Change the btree algorithm to split nodes bottom-up instead of top-down.

This results in an (IMO) simpler algorithm, results in fewer splits, and
is more amenable to delta handling (there is no reason to mutate the tree
at all when adding a delta to a position that already exists in the tree).


git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@49609 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Chris Lattner 2008-04-13 08:52:45 +00:00
Родитель b169e903c5
Коммит 3b7ff0d143
1 изменённых файлов: 121 добавлений и 83 удалений

Просмотреть файл

@ -14,6 +14,7 @@
#include "clang/Rewrite/DeltaTree.h"
#include "llvm/Support/Casting.h"
#include <cstring>
#include <cstdio>
using namespace clang;
using llvm::cast;
using llvm::dyn_cast;
@ -60,7 +61,6 @@ namespace {
struct InsertResult {
DeltaTreeNode *LHS, *RHS;
SourceDelta Split;
InsertResult() : LHS(0) {}
};
@ -109,13 +109,15 @@ namespace {
return Values[i];
}
/// DoInsertion - Do an insertion of the specified FileIndex/Delta pair into
/// this node. If insertion is easy, do it and return false. Otherwise,
/// split the node, populate InsertRes with info about the split, and return
/// true.
bool DoInsertion(unsigned FileIndex, int Delta, InsertResult *InsertRes);
void DoSplit(InsertResult &InsertRes);
/// AddDeltaNonFull - Add a delta to this tree and/or it's children, knowing
/// that this node is not currently full.
void AddDeltaNonFull(unsigned FileIndex, int Delta);
/// RecomputeFullDeltaLocally - Recompute the FullDelta field by doing a
/// local walk over our contained deltas.
void RecomputeFullDeltaLocally();
@ -145,6 +147,15 @@ namespace {
Children[0] = FirstChild;
}
DeltaTreeInteriorNode(const InsertResult &IR)
: DeltaTreeNode(false /*nonleaf*/) {
Children[0] = IR.LHS;
Children[1] = IR.RHS;
Values[0] = IR.Split;
FullDelta = IR.LHS->getFullDelta()+IR.RHS->getFullDelta()+IR.Split.Delta;
NumValuesUsed = 1;
}
const DeltaTreeNode *getChild(unsigned i) const {
assert(i < getNumValuesUsed()+1 && "Invalid child");
return Children[i];
@ -156,8 +167,6 @@ namespace {
static inline bool classof(const DeltaTreeInteriorNode *) { return true; }
static inline bool classof(const DeltaTreeNode *N) { return !N->isLeaf(); }
private:
void SplitChild(unsigned ChildNo);
};
}
@ -182,12 +191,12 @@ void DeltaTreeNode::RecomputeFullDeltaLocally() {
FullDelta = NewFullDelta;
}
/// AddDeltaNonFull - Add a delta to this tree and/or it's children, knowing
/// that this node is not currently full.
void DeltaTreeNode::AddDeltaNonFull(unsigned FileIndex, int Delta) {
assert(!isFull() && "AddDeltaNonFull on a full tree?");
/// DoInsertion - Do an insertion of the specified FileIndex/Delta pair into
/// this node. If insertion is easy, do it and return false. Otherwise,
/// split the node, populate InsertRes with info about the split, and return
/// true.
bool DeltaTreeNode::DoInsertion(unsigned FileIndex, int Delta,
InsertResult *InsertRes) {
// Maintain full delta for this node.
FullDelta += Delta;
@ -197,77 +206,109 @@ void DeltaTreeNode::AddDeltaNonFull(unsigned FileIndex, int Delta) {
++i;
// If we found an a record for exactly this file index, just merge this
// value into the preexisting record and finish early.
// value into the pre-existing record and finish early.
if (i != e && getValue(i).FileLoc == FileIndex) {
// NOTE: Delta could drop to zero here. This means that the next delta
// entry is useless and could be removed. Supporting erases is
// significantly more complex though, so we just leave an entry with
// Delta=0 in the tree.
// NOTE: Delta could drop to zero here. This means that the delta entry is
// useless and could be removed. Supporting erases is more complex than
// leaving an entry with Delta=0, so we just leave an entry with Delta=0 in
// the tree.
Values[i].Delta += Delta;
return;
return false;
}
if (DeltaTreeInteriorNode *IN = dyn_cast<DeltaTreeInteriorNode>(this)) {
// Insertion into an interior node propagates the value down to a child.
DeltaTreeNode *Child = IN->getChild(i);
// If the child tree is full, split it, pulling an element up into our
// node.
if (Child->isFull()) {
IN->SplitChild(i);
SourceDelta &MedianVal = getValue(i);
// If the median value we pulled up is exactly our insert position, add
// the delta and return.
if (MedianVal.FileLoc == FileIndex) {
MedianVal.Delta += Delta;
return;
}
// If the median value pulled up is less than our current search point,
// include those deltas and search down the RHS now.
if (MedianVal.FileLoc < FileIndex)
Child = IN->getChild(i+1);
// Otherwise, we found an insertion point, and we know that the value at the
// specified index is > FileIndex. Handle the leaf case first.
if (isLeaf()) {
if (!isFull()) {
// For an insertion into a non-full leaf node, just insert the value in
// its sorted position. This requires moving later values over.
if (i != e)
memmove(&Values[i+1], &Values[i], sizeof(Values[0])*(e-i));
Values[i] = SourceDelta::get(FileIndex, Delta);
++NumValuesUsed;
return false;
}
Child->AddDeltaNonFull(FileIndex, Delta);
} else {
// For an insertion into a non-full leaf node, just insert the value in
// its sorted position. This requires moving later values over.
if (i != e)
memmove(&Values[i+1], &Values[i], sizeof(Values[0])*(e-i));
Values[i] = SourceDelta::get(FileIndex, Delta);
++NumValuesUsed;
// Otherwise, if this is leaf is full, split the node at its median, insert
// the value into one of the children, and return the result.
assert(InsertRes && "No result location specified");
DoSplit(*InsertRes);
if (InsertRes->Split.FileLoc > FileIndex)
InsertRes->LHS->DoInsertion(FileIndex, Delta, 0 /*can't fail*/);
else
InsertRes->RHS->DoInsertion(FileIndex, Delta, 0 /*can't fail*/);
return true;
}
// Otherwise, this is an interior node. Send the request down the tree.
DeltaTreeInteriorNode *IN = cast<DeltaTreeInteriorNode>(this);
if (!IN->Children[i]->DoInsertion(FileIndex, Delta, InsertRes))
return false; // If there was space in the child, just return.
// Okay, this split the subtree, producing a new value and two children to
// insert here. If this node is non-full, we can just insert it directly.
if (!isFull()) {
// Now that we have two nodes and a new element, insert the perclated value
// into ourself by moving all the later values/children down, then inserting
// the new one.
if (i != e)
memmove(&IN->Children[i+2], &IN->Children[i+1],
(e-i)*sizeof(IN->Children[0]));
IN->Children[i] = InsertRes->LHS;
IN->Children[i+1] = InsertRes->RHS;
if (e != i)
memmove(&Values[i+1], &Values[i], (e-i)*sizeof(Values[0]));
Values[i] = InsertRes->Split;
++NumValuesUsed;
return false;
}
// Finally, if this interior node was full and a node is percolated up, split
// ourself and return that up the chain. Start by saving all our info to
// avoid having the split clobber it.
IN->Children[i] = InsertRes->LHS;
DeltaTreeNode *SubRHS = InsertRes->RHS;
SourceDelta SubSplit = InsertRes->Split;
// Do the split.
DoSplit(*InsertRes);
// Figure out where to insert SubRHS/NewSplit.
DeltaTreeInteriorNode *InsertSide;
if (SubSplit.FileLoc < InsertRes->Split.FileLoc)
InsertSide = cast<DeltaTreeInteriorNode>(InsertRes->LHS);
else
InsertSide = cast<DeltaTreeInteriorNode>(InsertRes->RHS);
// We now have a non-empty interior node 'InsertSide' to insert
// SubRHS/SubSplit into. Find out where to insert SubSplit.
// Find the insertion point, the first delta whose index is >SubSplit.FileLoc.
i = 0; e = InsertSide->getNumValuesUsed();
while (i != e && SubSplit.FileLoc > InsertSide->getValue(i).FileLoc)
++i;
// Now we know that i is the place to insert the split value into. Insert it
// and the child right after it.
if (i != e)
memmove(&InsertSide->Children[i+2], &InsertSide->Children[i+1],
(e-i)*sizeof(IN->Children[0]));
InsertSide->Children[i+1] = SubRHS;
if (e != i)
memmove(&InsertSide->Values[i+1], &InsertSide->Values[i],
(e-i)*sizeof(Values[0]));
InsertSide->Values[i] = SubSplit;
++InsertSide->NumValuesUsed;
InsertSide->FullDelta += SubSplit.Delta + SubRHS->getFullDelta();
return true;
}
/// SplitChild - At this point, we know that the current node is not full and
/// that the specified child of this node is. Split the child in half at its
/// median, propagating one value up into us. Child may be either an interior
/// or leaf node.
void DeltaTreeInteriorNode::SplitChild(unsigned ChildNo) {
DeltaTreeNode *Child = getChild(ChildNo);
assert(!isFull() && Child->isFull() && "Inconsistent constraints");
InsertResult SplitRes;
Child->DoSplit(SplitRes);
// Now that we have two nodes and a new element, insert the median value
// into ourself by moving all the later values/children down, then inserting
// the new one.
if (getNumValuesUsed() != ChildNo)
memmove(&Children[ChildNo+2], &Children[ChildNo+1],
(getNumValuesUsed()-ChildNo)*sizeof(Children[0]));
Children[ChildNo] = SplitRes.LHS;
Children[ChildNo+1] = SplitRes.RHS;
if (getNumValuesUsed() != ChildNo)
memmove(&Values[ChildNo+1], &Values[ChildNo],
(getNumValuesUsed()-ChildNo)*sizeof(Values[0]));
Values[ChildNo] = SplitRes.Split;
++NumValuesUsed;
}
/// DoSplit - Split the currently full node (which has 2*WidthFactor-1 values)
/// into two subtrees each with "WidthFactor-1" values and a pivot value.
/// Return the pieces in InsertRes.
void DeltaTreeNode::DoSplit(InsertResult &InsertRes) {
assert(isFull() && "Why split a non-full node?");
@ -423,7 +464,6 @@ int DeltaTree::getDeltaAt(unsigned FileIndex) const {
// NOT REACHED.
}
/// AddDelta - When a change is made that shifts around the text buffer,
/// this method is used to record that info. It inserts a delta of 'Delta'
/// into the current DeltaTree at offset FileIndex.
@ -431,12 +471,10 @@ void DeltaTree::AddDelta(unsigned FileIndex, int Delta) {
assert(Delta && "Adding a noop?");
DeltaTreeNode *MyRoot = getRoot(Root);
// If the root is full, create a new dummy (non-empty) interior node that
// points to it, allowing the old root to be split.
if (MyRoot->isFull())
Root = MyRoot = new DeltaTreeInteriorNode(MyRoot);
MyRoot->AddDeltaNonFull(FileIndex, Delta);
InsertResult InsertRes;
if (MyRoot->DoInsertion(FileIndex, Delta, &InsertRes)) {
Root = MyRoot = new DeltaTreeInteriorNode(InsertRes);
}
#ifdef VERIFY_TREE
VerifyTree(MyRoot);