//------------------------------compute_separating_interferences---------------
// Factored code from copy_copy that computes extra interferences from
// lengthening a live range by double-coalescing.
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 ) {
assert(!lrgs(lr1)._fat_proj, "cannot coalesce fat_proj");
assert(!lrgs(lr2)._fat_proj, "cannot coalesce fat_proj");
Node *prev_copy = dst_copy->in(dst_copy->is_Copy());
Block *b2 = b;
uint bindex2 = bindex;
while( 1 ) {
// Find previous instruction
bindex2--; // Chain backwards 1 instruction
while( bindex2 == 0 ) { // At block start, find prior block
assert( b2->num_preds() == 2, "cannot double coalesce across c-flow" );
b2 = _phc._cfg._bbs[b2->pred(1)->_idx];
bindex2 = b2->end_idx()-1;
}
// Get prior instruction
assert(bindex2 < b2->_nodes.size(), "index out of bounds");
Node *x = b2->_nodes[bindex2];
if( x == prev_copy ) { // Previous copy in copy chain?
if( prev_copy == src_copy)// Found end of chain and all interferences
break; // So break out of loop
// Else work back one in copy chain
prev_copy = prev_copy->in(prev_copy->is_Copy());
} else { // Else collect interferences
uint lidx = _phc.Find(x);
// Found another def of live-range being stretched?
if( lidx == lr1 ) return max_juint;
if( lidx == lr2 ) return max_juint;
// If we attempt to coalesce across a bound def
if( lrgs(lidx).is_bound() ) {
// Do not let the coalesced LRG expect to get the bound color
rm.SUBTRACT( lrgs(lidx).mask() );
// Recompute rm_size
rm_size = rm.Size();
//if( rm._flags ) rm_size += 1000000;
if( reg_degree >= rm_size ) return max_juint;
}
if( rm.overlap(lrgs(lidx).mask()) ) {
// Insert lidx into union LRG; returns TRUE if actually inserted
if( _ulr.insert(lidx) ) {
// Infinite-stack neighbors do not alter colorability, as they
// can always color to some other color.
if( !lrgs(lidx).mask().is_AllStack() ) {
// If this coalesce will make any new neighbor uncolorable,
// do not coalesce.
if( lrgs(lidx).just_lo_degree() )
return max_juint;
// Bump our degree
if( ++reg_degree >= rm_size )
return max_juint;
} // End of if not infinite-stack neighbor
} // End of if actually inserted
} // End of if live range overlaps
} // End of else collect interferences for 1 node
} // End of while forever, scan back for interferences
return reg_degree;
}
//------------------------------copy_copy--------------------------------------
// See if I can coalesce a series of multiple copies together. I need the
// final dest copy and the original src copy. They can be the same Node.
// Compute the compatible register masks.
bool PhaseConservativeCoalesce::copy_copy( Node *dst_copy, Node *src_copy, Block *b, uint bindex ) {
if( !dst_copy->is_SpillCopy() ) return false;
if( !src_copy->is_SpillCopy() ) return false;
Node *src_def = src_copy->in(src_copy->is_Copy());
uint lr1 = _phc.Find(dst_copy);
uint lr2 = _phc.Find(src_def );
// Same live ranges already?
if( lr1 == lr2 ) return false;
// Interfere?
if( _phc._ifg->test_edge_sq( lr1, lr2 ) ) return false;
// Not an oop->int cast; oop->oop, int->int, AND int->oop are OK.
if( !lrgs(lr1)._is_oop && lrgs(lr2)._is_oop ) // not an oop->int cast
return false;
// Coalescing between an aligned live range and a mis-aligned live range?
// No, no! Alignment changes how we count degree.
if( lrgs(lr1)._fat_proj != lrgs(lr2)._fat_proj )
return false;
// Sort; use smaller live-range number
Node *lr1_node = dst_copy;
Node *lr2_node = src_def;
if( lr1 > lr2 ) {
uint tmp = lr1; lr1 = lr2; lr2 = tmp;
lr1_node = src_def; lr2_node = dst_copy;
}
// Check for compatibility of the 2 live ranges by
// intersecting their allowed register sets.
RegMask rm = lrgs(lr1).mask();
rm.AND(lrgs(lr2).mask());
// Number of bits free
uint rm_size = rm.Size();
if (UseFPUForSpilling && rm.is_AllStack() ) {
// Don't coalesce when frequency difference is large
Block *dst_b = _phc._cfg._bbs[dst_copy->_idx];
Block *src_def_b = _phc._cfg._bbs[src_def->_idx];
if (src_def_b->_freq > 10*dst_b->_freq )
return false;
}
// If we can use any stack slot, then effective size is infinite
if( rm.is_AllStack() ) rm_size += 1000000;
// Incompatible masks, no way to coalesce
if( rm_size == 0 ) return false;
// Another early bail-out test is when we are double-coalescing and the
// 2 copies are separated by some control flow.
if( dst_copy != src_copy ) {
Block *src_b = _phc._cfg._bbs[src_copy->_idx];
Block *b2 = b;
while( b2 != src_b ) {
if( b2->num_preds() > 2 ){// Found merge-point
_phc._lost_opp_cflow_coalesce++;
// extra record_bias commented out because Chris believes it is not
// productive. Since we can record only 1 bias, we want to choose one
// that stands a chance of working and this one probably does not.
//record_bias( _phc._lrgs, lr1, lr2 );
return false; // To hard to find all interferences
}
b2 = _phc._cfg._bbs[b2->pred(1)->_idx];
}
}
// Union the two interference sets together into '_ulr'
uint reg_degree = _ulr.lrg_union( lr1, lr2, rm_size, _phc._ifg, rm );
if( reg_degree >= rm_size ) {
record_bias( _phc._ifg, lr1, lr2 );
return false;
}
// Now I need to compute all the interferences between dst_copy and
// src_copy. I'm not willing visit the entire interference graph, so
// I limit my search to things in dst_copy's block or in a straight
// line of previous blocks. I give up at merge points or when I get
// more interferences than my degree. I can stop when I find src_copy.
if( dst_copy != src_copy ) {
reg_degree = compute_separating_interferences(dst_copy, src_copy, b, bindex, rm, rm_size, reg_degree, lr1, lr2 );
if( reg_degree == max_juint ) {
record_bias( _phc._ifg, lr1, lr2 );
return false;
}
} // End of if dst_copy & src_copy are different
// ---- THE COMBINED LRG IS COLORABLE ----
// YEAH - Now coalesce this copy away
assert( lrgs(lr1).num_regs() == lrgs(lr2).num_regs(), "" );
IndexSet *n_lr1 = _phc._ifg->neighbors(lr1);
IndexSet *n_lr2 = _phc._ifg->neighbors(lr2);
// Update the interference graph
update_ifg(lr1, lr2, n_lr1, n_lr2);
_ulr.remove(lr1);
// Uncomment the following code to trace Coalescing in great detail.
//
//if (false) {
// tty->cr();
// tty->print_cr("#######################################");
// tty->print_cr("union %d and %d", lr1, lr2);
// n_lr1->dump();
// n_lr2->dump();
// tty->print_cr("resulting set is");
// _ulr.dump();
//}
// Replace n_lr1 with the new combined live range. _ulr will use
// n_lr1's old memory on the next iteration. n_lr2 is cleared to
// send its internal memory to the free list.
_ulr.swap(n_lr1);
_ulr.clear();
n_lr2->clear();
lrgs(lr1).set_degree( _phc._ifg->effective_degree(lr1) );
lrgs(lr2).set_degree( 0 );
// Join live ranges. Merge larger into smaller. Union lr2 into lr1 in the
// union-find tree
union_helper( lr1_node, lr2_node, lr1, lr2, src_def, dst_copy, src_copy, b, bindex );
// Combine register restrictions
lrgs(lr1).set_mask(rm);
lrgs(lr1).compute_set_mask_size();
lrgs(lr1)._cost += lrgs(lr2)._cost;
lrgs(lr1)._area += lrgs(lr2)._area;
// While its uncommon to successfully coalesce live ranges that started out
// being not-lo-degree, it can happen. In any case the combined coalesced
// live range better Simplify nicely.
lrgs(lr1)._was_lo = 1;
// kinda expensive to do all the time
//tty->print_cr("warning: slow verify happening");
//_phc._ifg->verify( &_phc );
return true;
}
