in src/hotspot/share/opto/reg_split.cpp [496:1460]
uint PhaseChaitin::Split(uint maxlrg, ResourceArea* split_arena) {
Compile::TracePhase tp(_t_regAllocSplit);
// Free thread local resources used by this method on exit.
ResourceMark rm(split_arena);
uint bidx, pidx, slidx, insidx, inpidx, twoidx;
uint non_phi = 1, spill_cnt = 0;
Node *n1, *n2, *n3;
bool *UPblock;
bool u1, u2, u3;
Block *b, *pred;
PhiNode *phi;
GrowableArray<uint> lidxs(split_arena, maxlrg, 0, 0);
// Array of counters to count splits per live range
GrowableArray<uint> splits(split_arena, maxlrg, 0, 0);
#define NEW_SPLIT_ARRAY(type, size)\
(type*) split_arena->allocate_bytes((size) * sizeof(type))
//----------Setup Code----------
// Create a convenient mapping from lrg numbers to reaches/leaves indices
uint *lrg2reach = NEW_SPLIT_ARRAY(uint, maxlrg);
// Gather info on which LRG's are spilling, and build maps
for (bidx = 1; bidx < maxlrg; bidx++) {
if (lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG) {
assert(!lrgs(bidx).mask().is_infinite_stack(), "Infinite stack mask should color");
lrg2reach[bidx] = spill_cnt;
spill_cnt++;
lidxs.append(bidx);
#ifdef ASSERT
// Initialize the split counts to zero
splits.append(0);
#endif
if (PrintOpto && WizardMode && lrgs(bidx)._was_spilled1) {
tty->print_cr("Warning, 2nd spill of L%d",bidx);
}
}
}
// Create side arrays for propagating reaching defs info.
// Each block needs a node pointer for each spilling live range for the
// Def which is live into the block. Phi nodes handle multiple input
// Defs by querying the output of their predecessor blocks and resolving
// them to a single Def at the phi. The pointer is updated for each
// Def in the block, and then becomes the output for the block when
// processing of the block is complete. We also need to track whether
// a Def is UP or DOWN. UP means that it should get a register (ie -
// it is always in LRP regions), and DOWN means that it is probably
// on the stack (ie - it crosses HRP regions).
Node ***Reaches = NEW_SPLIT_ARRAY( Node**, _cfg.number_of_blocks() + 1);
bool **UP = NEW_SPLIT_ARRAY( bool*, _cfg.number_of_blocks() + 1);
Node **debug_defs = NEW_SPLIT_ARRAY( Node*, spill_cnt );
VectorSet **UP_entry= NEW_SPLIT_ARRAY( VectorSet*, spill_cnt );
// Initialize Reaches & UP
for (bidx = 0; bidx < _cfg.number_of_blocks() + 1; bidx++) {
Reaches[bidx] = NEW_SPLIT_ARRAY( Node*, spill_cnt );
UP[bidx] = NEW_SPLIT_ARRAY( bool, spill_cnt );
Node **Reachblock = Reaches[bidx];
bool *UPblock = UP[bidx];
for( slidx = 0; slidx < spill_cnt; slidx++ ) {
UPblock[slidx] = true; // Assume they start in registers
Reachblock[slidx] = nullptr; // Assume that no def is present
}
}
#undef NEW_SPLIT_ARRAY
// Initialize to array of empty vectorsets
// Each containing at most spill_cnt * _cfg.number_of_blocks() entries.
for (slidx = 0; slidx < spill_cnt; slidx++) {
UP_entry[slidx] = new(split_arena) VectorSet(split_arena);
}
// Keep track of DEFS & Phis for later passes
Node_List defs(split_arena, 8);
Node_List phis(split_arena, 16);
//----------PASS 1----------
//----------Propagation & Node Insertion Code----------
// Walk the Blocks in RPO for DEF & USE info
for( bidx = 0; bidx < _cfg.number_of_blocks(); bidx++ ) {
if (C->check_node_count(spill_cnt, out_of_nodes)) {
return 0;
}
b = _cfg.get_block(bidx);
// Reaches & UP arrays for this block
Node** Reachblock = Reaches[b->_pre_order];
UPblock = UP[b->_pre_order];
// Reset counter of start of non-Phi nodes in block
non_phi = 1;
//----------Block Entry Handling----------
// Check for need to insert a new phi
// Cycle through this block's predecessors, collecting Reaches
// info for each spilled LRG. If they are identical, no phi is
// needed. If they differ, check for a phi, and insert if missing,
// or update edges if present. Set current block's Reaches set to
// be either the phi's or the reaching def, as appropriate.
// If no Phi is needed, check if the LRG needs to spill on entry
// to the block due to HRP.
for( slidx = 0; slidx < spill_cnt; slidx++ ) {
// Grab the live range number
uint lidx = lidxs.at(slidx);
// Do not bother splitting or putting in Phis for single-def
// rematerialized live ranges. This happens a lot to constants
// with long live ranges.
if( lrgs(lidx).is_singledef() &&
lrgs(lidx)._def->rematerialize() ) {
// reset the Reaches & UP entries
Reachblock[slidx] = lrgs(lidx)._def;
UPblock[slidx] = true;
// Record following instruction in case 'n' rematerializes and
// kills flags
Block *pred1 = _cfg.get_block_for_node(b->pred(1));
continue;
}
// Initialize needs_phi and needs_split
bool needs_phi = false;
bool needs_split = false;
bool has_phi = false;
// Walk the predecessor blocks to check inputs for that live range
// Grab predecessor block header
n1 = b->pred(1);
// Grab the appropriate reaching def info for inpidx
pred = _cfg.get_block_for_node(n1);
pidx = pred->_pre_order;
Node **Ltmp = Reaches[pidx];
bool *Utmp = UP[pidx];
n1 = Ltmp[slidx];
u1 = Utmp[slidx];
// Initialize node for saving type info
n3 = n1;
u3 = u1;
// Compare inputs to see if a Phi is needed
for( inpidx = 2; inpidx < b->num_preds(); inpidx++ ) {
// Grab predecessor block headers
n2 = b->pred(inpidx);
// Grab the appropriate reaching def info for inpidx
pred = _cfg.get_block_for_node(n2);
pidx = pred->_pre_order;
Ltmp = Reaches[pidx];
Utmp = UP[pidx];
n2 = Ltmp[slidx];
u2 = Utmp[slidx];
// For each LRG, decide if a phi is necessary
if( n1 != n2 ) {
needs_phi = true;
}
// See if the phi has mismatched inputs, UP vs. DOWN
if( n1 && n2 && (u1 != u2) ) {
needs_split = true;
}
// Move n2/u2 to n1/u1 for next iteration
n1 = n2;
u1 = u2;
// Preserve a non-null predecessor for later type referencing
if( (n3 == nullptr) && (n2 != nullptr) ){
n3 = n2;
u3 = u2;
}
} // End for all potential Phi inputs
// check block for appropriate phinode & update edges
for( insidx = 1; insidx <= b->end_idx(); insidx++ ) {
n1 = b->get_node(insidx);
// bail if this is not a phi
phi = n1->is_Phi() ? n1->as_Phi() : nullptr;
if( phi == nullptr ) {
// Keep track of index of first non-PhiNode instruction in block
non_phi = insidx;
// break out of the for loop as we have handled all phi nodes
break;
}
// must be looking at a phi
if (_lrg_map.find_id(n1) == lidxs.at(slidx)) {
// found the necessary phi
needs_phi = false;
has_phi = true;
// initialize the Reaches entry for this LRG
Reachblock[slidx] = phi;
break;
} // end if found correct phi
} // end for all phi's
// If a phi is needed or exist, check for it
if( needs_phi || has_phi ) {
// add new phinode if one not already found
if( needs_phi ) {
// create a new phi node and insert it into the block
// type is taken from left over pointer to a predecessor
guarantee(n3, "No non-null reaching DEF for a Phi");
phi = new PhiNode(b->head(), n3->bottom_type());
// initialize the Reaches entry for this LRG
Reachblock[slidx] = phi;
// add node to block & node_to_block mapping
insert_proj(b, insidx++, phi, maxlrg++);
non_phi++;
// Reset new phi's mapping to be the spilling live range
_lrg_map.map(phi->_idx, lidx);
assert(_lrg_map.find_id(phi) == lidx, "Bad update on Union-Find mapping");
} // end if not found correct phi
// Here you have either found or created the Phi, so record it
assert(phi != nullptr,"Must have a Phi Node here");
phis.push(phi);
// PhiNodes should either force the LRG UP or DOWN depending
// on its inputs and the register pressure in the Phi's block.
UPblock[slidx] = true; // Assume new DEF is UP
// If entering a high-pressure area with no immediate use,
// assume Phi is DOWN
if( is_high_pressure( b, &lrgs(lidx), b->end_idx()) && !prompt_use(b,lidx) )
UPblock[slidx] = false;
// If we are not split up/down and all inputs are down, then we
// are down
if( !needs_split && !u3 )
UPblock[slidx] = false;
} // end if phi is needed
// Do not need a phi, so grab the reaching DEF
else {
// Grab predecessor block header
n1 = b->pred(1);
// Grab the appropriate reaching def info for k
pred = _cfg.get_block_for_node(n1);
pidx = pred->_pre_order;
Node **Ltmp = Reaches[pidx];
bool *Utmp = UP[pidx];
// reset the Reaches & UP entries
Reachblock[slidx] = Ltmp[slidx];
UPblock[slidx] = Utmp[slidx];
} // end else no Phi is needed
} // end for all spilling live ranges
// DEBUG
#ifndef PRODUCT
if(trace_spilling()) {
tty->print("/`\nBlock %d: ", b->_pre_order);
tty->print("Reaching Definitions after Phi handling\n");
for( uint x = 0; x < spill_cnt; x++ ) {
tty->print("Spill Idx %d: UP %d: Node\n",x,UPblock[x]);
if( Reachblock[x] )
Reachblock[x]->dump();
else
tty->print("Undefined\n");
}
}
#endif
//----------Non-Phi Node Splitting----------
// Since phi-nodes have now been handled, the Reachblock array for this
// block is initialized with the correct starting value for the defs which
// reach non-phi instructions in this block. Thus, process non-phi
// instructions normally, inserting SpillCopy nodes for all spill
// locations.
// Memoize any DOWN reaching definitions for use as DEBUG info
for( insidx = 0; insidx < spill_cnt; insidx++ ) {
debug_defs[insidx] = (UPblock[insidx]) ? nullptr : Reachblock[insidx];
if( UPblock[insidx] ) // Memoize UP decision at block start
UP_entry[insidx]->set( b->_pre_order );
}
//----------Walk Instructions in the Block and Split----------
// For all non-phi instructions in the block
for( insidx = 1; insidx <= b->end_idx(); insidx++ ) {
Node *n = b->get_node(insidx);
// Find the defining Node's live range index
uint defidx = _lrg_map.find_id(n);
uint cnt = n->req();
if (n->is_Phi()) {
// Skip phi nodes after removing dead copies.
if (defidx < _lrg_map.max_lrg_id()) {
// Check for useless Phis. These appear if we spill, then
// coalesce away copies. Dont touch Phis in spilling live
// ranges; they are busy getting modified in this pass.
if( lrgs(defidx).reg() < LRG::SPILL_REG ) {
uint i;
Node *u = nullptr;
// Look for the Phi merging 2 unique inputs
for( i = 1; i < cnt; i++ ) {
// Ignore repeats and self
if( n->in(i) != u && n->in(i) != n ) {
// Found a unique input
if( u != nullptr ) // If it's the 2nd, bail out
break;
u = n->in(i); // Else record it
}
}
assert( u, "at least 1 valid input expected" );
if (i >= cnt) { // Found one unique input
assert(_lrg_map.find_id(n) == _lrg_map.find_id(u), "should be the same lrg");
n->replace_by(u); // Then replace with unique input
n->disconnect_inputs(C);
b->remove_node(insidx);
insidx--;
b->_ihrp_index--;
b->_fhrp_index--;
}
}
}
continue;
}
assert( insidx > b->_ihrp_index ||
(b->_reg_pressure < Matcher::int_pressure_limit()) ||
b->_ihrp_index > 4000000 ||
b->_ihrp_index >= b->end_idx() ||
!b->get_node(b->_ihrp_index)->is_Proj(), "" );
assert( insidx > b->_fhrp_index ||
(b->_freg_pressure < Matcher::float_pressure_limit()) ||
b->_fhrp_index > 4000000 ||
b->_fhrp_index >= b->end_idx() ||
!b->get_node(b->_fhrp_index)->is_Proj(), "" );
// ********** Handle Crossing HRP Boundary **********
if( (insidx == b->_ihrp_index) || (insidx == b->_fhrp_index) ) {
for( slidx = 0; slidx < spill_cnt; slidx++ ) {
// Check for need to split at HRP boundary - split if UP
n1 = Reachblock[slidx];
// bail out if no reaching DEF
if( n1 == nullptr ) continue;
// bail out if live range is 'isolated' around inner loop
uint lidx = lidxs.at(slidx);
// If live range is currently UP
if( UPblock[slidx] ) {
// set location to insert spills at
// SPLIT DOWN HERE - NO CISC SPILL
if( is_high_pressure( b, &lrgs(lidx), insidx ) &&
!n1->rematerialize() ) {
// If there is already a valid stack definition available, use it
if( debug_defs[slidx] != nullptr ) {
Reachblock[slidx] = debug_defs[slidx];
}
else {
// Insert point is just past last use or def in the block
int insert_point = insidx-1;
while( insert_point > 0 ) {
Node *n = b->get_node(insert_point);
// Hit top of block? Quit going backwards
if (n->is_Phi()) {
break;
}
// Found a def? Better split after it.
if (_lrg_map.live_range_id(n) == lidx) {
break;
}
// Look for a use
uint i;
for( i = 1; i < n->req(); i++ ) {
if (_lrg_map.live_range_id(n->in(i)) == lidx) {
break;
}
}
// Found a use? Better split after it.
if (i < n->req()) {
break;
}
insert_point--;
}
uint orig_eidx = b->end_idx();
maxlrg = split_DEF( n1, b, insert_point, maxlrg, Reachblock, debug_defs, splits, slidx);
// If it wasn't split bail
if (!maxlrg) {
return 0;
}
// Spill of null check mem op goes into the following block.
if (b->end_idx() > orig_eidx) {
insidx++;
}
}
// This is a new DEF, so update UP
UPblock[slidx] = false;
#ifndef PRODUCT
// DEBUG
if( trace_spilling() ) {
tty->print("\nNew Split DOWN DEF of Spill Idx ");
tty->print("%d, UP %d:\n",slidx,false);
n1->dump();
}
#endif
}
} // end if LRG is UP
} // end for all spilling live ranges
assert( b->get_node(insidx) == n, "got insidx set incorrectly" );
} // end if crossing HRP Boundary
// If the LRG index is oob, then this is a new spillcopy, skip it.
if (defidx >= _lrg_map.max_lrg_id()) {
continue;
}
LRG &deflrg = lrgs(defidx);
uint copyidx = n->is_Copy();
// Remove coalesced copy from CFG
if (copyidx && defidx == _lrg_map.live_range_id(n->in(copyidx))) {
n->replace_by( n->in(copyidx) );
n->set_req( copyidx, nullptr );
b->remove_node(insidx--);
b->_ihrp_index--; // Adjust the point where we go hi-pressure
b->_fhrp_index--;
continue;
}
#define DERIVED 0
// ********** Handle USES **********
bool nullcheck = false;
// Implicit null checks never use the spilled value
if( n->is_MachNullCheck() )
nullcheck = true;
if( !nullcheck ) {
// Search all inputs for a Spill-USE
JVMState* jvms = n->jvms();
uint oopoff = jvms ? jvms->oopoff() : cnt;
uint old_last = cnt - 1;
for( inpidx = 1; inpidx < cnt; inpidx++ ) {
// Derived/base pairs may be added to our inputs during this loop.
// If inpidx > old_last, then one of these new inputs is being
// handled. Skip the derived part of the pair, but process
// the base like any other input.
if (inpidx > old_last && ((inpidx - oopoff) & 1) == DERIVED) {
continue; // skip derived_debug added below
}
// Get lidx of input
uint useidx = _lrg_map.find_id(n->in(inpidx));
// Not a brand-new split, and it is a spill use
if (useidx < _lrg_map.max_lrg_id() && lrgs(useidx).reg() >= LRG::SPILL_REG) {
// Check for valid reaching DEF
slidx = lrg2reach[useidx];
Node *def = Reachblock[slidx];
assert( def != nullptr, "Using Undefined Value in Split()\n");
// (+++) %%%% remove this in favor of pre-pass in matcher.cpp
// monitor references do not care where they live, so just hook
if ( jvms && jvms->is_monitor_use(inpidx) ) {
// The effect of this clone is to drop the node out of the block,
// so that the allocator does not see it anymore, and therefore
// does not attempt to assign it a register.
def = clone_node(def, b, C);
if (def == nullptr || C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) {
return 0;
}
_lrg_map.extend(def->_idx, 0);
_cfg.map_node_to_block(def, b);
n->set_req(inpidx, def);
continue;
}
// Rematerializable? Then clone def at use site instead
// of store/load
if( def->rematerialize() ) {
int old_size = b->number_of_nodes();
def = split_Rematerialize( def, b, insidx, maxlrg, splits, slidx, lrg2reach, Reachblock, true );
if( !def ) return 0; // Bail out
insidx += b->number_of_nodes()-old_size;
}
MachNode *mach = n->is_Mach() ? n->as_Mach() : nullptr;
// Base pointers and oopmap references do not care where they live.
if ((inpidx >= oopoff) ||
(mach && mach->ideal_Opcode() == Op_AddP && inpidx == AddPNode::Base)) {
if (def->rematerialize() && lrgs(useidx)._was_spilled2) {
// This def has been rematerialized a couple of times without
// progress. It doesn't care if it lives UP or DOWN, so
// spill it down now.
int delta = split_USE(MachSpillCopyNode::BasePointerToMem, def,b,n,inpidx,maxlrg,false,false,splits,slidx);
// If it wasn't split bail
if (delta < 0) {
return 0;
}
maxlrg += delta;
insidx += delta; // Reset iterator to skip USE side split
} else {
// Just hook the def edge
n->set_req(inpidx, def);
}
if (inpidx >= oopoff) {
// After oopoff, we have derived/base pairs. We must mention all
// derived pointers here as derived/base pairs for GC. If the
// derived value is spilling and we have a copy both in Reachblock
// (called here 'def') and debug_defs[slidx] we need to mention
// both in derived/base pairs or kill one.
Node *derived_debug = debug_defs[slidx];
if( ((inpidx - oopoff) & 1) == DERIVED && // derived vs base?
mach && mach->ideal_Opcode() != Op_Halt &&
derived_debug != nullptr &&
derived_debug != def ) { // Actual 2nd value appears
// We have already set 'def' as a derived value.
// Also set debug_defs[slidx] as a derived value.
uint k;
for( k = oopoff; k < cnt; k += 2 )
if( n->in(k) == derived_debug )
break; // Found an instance of debug derived
if( k == cnt ) {// No instance of debug_defs[slidx]
// Add a derived/base pair to cover the debug info.
// We have to process the added base later since it is not
// handled yet at this point but skip derived part.
assert(((n->req() - oopoff) & 1) == DERIVED,
"must match skip condition above");
n->add_req( derived_debug ); // this will be skipped above
n->add_req( n->in(inpidx+1) ); // this will be processed
// Increment cnt to handle added input edges on
// subsequent iterations.
cnt += 2;
}
}
}
continue;
}
// Special logic for DEBUG info
if( jvms && b->_freq > BLOCK_FREQUENCY(0.5) ) {
uint debug_start = jvms->debug_start();
// If this is debug info use & there is a reaching DOWN def
if ((debug_start <= inpidx) && (debug_defs[slidx] != nullptr)) {
assert(inpidx < oopoff, "handle only debug info here");
// Just hook it in & move on
n->set_req(inpidx, debug_defs[slidx]);
// (Note that this can make two sides of a split live at the
// same time: The debug def on stack, and another def in a
// register. The GC needs to know about both of them, but any
// derived pointers after oopoff will refer to only one of the
// two defs and the GC would therefore miss the other. Thus
// this hack is only allowed for debug info which is Java state
// and therefore never a derived pointer.)
continue;
}
}
// Grab register mask info
const RegMask &dmask = def->out_RegMask();
const RegMask &umask = n->in_RegMask(inpidx);
bool is_vect = RegMask::is_vector(def->ideal_reg());
assert(inpidx < oopoff, "cannot use-split oop map info");
bool dup = UPblock[slidx];
bool uup = umask.is_UP();
// Need special logic to handle bound USES. Insert a split at this
// bound use if we can't rematerialize the def, or if we need the
// split to form a misaligned pair.
if (!umask.is_infinite_stack() &&
(int)umask.size() <= lrgs(useidx).num_regs() &&
(!def->rematerialize() ||
(!is_vect && umask.is_misaligned_pair()))) {
// These need a Split regardless of overlap or pressure
// SPLIT - NO DEF - NO CISC SPILL
int delta = split_USE(MachSpillCopyNode::Bound, def,b,n,inpidx,maxlrg,dup,false, splits,slidx);
// If it wasn't split bail
if (delta < 0) {
return 0;
}
maxlrg += delta;
insidx += delta; // Reset iterator to skip USE side split
continue;
}
if (UseFPUForSpilling && n->is_MachCall() && !uup && !dup ) {
// The use at the call can force the def down so insert
// a split before the use to allow the def more freedom.
int delta = split_USE(MachSpillCopyNode::CallUse, def,b,n,inpidx,maxlrg,dup,false, splits,slidx);
// If it wasn't split bail
if (delta < 0) {
return 0;
}
maxlrg += delta;
insidx += delta; // Reset iterator to skip USE side split
continue;
}
// Here is the logic chart which describes USE Splitting:
// 0 = false or DOWN, 1 = true or UP
//
// Overlap | DEF | USE | Action
//-------------------------------------------------------
// 0 | 0 | 0 | Copy - mem -> mem
// 0 | 0 | 1 | Split-UP - Check HRP
// 0 | 1 | 0 | Split-DOWN - Debug Info?
// 0 | 1 | 1 | Copy - reg -> reg
// 1 | 0 | 0 | Reset Input Edge (no Split)
// 1 | 0 | 1 | Split-UP - Check HRP
// 1 | 1 | 0 | Split-DOWN - Debug Info?
// 1 | 1 | 1 | Reset Input Edge (no Split)
//
// So, if (dup == uup), then overlap test determines action,
// with true being no split, and false being copy. Else,
// if DEF is DOWN, Split-UP, and check HRP to decide on
// resetting DEF. Finally if DEF is UP, Split-DOWN, with
// special handling for Debug Info.
if( dup == uup ) {
if( dmask.overlap(umask) ) {
// Both are either up or down, and there is overlap, No Split
n->set_req(inpidx, def);
}
else { // Both are either up or down, and there is no overlap
if( dup ) { // If UP, reg->reg copy
// COPY ACROSS HERE - NO DEF - NO CISC SPILL
int delta = split_USE(MachSpillCopyNode::RegToReg, def,b,n,inpidx,maxlrg,false,false, splits,slidx);
// If it wasn't split bail
if (delta < 0) {
return 0;
}
maxlrg += delta;
insidx += delta; // Reset iterator to skip USE side split
}
else { // DOWN, mem->mem copy
// COPY UP & DOWN HERE - NO DEF - NO CISC SPILL
// First Split-UP to move value into Register
uint def_ideal = def->ideal_reg();
const RegMask* tmp_rm = Matcher::idealreg2regmask[def_ideal];
Node *spill = new MachSpillCopyNode(MachSpillCopyNode::MemToReg, def, dmask, *tmp_rm);
insert_proj( b, insidx, spill, maxlrg );
maxlrg++; insidx++;
// Then Split-DOWN as if previous Split was DEF
int delta = split_USE(MachSpillCopyNode::RegToMem, spill,b,n,inpidx,maxlrg,false,false, splits,slidx);
// If it wasn't split bail
if (delta < 0) {
return 0;
}
maxlrg += delta;
insidx += delta; // Reset iterator to skip USE side splits
}
} // End else no overlap
} // End if dup == uup
// dup != uup, so check dup for direction of Split
else {
if( dup ) { // If UP, Split-DOWN and check Debug Info
// If this node is already a SpillCopy, just patch the edge
// except the case of spilling to stack.
if( n->is_SpillCopy() ) {
ResourceMark rm(C->regmask_arena());
RegMask tmp_rm(umask, C->regmask_arena());
tmp_rm.subtract(Matcher::STACK_ONLY_mask);
if( dmask.overlap(tmp_rm) ) {
if( def != n->in(inpidx) ) {
n->set_req(inpidx, def);
}
continue;
}
}
// COPY DOWN HERE - NO DEF - NO CISC SPILL
int delta = split_USE(MachSpillCopyNode::RegToMem, def,b,n,inpidx,maxlrg,false,false, splits,slidx);
// If it wasn't split bail
if (delta < 0) {
return 0;
}
maxlrg += delta;
insidx += delta; // Reset iterator to skip USE side split
// Check for debug-info split. Capture it for later
// debug splits of the same value
if (jvms && jvms->debug_start() <= inpidx && inpidx < oopoff)
debug_defs[slidx] = n->in(inpidx);
}
else { // DOWN, Split-UP and check register pressure
if( is_high_pressure( b, &lrgs(useidx), insidx ) ) {
// COPY UP HERE - NO DEF - CISC SPILL
int delta = split_USE(MachSpillCopyNode::MemToReg, def,b,n,inpidx,maxlrg,true,true, splits,slidx);
// If it wasn't split bail
if (delta < 0) {
return 0;
}
maxlrg += delta;
insidx += delta; // Reset iterator to skip USE side split
} else { // LRP
// COPY UP HERE - WITH DEF - NO CISC SPILL
int delta = split_USE(MachSpillCopyNode::MemToReg, def,b,n,inpidx,maxlrg,true,false, splits,slidx);
// If it wasn't split bail
if (delta < 0) {
return 0;
}
// Flag this lift-up in a low-pressure block as
// already-spilled, so if it spills again it will
// spill hard (instead of not spilling hard and
// coalescing away).
set_was_spilled(n->in(inpidx));
// Since this is a new DEF, update Reachblock & UP
Reachblock[slidx] = n->in(inpidx);
UPblock[slidx] = true;
maxlrg += delta;
insidx += delta; // Reset iterator to skip USE side split
}
} // End else DOWN
} // End dup != uup
} // End if Spill USE
} // End For All Inputs
} // End If not nullcheck
// ********** Handle DEFS **********
// DEFS either Split DOWN in HRP regions or when the LRG is bound, or
// just reset the Reaches info in LRP regions. DEFS must always update
// UP info.
if( deflrg.reg() >= LRG::SPILL_REG ) { // Spilled?
uint slidx = lrg2reach[defidx];
// Add to defs list for later assignment of new live range number
defs.push(n);
// Set a flag on the Node indicating it has already spilled.
// Only do it for capacity spills not conflict spills.
if( !deflrg._direct_conflict )
set_was_spilled(n);
assert(!n->is_Phi(),"Cannot insert Phi into DEFS list");
// Grab UP info for DEF
const RegMask &dmask = n->out_RegMask();
bool defup = dmask.is_UP();
uint ireg = n->ideal_reg();
bool is_vect = RegMask::is_vector(ireg);
// Only split at Def if this is a HRP block or bound (and spilled once)
if( !n->rematerialize() &&
(((dmask.is_bound(ireg) || (!is_vect && dmask.is_misaligned_pair())) &&
(deflrg._direct_conflict || deflrg._must_spill)) ||
// Check for LRG being up in a register and we are inside a high
// pressure area. Spill it down immediately.
(defup && is_high_pressure(b,&deflrg,insidx) && !n->is_SpillCopy())) ) {
assert( !n->rematerialize(), "" );
// Do a split at the def site.
maxlrg = split_DEF( n, b, insidx, maxlrg, Reachblock, debug_defs, splits, slidx );
// If it wasn't split bail
if (!maxlrg) {
return 0;
}
// Split DEF's Down
UPblock[slidx] = 0;
#ifndef PRODUCT
// DEBUG
if( trace_spilling() ) {
tty->print("\nNew Split DOWN DEF of Spill Idx ");
tty->print("%d, UP %d:\n",slidx,false);
n->dump();
}
#endif
}
else { // Neither bound nor HRP, must be LRP
// otherwise, just record the def
Reachblock[slidx] = n;
// UP should come from the outRegmask() of the DEF
UPblock[slidx] = defup;
// Update debug list of reaching down definitions, kill if DEF is UP
debug_defs[slidx] = defup ? nullptr : n;
#ifndef PRODUCT
// DEBUG
if( trace_spilling() ) {
tty->print("\nNew DEF of Spill Idx ");
tty->print("%d, UP %d:\n",slidx,defup);
n->dump();
}
#endif
} // End else LRP
} // End if spill def
// ********** Split Left Over Mem-Mem Moves **********
// Check for mem-mem copies and split them now. Do not do this
// to copies about to be spilled; they will be Split shortly.
if (copyidx) {
Node *use = n->in(copyidx);
uint useidx = _lrg_map.find_id(use);
if (useidx < _lrg_map.max_lrg_id() && // This is not a new split
OptoReg::is_stack(deflrg.reg()) &&
deflrg.reg() < LRG::SPILL_REG ) { // And DEF is from stack
LRG &uselrg = lrgs(useidx);
if( OptoReg::is_stack(uselrg.reg()) &&
uselrg.reg() < LRG::SPILL_REG && // USE is from stack
deflrg.reg() != uselrg.reg() ) { // Not trivially removed
uint def_ideal_reg = n->bottom_type()->ideal_reg();
const RegMask &def_rm = *Matcher::idealreg2regmask[def_ideal_reg];
const RegMask &use_rm = n->in_RegMask(copyidx);
if( def_rm.overlap(use_rm) && n->is_SpillCopy() ) { // Bug 4707800, 'n' may be a storeSSL
if (C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { // Check when generating nodes
return 0;
}
Node *spill = new MachSpillCopyNode(MachSpillCopyNode::MemToReg, use,use_rm,def_rm);
n->set_req(copyidx,spill);
n->as_MachSpillCopy()->set_in_RegMask(def_rm);
// Put the spill just before the copy
insert_proj( b, insidx++, spill, maxlrg++ );
}
}
}
}
} // End For All Instructions in Block - Non-PHI Pass
// Check if each LRG is live out of this block so as not to propagate
// beyond the last use of a LRG.
for( slidx = 0; slidx < spill_cnt; slidx++ ) {
uint defidx = lidxs.at(slidx);
IndexSet *liveout = _live->live(b);
if( !liveout->member(defidx) ) {
#ifdef ASSERT
if (VerifyRegisterAllocator) {
// The index defidx is not live. Check the liveout array to ensure that
// it contains no members which compress to defidx. Finding such an
// instance may be a case to add liveout adjustment in compress_uf_map().
// See 5063219.
if (!liveout->is_empty()) {
uint member;
IndexSetIterator isi(liveout);
while ((member = isi.next()) != 0) {
assert(defidx != _lrg_map.find_const(member), "Live out member has not been compressed");
}
}
}
#endif
Reachblock[slidx] = nullptr;
} else {
assert(Reachblock[slidx] != nullptr,"No reaching definition for liveout value");
}
}
#ifndef PRODUCT
if( trace_spilling() )
b->dump();
#endif
} // End For All Blocks
//----------PASS 2----------
// Reset all DEF live range numbers here
for( insidx = 0; insidx < defs.size(); insidx++ ) {
// Grab the def
n1 = defs.at(insidx);
// Set new lidx for DEF
new_lrg(n1, maxlrg++);
}
//----------Phi Node Splitting----------
// Clean up a phi here, and assign a new live range number
// Cycle through this block's predecessors, collecting Reaches
// info for each spilled LRG and update edges.
// Walk the phis list to patch inputs, split phis, and name phis
uint lrgs_before_phi_split = maxlrg;
for( insidx = 0; insidx < phis.size(); insidx++ ) {
Node *phi = phis.at(insidx);
assert(phi->is_Phi(),"This list must only contain Phi Nodes");
Block *b = _cfg.get_block_for_node(phi);
// Grab the live range number
uint lidx = _lrg_map.find_id(phi);
uint slidx = lrg2reach[lidx];
// Update node to lidx map
new_lrg(phi, maxlrg++);
// Get PASS1's up/down decision for the block.
int phi_up = !!UP_entry[slidx]->test(b->_pre_order);
// Force down if double-spilling live range
if( lrgs(lidx)._was_spilled1 )
phi_up = false;
// When splitting a Phi we an split it normal or "inverted".
// An inverted split makes the splits target the Phi's UP/DOWN
// sense inverted; then the Phi is followed by a final def-side
// split to invert back. It changes which blocks the spill code
// goes in.
// Walk the predecessor blocks and assign the reaching def to the Phi.
// Split Phi nodes by placing USE side splits wherever the reaching
// DEF has the wrong UP/DOWN value.
for( uint i = 1; i < b->num_preds(); i++ ) {
// Get predecessor block pre-order number
Block *pred = _cfg.get_block_for_node(b->pred(i));
pidx = pred->_pre_order;
// Grab reaching def
Node *def = Reaches[pidx][slidx];
Node** Reachblock = Reaches[pidx];
assert( def, "must have reaching def" );
// If input up/down sense and reg-pressure DISagree
if (def->rematerialize()) {
// Place the rematerialized node above any MSCs created during
// phi node splitting. end_idx points at the insertion point
// so look at the node before it.
int insert = pred->end_idx();
while (insert >= 1 &&
pred->get_node(insert - 1)->is_SpillCopy() &&
_lrg_map.find(pred->get_node(insert - 1)) >= lrgs_before_phi_split) {
insert--;
}
def = split_Rematerialize(def, pred, insert, maxlrg, splits, slidx, lrg2reach, Reachblock, false);
if (!def) {
return 0; // Bail out
}
}
// Update the Phi's input edge array
phi->set_req(i,def);
// Grab the UP/DOWN sense for the input
u1 = UP[pidx][slidx];
if( u1 != (phi_up != 0)) {
int delta = split_USE(MachSpillCopyNode::PhiLocationDifferToInputLocation, def, b, phi, i, maxlrg, !u1, false, splits,slidx);
// If it wasn't split bail
if (delta < 0) {
return 0;
}
maxlrg += delta;
}
} // End for all inputs to the Phi
} // End for all Phi Nodes
// Update _maxlrg to save Union asserts
_lrg_map.set_max_lrg_id(maxlrg);
//----------PASS 3----------
// Pass over all Phi's to union the live ranges
for( insidx = 0; insidx < phis.size(); insidx++ ) {
Node *phi = phis.at(insidx);
assert(phi->is_Phi(),"This list must only contain Phi Nodes");
// Walk all inputs to Phi and Union input live range with Phi live range
for( uint i = 1; i < phi->req(); i++ ) {
// Grab the input node
Node *n = phi->in(i);
assert(n, "node should exist");
uint lidx = _lrg_map.find(n);
uint pidx = _lrg_map.find(phi);
if (lidx < pidx) {
Union(n, phi);
}
else if(lidx > pidx) {
Union(phi, n);
}
} // End for all inputs to the Phi Node
} // End for all Phi Nodes
// Now union all two address instructions
for (insidx = 0; insidx < defs.size(); insidx++) {
// Grab the def
n1 = defs.at(insidx);
// Set new lidx for DEF & handle 2-addr instructions
if (n1->is_Mach() && ((twoidx = n1->as_Mach()->two_adr()) != 0)) {
assert(_lrg_map.find(n1->in(twoidx)) < maxlrg,"Assigning bad live range index");
// Union the input and output live ranges
uint lr1 = _lrg_map.find(n1);
uint lr2 = _lrg_map.find(n1->in(twoidx));
if (lr1 < lr2) {
Union(n1, n1->in(twoidx));
}
else if (lr1 > lr2) {
Union(n1->in(twoidx), n1);
}
} // End if two address
} // End for all defs
// DEBUG
#ifdef ASSERT
// Validate all live range index assignments
for (bidx = 0; bidx < _cfg.number_of_blocks(); bidx++) {
b = _cfg.get_block(bidx);
for (insidx = 0; insidx <= b->end_idx(); insidx++) {
Node *n = b->get_node(insidx);
uint defidx = _lrg_map.find(n);
assert(defidx < _lrg_map.max_lrg_id(), "Bad live range index in Split");
assert(defidx < maxlrg,"Bad live range index in Split");
}
}
// Issue a warning if splitting made no progress
int noprogress = 0;
for (slidx = 0; slidx < spill_cnt; slidx++) {
if (PrintOpto && WizardMode && splits.at(slidx) == 0) {
tty->print_cr("Failed to split live range %d", lidxs.at(slidx));
//BREAKPOINT;
}
else {
noprogress++;
}
}
if(!noprogress) {
tty->print_cr("Failed to make progress in Split");
//BREAKPOINT;
}
#endif
// Return updated count of live ranges
return maxlrg;
}