/*
* Find the immediate dominator of each block using Cooper, Harvey, and
* Kennedy's "A Simple, Fast Dominance Algorithm", returned as a vector
* of postorder ids, indexed by postorder id.
*/
IdomVector findDominators(const BlockList& blocks) {
assert(isRPOSorted(blocks));
// Calculate immediate dominators with the iterative two-finger algorithm.
// When it terminates, idom[post-id] will contain the post-id of the
// immediate dominator of each block. idom[start] will be -1. This is
// the general algorithm but it will only loop twice for loop-free graphs.
auto const num_blocks = blocks.size();
IdomVector idom(num_blocks, -1);
auto start = blocks.begin();
int start_id = (*start)->postId();
idom[start_id] = start_id;
start++;
for (bool changed = true; changed; ) {
changed = false;
// for each block after start, in reverse postorder
for (auto it = start; it != blocks.end(); it++) {
Block* block = *it;
int b = block->postId();
// new_idom = any already-processed predecessor
auto edge_it = block->preds().begin();
int new_idom = edge_it->from()->postId();
while (idom[new_idom] == -1) new_idom = (++edge_it)->from()->postId();
// for all other already-processed predecessors p of b
for (auto& edge : block->preds()) {
auto p = edge.from()->postId();
if (p != new_idom && idom[p] != -1) {
// find earliest common predecessor of p and new_idom
// (higher postIds are earlier in flow and in dom-tree).
int b1 = p, b2 = new_idom;
do {
while (b1 < b2) b1 = idom[b1];
while (b2 < b1) b2 = idom[b2];
} while (b1 != b2);
new_idom = b1;
}
}
if (idom[b] != new_idom) {
idom[b] = new_idom;
changed = true;
}
}
}
idom[start_id] = -1; // start has no idom.
return idom;
}
/*
* Find the immediate dominator of each block using Cooper, Harvey, and
* Kennedy's "A Simple, Fast Dominance Algorithm", returned as a vector
* of Block*, indexed by block. IdomVector[b] == nullptr if b has no
* dominator. This is the case for the entry block and any blocks not
* reachable from the entry block.
*/
IdomVector findDominators(const IRUnit& unit, const BlocksWithIds& blockIds) {
auto& blocks = blockIds.blocks;
auto& postIds = blockIds.ids;
// Calculate immediate dominators with the iterative two-finger algorithm.
// When it terminates, idom[post-id] will contain the post-id of the
// immediate dominator of each block. idom[start] will be -1. This is
// the general algorithm but it will only loop twice for loop-free graphs.
IdomVector idom(unit, nullptr);
auto start = blocks.begin();
auto entry = *start;
idom[entry] = entry;
start++;
for (bool changed = true; changed; ) {
changed = false;
// for each block after start, in reverse postorder
for (auto it = start; it != blocks.end(); it++) {
Block* block = *it;
// p1 = any already-processed predecessor
auto predIter = block->preds().begin();
auto predEnd = block->preds().end();
auto p1 = predIter->inst()->block();
while (!idom[p1]) p1 = (++predIter)->inst()->block();
// for all other already-processed predecessors p2 of block
for (++predIter; predIter != predEnd; ++predIter) {
auto p2 = predIter->inst()->block();
if (p2 == p1 || !idom[p2]) continue;
// find earliest common predecessor of p1 and p2
// (higher postIds are earlier in flow and in dom-tree).
do {
while (postIds[p1] < postIds[p2]) p1 = idom[p1];
while (postIds[p2] < postIds[p1]) p2 = idom[p2];
} while (p1 != p2);
}
if (idom[block] != p1) {
idom[block] = p1;
changed = true;
}
}
}
idom[entry] = nullptr; // entry has no dominator.
return idom;
}
//------------------------------create_new_if_for_predicate------------------------
// create a new if above the uct_if_pattern for the predicate to be promoted.
//
// before after
// ---------- ----------
// ctrl ctrl
// | |
// | |
// v v
// iff new_iff
// / \ / \
// / \ / \
// v v v v
// uncommon_proj cont_proj if_uct if_cont
// \ | | | |
// \ | | | |
// v v v | v
// rgn loop | iff
// | | / \
// | | / \
// v | v v
// uncommon_trap | uncommon_proj cont_proj
// \ \ | |
// \ \ | |
// v v v v
// rgn loop
// |
// |
// v
// uncommon_trap
//
//
// We will create a region to guard the uct call if there is no one there.
// The true projecttion (if_cont) of the new_iff is returned.
// This code is also used to clone predicates to clonned loops.
ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
Deoptimization::DeoptReason reason) {
assert(is_uncommon_trap_if_pattern(cont_proj, reason), "must be a uct if pattern!");
IfNode* iff = cont_proj->in(0)->as_If();
ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
Node *rgn = uncommon_proj->unique_ctrl_out();
assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
uint proj_index = 1; // region's edge corresponding to uncommon_proj
if (!rgn->is_Region()) { // create a region to guard the call
assert(rgn->is_Call(), "must be call uct");
CallNode* call = rgn->as_Call();
IdealLoopTree* loop = get_loop(call);
rgn = new (C) RegionNode(1);
rgn->add_req(uncommon_proj);
register_control(rgn, loop, uncommon_proj);
_igvn.hash_delete(call);
call->set_req(0, rgn);
// When called from beautify_loops() idom is not constructed yet.
if (_idom != NULL) {
set_idom(call, rgn, dom_depth(rgn));
}
} else {
// Find region's edge corresponding to uncommon_proj
for (; proj_index < rgn->req(); proj_index++)
if (rgn->in(proj_index) == uncommon_proj) break;
assert(proj_index < rgn->req(), "sanity");
}
Node* entry = iff->in(0);
if (new_entry != NULL) {
// Clonning the predicate to new location.
entry = new_entry;
}
// Create new_iff
IdealLoopTree* lp = get_loop(entry);
IfNode *new_iff = iff->clone()->as_If();
new_iff->set_req(0, entry);
register_control(new_iff, lp, entry);
Node *if_cont = new (C) IfTrueNode(new_iff);
Node *if_uct = new (C) IfFalseNode(new_iff);
if (cont_proj->is_IfFalse()) {
// Swap
Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp;
}
register_control(if_cont, lp, new_iff);
register_control(if_uct, get_loop(rgn), new_iff);
// if_uct to rgn
_igvn.hash_delete(rgn);
rgn->add_req(if_uct);
// When called from beautify_loops() idom is not constructed yet.
if (_idom != NULL) {
Node* ridom = idom(rgn);
Node* nrdom = dom_lca(ridom, new_iff);
set_idom(rgn, nrdom, dom_depth(rgn));
}
// If rgn has phis add new edges which has the same
// value as on original uncommon_proj pass.
assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last");
bool has_phi = false;
for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) {
Node* use = rgn->fast_out(i);
if (use->is_Phi() && use->outcnt() > 0) {
assert(use->in(0) == rgn, "");
_igvn.rehash_node_delayed(use);
use->add_req(use->in(proj_index));
has_phi = true;
}
}
assert(!has_phi || rgn->req() > 3, "no phis when region is created");
if (new_entry == NULL) {
// Attach if_cont to iff
_igvn.hash_delete(iff);
iff->set_req(0, if_cont);
if (_idom != NULL) {
set_idom(iff, if_cont, dom_depth(iff));
}
}
return if_cont->as_Proj();
}
//------------------------------do_split_if------------------------------------
// Found an If getting its condition-code input from a Phi in the same block.
// Split thru the Region.
void PhaseIdealLoop::do_split_if( Node *iff ) {
#ifndef PRODUCT
if( PrintOpto && VerifyLoopOptimizations )
tty->print_cr("Split-if");
#endif
C->set_major_progress();
Node *region = iff->in(0);
Node *region_dom = idom(region);
// We are going to clone this test (and the control flow with it) up through
// the incoming merge point. We need to empty the current basic block.
// Clone any instructions which must be in this block up through the merge
// point.
DUIterator i, j;
bool progress = true;
while (progress) {
progress = false;
for (i = region->outs(); region->has_out(i); i++) {
Node* n = region->out(i);
if( n == region ) continue;
// The IF to be split is OK.
if( n == iff ) continue;
if( !n->is_Phi() ) { // Found pinned memory op or such
if (split_up(n, region, iff)) {
i = region->refresh_out_pos(i);
progress = true;
}
continue;
}
assert( n->in(0) == region, "" );
// Recursively split up all users of a Phi
for (j = n->outs(); n->has_out(j); j++) {
Node* m = n->out(j);
// If m is dead, throw it away, and declare progress
if (_nodes[m->_idx] == NULL) {
_igvn.remove_dead_node(m);
// fall through
}
else if (m != iff && split_up(m, region, iff)) {
// fall through
} else {
continue;
}
// Something unpredictable changed.
// Tell the iterators to refresh themselves, and rerun the loop.
i = region->refresh_out_pos(i);
j = region->refresh_out_pos(j);
progress = true;
}
}
}
// Now we have no instructions in the block containing the IF.
// Split the IF.
Node *new_iff = split_thru_region( iff, region );
// Replace both uses of 'new_iff' with Regions merging True/False
// paths. This makes 'new_iff' go dead.
Node *old_false, *old_true;
Node *new_false, *new_true;
for (DUIterator_Last j2min, j2 = iff->last_outs(j2min); j2 >= j2min; --j2) {
Node *ifp = iff->last_out(j2);
assert( ifp->Opcode() == Op_IfFalse || ifp->Opcode() == Op_IfTrue, "" );
ifp->set_req(0, new_iff);
Node *ifpx = split_thru_region( ifp, region );
// Replace 'If' projection of a Region with a Region of
// 'If' projections.
ifpx->set_req(0, ifpx); // A TRUE RegionNode
// Setup dominator info
set_idom(ifpx, region_dom, dom_depth(region_dom) + 1);
// Check for splitting loop tails
if( get_loop(iff)->tail() == ifp )
get_loop(iff)->_tail = ifpx;
// Replace in the graph with lazy-update mechanism
new_iff->set_req(0, new_iff); // hook self so it does not go dead
lazy_replace_proj( ifp, ifpx );
new_iff->set_req(0, region);
// Record bits for later xforms
if( ifp->Opcode() == Op_IfFalse ) {
old_false = ifp;
new_false = ifpx;
} else {
old_true = ifp;
new_true = ifpx;
}
}
_igvn.remove_dead_node(new_iff);
// Lazy replace IDOM info with the region's dominator
lazy_replace( iff, region_dom );
// Now make the original merge point go dead, by handling all its uses.
small_cache region_cache;
// Preload some control flow in region-cache
region_cache.lru_insert( new_false, new_false );
region_cache.lru_insert( new_true , new_true );
// Now handle all uses of the splitting block
for (DUIterator_Last kmin, k = region->last_outs(kmin); k >= kmin; --k) {
Node* phi = region->last_out(k);
if( !phi->in(0) ) { // Dead phi? Remove it
_igvn.remove_dead_node(phi);
continue;
}
assert( phi->in(0) == region, "" );
if( phi == region ) { // Found the self-reference
phi->set_req(0, NULL);
continue; // Break the self-cycle
}
// Expected common case: Phi hanging off of Region
if( phi->is_Phi() ) {
// Need a per-def cache. Phi represents a def, so make a cache
small_cache phi_cache;
// Inspect all Phi uses to make the Phi go dead
for (DUIterator_Last lmin, l = phi->last_outs(lmin); l >= lmin; --l) {
Node* use = phi->last_out(l);
// Compute the new DEF for this USE. New DEF depends on the path
// taken from the original DEF to the USE. The new DEF may be some
// collection of PHI's merging values from different paths. The Phis
// inserted depend only on the location of the USE. We use a
// 2-element cache to handle multiple uses from the same block.
handle_use( use, phi, &phi_cache, region_dom, new_false, new_true, old_false, old_true );
} // End of while phi has uses
// Because handle_use might relocate region->_out,
// we must refresh the iterator.
k = region->last_outs(kmin);
// Remove the dead Phi
_igvn.remove_dead_node( phi );
} else {
// Random memory op guarded by Region. Compute new DEF for USE.
handle_use( phi, region, ®ion_cache, region_dom, new_false, new_true, old_false, old_true );
}
} // End of while merge point has phis
// Any leftover bits in the splitting block must not have depended on local
// Phi inputs (these have already been split-up). Hence it's safe to hoist
// these guys to the dominating point.
lazy_replace( region, region_dom );
#ifndef PRODUCT
if( VerifyLoopOptimizations ) verify();
#endif
}
// We must be at the merge point which post-dominates 'new_false' and
// 'new_true'. Figure out which edges into the RegionNode eventually lead up
// to false and which to true. Put in a PhiNode to merge values; plug in
// the appropriate false-arm or true-arm values. If some path leads to the
// original IF, then insert a Phi recursively.
Node *PhaseIdealLoop::spinup( Node *iff_dom, Node *new_false, Node *new_true, Node *use_blk, Node *def, small_cache *cache ) {
if (use_blk->is_top()) // Handle dead uses
return use_blk;
Node *prior_n = (Node*)0xdeadbeef;
Node *n = use_blk; // Get path input
assert( use_blk != iff_dom, "" );
// Here's the "spinup" the dominator tree loop. Do a cache-check
// along the way, in case we've come this way before.
while( n != iff_dom ) { // Found post-dominating point?
prior_n = n;
n = idom(n); // Search higher
Node *s = cache->probe( prior_n ); // Check cache
if( s ) return s; // Cache hit!
}
Node *phi_post;
if( prior_n == new_false || prior_n == new_true ) {
phi_post = def->clone();
phi_post->set_req(0, prior_n );
register_new_node(phi_post, prior_n);
} else {
// This method handles both control uses (looking for Regions) or data
// uses (looking for Phis). If looking for a control use, then we need
// to insert a Region instead of a Phi; however Regions always exist
// previously (the hash_find_insert below would always hit) so we can
// return the existing Region.
if( def->is_CFG() ) {
phi_post = prior_n; // If looking for CFG, return prior
} else {
assert( def->is_Phi(), "" );
assert( prior_n->is_Region(), "must be a post-dominating merge point" );
// Need a Phi here
phi_post = PhiNode::make_blank(prior_n, def);
// Search for both true and false on all paths till find one.
for( uint i = 1; i < phi_post->req(); i++ ) // For all paths
phi_post->init_req( i, spinup( iff_dom, new_false, new_true, prior_n->in(i), def, cache ) );
Node *t = _igvn.hash_find_insert(phi_post);
if( t ) { // See if we already have this one
// phi_post will not be used, so kill it
_igvn.remove_dead_node(phi_post);
phi_post->destruct();
phi_post = t;
} else {
register_new_node( phi_post, prior_n );
}
}
}
// Update cache everywhere
prior_n = (Node*)0xdeadbeef; // Reset IDOM walk
n = use_blk; // Get path input
// Spin-up the idom tree again, basically doing path-compression.
// Insert cache entries along the way, so that if we ever hit this
// point in the IDOM tree again we'll stop immediately on a cache hit.
while( n != iff_dom ) { // Found post-dominating point?
prior_n = n;
n = idom(n); // Search higher
cache->lru_insert( prior_n, phi_post ); // Fill cache
} // End of while not gone high enough
return phi_post;
}
