chaitin.cpp revision 295
0N/A/*
0N/A * Copyright 2000-2008 Sun Microsystems, Inc. All Rights Reserved.
0N/A * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
0N/A *
0N/A * This code is free software; you can redistribute it and/or modify it
0N/A * under the terms of the GNU General Public License version 2 only, as
0N/A * published by the Free Software Foundation.
0N/A *
0N/A * This code is distributed in the hope that it will be useful, but WITHOUT
0N/A * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
0N/A * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
0N/A * version 2 for more details (a copy is included in the LICENSE file that
0N/A * accompanied this code).
0N/A *
0N/A * You should have received a copy of the GNU General Public License version
0N/A * 2 along with this work; if not, write to the Free Software Foundation,
0N/A * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
0N/A *
0N/A * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
0N/A * CA 95054 USA or visit www.sun.com if you need additional information or
0N/A * have any questions.
0N/A *
0N/A */
0N/A
0N/A#include "incls/_precompiled.incl"
0N/A#include "incls/_chaitin.cpp.incl"
0N/A
0N/A//=============================================================================
0N/A
0N/A#ifndef PRODUCT
0N/Avoid LRG::dump( ) const {
0N/A ttyLocker ttyl;
0N/A tty->print("%d ",num_regs());
0N/A _mask.dump();
0N/A if( _msize_valid ) {
0N/A if( mask_size() == compute_mask_size() ) tty->print(", #%d ",_mask_size);
233N/A else tty->print(", #!!!_%d_vs_%d ",_mask_size,_mask.Size());
233N/A } else {
233N/A tty->print(", #?(%d) ",_mask.Size());
233N/A }
233N/A
233N/A tty->print("EffDeg: ");
233N/A if( _degree_valid ) tty->print( "%d ", _eff_degree );
233N/A else tty->print("? ");
233N/A
233N/A if( is_multidef() ) {
233N/A tty->print("MultiDef ");
233N/A if (_defs != NULL) {
233N/A tty->print("(");
233N/A for (int i = 0; i < _defs->length(); i++) {
233N/A tty->print("N%d ", _defs->at(i)->_idx);
233N/A }
0N/A tty->print(") ");
0N/A }
0N/A }
0N/A else if( _def == 0 ) tty->print("Dead ");
0N/A else tty->print("Def: N%d ",_def->_idx);
0N/A
0N/A tty->print("Cost:%4.2g Area:%4.2g Score:%4.2g ",_cost,_area, score());
0N/A // Flags
0N/A if( _is_oop ) tty->print("Oop ");
0N/A if( _is_float ) tty->print("Float ");
233N/A if( _was_spilled1 ) tty->print("Spilled ");
233N/A if( _was_spilled2 ) tty->print("Spilled2 ");
0N/A if( _direct_conflict ) tty->print("Direct_conflict ");
114N/A if( _fat_proj ) tty->print("Fat ");
114N/A if( _was_lo ) tty->print("Lo ");
114N/A if( _has_copy ) tty->print("Copy ");
114N/A if( _at_risk ) tty->print("Risk ");
114N/A
114N/A if( _must_spill ) tty->print("Must_spill ");
0N/A if( _is_bound ) tty->print("Bound ");
0N/A if( _msize_valid ) {
0N/A if( _degree_valid && lo_degree() ) tty->print("Trivial ");
0N/A }
0N/A
0N/A tty->cr();
0N/A}
563N/A#endif
563N/A
563N/A//------------------------------score------------------------------------------
563N/A// Compute score from cost and area. Low score is best to spill.
0N/Astatic double raw_score( double cost, double area ) {
422N/A return cost - (area*RegisterCostAreaRatio) * 1.52588e-5;
422N/A}
422N/A
0N/Adouble LRG::score() const {
422N/A // Scale _area by RegisterCostAreaRatio/64K then subtract from cost.
0N/A // Bigger area lowers score, encourages spilling this live range.
422N/A // Bigger cost raise score, prevents spilling this live range.
0N/A // (Note: 1/65536 is the magic constant below; I dont trust the C optimizer
0N/A // to turn a divide by a constant into a multiply by the reciprical).
0N/A double score = raw_score( _cost, _area);
0N/A
0N/A // Account for area. Basically, LRGs covering large areas are better
0N/A // to spill because more other LRGs get freed up.
0N/A if( _area == 0.0 ) // No area? Then no progress to spill
0N/A return 1e35;
233N/A
233N/A if( _was_spilled2 ) // If spilled once before, we are unlikely
233N/A return score + 1e30; // to make progress again.
233N/A
0N/A if( _cost >= _area*3.0 ) // Tiny area relative to cost
0N/A return score + 1e17; // Probably no progress to spill
0N/A
0N/A if( (_cost+_cost) >= _area*3.0 ) // Small area relative to cost
0N/A return score + 1e10; // Likely no progress to spill
0N/A
110N/A return score;
110N/A}
121N/A
0N/A//------------------------------LRG_List---------------------------------------
0N/ALRG_List::LRG_List( uint max ) : _cnt(max), _max(max), _lidxs(NEW_RESOURCE_ARRAY(uint,max)) {
0N/A memset( _lidxs, 0, sizeof(uint)*max );
0N/A}
0N/A
0N/Avoid LRG_List::extend( uint nidx, uint lidx ) {
0N/A _nesting.check();
0N/A if( nidx >= _max ) {
0N/A uint size = 16;
0N/A while( size <= nidx ) size <<=1;
0N/A _lidxs = REALLOC_RESOURCE_ARRAY( uint, _lidxs, _max, size );
0N/A _max = size;
0N/A }
0N/A while( _cnt <= nidx )
121N/A _lidxs[_cnt++] = 0;
0N/A _lidxs[nidx] = lidx;
0N/A}
0N/A
0N/A#define NUMBUCKS 3
0N/A
0N/A//------------------------------Chaitin----------------------------------------
0N/APhaseChaitin::PhaseChaitin(uint unique, PhaseCFG &cfg, Matcher &matcher)
511N/A : PhaseRegAlloc(unique, cfg, matcher,
0N/A#ifndef PRODUCT
0N/A print_chaitin_statistics
0N/A#else
0N/A NULL
0N/A#endif
0N/A ),
0N/A _names(unique), _uf_map(unique),
0N/A _maxlrg(0), _live(0),
0N/A _spilled_once(Thread::current()->resource_area()),
0N/A _spilled_twice(Thread::current()->resource_area()),
0N/A _lo_degree(0), _lo_stk_degree(0), _hi_degree(0), _simplified(0),
0N/A _oldphi(unique)
0N/A#ifndef PRODUCT
0N/A , _trace_spilling(TraceSpilling || C->method_has_option("TraceSpilling"))
185N/A#endif
0N/A{
0N/A NOT_PRODUCT( Compile::TracePhase t3("ctorChaitin", &_t_ctorChaitin, TimeCompiler); )
0N/A uint i,j;
0N/A // Build a list of basic blocks, sorted by frequency
0N/A _blks = NEW_RESOURCE_ARRAY( Block *, _cfg._num_blocks );
0N/A // Experiment with sorting strategies to speed compilation
0N/A double cutoff = BLOCK_FREQUENCY(1.0); // Cutoff for high frequency bucket
0N/A Block **buckets[NUMBUCKS]; // Array of buckets
0N/A uint buckcnt[NUMBUCKS]; // Array of bucket counters
0N/A double buckval[NUMBUCKS]; // Array of bucket value cutoffs
233N/A for( i = 0; i < NUMBUCKS; i++ ) {
233N/A buckets[i] = NEW_RESOURCE_ARRAY( Block *, _cfg._num_blocks );
233N/A buckcnt[i] = 0;
0N/A // Bump by three orders of magnitude each time
0N/A cutoff *= 0.001;
0N/A buckval[i] = cutoff;
121N/A for( j = 0; j < _cfg._num_blocks; j++ ) {
121N/A buckets[i][j] = NULL;
0N/A }
0N/A }
0N/A // Sort blocks into buckets
121N/A for( i = 0; i < _cfg._num_blocks; i++ ) {
0N/A for( j = 0; j < NUMBUCKS; j++ ) {
0N/A if( (j == NUMBUCKS-1) || (_cfg._blocks[i]->_freq > buckval[j]) ) {
0N/A // Assign block to end of list for appropriate bucket
0N/A buckets[j][buckcnt[j]++] = _cfg._blocks[i];
0N/A break; // kick out of inner loop
0N/A }
0N/A }
0N/A }
0N/A // Dump buckets into final block array
0N/A uint blkcnt = 0;
0N/A for( i = 0; i < NUMBUCKS; i++ ) {
0N/A for( j = 0; j < buckcnt[i]; j++ ) {
0N/A _blks[blkcnt++] = buckets[i][j];
0N/A }
0N/A }
0N/A
0N/A assert(blkcnt == _cfg._num_blocks, "Block array not totally filled");
0N/A}
0N/A
0N/Avoid PhaseChaitin::Register_Allocate() {
0N/A
0N/A // Above the OLD FP (and in registers) are the incoming arguments. Stack
0N/A // slots in this area are called "arg_slots". Above the NEW FP (and in
0N/A // registers) is the outgoing argument area; above that is the spill/temp
0N/A // area. These are all "frame_slots". Arg_slots start at the zero
0N/A // stack_slots and count up to the known arg_size. Frame_slots start at
0N/A // the stack_slot #arg_size and go up. After allocation I map stack
0N/A // slots to actual offsets. Stack-slots in the arg_slot area are biased
0N/A // by the frame_size; stack-slots in the frame_slot area are biased by 0.
0N/A
0N/A _trip_cnt = 0;
0N/A _alternate = 0;
0N/A _matcher._allocation_started = true;
0N/A
0N/A ResourceArea live_arena; // Arena for liveness & IFG info
109N/A ResourceMark rm(&live_arena);
109N/A
109N/A // Need live-ness for the IFG; need the IFG for coalescing. If the
109N/A // liveness is JUST for coalescing, then I can get some mileage by renaming
109N/A // all copy-related live ranges low and then using the max copy-related
109N/A // live range as a cut-off for LIVE and the IFG. In other words, I can
109N/A // build a subset of LIVE and IFG just for copies.
109N/A PhaseLive live(_cfg,_names,&live_arena);
110N/A
109N/A // Need IFG for coalescing and coloring
109N/A PhaseIFG ifg( &live_arena );
109N/A _ifg = &ifg;
109N/A
109N/A if (C->unique() > _names.Size()) _names.extend(C->unique()-1, 0);
109N/A
0N/A // Come out of SSA world to the Named world. Assign (virtual) registers to
0N/A // Nodes. Use the same register for all inputs and the output of PhiNodes
0N/A // - effectively ending SSA form. This requires either coalescing live
0N/A // ranges or inserting copies. For the moment, we insert "virtual copies"
0N/A // - we pretend there is a copy prior to each Phi in predecessor blocks.
0N/A // We will attempt to coalesce such "virtual copies" before we manifest
0N/A // them for real.
0N/A de_ssa();
0N/A
0N/A {
0N/A NOT_PRODUCT( Compile::TracePhase t3("computeLive", &_t_computeLive, TimeCompiler); )
563N/A _live = NULL; // Mark live as being not available
0N/A rm.reset_to_mark(); // Reclaim working storage
0N/A IndexSet::reset_memory(C, &live_arena);
0N/A ifg.init(_maxlrg); // Empty IFG
0N/A gather_lrg_masks( false ); // Collect LRG masks
0N/A live.compute( _maxlrg ); // Compute liveness
563N/A _live = &live; // Mark LIVE as being available
563N/A }
563N/A
563N/A // Base pointers are currently "used" by instructions which define new
563N/A // derived pointers. This makes base pointers live up to the where the
563N/A // derived pointer is made, but not beyond. Really, they need to be live
0N/A // across any GC point where the derived value is live. So this code looks
0N/A // at all the GC points, and "stretches" the live range of any base pointer
0N/A // to the GC point.
0N/A if( stretch_base_pointer_live_ranges(&live_arena) ) {
0N/A NOT_PRODUCT( Compile::TracePhase t3("computeLive (sbplr)", &_t_computeLive, TimeCompiler); )
0N/A // Since some live range stretched, I need to recompute live
0N/A _live = NULL;
0N/A rm.reset_to_mark(); // Reclaim working storage
0N/A IndexSet::reset_memory(C, &live_arena);
0N/A ifg.init(_maxlrg);
114N/A gather_lrg_masks( false );
114N/A live.compute( _maxlrg );
114N/A _live = &live;
114N/A }
114N/A // Create the interference graph using virtual copies
114N/A build_ifg_virtual( ); // Include stack slots this time
114N/A
114N/A // Aggressive (but pessimistic) copy coalescing.
114N/A // This pass works on virtual copies. Any virtual copies which are not
114N/A // coalesced get manifested as actual copies
114N/A {
114N/A // The IFG is/was triangular. I am 'squaring it up' so Union can run
0N/A // faster. Union requires a 'for all' operation which is slow on the
0N/A // triangular adjacency matrix (quick reminder: the IFG is 'sparse' -
0N/A // meaning I can visit all the Nodes neighbors less than a Node in time
0N/A // O(# of neighbors), but I have to visit all the Nodes greater than a
121N/A // given Node and search them for an instance, i.e., time O(#MaxLRG)).
563N/A _ifg->SquareUp();
121N/A
121N/A PhaseAggressiveCoalesce coalesce( *this );
0N/A coalesce.coalesce_driver( );
0N/A // Insert un-coalesced copies. Visit all Phis. Where inputs to a Phi do
121N/A // not match the Phi itself, insert a copy.
121N/A coalesce.insert_copies(_matcher);
0N/A }
0N/A
0N/A // After aggressive coalesce, attempt a first cut at coloring.
0N/A // To color, we need the IFG and for that we need LIVE.
0N/A {
0N/A NOT_PRODUCT( Compile::TracePhase t3("computeLive", &_t_computeLive, TimeCompiler); )
0N/A _live = NULL;
0N/A rm.reset_to_mark(); // Reclaim working storage
0N/A IndexSet::reset_memory(C, &live_arena);
0N/A ifg.init(_maxlrg);
0N/A gather_lrg_masks( true );
0N/A live.compute( _maxlrg );
0N/A _live = &live;
0N/A }
0N/A
0N/A // Build physical interference graph
0N/A uint must_spill = 0;
0N/A must_spill = build_ifg_physical( &live_arena );
563N/A // If we have a guaranteed spill, might as well spill now
563N/A if( must_spill ) {
563N/A if( !_maxlrg ) return;
563N/A // Bail out if unique gets too large (ie - unique > MaxNodeLimit)
563N/A C->check_node_count(10*must_spill, "out of nodes before split");
563N/A if (C->failing()) return;
563N/A _maxlrg = Split( _maxlrg ); // Split spilling LRG everywhere
563N/A // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor)
563N/A // or we failed to split
563N/A C->check_node_count(2*NodeLimitFudgeFactor, "out of nodes after physical split");
563N/A if (C->failing()) return;
0N/A
563N/A#ifdef ASSERT
563N/A if( VerifyOpto ) {
563N/A _cfg.verify();
563N/A verify_base_ptrs(&live_arena);
563N/A }
563N/A#endif
563N/A NOT_PRODUCT( C->verify_graph_edges(); )
563N/A
563N/A compact(); // Compact LRGs; return new lower max lrg
563N/A
563N/A {
563N/A NOT_PRODUCT( Compile::TracePhase t3("computeLive", &_t_computeLive, TimeCompiler); )
563N/A _live = NULL;
563N/A rm.reset_to_mark(); // Reclaim working storage
563N/A IndexSet::reset_memory(C, &live_arena);
563N/A ifg.init(_maxlrg); // Build a new interference graph
0N/A gather_lrg_masks( true ); // Collect intersect mask
563N/A live.compute( _maxlrg ); // Compute LIVE
0N/A _live = &live;
0N/A }
0N/A build_ifg_physical( &live_arena );
0N/A _ifg->SquareUp();
0N/A _ifg->Compute_Effective_Degree();
0N/A // Only do conservative coalescing if requested
0N/A if( OptoCoalesce ) {
0N/A // Conservative (and pessimistic) copy coalescing of those spills
0N/A PhaseConservativeCoalesce coalesce( *this );
0N/A // If max live ranges greater than cutoff, don't color the stack.
0N/A // This cutoff can be larger than below since it is only done once.
0N/A coalesce.coalesce_driver( );
0N/A }
0N/A compress_uf_map_for_nodes();
0N/A
0N/A#ifdef ASSERT
0N/A if( VerifyOpto ) _ifg->verify(this);
0N/A#endif
0N/A } else {
0N/A ifg.SquareUp();
0N/A ifg.Compute_Effective_Degree();
0N/A#ifdef ASSERT
0N/A set_was_low();
0N/A#endif
563N/A }
563N/A
563N/A // Prepare for Simplify & Select
0N/A cache_lrg_info(); // Count degree of LRGs
0N/A
0N/A // Simplify the InterFerence Graph by removing LRGs of low degree.
0N/A // LRGs of low degree are trivially colorable.
0N/A Simplify();
0N/A
0N/A // Select colors by re-inserting LRGs back into the IFG in reverse order.
0N/A // Return whether or not something spills.
0N/A uint spills = Select( );
0N/A
0N/A // If we spill, split and recycle the entire thing
0N/A while( spills ) {
0N/A if( _trip_cnt++ > 24 ) {
0N/A DEBUG_ONLY( dump_for_spill_split_recycle(); )
0N/A if( _trip_cnt > 27 ) {
0N/A C->record_method_not_compilable("failed spill-split-recycle sanity check");
0N/A return;
0N/A }
563N/A }
563N/A
563N/A if( !_maxlrg ) return;
563N/A _maxlrg = Split( _maxlrg ); // Split spilling LRG everywhere
563N/A // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor)
563N/A C->check_node_count(2*NodeLimitFudgeFactor, "out of nodes after split");
563N/A if (C->failing()) return;
563N/A#ifdef ASSERT
563N/A if( VerifyOpto ) {
563N/A _cfg.verify();
563N/A verify_base_ptrs(&live_arena);
563N/A }
563N/A#endif
563N/A
563N/A compact(); // Compact LRGs; return new lower max lrg
563N/A
563N/A // Nuke the live-ness and interference graph and LiveRanGe info
563N/A {
563N/A NOT_PRODUCT( Compile::TracePhase t3("computeLive", &_t_computeLive, TimeCompiler); )
563N/A _live = NULL;
563N/A rm.reset_to_mark(); // Reclaim working storage
563N/A IndexSet::reset_memory(C, &live_arena);
563N/A ifg.init(_maxlrg);
563N/A
563N/A // Create LiveRanGe array.
563N/A // Intersect register masks for all USEs and DEFs
563N/A gather_lrg_masks( true );
563N/A live.compute( _maxlrg );
563N/A _live = &live;
563N/A }
563N/A must_spill = build_ifg_physical( &live_arena );
0N/A _ifg->SquareUp();
0N/A _ifg->Compute_Effective_Degree();
0N/A
0N/A // Only do conservative coalescing if requested
0N/A if( OptoCoalesce ) {
0N/A // Conservative (and pessimistic) copy coalescing
121N/A PhaseConservativeCoalesce coalesce( *this );
0N/A // Check for few live ranges determines how aggressive coalesce is.
0N/A coalesce.coalesce_driver( );
0N/A }
0N/A compress_uf_map_for_nodes();
0N/A#ifdef ASSERT
0N/A if( VerifyOpto ) _ifg->verify(this);
0N/A#endif
0N/A cache_lrg_info(); // Count degree of LRGs
0N/A
0N/A // Simplify the InterFerence Graph by removing LRGs of low degree.
0N/A // LRGs of low degree are trivially colorable.
0N/A Simplify();
0N/A
0N/A // Select colors by re-inserting LRGs back into the IFG in reverse order.
107N/A // Return whether or not something spills.
0N/A spills = Select( );
0N/A }
0N/A
29N/A // Count number of Simplify-Select trips per coloring success.
0N/A _allocator_attempts += _trip_cnt + 1;
0N/A _allocator_successes += 1;
0N/A
0N/A // Peephole remove copies
0N/A post_allocate_copy_removal();
0N/A
0N/A // max_reg is past the largest *register* used.
563N/A // Convert that to a frame_slot number.
563N/A if( _max_reg <= _matcher._new_SP )
563N/A _framesize = C->out_preserve_stack_slots();
563N/A else _framesize = _max_reg -_matcher._new_SP;
0N/A assert((int)(_matcher._new_SP+_framesize) >= (int)_matcher._out_arg_limit, "framesize must be large enough");
0N/A
0N/A // This frame must preserve the required fp alignment
0N/A const int stack_alignment_in_words = Matcher::stack_alignment_in_slots();
547N/A if (stack_alignment_in_words > 0)
547N/A _framesize = round_to(_framesize, Matcher::stack_alignment_in_bytes());
547N/A assert( _framesize >= 0 && _framesize <= 1000000, "sanity check" );
547N/A#ifndef PRODUCT
0N/A _total_framesize += _framesize;
0N/A if( (int)_framesize > _max_framesize )
0N/A _max_framesize = _framesize;
0N/A#endif
0N/A
0N/A // Convert CISC spills
0N/A fixup_spills();
0N/A
0N/A // Log regalloc results
0N/A CompileLog* log = Compile::current()->log();
0N/A if (log != NULL) {
0N/A log->elem("regalloc attempts='%d' success='%d'", _trip_cnt, !C->failing());
0N/A }
107N/A
107N/A if (C->failing()) return;
107N/A
107N/A NOT_PRODUCT( C->verify_graph_edges(); )
29N/A
29N/A // Move important info out of the live_arena to longer lasting storage.
29N/A alloc_node_regs(_names.Size());
29N/A for( uint i=0; i < _names.Size(); i++ ) {
29N/A if( _names[i] ) { // Live range associated with Node?
29N/A LRG &lrg = lrgs( _names[i] );
29N/A if( lrg.num_regs() == 1 ) {
29N/A _node_regs[i].set1( lrg.reg() );
547N/A } else { // Must be a register-pair
547N/A if( !lrg._fat_proj ) { // Must be aligned adjacent register pair
547N/A // Live ranges record the highest register in their mask.
547N/A // We want the low register for the AD file writer's convenience.
84N/A _node_regs[i].set2( OptoReg::add(lrg.reg(),-1) );
84N/A } else { // Misaligned; extract 2 bits
485N/A OptoReg::Name hi = lrg.reg(); // Get hi register
485N/A lrg.Remove(hi); // Yank from mask
485N/A int lo = lrg.mask().find_first_elem(); // Find lo
485N/A _node_regs[i].set_pair( hi, lo );
485N/A }
0N/A }
0N/A if( lrg._is_oop ) _node_oops.set(i);
0N/A } else {
0N/A _node_regs[i].set_bad();
0N/A }
0N/A }
0N/A
0N/A // Done!
0N/A _live = NULL;
0N/A _ifg = NULL;
485N/A C->set_indexSet_arena(NULL); // ResourceArea is at end of scope
499N/A}
0N/A
499N/A//------------------------------de_ssa-----------------------------------------
499N/Avoid PhaseChaitin::de_ssa() {
499N/A // Set initial Names for all Nodes. Most Nodes get the virtual register
499N/A // number. A few get the ZERO live range number. These do not
499N/A // get allocated, but instead rely on correct scheduling to ensure that
499N/A // only one instance is simultaneously live at a time.
499N/A uint lr_counter = 1;
499N/A for( uint i = 0; i < _cfg._num_blocks; i++ ) {
0N/A Block *b = _cfg._blocks[i];
0N/A uint cnt = b->_nodes.size();
0N/A
0N/A // Handle all the normal Nodes in the block
0N/A for( uint j = 0; j < cnt; j++ ) {
0N/A Node *n = b->_nodes[j];
0N/A // Pre-color to the zero live range, or pick virtual register
0N/A const RegMask &rm = n->out_RegMask();
0N/A _names.map( n->_idx, rm.is_NotEmpty() ? lr_counter++ : 0 );
0N/A }
0N/A }
0N/A // Reset the Union-Find mapping to be identity
0N/A reset_uf_map(lr_counter);
0N/A}
0N/A
0N/A
0N/A//------------------------------gather_lrg_masks-------------------------------
0N/A// Gather LiveRanGe information, including register masks. Modification of
0N/A// cisc spillable in_RegMasks should not be done before AggressiveCoalesce.
0N/Avoid PhaseChaitin::gather_lrg_masks( bool after_aggressive ) {
0N/A
0N/A // Nail down the frame pointer live range
0N/A uint fp_lrg = n2lidx(_cfg._root->in(1)->in(TypeFunc::FramePtr));
0N/A lrgs(fp_lrg)._cost += 1e12; // Cost is infinite
0N/A
0N/A // For all blocks
0N/A for( uint i = 0; i < _cfg._num_blocks; i++ ) {
0N/A Block *b = _cfg._blocks[i];
0N/A
0N/A // For all instructions
0N/A for( uint j = 1; j < b->_nodes.size(); j++ ) {
0N/A Node *n = b->_nodes[j];
0N/A uint input_edge_start =1; // Skip control most nodes
0N/A if( n->is_Mach() ) input_edge_start = n->as_Mach()->oper_input_base();
0N/A uint idx = n->is_Copy();
0N/A
0N/A // Get virtual register number, same as LiveRanGe index
0N/A uint vreg = n2lidx(n);
0N/A LRG &lrg = lrgs(vreg);
0N/A if( vreg ) { // No vreg means un-allocable (e.g. memory)
0N/A
0N/A // Collect has-copy bit
0N/A if( idx ) {
0N/A lrg._has_copy = 1;
0N/A uint clidx = n2lidx(n->in(idx));
0N/A LRG &copy_src = lrgs(clidx);
0N/A copy_src._has_copy = 1;
0N/A }
0N/A
0N/A // Check for float-vs-int live range (used in register-pressure
0N/A // calculations)
0N/A const Type *n_type = n->bottom_type();
0N/A if( n_type->is_floatingpoint() )
0N/A lrg._is_float = 1;
0N/A
0N/A // Check for twice prior spilling. Once prior spilling might have
0N/A // spilled 'soft', 2nd prior spill should have spilled 'hard' and
0N/A // further spilling is unlikely to make progress.
0N/A if( _spilled_once.test(n->_idx) ) {
0N/A lrg._was_spilled1 = 1;
0N/A if( _spilled_twice.test(n->_idx) )
0N/A lrg._was_spilled2 = 1;
0N/A }
0N/A
0N/A#ifndef PRODUCT
0N/A if (trace_spilling() && lrg._def != NULL) {
0N/A // collect defs for MultiDef printing
0N/A if (lrg._defs == NULL) {
0N/A lrg._defs = new (_ifg->_arena) GrowableArray<Node*>();
0N/A lrg._defs->append(lrg._def);
0N/A }
0N/A lrg._defs->append(n);
0N/A }
0N/A#endif
0N/A
0N/A // Check for a single def LRG; these can spill nicely
0N/A // via rematerialization. Flag as NULL for no def found
0N/A // yet, or 'n' for single def or -1 for many defs.
0N/A lrg._def = lrg._def ? NodeSentinel : n;
0N/A
1N/A // Limit result register mask to acceptable registers
0N/A const RegMask &rm = n->out_RegMask();
0N/A lrg.AND( rm );
0N/A // Check for bound register masks
0N/A const RegMask &lrgmask = lrg.mask();
0N/A if( lrgmask.is_bound1() || lrgmask.is_bound2() )
0N/A lrg._is_bound = 1;
0N/A
0N/A // Check for maximum frequency value
0N/A if( lrg._maxfreq < b->_freq )
0N/A lrg._maxfreq = b->_freq;
0N/A
0N/A int ireg = n->ideal_reg();
0N/A assert( !n->bottom_type()->isa_oop_ptr() || ireg == Op_RegP,
0N/A "oops must be in Op_RegP's" );
0N/A // Check for oop-iness, or long/double
0N/A // Check for multi-kill projection
0N/A switch( ireg ) {
0N/A case MachProjNode::fat_proj:
0N/A // Fat projections have size equal to number of registers killed
0N/A lrg.set_num_regs(rm.Size());
0N/A lrg.set_reg_pressure(lrg.num_regs());
0N/A lrg._fat_proj = 1;
0N/A lrg._is_bound = 1;
0N/A break;
0N/A case Op_RegP:
0N/A#ifdef _LP64
0N/A lrg.set_num_regs(2); // Size is 2 stack words
121N/A#else
121N/A lrg.set_num_regs(1); // Size is 1 stack word
0N/A#endif
0N/A // Register pressure is tracked relative to the maximum values
0N/A // suggested for that platform, INTPRESSURE and FLOATPRESSURE,
0N/A // and relative to other types which compete for the same regs.
0N/A //
0N/A // The following table contains suggested values based on the
0N/A // architectures as defined in each .ad file.
0N/A // INTPRESSURE and FLOATPRESSURE may be tuned differently for
0N/A // compile-speed or performance.
0N/A // Note1:
0N/A // SPARC and SPARCV9 reg_pressures are at 2 instead of 1
0N/A // since .ad registers are defined as high and low halves.
0N/A // These reg_pressure values remain compatible with the code
0N/A // in is_high_pressure() which relates get_invalid_mask_size(),
0N/A // Block::_reg_pressure and INTPRESSURE, FLOATPRESSURE.
21N/A // Note2:
0N/A // SPARC -d32 has 24 registers available for integral values,
0N/A // but only 10 of these are safe for 64-bit longs.
0N/A // Using set_reg_pressure(2) for both int and long means
0N/A // the allocator will believe it can fit 26 longs into
0N/A // registers. Using 2 for longs and 1 for ints means the
0N/A // allocator will attempt to put 52 integers into registers.
0N/A // The settings below limit this problem to methods with
0N/A // many long values which are being run on 32-bit SPARC.
0N/A //
0N/A // ------------------- reg_pressure --------------------
0N/A // Each entry is reg_pressure_per_value,number_of_regs
0N/A // RegL RegI RegFlags RegF RegD INTPRESSURE FLOATPRESSURE
0N/A // IA32 2 1 1 1 1 6 6
0N/A // IA64 1 1 1 1 1 50 41
0N/A // SPARC 2 2 2 2 2 48 (24) 52 (26)
0N/A // SPARCV9 2 2 2 2 2 48 (24) 52 (26)
0N/A // AMD64 1 1 1 1 1 14 15
0N/A // -----------------------------------------------------
0N/A#if defined(SPARC)
0N/A lrg.set_reg_pressure(2); // use for v9 as well
0N/A#else
0N/A lrg.set_reg_pressure(1); // normally one value per register
0N/A#endif
0N/A if( n_type->isa_oop_ptr() ) {
0N/A lrg._is_oop = 1;
0N/A }
0N/A break;
0N/A case Op_RegL: // Check for long or double
0N/A case Op_RegD:
0N/A lrg.set_num_regs(2);
0N/A // Define platform specific register pressure
0N/A#ifdef SPARC
0N/A lrg.set_reg_pressure(2);
0N/A#elif defined(IA32)
0N/A if( ireg == Op_RegL ) {
0N/A lrg.set_reg_pressure(2);
233N/A } else {
233N/A lrg.set_reg_pressure(1);
233N/A }
233N/A#else
233N/A lrg.set_reg_pressure(1); // normally one value per register
233N/A#endif
233N/A // If this def of a double forces a mis-aligned double,
233N/A // flag as '_fat_proj' - really flag as allowing misalignment
233N/A // AND changes how we count interferences. A mis-aligned
318N/A // double can interfere with TWO aligned pairs, or effectively
318N/A // FOUR registers!
318N/A if( rm.is_misaligned_Pair() ) {
318N/A lrg._fat_proj = 1;
318N/A lrg._is_bound = 1;
233N/A }
233N/A break;
0N/A case Op_RegF:
0N/A case Op_RegI:
21N/A case Op_RegN:
21N/A case Op_RegFlags:
0N/A case 0: // not an ideal register
185N/A lrg.set_num_regs(1);
185N/A#ifdef SPARC
185N/A lrg.set_reg_pressure(2);
185N/A#else
318N/A lrg.set_reg_pressure(1);
318N/A#endif
318N/A break;
318N/A default:
318N/A ShouldNotReachHere();
185N/A }
185N/A }
185N/A
318N/A // Now do the same for inputs
185N/A uint cnt = n->req();
185N/A // Setup for CISC SPILLING
185N/A uint inp = (uint)AdlcVMDeps::Not_cisc_spillable;
185N/A if( UseCISCSpill && after_aggressive ) {
185N/A inp = n->cisc_operand();
21N/A if( inp != (uint)AdlcVMDeps::Not_cisc_spillable )
233N/A // Convert operand number to edge index number
121N/A inp = n->as_Mach()->operand_index(inp);
121N/A }
121N/A // Prepare register mask for each input
121N/A for( uint k = input_edge_start; k < cnt; k++ ) {
121N/A uint vreg = n2lidx(n->in(k));
121N/A if( !vreg ) continue;
121N/A
121N/A // If this instruction is CISC Spillable, add the flags
121N/A // bit to its appropriate input
233N/A if( UseCISCSpill && after_aggressive && inp == k ) {
233N/A#ifndef PRODUCT
233N/A if( TraceCISCSpill ) {
233N/A tty->print(" use_cisc_RegMask: ");
233N/A n->dump();
233N/A }
318N/A#endif
318N/A n->as_Mach()->use_cisc_RegMask();
318N/A }
318N/A
318N/A LRG &lrg = lrgs(vreg);
318N/A // // Testing for floating point code shape
121N/A // Node *test = n->in(k);
121N/A // if( test->is_Mach() ) {
0N/A // MachNode *m = test->as_Mach();
0N/A // int op = m->ideal_Opcode();
0N/A // if (n->is_Call() && (op == Op_AddF || op == Op_MulF) ) {
0N/A // int zzz = 1;
0N/A // }
0N/A // }
0N/A
0N/A // Limit result register mask to acceptable registers.
0N/A // Do not limit registers from uncommon uses before
0N/A // AggressiveCoalesce. This effectively pre-virtual-splits
185N/A // around uncommon uses of common defs.
161N/A const RegMask &rm = n->in_RegMask(k);
161N/A if( !after_aggressive &&
161N/A _cfg._bbs[n->in(k)->_idx]->_freq > 1000*b->_freq ) {
161N/A // Since we are BEFORE aggressive coalesce, leave the register
161N/A // mask untrimmed by the call. This encourages more coalescing.
503N/A // Later, AFTER aggressive, this live range will have to spill
503N/A // but the spiller handles slow-path calls very nicely.
503N/A } else {
503N/A lrg.AND( rm );
503N/A }
503N/A // Check for bound register masks
503N/A const RegMask &lrgmask = lrg.mask();
503N/A if( lrgmask.is_bound1() || lrgmask.is_bound2() )
185N/A lrg._is_bound = 1;
0N/A // If this use of a double forces a mis-aligned double,
161N/A // flag as '_fat_proj' - really flag as allowing misalignment
0N/A // AND changes how we count interferences. A mis-aligned
0N/A // double can interfere with TWO aligned pairs, or effectively
151N/A // FOUR registers!
151N/A if( lrg.num_regs() == 2 && !lrg._fat_proj && rm.is_misaligned_Pair() ) {
0N/A lrg._fat_proj = 1;
0N/A lrg._is_bound = 1;
42N/A }
0N/A // if the LRG is an unaligned pair, we will have to spill
151N/A // so clear the LRG's register mask if it is not already spilled
0N/A if ( !n->is_SpillCopy() &&
0N/A (lrg._def == NULL || lrg.is_multidef() || !lrg._def->is_SpillCopy()) &&
0N/A lrgmask.is_misaligned_Pair()) {
0N/A lrg.Clear();
0N/A }
0N/A
0N/A // Check for maximum frequency value
422N/A if( lrg._maxfreq < b->_freq )
42N/A lrg._maxfreq = b->_freq;
151N/A
503N/A } // End for all allocated inputs
371N/A } // end for all instructions
371N/A } // end for all blocks
233N/A
0N/A // Final per-liverange setup
121N/A for( uint i2=0; i2<_maxlrg; i2++ ) {
422N/A LRG &lrg = lrgs(i2);
42N/A if( lrg.num_regs() == 2 && !lrg._fat_proj )
95N/A lrg.ClearToPairs();
422N/A lrg.compute_set_mask_size();
422N/A if( lrg.not_free() ) { // Handle case where we lose from the start
422N/A lrg.set_reg(OptoReg::Name(LRG::SPILL_REG));
422N/A lrg._direct_conflict = 1;
422N/A }
42N/A lrg.set_degree(0); // no neighbors in IFG yet
422N/A }
422N/A}
42N/A
42N/A//------------------------------set_was_low------------------------------------
121N/A// Set the was-lo-degree bit. Conservative coalescing should not change the
0N/A// colorability of the graph. If any live range was of low-degree before
0N/A// coalescing, it should Simplify. This call sets the was-lo-degree bit.
318N/A// The bit is checked in Simplify.
233N/Avoid PhaseChaitin::set_was_low() {
233N/A#ifdef ASSERT
318N/A for( uint i = 1; i < _maxlrg; i++ ) {
318N/A int size = lrgs(i).num_regs();
318N/A uint old_was_lo = lrgs(i)._was_lo;
318N/A lrgs(i)._was_lo = 0;
318N/A if( lrgs(i).lo_degree() ) {
318N/A lrgs(i)._was_lo = 1; // Trivially of low degree
318N/A } else { // Else check the Brigg's assertion
318N/A // Brigg's observation is that the lo-degree neighbors of a
318N/A // hi-degree live range will not interfere with the color choices
318N/A // of said hi-degree live range. The Simplify reverse-stack-coloring
233N/A // order takes care of the details. Hence you do not have to count
21N/A // low-degree neighbors when determining if this guy colors.
21N/A int briggs_degree = 0;
21N/A IndexSet *s = _ifg->neighbors(i);
0N/A IndexSetIterator elements(s);
0N/A uint lidx;
0N/A while((lidx = elements.next()) != 0) {
0N/A if( !lrgs(lidx).lo_degree() )
0N/A briggs_degree += MAX2(size,lrgs(lidx).num_regs());
0N/A }
0N/A if( briggs_degree < lrgs(i).degrees_of_freedom() )
0N/A lrgs(i)._was_lo = 1; // Low degree via the briggs assertion
0N/A }
0N/A assert(old_was_lo <= lrgs(i)._was_lo, "_was_lo may not decrease");
121N/A }
121N/A#endif
422N/A}
423N/A
423N/A#define REGISTER_CONSTRAINED 16
51N/A
233N/A//------------------------------cache_lrg_info---------------------------------
423N/A// Compute cost/area ratio, in case we spill. Build the lo-degree list.
233N/Avoid PhaseChaitin::cache_lrg_info( ) {
233N/A
233N/A for( uint i = 1; i < _maxlrg; i++ ) {
233N/A LRG &lrg = lrgs(i);
233N/A
233N/A // Check for being of low degree: means we can be trivially colored.
233N/A // Low degree, dead or must-spill guys just get to simplify right away
233N/A if( lrg.lo_degree() ||
233N/A !lrg.alive() ||
233N/A lrg._must_spill ) {
233N/A // Split low degree list into those guys that must get a
233N/A // register and those that can go to register or stack.
233N/A // The idea is LRGs that can go register or stack color first when
233N/A // they have a good chance of getting a register. The register-only
233N/A // lo-degree live ranges always get a register.
233N/A OptoReg::Name hi_reg = lrg.mask().find_last_elem();
233N/A if( OptoReg::is_stack(hi_reg)) { // Can go to stack?
233N/A lrg._next = _lo_stk_degree;
233N/A _lo_stk_degree = i;
233N/A } else {
233N/A lrg._next = _lo_degree;
233N/A _lo_degree = i;
233N/A }
233N/A } else { // Else high degree
233N/A lrgs(_hi_degree)._prev = i;
233N/A lrg._next = _hi_degree;
233N/A lrg._prev = 0;
233N/A _hi_degree = i;
233N/A }
51N/A }
0N/A}
//------------------------------Pre-Simplify-----------------------------------
// Simplify the IFG by removing LRGs of low degree that have NO copies
void PhaseChaitin::Pre_Simplify( ) {
// Warm up the lo-degree no-copy list
int lo_no_copy = 0;
for( uint i = 1; i < _maxlrg; i++ ) {
if( (lrgs(i).lo_degree() && !lrgs(i)._has_copy) ||
!lrgs(i).alive() ||
lrgs(i)._must_spill ) {
lrgs(i)._next = lo_no_copy;
lo_no_copy = i;
}
}
while( lo_no_copy ) {
uint lo = lo_no_copy;
lo_no_copy = lrgs(lo)._next;
int size = lrgs(lo).num_regs();
// Put the simplified guy on the simplified list.
lrgs(lo)._next = _simplified;
_simplified = lo;
// Yank this guy from the IFG.
IndexSet *adj = _ifg->remove_node( lo );
// If any neighbors' degrees fall below their number of
// allowed registers, then put that neighbor on the low degree
// list. Note that 'degree' can only fall and 'numregs' is
// unchanged by this action. Thus the two are equal at most once,
// so LRGs hit the lo-degree worklists at most once.
IndexSetIterator elements(adj);
uint neighbor;
while ((neighbor = elements.next()) != 0) {
LRG *n = &lrgs(neighbor);
assert( _ifg->effective_degree(neighbor) == n->degree(), "" );
// Check for just becoming of-low-degree
if( n->just_lo_degree() && !n->_has_copy ) {
assert(!(*_ifg->_yanked)[neighbor],"Cannot move to lo degree twice");
// Put on lo-degree list
n->_next = lo_no_copy;
lo_no_copy = neighbor;
}
}
} // End of while lo-degree no_copy worklist not empty
// No more lo-degree no-copy live ranges to simplify
}
//------------------------------Simplify---------------------------------------
// Simplify the IFG by removing LRGs of low degree.
void PhaseChaitin::Simplify( ) {
while( 1 ) { // Repeat till simplified it all
// May want to explore simplifying lo_degree before _lo_stk_degree.
// This might result in more spills coloring into registers during
// Select().
while( _lo_degree || _lo_stk_degree ) {
// If possible, pull from lo_stk first
uint lo;
if( _lo_degree ) {
lo = _lo_degree;
_lo_degree = lrgs(lo)._next;
} else {
lo = _lo_stk_degree;
_lo_stk_degree = lrgs(lo)._next;
}
// Put the simplified guy on the simplified list.
lrgs(lo)._next = _simplified;
_simplified = lo;
// If this guy is "at risk" then mark his current neighbors
if( lrgs(lo)._at_risk ) {
IndexSetIterator elements(_ifg->neighbors(lo));
uint datum;
while ((datum = elements.next()) != 0) {
lrgs(datum)._risk_bias = lo;
}
}
// Yank this guy from the IFG.
IndexSet *adj = _ifg->remove_node( lo );
// If any neighbors' degrees fall below their number of
// allowed registers, then put that neighbor on the low degree
// list. Note that 'degree' can only fall and 'numregs' is
// unchanged by this action. Thus the two are equal at most once,
// so LRGs hit the lo-degree worklist at most once.
IndexSetIterator elements(adj);
uint neighbor;
while ((neighbor = elements.next()) != 0) {
LRG *n = &lrgs(neighbor);
#ifdef ASSERT
if( VerifyOpto ) {
assert( _ifg->effective_degree(neighbor) == n->degree(), "" );
}
#endif
// Check for just becoming of-low-degree just counting registers.
// _must_spill live ranges are already on the low degree list.
if( n->just_lo_degree() && !n->_must_spill ) {
assert(!(*_ifg->_yanked)[neighbor],"Cannot move to lo degree twice");
// Pull from hi-degree list
uint prev = n->_prev;
uint next = n->_next;
if( prev ) lrgs(prev)._next = next;
else _hi_degree = next;
lrgs(next)._prev = prev;
n->_next = _lo_degree;
_lo_degree = neighbor;
}
}
} // End of while lo-degree/lo_stk_degree worklist not empty
// Check for got everything: is hi-degree list empty?
if( !_hi_degree ) break;
// Time to pick a potential spill guy
uint lo_score = _hi_degree;
double score = lrgs(lo_score).score();
double area = lrgs(lo_score)._area;
// Find cheapest guy
debug_only( int lo_no_simplify=0; );
for( uint i = _hi_degree; i; i = lrgs(i)._next ) {
assert( !(*_ifg->_yanked)[i], "" );
// It's just vaguely possible to move hi-degree to lo-degree without
// going through a just-lo-degree stage: If you remove a double from
// a float live range it's degree will drop by 2 and you can skip the
// just-lo-degree stage. It's very rare (shows up after 5000+ methods
// in -Xcomp of Java2Demo). So just choose this guy to simplify next.
if( lrgs(i).lo_degree() ) {
lo_score = i;
break;
}
debug_only( if( lrgs(i)._was_lo ) lo_no_simplify=i; );
double iscore = lrgs(i).score();
double iarea = lrgs(i)._area;
// Compare cost/area of i vs cost/area of lo_score. Smaller cost/area
// wins. Ties happen because all live ranges in question have spilled
// a few times before and the spill-score adds a huge number which
// washes out the low order bits. We are choosing the lesser of 2
// evils; in this case pick largest area to spill.
if( iscore < score ||
(iscore == score && iarea > area && lrgs(lo_score)._was_spilled2) ) {
lo_score = i;
score = iscore;
area = iarea;
}
}
LRG *lo_lrg = &lrgs(lo_score);
// The live range we choose for spilling is either hi-degree, or very
// rarely it can be low-degree. If we choose a hi-degree live range
// there better not be any lo-degree choices.
assert( lo_lrg->lo_degree() || !lo_no_simplify, "Live range was lo-degree before coalesce; should simplify" );
// Pull from hi-degree list
uint prev = lo_lrg->_prev;
uint next = lo_lrg->_next;
if( prev ) lrgs(prev)._next = next;
else _hi_degree = next;
lrgs(next)._prev = prev;
// Jam him on the lo-degree list, despite his high degree.
// Maybe he'll get a color, and maybe he'll spill.
// Only Select() will know.
lrgs(lo_score)._at_risk = true;
_lo_degree = lo_score;
lo_lrg->_next = 0;
} // End of while not simplified everything
}
//------------------------------bias_color-------------------------------------
// Choose a color using the biasing heuristic
OptoReg::Name PhaseChaitin::bias_color( LRG &lrg, int chunk ) {
// Check for "at_risk" LRG's
uint risk_lrg = Find(lrg._risk_bias);
if( risk_lrg != 0 ) {
// Walk the colored neighbors of the "at_risk" candidate
// Choose a color which is both legal and already taken by a neighbor
// of the "at_risk" candidate in order to improve the chances of the
// "at_risk" candidate of coloring
IndexSetIterator elements(_ifg->neighbors(risk_lrg));
uint datum;
while ((datum = elements.next()) != 0) {
OptoReg::Name reg = lrgs(datum).reg();
// If this LRG's register is legal for us, choose it
if( reg >= chunk && reg < chunk + RegMask::CHUNK_SIZE &&
lrg.mask().Member(OptoReg::add(reg,-chunk)) &&
(lrg.num_regs()==1 || // either size 1
(reg&1) == 1) ) // or aligned (adjacent reg is available since we already cleared-to-pairs)
return reg;
}
}
uint copy_lrg = Find(lrg._copy_bias);
if( copy_lrg != 0 ) {
// If he has a color,
if( !(*(_ifg->_yanked))[copy_lrg] ) {
OptoReg::Name reg = lrgs(copy_lrg).reg();
// And it is legal for you,
if( reg >= chunk && reg < chunk + RegMask::CHUNK_SIZE &&
lrg.mask().Member(OptoReg::add(reg,-chunk)) &&
(lrg.num_regs()==1 || // either size 1
(reg&1) == 1) ) // or aligned (adjacent reg is available since we already cleared-to-pairs)
return reg;
} else if( chunk == 0 ) {
// Choose a color which is legal for him
RegMask tempmask = lrg.mask();
tempmask.AND(lrgs(copy_lrg).mask());
OptoReg::Name reg;
if( lrg.num_regs() == 1 ) {
reg = tempmask.find_first_elem();
} else {
tempmask.ClearToPairs();
reg = tempmask.find_first_pair();
}
if( OptoReg::is_valid(reg) )
return reg;
}
}
// If no bias info exists, just go with the register selection ordering
if( lrg.num_regs() == 2 ) {
// Find an aligned pair
return OptoReg::add(lrg.mask().find_first_pair(),chunk);
}
// CNC - Fun hack. Alternate 1st and 2nd selection. Enables post-allocate
// copy removal to remove many more copies, by preventing a just-assigned
// register from being repeatedly assigned.
OptoReg::Name reg = lrg.mask().find_first_elem();
if( (++_alternate & 1) && OptoReg::is_valid(reg) ) {
// This 'Remove; find; Insert' idiom is an expensive way to find the
// SECOND element in the mask.
lrg.Remove(reg);
OptoReg::Name reg2 = lrg.mask().find_first_elem();
lrg.Insert(reg);
if( OptoReg::is_reg(reg2))
reg = reg2;
}
return OptoReg::add( reg, chunk );
}
//------------------------------choose_color-----------------------------------
// Choose a color in the current chunk
OptoReg::Name PhaseChaitin::choose_color( LRG &lrg, int chunk ) {
assert( C->in_preserve_stack_slots() == 0 || chunk != 0 || lrg._is_bound || lrg.mask().is_bound1() || !lrg.mask().Member(OptoReg::Name(_matcher._old_SP-1)), "must not allocate stack0 (inside preserve area)");
assert(C->out_preserve_stack_slots() == 0 || chunk != 0 || lrg._is_bound || lrg.mask().is_bound1() || !lrg.mask().Member(OptoReg::Name(_matcher._old_SP+0)), "must not allocate stack0 (inside preserve area)");
if( lrg.num_regs() == 1 || // Common Case
!lrg._fat_proj ) // Aligned+adjacent pairs ok
// Use a heuristic to "bias" the color choice
return bias_color(lrg, chunk);
assert( lrg.num_regs() >= 2, "dead live ranges do not color" );
// Fat-proj case or misaligned double argument.
assert(lrg.compute_mask_size() == lrg.num_regs() ||
lrg.num_regs() == 2,"fat projs exactly color" );
assert( !chunk, "always color in 1st chunk" );
// Return the highest element in the set.
return lrg.mask().find_last_elem();
}
//------------------------------Select-----------------------------------------
// Select colors by re-inserting LRGs back into the IFG. LRGs are re-inserted
// in reverse order of removal. As long as nothing of hi-degree was yanked,
// everything going back is guaranteed a color. Select that color. If some
// hi-degree LRG cannot get a color then we record that we must spill.
uint PhaseChaitin::Select( ) {
uint spill_reg = LRG::SPILL_REG;
_max_reg = OptoReg::Name(0); // Past max register used
while( _simplified ) {
// Pull next LRG from the simplified list - in reverse order of removal
uint lidx = _simplified;
LRG *lrg = &lrgs(lidx);
_simplified = lrg->_next;
#ifndef PRODUCT
if (trace_spilling()) {
ttyLocker ttyl;
tty->print_cr("L%d selecting degree %d degrees_of_freedom %d", lidx, lrg->degree(),
lrg->degrees_of_freedom());
lrg->dump();
}
#endif
// Re-insert into the IFG
_ifg->re_insert(lidx);
if( !lrg->alive() ) continue;
// capture allstackedness flag before mask is hacked
const int is_allstack = lrg->mask().is_AllStack();
// Yeah, yeah, yeah, I know, I know. I can refactor this
// to avoid the GOTO, although the refactored code will not
// be much clearer. We arrive here IFF we have a stack-based
// live range that cannot color in the current chunk, and it
// has to move into the next free stack chunk.
int chunk = 0; // Current chunk is first chunk
retry_next_chunk:
// Remove neighbor colors
IndexSet *s = _ifg->neighbors(lidx);
debug_only(RegMask orig_mask = lrg->mask();)
IndexSetIterator elements(s);
uint neighbor;
while ((neighbor = elements.next()) != 0) {
// Note that neighbor might be a spill_reg. In this case, exclusion
// of its color will be a no-op, since the spill_reg chunk is in outer
// space. Also, if neighbor is in a different chunk, this exclusion
// will be a no-op. (Later on, if lrg runs out of possible colors in
// its chunk, a new chunk of color may be tried, in which case
// examination of neighbors is started again, at retry_next_chunk.)
LRG &nlrg = lrgs(neighbor);
OptoReg::Name nreg = nlrg.reg();
// Only subtract masks in the same chunk
if( nreg >= chunk && nreg < chunk + RegMask::CHUNK_SIZE ) {
#ifndef PRODUCT
uint size = lrg->mask().Size();
RegMask rm = lrg->mask();
#endif
lrg->SUBTRACT(nlrg.mask());
#ifndef PRODUCT
if (trace_spilling() && lrg->mask().Size() != size) {
ttyLocker ttyl;
tty->print("L%d ", lidx);
rm.dump();
tty->print(" intersected L%d ", neighbor);
nlrg.mask().dump();
tty->print(" removed ");
rm.SUBTRACT(lrg->mask());
rm.dump();
tty->print(" leaving ");
lrg->mask().dump();
tty->cr();
}
#endif
}
}
//assert(is_allstack == lrg->mask().is_AllStack(), "nbrs must not change AllStackedness");
// Aligned pairs need aligned masks
if( lrg->num_regs() == 2 && !lrg->_fat_proj )
lrg->ClearToPairs();
// Check if a color is available and if so pick the color
OptoReg::Name reg = choose_color( *lrg, chunk );
#ifdef SPARC
debug_only(lrg->compute_set_mask_size());
assert(lrg->num_regs() != 2 || lrg->is_bound() || is_even(reg-1), "allocate all doubles aligned");
#endif
//---------------
// If we fail to color and the AllStack flag is set, trigger
// a chunk-rollover event
if(!OptoReg::is_valid(OptoReg::add(reg,-chunk)) && is_allstack) {
// Bump register mask up to next stack chunk
chunk += RegMask::CHUNK_SIZE;
lrg->Set_All();
goto retry_next_chunk;
}
//---------------
// Did we get a color?
else if( OptoReg::is_valid(reg)) {
#ifndef PRODUCT
RegMask avail_rm = lrg->mask();
#endif
// Record selected register
lrg->set_reg(reg);
if( reg >= _max_reg ) // Compute max register limit
_max_reg = OptoReg::add(reg,1);
// Fold reg back into normal space
reg = OptoReg::add(reg,-chunk);
// If the live range is not bound, then we actually had some choices
// to make. In this case, the mask has more bits in it than the colors
// choosen. Restrict the mask to just what was picked.
if( lrg->num_regs() == 1 ) { // Size 1 live range
lrg->Clear(); // Clear the mask
lrg->Insert(reg); // Set regmask to match selected reg
lrg->set_mask_size(1);
} else if( !lrg->_fat_proj ) {
// For pairs, also insert the low bit of the pair
assert( lrg->num_regs() == 2, "unbound fatproj???" );
lrg->Clear(); // Clear the mask
lrg->Insert(reg); // Set regmask to match selected reg
lrg->Insert(OptoReg::add(reg,-1));
lrg->set_mask_size(2);
} else { // Else fatproj
// mask must be equal to fatproj bits, by definition
}
#ifndef PRODUCT
if (trace_spilling()) {
ttyLocker ttyl;
tty->print("L%d selected ", lidx);
lrg->mask().dump();
tty->print(" from ");
avail_rm.dump();
tty->cr();
}
#endif
// Note that reg is the highest-numbered register in the newly-bound mask.
} // end color available case
//---------------
// Live range is live and no colors available
else {
assert( lrg->alive(), "" );
assert( !lrg->_fat_proj || lrg->is_multidef() ||
lrg->_def->outcnt() > 0, "fat_proj cannot spill");
assert( !orig_mask.is_AllStack(), "All Stack does not spill" );
// Assign the special spillreg register
lrg->set_reg(OptoReg::Name(spill_reg++));
// Do not empty the regmask; leave mask_size lying around
// for use during Spilling
#ifndef PRODUCT
if( trace_spilling() ) {
ttyLocker ttyl;
tty->print("L%d spilling with neighbors: ", lidx);
s->dump();
debug_only(tty->print(" original mask: "));
debug_only(orig_mask.dump());
dump_lrg(lidx);
}
#endif
} // end spill case
}
return spill_reg-LRG::SPILL_REG; // Return number of spills
}
//------------------------------copy_was_spilled-------------------------------
// Copy 'was_spilled'-edness from the source Node to the dst Node.
void PhaseChaitin::copy_was_spilled( Node *src, Node *dst ) {
if( _spilled_once.test(src->_idx) ) {
_spilled_once.set(dst->_idx);
lrgs(Find(dst))._was_spilled1 = 1;
if( _spilled_twice.test(src->_idx) ) {
_spilled_twice.set(dst->_idx);
lrgs(Find(dst))._was_spilled2 = 1;
}
}
}
//------------------------------set_was_spilled--------------------------------
// Set the 'spilled_once' or 'spilled_twice' flag on a node.
void PhaseChaitin::set_was_spilled( Node *n ) {
if( _spilled_once.test_set(n->_idx) )
_spilled_twice.set(n->_idx);
}
//------------------------------fixup_spills-----------------------------------
// Convert Ideal spill instructions into proper FramePtr + offset Loads and
// Stores. Use-def chains are NOT preserved, but Node->LRG->reg maps are.
void PhaseChaitin::fixup_spills() {
// This function does only cisc spill work.
if( !UseCISCSpill ) return;
NOT_PRODUCT( Compile::TracePhase t3("fixupSpills", &_t_fixupSpills, TimeCompiler); )
// Grab the Frame Pointer
Node *fp = _cfg._broot->head()->in(1)->in(TypeFunc::FramePtr);
// For all blocks
for( uint i = 0; i < _cfg._num_blocks; i++ ) {
Block *b = _cfg._blocks[i];
// For all instructions in block
uint last_inst = b->end_idx();
for( uint j = 1; j <= last_inst; j++ ) {
Node *n = b->_nodes[j];
// Dead instruction???
assert( n->outcnt() != 0 ||// Nothing dead after post alloc
C->top() == n || // Or the random TOP node
n->is_Proj(), // Or a fat-proj kill node
"No dead instructions after post-alloc" );
int inp = n->cisc_operand();
if( inp != AdlcVMDeps::Not_cisc_spillable ) {
// Convert operand number to edge index number
MachNode *mach = n->as_Mach();
inp = mach->operand_index(inp);
Node *src = n->in(inp); // Value to load or store
LRG &lrg_cisc = lrgs( Find_const(src) );
OptoReg::Name src_reg = lrg_cisc.reg();
// Doubles record the HIGH register of an adjacent pair.
src_reg = OptoReg::add(src_reg,1-lrg_cisc.num_regs());
if( OptoReg::is_stack(src_reg) ) { // If input is on stack
// This is a CISC Spill, get stack offset and construct new node
#ifndef PRODUCT
if( TraceCISCSpill ) {
tty->print(" reg-instr: ");
n->dump();
}
#endif
int stk_offset = reg2offset(src_reg);
// Bailout if we might exceed node limit when spilling this instruction
C->check_node_count(0, "out of nodes fixing spills");
if (C->failing()) return;
// Transform node
MachNode *cisc = mach->cisc_version(stk_offset, C)->as_Mach();
cisc->set_req(inp,fp); // Base register is frame pointer
if( cisc->oper_input_base() > 1 && mach->oper_input_base() <= 1 ) {
assert( cisc->oper_input_base() == 2, "Only adding one edge");
cisc->ins_req(1,src); // Requires a memory edge
}
b->_nodes.map(j,cisc); // Insert into basic block
n->subsume_by(cisc); // Correct graph
//
++_used_cisc_instructions;
#ifndef PRODUCT
if( TraceCISCSpill ) {
tty->print(" cisc-instr: ");
cisc->dump();
}
#endif
} else {
#ifndef PRODUCT
if( TraceCISCSpill ) {
tty->print(" using reg-instr: ");
n->dump();
}
#endif
++_unused_cisc_instructions; // input can be on stack
}
}
} // End of for all instructions
} // End of for all blocks
}
//------------------------------find_base_for_derived--------------------------
// Helper to stretch above; recursively discover the base Node for a
// given derived Node. Easy for AddP-related machine nodes, but needs
// to be recursive for derived Phis.
Node *PhaseChaitin::find_base_for_derived( Node **derived_base_map, Node *derived, uint &maxlrg ) {
// See if already computed; if so return it
if( derived_base_map[derived->_idx] )
return derived_base_map[derived->_idx];
// See if this happens to be a base.
// NOTE: we use TypePtr instead of TypeOopPtr because we can have
// pointers derived from NULL! These are always along paths that
// can't happen at run-time but the optimizer cannot deduce it so
// we have to handle it gracefully.
const TypePtr *tj = derived->bottom_type()->isa_ptr();
// If its an OOP with a non-zero offset, then it is derived.
if( tj->_offset == 0 ) {
derived_base_map[derived->_idx] = derived;
return derived;
}
// Derived is NULL+offset? Base is NULL!
if( derived->is_Con() ) {
Node *base = new (C, 1) ConPNode( TypePtr::NULL_PTR );
uint no_lidx = 0; // an unmatched constant in debug info has no LRG
_names.extend(base->_idx, no_lidx);
derived_base_map[derived->_idx] = base;
return base;
}
// Check for AddP-related opcodes
if( !derived->is_Phi() ) {
assert( derived->as_Mach()->ideal_Opcode() == Op_AddP, "" );
Node *base = derived->in(AddPNode::Base);
derived_base_map[derived->_idx] = base;
return base;
}
// Recursively find bases for Phis.
// First check to see if we can avoid a base Phi here.
Node *base = find_base_for_derived( derived_base_map, derived->in(1),maxlrg);
uint i;
for( i = 2; i < derived->req(); i++ )
if( base != find_base_for_derived( derived_base_map,derived->in(i),maxlrg))
break;
// Went to the end without finding any different bases?
if( i == derived->req() ) { // No need for a base Phi here
derived_base_map[derived->_idx] = base;
return base;
}
// Now we see we need a base-Phi here to merge the bases
base = new (C, derived->req()) PhiNode( derived->in(0), base->bottom_type() );
for( i = 1; i < derived->req(); i++ )
base->init_req(i, find_base_for_derived(derived_base_map, derived->in(i), maxlrg));
// Search the current block for an existing base-Phi
Block *b = _cfg._bbs[derived->_idx];
for( i = 1; i <= b->end_idx(); i++ ) {// Search for matching Phi
Node *phi = b->_nodes[i];
if( !phi->is_Phi() ) { // Found end of Phis with no match?
b->_nodes.insert( i, base ); // Must insert created Phi here as base
_cfg._bbs.map( base->_idx, b );
new_lrg(base,maxlrg++);
break;
}
// See if Phi matches.
uint j;
for( j = 1; j < base->req(); j++ )
if( phi->in(j) != base->in(j) &&
!(phi->in(j)->is_Con() && base->in(j)->is_Con()) ) // allow different NULLs
break;
if( j == base->req() ) { // All inputs match?
base = phi; // Then use existing 'phi' and drop 'base'
break;
}
}
// Cache info for later passes
derived_base_map[derived->_idx] = base;
return base;
}
//------------------------------stretch_base_pointer_live_ranges---------------
// At each Safepoint, insert extra debug edges for each pair of derived value/
// base pointer that is live across the Safepoint for oopmap building. The
// edge pairs get added in after sfpt->jvmtail()->oopoff(), but are in the
// required edge set.
bool PhaseChaitin::stretch_base_pointer_live_ranges( ResourceArea *a ) {
int must_recompute_live = false;
uint maxlrg = _maxlrg;
Node **derived_base_map = (Node**)a->Amalloc(sizeof(Node*)*C->unique());
memset( derived_base_map, 0, sizeof(Node*)*C->unique() );
// For all blocks in RPO do...
for( uint i=0; i<_cfg._num_blocks; i++ ) {
Block *b = _cfg._blocks[i];
// Note use of deep-copy constructor. I cannot hammer the original
// liveout bits, because they are needed by the following coalesce pass.
IndexSet liveout(_live->live(b));
for( uint j = b->end_idx() + 1; j > 1; j-- ) {
Node *n = b->_nodes[j-1];
// Pre-split compares of loop-phis. Loop-phis form a cycle we would
// like to see in the same register. Compare uses the loop-phi and so
// extends its live range BUT cannot be part of the cycle. If this
// extended live range overlaps with the update of the loop-phi value
// we need both alive at the same time -- which requires at least 1
// copy. But because Intel has only 2-address registers we end up with
// at least 2 copies, one before the loop-phi update instruction and
// one after. Instead we split the input to the compare just after the
// phi.
if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CmpI ) {
Node *phi = n->in(1);
if( phi->is_Phi() && phi->as_Phi()->region()->is_Loop() ) {
Block *phi_block = _cfg._bbs[phi->_idx];
if( _cfg._bbs[phi_block->pred(2)->_idx] == b ) {
const RegMask *mask = C->matcher()->idealreg2spillmask[Op_RegI];
Node *spill = new (C) MachSpillCopyNode( phi, *mask, *mask );
insert_proj( phi_block, 1, spill, maxlrg++ );
n->set_req(1,spill);
must_recompute_live = true;
}
}
}
// Get value being defined
uint lidx = n2lidx(n);
if( lidx && lidx < _maxlrg /* Ignore the occasional brand-new live range */) {
// Remove from live-out set
liveout.remove(lidx);
// Copies do not define a new value and so do not interfere.
// Remove the copies source from the liveout set before interfering.
uint idx = n->is_Copy();
if( idx ) liveout.remove( n2lidx(n->in(idx)) );
}
// Found a safepoint?
JVMState *jvms = n->jvms();
if( jvms ) {
// Now scan for a live derived pointer
IndexSetIterator elements(&liveout);
uint neighbor;
while ((neighbor = elements.next()) != 0) {
// Find reaching DEF for base and derived values
// This works because we are still in SSA during this call.
Node *derived = lrgs(neighbor)._def;
const TypePtr *tj = derived->bottom_type()->isa_ptr();
// If its an OOP with a non-zero offset, then it is derived.
if( tj && tj->_offset != 0 && tj->isa_oop_ptr() ) {
Node *base = find_base_for_derived( derived_base_map, derived, maxlrg );
assert( base->_idx < _names.Size(), "" );
// Add reaching DEFs of derived pointer and base pointer as a
// pair of inputs
n->add_req( derived );
n->add_req( base );
// See if the base pointer is already live to this point.
// Since I'm working on the SSA form, live-ness amounts to
// reaching def's. So if I find the base's live range then
// I know the base's def reaches here.
if( (n2lidx(base) >= _maxlrg ||// (Brand new base (hence not live) or
!liveout.member( n2lidx(base) ) ) && // not live) AND
(n2lidx(base) > 0) && // not a constant
_cfg._bbs[base->_idx] != b ) { // base not def'd in blk)
// Base pointer is not currently live. Since I stretched
// the base pointer to here and it crosses basic-block
// boundaries, the global live info is now incorrect.
// Recompute live.
must_recompute_live = true;
} // End of if base pointer is not live to debug info
}
} // End of scan all live data for derived ptrs crossing GC point
} // End of if found a GC point
// Make all inputs live
if( !n->is_Phi() ) { // Phi function uses come from prior block
for( uint k = 1; k < n->req(); k++ ) {
uint lidx = n2lidx(n->in(k));
if( lidx < _maxlrg )
liveout.insert( lidx );
}
}
} // End of forall instructions in block
liveout.clear(); // Free the memory used by liveout.
} // End of forall blocks
_maxlrg = maxlrg;
// If I created a new live range I need to recompute live
if( maxlrg != _ifg->_maxlrg )
must_recompute_live = true;
return must_recompute_live != 0;
}
//------------------------------add_reference----------------------------------
// Extend the node to LRG mapping
void PhaseChaitin::add_reference( const Node *node, const Node *old_node ) {
_names.extend( node->_idx, n2lidx(old_node) );
}
//------------------------------dump-------------------------------------------
#ifndef PRODUCT
void PhaseChaitin::dump( const Node *n ) const {
uint r = (n->_idx < _names.Size() ) ? Find_const(n) : 0;
tty->print("L%d",r);
if( r && n->Opcode() != Op_Phi ) {
if( _node_regs ) { // Got a post-allocation copy of allocation?
tty->print("[");
OptoReg::Name second = get_reg_second(n);
if( OptoReg::is_valid(second) ) {
if( OptoReg::is_reg(second) )
tty->print("%s:",Matcher::regName[second]);
else
tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(second));
}
OptoReg::Name first = get_reg_first(n);
if( OptoReg::is_reg(first) )
tty->print("%s]",Matcher::regName[first]);
else
tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(first));
} else
n->out_RegMask().dump();
}
tty->print("/N%d\t",n->_idx);
tty->print("%s === ", n->Name());
uint k;
for( k = 0; k < n->req(); k++) {
Node *m = n->in(k);
if( !m ) tty->print("_ ");
else {
uint r = (m->_idx < _names.Size() ) ? Find_const(m) : 0;
tty->print("L%d",r);
// Data MultiNode's can have projections with no real registers.
// Don't die while dumping them.
int op = n->Opcode();
if( r && op != Op_Phi && op != Op_Proj && op != Op_SCMemProj) {
if( _node_regs ) {
tty->print("[");
OptoReg::Name second = get_reg_second(n->in(k));
if( OptoReg::is_valid(second) ) {
if( OptoReg::is_reg(second) )
tty->print("%s:",Matcher::regName[second]);
else
tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer),
reg2offset_unchecked(second));
}
OptoReg::Name first = get_reg_first(n->in(k));
if( OptoReg::is_reg(first) )
tty->print("%s]",Matcher::regName[first]);
else
tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer),
reg2offset_unchecked(first));
} else
n->in_RegMask(k).dump();
}
tty->print("/N%d ",m->_idx);
}
}
if( k < n->len() && n->in(k) ) tty->print("| ");
for( ; k < n->len(); k++ ) {
Node *m = n->in(k);
if( !m ) break;
uint r = (m->_idx < _names.Size() ) ? Find_const(m) : 0;
tty->print("L%d",r);
tty->print("/N%d ",m->_idx);
}
if( n->is_Mach() ) n->as_Mach()->dump_spec(tty);
else n->dump_spec(tty);
if( _spilled_once.test(n->_idx ) ) {
tty->print(" Spill_1");
if( _spilled_twice.test(n->_idx ) )
tty->print(" Spill_2");
}
tty->print("\n");
}
void PhaseChaitin::dump( const Block * b ) const {
b->dump_head( &_cfg._bbs );
// For all instructions
for( uint j = 0; j < b->_nodes.size(); j++ )
dump(b->_nodes[j]);
// Print live-out info at end of block
if( _live ) {
tty->print("Liveout: ");
IndexSet *live = _live->live(b);
IndexSetIterator elements(live);
tty->print("{");
uint i;
while ((i = elements.next()) != 0) {
tty->print("L%d ", Find_const(i));
}
tty->print_cr("}");
}
tty->print("\n");
}
void PhaseChaitin::dump() const {
tty->print( "--- Chaitin -- argsize: %d framesize: %d ---\n",
_matcher._new_SP, _framesize );
// For all blocks
for( uint i = 0; i < _cfg._num_blocks; i++ )
dump(_cfg._blocks[i]);
// End of per-block dump
tty->print("\n");
if (!_ifg) {
tty->print("(No IFG.)\n");
return;
}
// Dump LRG array
tty->print("--- Live RanGe Array ---\n");
for(uint i2 = 1; i2 < _maxlrg; i2++ ) {
tty->print("L%d: ",i2);
if( i2 < _ifg->_maxlrg ) lrgs(i2).dump( );
else tty->print("new LRG");
}
tty->print_cr("");
// Dump lo-degree list
tty->print("Lo degree: ");
for(uint i3 = _lo_degree; i3; i3 = lrgs(i3)._next )
tty->print("L%d ",i3);
tty->print_cr("");
// Dump lo-stk-degree list
tty->print("Lo stk degree: ");
for(uint i4 = _lo_stk_degree; i4; i4 = lrgs(i4)._next )
tty->print("L%d ",i4);
tty->print_cr("");
// Dump lo-degree list
tty->print("Hi degree: ");
for(uint i5 = _hi_degree; i5; i5 = lrgs(i5)._next )
tty->print("L%d ",i5);
tty->print_cr("");
}
//------------------------------dump_degree_lists------------------------------
void PhaseChaitin::dump_degree_lists() const {
// Dump lo-degree list
tty->print("Lo degree: ");
for( uint i = _lo_degree; i; i = lrgs(i)._next )
tty->print("L%d ",i);
tty->print_cr("");
// Dump lo-stk-degree list
tty->print("Lo stk degree: ");
for(uint i2 = _lo_stk_degree; i2; i2 = lrgs(i2)._next )
tty->print("L%d ",i2);
tty->print_cr("");
// Dump lo-degree list
tty->print("Hi degree: ");
for(uint i3 = _hi_degree; i3; i3 = lrgs(i3)._next )
tty->print("L%d ",i3);
tty->print_cr("");
}
//------------------------------dump_simplified--------------------------------
void PhaseChaitin::dump_simplified() const {
tty->print("Simplified: ");
for( uint i = _simplified; i; i = lrgs(i)._next )
tty->print("L%d ",i);
tty->print_cr("");
}
static char *print_reg( OptoReg::Name reg, const PhaseChaitin *pc, char *buf ) {
if ((int)reg < 0)
sprintf(buf, "<OptoReg::%d>", (int)reg);
else if (OptoReg::is_reg(reg))
strcpy(buf, Matcher::regName[reg]);
else
sprintf(buf,"%s + #%d",OptoReg::regname(OptoReg::c_frame_pointer),
pc->reg2offset(reg));
return buf+strlen(buf);
}
//------------------------------dump_register----------------------------------
// Dump a register name into a buffer. Be intelligent if we get called
// before allocation is complete.
char *PhaseChaitin::dump_register( const Node *n, char *buf ) const {
if( !this ) { // Not got anything?
sprintf(buf,"N%d",n->_idx); // Then use Node index
} else if( _node_regs ) {
// Post allocation, use direct mappings, no LRG info available
print_reg( get_reg_first(n), this, buf );
} else {
uint lidx = Find_const(n); // Grab LRG number
if( !_ifg ) {
sprintf(buf,"L%d",lidx); // No register binding yet
} else if( !lidx ) { // Special, not allocated value
strcpy(buf,"Special");
} else if( (lrgs(lidx).num_regs() == 1)
? !lrgs(lidx).mask().is_bound1()
: !lrgs(lidx).mask().is_bound2() ) {
sprintf(buf,"L%d",lidx); // No register binding yet
} else { // Hah! We have a bound machine register
print_reg( lrgs(lidx).reg(), this, buf );
}
}
return buf+strlen(buf);
}
//----------------------dump_for_spill_split_recycle--------------------------
void PhaseChaitin::dump_for_spill_split_recycle() const {
if( WizardMode && (PrintCompilation || PrintOpto) ) {
// Display which live ranges need to be split and the allocator's state
tty->print_cr("Graph-Coloring Iteration %d will split the following live ranges", _trip_cnt);
for( uint bidx = 1; bidx < _maxlrg; bidx++ ) {
if( lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG ) {
tty->print("L%d: ", bidx);
lrgs(bidx).dump();
}
}
tty->cr();
dump();
}
}
//------------------------------dump_frame------------------------------------
void PhaseChaitin::dump_frame() const {
const char *fp = OptoReg::regname(OptoReg::c_frame_pointer);
const TypeTuple *domain = C->tf()->domain();
const int argcnt = domain->cnt() - TypeFunc::Parms;
// Incoming arguments in registers dump
for( int k = 0; k < argcnt; k++ ) {
OptoReg::Name parmreg = _matcher._parm_regs[k].first();
if( OptoReg::is_reg(parmreg)) {
const char *reg_name = OptoReg::regname(parmreg);
tty->print("#r%3.3d %s", parmreg, reg_name);
parmreg = _matcher._parm_regs[k].second();
if( OptoReg::is_reg(parmreg)) {
tty->print(":%s", OptoReg::regname(parmreg));
}
tty->print(" : parm %d: ", k);
domain->field_at(k + TypeFunc::Parms)->dump();
tty->print_cr("");
}
}
// Check for un-owned padding above incoming args
OptoReg::Name reg = _matcher._new_SP;
if( reg > _matcher._in_arg_limit ) {
reg = OptoReg::add(reg, -1);
tty->print_cr("#r%3.3d %s+%2d: pad0, owned by CALLER", reg, fp, reg2offset_unchecked(reg));
}
// Incoming argument area dump
OptoReg::Name begin_in_arg = OptoReg::add(_matcher._old_SP,C->out_preserve_stack_slots());
while( reg > begin_in_arg ) {
reg = OptoReg::add(reg, -1);
tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg));
int j;
for( j = 0; j < argcnt; j++) {
if( _matcher._parm_regs[j].first() == reg ||
_matcher._parm_regs[j].second() == reg ) {
tty->print("parm %d: ",j);
domain->field_at(j + TypeFunc::Parms)->dump();
tty->print_cr("");
break;
}
}
if( j >= argcnt )
tty->print_cr("HOLE, owned by SELF");
}
// Old outgoing preserve area
while( reg > _matcher._old_SP ) {
reg = OptoReg::add(reg, -1);
tty->print_cr("#r%3.3d %s+%2d: old out preserve",reg,fp,reg2offset_unchecked(reg));
}
// Old SP
tty->print_cr("# -- Old %s -- Framesize: %d --",fp,
reg2offset_unchecked(OptoReg::add(_matcher._old_SP,-1)) - reg2offset_unchecked(_matcher._new_SP)+jintSize);
// Preserve area dump
reg = OptoReg::add(reg, -1);
while( OptoReg::is_stack(reg)) {
tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg));
if( _matcher.return_addr() == reg )
tty->print_cr("return address");
else if( _matcher.return_addr() == OptoReg::add(reg,1) &&
VerifyStackAtCalls )
tty->print_cr("0xBADB100D +VerifyStackAtCalls");
else if ((int)OptoReg::reg2stack(reg) < C->fixed_slots())
tty->print_cr("Fixed slot %d", OptoReg::reg2stack(reg));
else
tty->print_cr("pad2, in_preserve");
reg = OptoReg::add(reg, -1);
}
// Spill area dump
reg = OptoReg::add(_matcher._new_SP, _framesize );
while( reg > _matcher._out_arg_limit ) {
reg = OptoReg::add(reg, -1);
tty->print_cr("#r%3.3d %s+%2d: spill",reg,fp,reg2offset_unchecked(reg));
}
// Outgoing argument area dump
while( reg > OptoReg::add(_matcher._new_SP, C->out_preserve_stack_slots()) ) {
reg = OptoReg::add(reg, -1);
tty->print_cr("#r%3.3d %s+%2d: outgoing argument",reg,fp,reg2offset_unchecked(reg));
}
// Outgoing new preserve area
while( reg > _matcher._new_SP ) {
reg = OptoReg::add(reg, -1);
tty->print_cr("#r%3.3d %s+%2d: new out preserve",reg,fp,reg2offset_unchecked(reg));
}
tty->print_cr("#");
}
//------------------------------dump_bb----------------------------------------
void PhaseChaitin::dump_bb( uint pre_order ) const {
tty->print_cr("---dump of B%d---",pre_order);
for( uint i = 0; i < _cfg._num_blocks; i++ ) {
Block *b = _cfg._blocks[i];
if( b->_pre_order == pre_order )
dump(b);
}
}
//------------------------------dump_lrg---------------------------------------
void PhaseChaitin::dump_lrg( uint lidx ) const {
tty->print_cr("---dump of L%d---",lidx);
if( _ifg ) {
if( lidx >= _maxlrg ) {
tty->print("Attempt to print live range index beyond max live range.\n");
return;
}
tty->print("L%d: ",lidx);
lrgs(lidx).dump( );
}
if( _ifg ) { tty->print("Neighbors: %d - ", _ifg->neighbor_cnt(lidx));
_ifg->neighbors(lidx)->dump();
tty->cr();
}
// For all blocks
for( uint i = 0; i < _cfg._num_blocks; i++ ) {
Block *b = _cfg._blocks[i];
int dump_once = 0;
// For all instructions
for( uint j = 0; j < b->_nodes.size(); j++ ) {
Node *n = b->_nodes[j];
if( Find_const(n) == lidx ) {
if( !dump_once++ ) {
tty->cr();
b->dump_head( &_cfg._bbs );
}
dump(n);
continue;
}
uint cnt = n->req();
for( uint k = 1; k < cnt; k++ ) {
Node *m = n->in(k);
if (!m) continue; // be robust in the dumper
if( Find_const(m) == lidx ) {
if( !dump_once++ ) {
tty->cr();
b->dump_head( &_cfg._bbs );
}
dump(n);
}
}
}
} // End of per-block dump
tty->cr();
}
#endif // not PRODUCT
//------------------------------print_chaitin_statistics-------------------------------
int PhaseChaitin::_final_loads = 0;
int PhaseChaitin::_final_stores = 0;
int PhaseChaitin::_final_memoves= 0;
int PhaseChaitin::_final_copies = 0;
double PhaseChaitin::_final_load_cost = 0;
double PhaseChaitin::_final_store_cost = 0;
double PhaseChaitin::_final_memove_cost= 0;
double PhaseChaitin::_final_copy_cost = 0;
int PhaseChaitin::_conserv_coalesce = 0;
int PhaseChaitin::_conserv_coalesce_pair = 0;
int PhaseChaitin::_conserv_coalesce_trie = 0;
int PhaseChaitin::_conserv_coalesce_quad = 0;
int PhaseChaitin::_post_alloc = 0;
int PhaseChaitin::_lost_opp_pp_coalesce = 0;
int PhaseChaitin::_lost_opp_cflow_coalesce = 0;
int PhaseChaitin::_used_cisc_instructions = 0;
int PhaseChaitin::_unused_cisc_instructions = 0;
int PhaseChaitin::_allocator_attempts = 0;
int PhaseChaitin::_allocator_successes = 0;
#ifndef PRODUCT
uint PhaseChaitin::_high_pressure = 0;
uint PhaseChaitin::_low_pressure = 0;
void PhaseChaitin::print_chaitin_statistics() {
tty->print_cr("Inserted %d spill loads, %d spill stores, %d mem-mem moves and %d copies.", _final_loads, _final_stores, _final_memoves, _final_copies);
tty->print_cr("Total load cost= %6.0f, store cost = %6.0f, mem-mem cost = %5.2f, copy cost = %5.0f.", _final_load_cost, _final_store_cost, _final_memove_cost, _final_copy_cost);
tty->print_cr("Adjusted spill cost = %7.0f.",
_final_load_cost*4.0 + _final_store_cost * 2.0 +
_final_copy_cost*1.0 + _final_memove_cost*12.0);
tty->print("Conservatively coalesced %d copies, %d pairs",
_conserv_coalesce, _conserv_coalesce_pair);
if( _conserv_coalesce_trie || _conserv_coalesce_quad )
tty->print(", %d tries, %d quads", _conserv_coalesce_trie, _conserv_coalesce_quad);
tty->print_cr(", %d post alloc.", _post_alloc);
if( _lost_opp_pp_coalesce || _lost_opp_cflow_coalesce )
tty->print_cr("Lost coalesce opportunity, %d private-private, and %d cflow interfered.",
_lost_opp_pp_coalesce, _lost_opp_cflow_coalesce );
if( _used_cisc_instructions || _unused_cisc_instructions )
tty->print_cr("Used cisc instruction %d, remained in register %d",
_used_cisc_instructions, _unused_cisc_instructions);
if( _allocator_successes != 0 )
tty->print_cr("Average allocation trips %f", (float)_allocator_attempts/(float)_allocator_successes);
tty->print_cr("High Pressure Blocks = %d, Low Pressure Blocks = %d", _high_pressure, _low_pressure);
}
#endif // not PRODUCT