc.Union( n1, n2 ); 1.119 + if (lrg1->_maxfreq < lrg2->_maxfreq) 1.120 + lrg1->_maxfreq = lrg2->_maxfreq; 1.121 + // Merge in the IFG 1.122 + _phc._ifg->Union( lr1, lr2 ); 1.123 + // Combine register restrictions 1.124 + lrg1->AND(lrg2->mask()); 1.125 + } 1.126 + } 1.127 +} 1.128 + 1.129 +// Copy coalescing 1.130 +void PhaseCoalesce::coalesce_driver() { 1.131 + verify(); 1.132 + // Coalesce from high frequency to low 1.133 + for (uint i = 0; i < _phc._cfg.number_of_blocks(); i++) { 1.134 + coalesce(_phc._blks[i]); 1.135 + } 1.136 +} 1.137 + 1.138 +// I am inserting copies to come out of SSA form. In the general case, I am 1.139 +// doing a parallel renaming. I'm in the Named world now, so I can't do a 1.140 +// general parallel renaming. All the copies now use "names" (live-ranges) 1.141 +// to carry values instead of the explicit use-def chains. Suppose I need to 1.142 +// insert 2 copies into the same block. They copy L161->L128 and L128->L132. 1.143 +// If I insert them in the wrong order then L128 will get clobbered before it 1.144 +// can get used by the second copy. This cannot happen in the SSA model; 1.145 +// direct use-def chains get me the right value. It DOES happen in the named 1.146 +// model so I have to handle the reordering of copies. 1.147 +// 1.148 +// In general, I need to topo-sort the placed copies to avoid conflicts. 1.149 +// Its possible to have a closed cycle of copies (e.g., recirculating the same 1.150 +// values around a loop). In this case I need a temp to break the cycle. 1.151 +void PhaseAggressiveCoalesce::insert_copy_with_overlap( Block *b, Node *copy, uint dst_name, uint src_name ) { 1.152 + 1.153 + // Scan backwards for the locations of the last use of the dst_name. 1.154 + // I am about to clobber the dst_name, so the copy must be inserted 1.155 + // after the last use. Last use is really first-use on a backwards scan. 1.156 + uint i = b->end_idx()-1; 1.157 + while(1) { 1.158 + Node *n = b->get_node(i); 1.159 + // Check for end of virtual copies; this is also the end of the 1.160 + // parallel renaming effort. 1.161 + if (n->_idx < _unique) { 1.162 + break; 1.163 + } 1.164 + uint idx = n->is_Copy(); 1.165 + assert( idx || n->is_Con() || n->is_MachProj(), "Only copies during parallel renaming" ); 1.166 + if (idx && _phc._lrg_map.find(n->in(idx)) == dst_name) { 1.167 + break; 1.168 + } 1.169 + i--; 1.170 + } 1.171 + uint last_use_idx = i; 1.172 + 1.173 + // Also search for any kill of src_name that exits the block. 1.174 + // Since the copy uses src_name, I have to come before any kill. 1.175 + uint kill_src_idx = b->end_idx(); 1.176 + // There can be only 1 kill that exits any block and that is 1.177 + // the last kill. Thus it is the first kill on a backwards scan. 1.178 + i = b->end_idx()-1; 1.179 + while (1) { 1.180 + Node *n = b->get_node(i); 1.181 + // Check for end of virtual copies; this is also the end of the 1.182 + // parallel renaming effort. 1.183 + if (n->_idx < _unique) { 1.184 + break; 1.185 + } 1.186 + assert( n->is_Copy() || n->is_Con() || n->is_MachProj(), "Only copies during parallel renaming" ); 1.187 + if (_phc._lrg_map.find(n) == src_name) { 1.188 + kill_src_idx = i; 1.189 + break; 1.190 + } 1.191 + i--; 1.192 + } 1.193 + // Need a temp? Last use of dst comes after the kill of src? 1.194 + if (last_use_idx >= kill_src_idx) { 1.195 + // Need to break a cycle with a temp 1.196 + uint idx = copy->is_Copy(); 1.197 + Node *tmp = copy->clone(); 1.198 + uint max_lrg_id = _phc._lrg_map.max_lrg_id(); 1.199 + _phc.new_lrg(tmp, max_lrg_id); 1.200 + _phc._lrg_map.set_max_lrg_id(max_lrg_id + 1); 1.201 + 1.202 + // Insert new temp between copy and source 1.203 + tmp ->set_req(idx,copy->in(idx)); 1.204 + copy->set_req(idx,tmp); 1.205 + // Save source in temp early, before source is killed 1.206 + b->insert_node(tmp, kill_src_idx); 1.207 + _phc._cfg.map_node_to_block(tmp, b); 1.208 + last_use_idx++; 1.209 + } 1.210 + 1.211 + // Insert just after last use 1.212 + b->insert_node(copy, last_use_idx + 1); 1.213 +} 1.214 + 1.215 +void PhaseAggressiveCoalesce::insert_copies( Matcher &matcher ) { 1.216 + // We do LRGs compressing and fix a liveout data only here since the other 1.217 + // place in Split() is guarded by the assert which we never hit. 1.218 + _phc._lrg_map.compress_uf_map_for_nodes(); 1.219 + // Fix block's liveout data for compressed live ranges. 1.220 + for (uint lrg = 1; lrg < _phc._lrg_map.max_lrg_id(); lrg++) { 1.221 + uint compressed_lrg = _phc._lrg_map.find(lrg); 1.222 + if (lrg != compressed_lrg) { 1.223 + for (uint bidx = 0; bidx < _phc._cfg.number_of_blocks(); bidx++) { 1.224 + IndexSet *liveout = _phc._live->live(_phc._cfg.get_block(bidx)); 1.225 + if (liveout->member(lrg)) { 1.226 + liveout->remove(lrg); 1.227 + liveout->insert(compressed_lrg); 1.228 + } 1.229 + } 1.230 + } 1.231 + } 1.232 + 1.233 + // All new nodes added are actual copies to replace virtual copies. 1.234 + // Nodes with index less than '_unique' are original, non-virtual Nodes. 1.235 + _unique = C->unique(); 1.236 + 1.237 + for (uint i = 0; i < _phc._cfg.number_of_blocks(); i++) { 1.238 + C->check_node_count(NodeLimitFudgeFactor, "out of nodes in coalesce"); 1.239 + if (C->failing()) return; 1.240 + Block *b = _phc._cfg.get_block(i); 1.241 + uint cnt = b->num_preds(); // Number of inputs to the Phi 1.242 + 1.243 + for( uint l = 1; l<b->number_of_nodes(); l++ ) { 1.244 + Node *n = b->get_node(l); 1.245 + 1.246 + // Do not use removed-copies, use copied value instead 1.247 + uint ncnt = n->req(); 1.248 + for( uint k = 1; k<ncnt; k++ ) { 1.249 + Node *copy = n->in(k); 1.250 + uint cidx = copy->is_Copy(); 1.251 + if( cidx ) { 1.252 + Node *def = copy->in(cidx); 1.253 + if (_phc._lrg_map.find(copy) == _phc._lrg_map.find(def)) { 1.254 + n->set_req(k, def); 1.255 + } 1.256 + } 1.257 + } 1.258 + 1.259 + // Remove any explicit copies that get coalesced. 1.260 + uint cidx = n->is_Copy(); 1.261 + if( cidx ) { 1.262 + Node *def = n->in(cidx); 1.263 + if (_phc._lrg_map.find(n) == _phc._lrg_map.find(def)) { 1.264 + n->replace_by(def); 1.265 + n->set_req(cidx,NULL); 1.266 + b->remove_node(l); 1.267 + l--; 1.268 + continue; 1.269 + } 1.270 + } 1.271 + 1.272 + if (n->is_Phi()) { 1.273 + // Get the chosen name for the Phi 1.274 + uint phi_name = _phc._lrg_map.find(n); 1.275 + // Ignore the pre-allocated specials 1.276 + if (!phi_name) { 1.277 + continue; 1.278 + } 1.279 + // Check for mismatch inputs to Phi 1.280 + for (uint j = 1; j < cnt; j++) { 1.281 + Node *m = n->in(j); 1.282 + uint src_name = _phc._lrg_map.find(m); 1.283 + if (src_name != phi_name) { 1.284 + Block *pred = _phc._cfg.get_block_for_node(b->pred(j)); 1.285 + Node *copy; 1.286 + assert(!m->is_Con() || m->is_Mach(), "all Con must be Mach"); 1.287 + // Rematerialize constants instead of copying them 1.288 + if( m->is_Mach() && m->as_Mach()->is_Con() && 1.289 + m->as_Mach()->rematerialize() ) { 1.290 + copy = m->clone(); 1.291 + // Insert the copy in the predecessor basic block 1.292 + pred->add_inst(copy); 1.293 + // Copy any flags as well 1.294 + _phc.clone_projs(pred, pred->end_idx(), m, copy, _phc._lrg_map); 1.295 + } else { 1.296 + const RegMask *rm = C->matcher()->idealreg2spillmask[m->ideal_reg()]; 1.297 + copy = new (C) MachSpillCopyNode(m, *rm, *rm); 1.298 + // Find a good place to insert. Kinda tricky, use a subroutine 1.299 + insert_copy_with_overlap(pred,copy,phi_name,src_name); 1.300 + } 1.301 + // Insert the copy in the use-def chain 1.302 + n->set_req(j, copy); 1.303 + _phc._cfg.map_node_to_block(copy, pred); 1.304 + // Extend ("register allocate") the names array for the copy. 1.305 + _phc._lrg_map.extend(copy->_idx, phi_name); 1.306 + } // End of if Phi names do not match 1.307 + } // End of for all inputs to Phi 1.308 + } else { // End of if Phi 1.309 + 1.310 + // Now check for 2-address instructions 1.311 + uint idx; 1.312 + if( n->is_Mach() && (idx=n->as_Mach()->two_adr()) ) { 1.313 + // Get the chosen name for the Node 1.314 + uint name = _phc._lrg_map.find(n); 1.315 + assert (name, "no 2-address specials"); 1.316 + // Check for name mis-match on the 2-address input 1.317 + Node *m = n->in(idx); 1.318 + if (_phc._lrg_map.find(m) != name) { 1.319 + Node *copy; 1.320 + assert(!m->is_Con() || m->is_Mach(), "all Con must be Mach"); 1.321 + // At this point it is unsafe to extend live ranges (6550579). 1.322 + // Rematerialize only constants as we do for Phi above. 1.323 + if(m->is_Mach() && m->as_Mach()->is_Con() && 1.324 + m->as_Mach()->rematerialize()) { 1.325 + copy = m->clone(); 1.326 + // Insert the copy in the basic block, just before us 1.327 + b->insert_node(copy, l++); 1.328 + l += _phc.clone_projs(b, l, m, copy, _phc._lrg_map); 1.329 + } else { 1.330 + const RegMask *rm = C->matcher()->idealreg2spillmask[m->ideal_reg()]; 1.331 + copy = new (C) MachSpillCopyNode(m, *rm, *rm); 1.332 + // Insert the copy in the basic block, just before us 1.333 + b->insert_node(copy, l++); 1.334 + } 1.335 + // Insert the copy in the use-def chain 1.336 + n->set_req(idx, copy); 1.337 + // Extend ("register allocate") the names array for the copy. 1.338 + _phc._lrg_map.extend(copy->_idx, name); 1.339 + _phc._cfg.map_node_to_block(copy, b); 1.340 + } 1.341 + 1.342 + } // End of is two-adr 1.343 + 1.344 + // Insert a copy at a debug use for a lrg which has high frequency 1.345 + if (b->_freq < OPTO_DEBUG_SPLIT_FREQ || _phc._cfg.is_uncommon(b)) { 1.346 + // Walk the debug inputs to the node and check for lrg freq 1.347 + JVMState* jvms = n->jvms(); 1.348 + uint debug_start = jvms ? jvms->debug_start() : 999999; 1.349 + uint debug_end = jvms ? jvms->debug_end() : 999999; 1.350 + for(uint inpidx = debug_start; inpidx < debug_end; inpidx++) { 1.351 + // Do not split monitors; they are only needed for debug table 1.352 + // entries and need no code. 1.353 + if (jvms->is_monitor_use(inpidx)) { 1.354 + continue; 1.355 + } 1.356 + Node *inp = n->in(inpidx); 1.357 + uint nidx = _phc._lrg_map.live_range_id(inp); 1.358 + LRG &lrg = lrgs(nidx); 1.359 + 1.360 + // If this lrg has a high frequency use/def 1.361 + if( lrg._maxfreq >= _phc.high_frequency_lrg() ) { 1.362 + // If the live range is also live out of this block (like it 1.363 + // would be for a fast/slow idiom), the normal spill mechanism 1.364 + // does an excellent job. If it is not live out of this block 1.365 + // (like it would be for debug info to uncommon trap) splitting 1.366 + // the live range now allows a better allocation in the high 1.367 + // frequency blocks. 1.368 + // Build_IFG_virtual has converted the live sets to 1.369 + // live-IN info, not live-OUT info. 1.370 + uint k; 1.371 + for( k=0; k < b->_num_succs; k++ ) 1.372 + if( _phc._live->live(b->_succs[k])->member( nidx ) ) 1.373 + break; // Live in to some successor block? 1.374 + if( k < b->_num_succs ) 1.375 + continue; // Live out; do not pre-split 1.376 + // Split the lrg at this use 1.377 + const RegMask *rm = C->matcher()->idealreg2spillmask[inp->ideal_reg()]; 1.378 + Node *copy = new (C) MachSpillCopyNode( inp, *rm, *rm ); 1.379 + // Insert the copy in the use-def chain 1.380 + n->set_req(inpidx, copy ); 1.381 + // Insert the copy in the basic block, just before us 1.382 + b->insert_node(copy, l++); 1.383 + // Extend ("register allocate") the names array for the copy. 1.384 + uint max_lrg_id = _phc._lrg_map.max_lrg_id(); 1.385 + _phc.new_lrg(copy, max_lrg_id); 1.386 + _phc._lrg_map.set_max_lrg_id(max_lrg_id + 1); 1.387 + _phc._cfg.map_node_to_block(copy, b); 1.388 + //tty->print_cr("Split a debug use in Aggressive Coalesce"); 1.389 + } // End of if high frequency use/def 1.390 + } // End of for all debug inputs 1.391 + } // End of if low frequency safepoint 1.392 + 1.393 + } // End of if Phi 1.394 + 1.395 + } // End of for all instructions 1.396 + } // End of for all blocks 1.397 +} 1.398 + 1.399 + 1.400 +// Aggressive (but pessimistic) copy coalescing of a single block 1.401 + 1.402 +// The following coalesce pass represents a single round of aggressive 1.403 +// pessimistic coalesce. "Aggressive" means no attempt to preserve 1.404 +// colorability when coalescing. This occasionally means more spills, but 1.405 +// it also means fewer rounds of coalescing for better code - and that means 1.406 +// faster compiles. 1.407 + 1.408 +// "Pessimistic" means we do not hit the fixed point in one pass (and we are 1.409 +// reaching for the least fixed point to boot). This is typically solved 1.410 +// with a few more rounds of coalescing, but the compiler must run fast. We 1.411 +// could optimistically coalescing everything touching PhiNodes together 1.412 +// into one big live range, then check for self-interference. Everywhere 1.413 +// the live range interferes with self it would have to be split. Finding 1.414 +// the right split points can be done with some heuristics (based on 1.415 +// expected frequency of edges in the live range). In short, it's a real 1.416 +// research problem and the timeline is too short to allow such research. 1.417 +// Further thoughts: (1) build the LR in a pass, (2) find self-interference 1.418 +// in another pass, (3) per each self-conflict, split, (4) split by finding 1.419 +// the low-cost cut (min-cut) of the LR, (5) edges in the LR are weighted 1.420 +// according to the GCM algorithm (or just exec freq on CFG edges). 1.421 + 1.422 +void PhaseAggressiveCoalesce::coalesce( Block *b ) { 1.423 + // Copies are still "virtual" - meaning we have not made them explicitly 1.424 + // copies. Instead, Phi functions of successor blocks have mis-matched 1.425 + // live-ranges. If I fail to coalesce, I'll have to insert a copy to line 1.426 + // up the live-ranges. Check for Phis in successor blocks. 1.427 + uint i; 1.428 + for( i=0; i<b->_num_succs; i++ ) { 1.429 + Block *bs = b->_succs[i]; 1.430 + // Find index of 'b' in 'bs' predecessors 1.431 + uint j=1; 1.432 + while (_phc._cfg.get_block_for_node(bs->pred(j)) != b) { 1.433 + j++; 1.434 + } 1.435 + 1.436 + // Visit all the Phis in successor block 1.437 + for( uint k = 1; k<bs->number_of_nodes(); k++ ) { 1.438 + Node *n = bs->get_node(k); 1.439 + if( !n->is_Phi() ) break; 1.440 + combine_these_two( n, n->in(j) ); 1.441 + } 1.442 + } // End of for all successor blocks 1.443 + 1.444 + 1.445 + // Check _this_ block for 2-address instructions and copies. 1.446 + uint cnt = b->end_idx(); 1.447 + for( i = 1; i<cnt; i++ ) { 1.448 + Node *n = b->get_node(i); 1.449 + uint idx; 1.450 + // 2-address instructions have a virtual Copy matching their input 1.451 + // to their output 1.452 + if (n->is_Mach() && (idx = n->as_Mach()->two_adr())) { 1.453 + MachNode *mach = n->as_Mach(); 1.454 + combine_these_two(mach, mach->in(idx)); 1.455 + } 1.456 + } // End of for all instructions in block 1.457 +} 1.458 + 1.459 +PhaseConservativeCoalesce::PhaseConservativeCoalesce(PhaseChaitin &chaitin) : PhaseCoalesce(chaitin) { 1.460 + _ulr.initialize(_phc._lrg_map.max_lrg_id()); 1.461 +} 1.462 + 1.463 +void PhaseConservativeCoalesce::verify() { 1.464 +#ifdef ASSERT 1.465 + _phc.set_was_low(); 1.466 +#endif 1.467 +} 1.468 + 1.469 +void PhaseConservativeCoalesce::union_helper( Node *lr1_node, Node *lr2_node, uint lr1, uint lr2, Node *src_def, Node *dst_copy, Node *src_copy, Block *b, uint bindex ) { 1.470 + // Join live ranges. Merge larger into smaller. Union lr2 into lr1 in the 1.471 + // union-find tree 1.472 + _phc.Union( lr1_node, lr2_node ); 1.473 + 1.474 + // Single-def live range ONLY if both live ranges are single-def. 1.475 + // If both are single def, then src_def powers one live range 1.476 + // and def_copy powers the other. After merging, src_def powers 1.477 + // the combined live range. 1.478 + lrgs(lr1)._def = (lrgs(lr1).is_multidef() || 1.479 + lrgs(lr2).is_multidef() ) 1.480 + ? NodeSentinel : src_def; 1.481 + lrgs(lr2)._def = NULL; // No def for lrg 2 1.482 + lrgs(lr2).Clear(); // Force empty mask for LRG 2 1.483 + //lrgs(lr2)._size = 0; // Live-range 2 goes dead 1.484 + lrgs(lr1)._is_oop |= lrgs(lr2)._is_oop; 1.485 + lrgs(lr2)._is_oop = 0; // In particular, not an oop for GC info 1.486 + 1.487 + if (lrgs(lr1)._maxfreq < lrgs(lr2)._maxfreq) 1.488 + lrgs(lr1)._maxfreq = lrgs(lr2)._maxfreq; 1.489 + 1.490 + // Copy original value instead. Intermediate copies go dead, and 1.491 + // the dst_copy becomes useless. 1.492 + int didx = dst_copy->is_Copy(); 1.493 + dst_copy->set_req( didx, src_def ); 1.494 + // Add copy to free list 1.495 + // _phc.free_spillcopy(b->_nodes[bindex]); 1.496 + assert( b->get_node(bindex) == dst_copy, "" ); 1.497 + dst_copy->replace_by( dst_copy->in(didx) ); 1.498 + dst_copy->set_req( didx, NULL); 1.499 + b->remove_node(bindex); 1.500 + if( bindex < b->_ihrp_index ) b->_ihrp_index--; 1.501 + if( bindex < b->_fhrp_index ) b->_fhrp_index--; 1.502 + 1.503 + // Stretched lr1; add it to liveness of intermediate blocks 1.504 + Block *b2 = _phc._cfg.get_block_for_node(src_copy); 1.505 + while( b != b2 ) { 1.506 + b = _phc._cfg.get_block_for_node(b->pred(1)); 1.507 + _phc._live->live(b)->insert(lr1); 1.508 + } 1.509 +} 1.510 + 1.511 +// Factored code from copy_copy that computes extra interferences from 1.512 +// lengthening a live range by double-coalescing. 1.513 +uint PhaseConservativeCoalesce::compute_separating_interferences(Node *dst_copy, Node *src_copy, Block *b, uint bindex, RegMask &rm, uint reg_degree, uint rm_size, uint lr1, uint lr2 ) { 1.514 + 1.515 + assert(!lrgs(lr1)._fat_proj, "cannot coalesce fat_proj"); 1.516 + assert(!lrgs(lr2)._fat_proj, "cannot coalesce fat_proj"); 1.517 + Node *prev_copy = dst_copy->in(dst_copy->is_Copy()); 1.518 + Block *b2 = b; 1.519 + uint bindex2 = bindex; 1.520 + while( 1 ) { 1.521 + // Find previous instruction 1.522 + bindex2--; // Chain backwards 1 instruction 1.523 + while( bindex2 == 0 ) { // At block start, find prior block 1.524 + assert( b2->num_preds() == 2, "cannot double coalesce across c-flow" ); 1.525 + b2 = _phc._cfg.get_block_for_node(b2->pred(1)); 1.526 + bindex2 = b2->end_idx()-1; 1.527 + } 1.528 + // Get prior instruction 1.529 + assert(bindex2 < b2->number_of_nodes(), "index out of bounds"); 1.530 + Node *x = b2->get_node(bindex2); 1.531 + if( x == prev_copy ) { // Previous copy in copy chain? 1.532 + if( prev_copy == src_copy)// Found end of chain and all interferences 1.533 + break; // So break out of loop 1.534 + // Else work back one in copy chain 1.535 + prev_copy = prev_copy->in(prev_copy->is_Copy()); 1.536 + } else { // Else collect interferences 1.537 + uint lidx = _phc._lrg_map.find(x); 1.538 + // Found another def of live-range being stretched? 1.539 + if(lidx == lr1) { 1.540 + return max_juint; 1.541 + } 1.542 + if(lidx == lr2) { 1.543 + return max_juint; 1.544 + } 1.545 + 1.546 + // If we attempt to coalesce across a bound def 1.547 + if( lrgs(lidx).is_bound() ) { 1.548 + // Do not let the coalesced LRG expect to get the bound color 1.549 + rm.SUBTRACT( lrgs(lidx).mask() ); 1.550 + // Recompute rm_size 1.551 + rm_size = rm.Size(); 1.552 + //if( rm._flags ) rm_size += 1000000; 1.553 + if( reg_degree >= rm_size ) return max_juint; 1.554 + } 1.555 + if( rm.overlap(lrgs(lidx).mask()) ) { 1.556 + // Insert lidx into union LRG; returns TRUE if actually inserted 1.557 + if( _ulr.insert(lidx) ) { 1.558 + // Infinite-stack neighbors do not alter colorability, as they 1.559 + // can always color to some other color. 1.560 + if( !lrgs(lidx).mask().is_AllStack() ) { 1.561 + // If this coalesce will make any new neighbor uncolorable, 1.562 + // do not coalesce. 1.563 + if( lrgs(lidx).just_lo_degree() ) 1.564 + return max_juint; 1.565 + // Bump our degree 1.566 + if( ++reg_degree >= rm_size ) 1.567 + return max_juint; 1.568 + } // End of if not infinite-stack neighbor 1.569 + } // End of if actually inserted 1.570 + } // End of if live range overlaps 1.571 + } // End of else collect interferences for 1 node 1.572 + } // End of while forever, scan back for interferences 1.573 + return reg_degree; 1.574 +} 1.575 + 1.576 +void PhaseConservativeCoalesce::update_ifg(uint lr1, uint lr2, IndexSet *n_lr1, IndexSet *n_lr2) { 1.577 + // Some original neighbors of lr1 might have gone away 1.578 + // because the constrained register mask prevented them. 1.579 + // Remove lr1 from such neighbors. 1.580 + IndexSetIterator one(n_lr1); 1.581 + uint neighbor; 1.582 + LRG &lrg1 = lrgs(lr1); 1.583 + while ((neighbor = one.next()) != 0) 1.584 + if( !_ulr.member(neighbor) ) 1.585 + if( _phc._ifg->neighbors(neighbor)->remove(lr1) ) 1.586 + lrgs(neighbor).inc_degree( -lrg1.compute_degree(lrgs(neighbor)) ); 1.587 + 1.588 + 1.589 + // lr2 is now called (coalesced into) lr1. 1.590 + // Remove lr2 from the IFG. 1.591 + IndexSetIterator two(n_lr2); 1.592 + LRG &lrg2 = lrgs(lr2); 1.593 + while ((neighbor = two.next()) != 0) 1.594 + if( _phc._ifg->neighbors(neighbor)->remove(lr2) ) 1.595 + lrgs(neighbor).inc_degree( -lrg2.compute_degree(lrgs(neighbor)) ); 1.596 + 1.597 + // Some neighbors of intermediate copies now interfere with the 1.598 + // combined live range. 1.599 + IndexSetIterator three(&_ulr); 1.600 + while ((neighbor = three.next()) != 0) 1.601 + if( _phc._ifg->neighbors(neighbor)->insert(lr1) ) 1.602 + lrgs(neighbor).inc_degree( lrg1.compute_degree(lrgs(neighbor)) ); 1.603 +} 1.604 + 1.605 +static void record_bias( const PhaseIFG *ifg, int lr1, int lr2 ) { 1.606 + // Tag copy bias here 1.607 + if( !ifg->lrgs(lr1)._copy_bias ) 1.608 + ifg->lrgs(lr1)._copy_bias = lr2; 1.609 + if( !ifg->lrgs(lr2)._copy_bias ) 1.610 + ifg->lrgs(lr2)._copy_bias = lr1; 1.611 +} 1.612 + 1.613 +// See if I can coalesce a series of multiple copies together. I need the 1.614 +// final dest copy and the original src copy. They can be the same Node. 1.615 +// Compute the compatible register masks. 1.616 +bool PhaseConservativeCoalesce::copy_copy(Node *dst_copy, Node *src_copy, Block *b, uint bindex) { 1.617 + 1.618 + if (!dst_copy->is_SpillCopy()) { 1.619 + return false; 1.620 + } 1.621 + if (!src_copy->is_SpillCopy()) { 1.622 + return false; 1.623 + } 1.624 + Node *src_def = src_copy->in(src_copy->is_Copy()); 1.625 + uint lr1 = _phc._lrg_map.find(dst_copy); 1.626 + uint lr2 = _phc._lrg_map.find(src_def); 1.627 + 1.628 + // Same live ranges already? 1.629 + if (lr1 == lr2) { 1.630 + return false; 1.631 + } 1.632 + 1.633 + // Interfere? 1.634 + if (_phc._ifg->test_edge_sq(lr1, lr2)) { 1.635 + return false; 1.636 + } 1.637 + 1.638 + // Not an oop->int cast; oop->oop, int->int, AND int->oop are OK. 1.639 + if (!lrgs(lr1)._is_oop && lrgs(lr2)._is_oop) { // not an oop->int cast 1.640 + return false; 1.641 + } 1.642 + 1.643 + // Coalescing between an aligned live range and a mis-aligned live range? 1.644 + // No, no! Alignment changes how we count degree. 1.645 + if (lrgs(lr1)._fat_proj != lrgs(lr2)._fat_proj) { 1.646 + return false; 1.647 + } 1.648 + 1.649 + // Sort; use smaller live-range number 1.650 + Node *lr1_node = dst_copy; 1.651 + Node *lr2_node = src_def; 1.652 + if (lr1 > lr2) { 1.653 + uint tmp = lr1; lr1 = lr2; lr2 = tmp; 1.654 + lr1_node = src_def; lr2_node = dst_copy; 1.655 + } 1.656 + 1.657 + // Check for compatibility of the 2 live ranges by 1.658 + // intersecting their allowed register sets. 1.659 + RegMask rm = lrgs(lr1).mask(); 1.660 + rm.AND(lrgs(lr2).mask()); 1.661 + // Number of bits free 1.662 + uint rm_size = rm.Size(); 1.663 + 1.664 + if (UseFPUForSpilling && rm.is_AllStack() ) { 1.665 + // Don't coalesce when frequency difference is large 1.666 + Block *dst_b = _phc._cfg.get_block_for_node(dst_copy); 1.667 + Block *src_def_b = _phc._cfg.get_block_for_node(src_def); 1.668 + if (src_def_b->_freq > 10*dst_b->_freq ) 1.669 + return false; 1.670 + } 1.671 + 1.672 + // If we can use any stack slot, then effective size is infinite 1.673 + if( rm.is_AllStack() ) rm_size += 1000000; 1.674 + // Incompatible masks, no way to coalesce 1.675 + if( rm_size == 0 ) return false; 1.676 + 1.677 + // Another early bail-out test is when we are double-coalescing and the 1.678 + // 2 copies are separated by some control flow. 1.679 + if( dst_copy != src_copy ) { 1.680 + Block *src_b = _phc._cfg.get_block_for_node(src_copy); 1.681 + Block *b2 = b; 1.682 + while( b2 != src_b ) { 1.683 + if( b2->num_preds() > 2 ){// Found merge-point 1.684 + _phc._lost_opp_cflow_coalesce++; 1.685 + // extra record_bias commented out because Chris believes it is not 1.686 + // productive. Since we can record only 1 bias, we want to choose one 1.687 + // that stands a chance of working and this one probably does not. 1.688 + //record_bias( _phc._lrgs, lr1, lr2 ); 1.689 + return false; // To hard to find all interferences 1.690 + } 1.691 + b2 = _phc._cfg.get_block_for_node(b2->pred(1)); 1.692 + } 1.693 + } 1.694 + 1.695 + // Union the two interference sets together into '_ulr' 1.696 + uint reg_degree = _ulr.lrg_union( lr1, lr2, rm_size, _phc._ifg, rm ); 1.697 + 1.698 + if( reg_degree >= rm_size ) { 1.699 + record_bias( _phc._ifg, lr1, lr2 ); 1.700 + return false; 1.701 + } 1.702 + 1.703 + // Now I need to compute all the interferences between dst_copy and 1.704 + // src_copy. I'm not willing visit the entire interference graph, so 1.705 + // I limit my search to things in dst_copy's block or in a straight 1.706 + // line of previous blocks. I give up at merge points or when I get 1.707 + // more interferences than my degree. I can stop when I find src_copy. 1.708 + if( dst_copy != src_copy ) { 1.709 + reg_degree = compute_separating_interferences(dst_copy, src_copy, b, bindex, rm, rm_size, reg_degree, lr1, lr2 ); 1.710 + if( reg_degree == max_juint ) { 1.711 + record_bias( _phc._ifg, lr1, lr2 ); 1.712 + return false; 1.713 + } 1.714 + } // End of if dst_copy & src_copy are different 1.715 + 1.716 + 1.717 + // ---- THE COMBINED LRG IS COLORABLE ---- 1.718 + 1.719 + // YEAH - Now coalesce this copy away 1.720 + assert( lrgs(lr1).num_regs() == lrgs(lr2).num_regs(), "" ); 1.721 + 1.722 + IndexSet *n_lr1 = _phc._ifg->neighbors(lr1); 1.723 + IndexSet *n_lr2 = _phc._ifg->neighbors(lr2); 1.724 + 1.725 + // Update the interference graph 1.726 + update_ifg(lr1, lr2, n_lr1, n_lr2); 1.727 + 1.728 + _ulr.remove(lr1); 1.729 + 1.730 + // Uncomment the following code to trace Coalescing in great detail. 1.731 + // 1.732 + //if (false) { 1.733 + // tty->cr(); 1.734 + // tty->print_cr("#######################################"); 1.735 + // tty->print_cr("union %d and %d", lr1, lr2); 1.736 + // n_lr1->dump(); 1.737 + // n_lr2->dump(); 1.738 + // tty->print_cr("resulting set is"); 1.739 + // _ulr.dump(); 1.740 + //} 1.741 + 1.742 + // Replace n_lr1 with the new combined live range. _ulr will use 1.743 + // n_lr1's old memory on the next iteration. n_lr2 is cleared to 1.744 + // send its internal memory to the free list. 1.745 + _ulr.swap(n_lr1); 1.746 + _ulr.clear(); 1.747 + n_lr2->clear(); 1.748 + 1.749 + lrgs(lr1).set_degree( _phc._ifg->effective_degree(lr1) ); 1.750 + lrgs(lr2).set_degree( 0 ); 1.751 + 1.752 + // Join live ranges. Merge larger into smaller. Union lr2 into lr1 in the 1.753 + // union-find tree 1.754 + union_helper( lr1_node, lr2_node, lr1, lr2, src_def, dst_copy, src_copy, b, bindex ); 1.755 + // Combine register restrictions 1.756 + lrgs(lr1).set_mask(rm); 1.757 + lrgs(lr1).compute_set_mask_size(); 1.758 + lrgs(lr1)._cost += lrgs(lr2)._cost; 1.759 + lrgs(lr1)._area += lrgs(lr2)._area; 1.760 + 1.761 + // While its uncommon to successfully coalesce live ranges that started out 1.762 + // being not-lo-degree, it can happen. In any case the combined coalesced 1.763 + // live range better Simplify nicely. 1.764 + lrgs(lr1)._was_lo = 1; 1.765 + 1.766 + // kinda expensive to do all the time 1.767 + //tty->print_cr("warning: slow verify happening"); 1.768 + //_phc._ifg->verify( &_phc ); 1.769 + return true; 1.770 +} 1.771 + 1.772 +// Conservative (but pessimistic) copy coalescing of a single block 1.773 +void PhaseConservativeCoalesce::coalesce( Block *b ) { 1.774 + // Bail out on infrequent blocks 1.775 + if (_phc._cfg.is_uncommon(b)) { 1.776 + return; 1.777 + } 1.778 + // Check this block for copies. 1.779 + for( uint i = 1; i<b->end_idx(); i++ ) { 1.780 + // Check for actual copies on inputs. Coalesce a copy into its 1.781 + // input if use and copy's input are compatible. 1.782 + Node *copy1 = b->get_node(i); 1.783 + uint idx1 = copy1->is_Copy(); 1.784 + if( !idx1 ) continue; // Not a copy 1.785 + 1.786 + if( copy_copy(copy1,copy1,b,i) ) { 1.787 + i--; // Retry, same location in block 1.788 + PhaseChaitin::_conserv_coalesce++; // Collect stats on success 1.789 + continue; 1.790 + } 1.791 + } 1.792 +}