[ADT] Simplify SmallDenseMap::grow (NFC) #167829
Conversation
Without this patch, SmallDenseMap::grow has two separate code paths to grow the bucket array. The code path to handle the small mode has its own traversal over the bucket array. This patch simplifies this logic as follows: 1. Allocate a temporary instance of SmallDenseMap. 2. Move valid key/value pairs to the temporary instance. 3. Move LargeRep to *this. Remarks: - This patch adds moveFromImpl to move key/value pairs. moveFromOldBuckets is updated to use the new helper function. - This patch adds a private constructor to SmallDenseMap that takes an exact number of buckets, accompanied by tag ExactBucketCount. - This patch adds a fast path to deallocateBuckets in case getLargeRep()->NumBuckets == 0, just like destroyAll. This path is used to destruct zombie instances after moves. - In somewhat rare cases, we "grow" from the small mode to the small mode when there are many tombstones in the inline storage. This is handled with another call to moveFrom.
|
@llvm/pr-subscribers-llvm-adt Author: Kazu Hirata (kazutakahirata) ChangesWithout this patch, SmallDenseMap::grow has two separate code paths to
Remarks:
Full diff: https://github.com/llvm/llvm-project/pull/167829.diff 1 Files Affected:
diff --git a/llvm/include/llvm/ADT/DenseMap.h b/llvm/include/llvm/ADT/DenseMap.h
index d5b13e7731550..c4d139411503f 100644
--- a/llvm/include/llvm/ADT/DenseMap.h
+++ b/llvm/include/llvm/ADT/DenseMap.h
@@ -413,9 +413,7 @@ class DenseMapBase : public DebugEpochBase {
return NextPowerOf2(NumEntries * 4 / 3 + 1);
}
- void moveFromOldBuckets(iterator_range<BucketT *> OldBuckets) {
- initEmpty();
-
+ void moveFromImpl(iterator_range<BucketT *> OldBuckets) {
// Insert all the old elements.
const KeyT EmptyKey = KeyInfoT::getEmptyKey();
const KeyT TombstoneKey = KeyInfoT::getTombstoneKey();
@@ -438,6 +436,14 @@ class DenseMapBase : public DebugEpochBase {
}
}
+ void moveFromOldBuckets(iterator_range<BucketT *> OldBuckets) {
+ initEmpty();
+ moveFromImpl(OldBuckets);
+ }
+
+ // Move key/value from Other to *this. Other will be in a zombie state.
+ void moveFrom(DerivedT &Other) { moveFromImpl(Other.buckets()); }
+
void copyFrom(const DerivedT &other) {
this->destroyAll();
derived().deallocateBuckets();
@@ -889,6 +895,12 @@ class SmallDenseMap
/// a large bucket. This union will be discriminated by the 'Small' bit.
AlignedCharArrayUnion<BucketT[InlineBuckets], LargeRep> storage;
+ struct ExactBucketCount {};
+ SmallDenseMap(unsigned NumBuckets, ExactBucketCount) {
+ allocateBuckets(NumBuckets);
+ this->BaseT::initEmpty();
+ }
+
public:
explicit SmallDenseMap(unsigned NumElementsToReserve = 0) {
init(NumElementsToReserve);
@@ -1065,7 +1077,7 @@ class SmallDenseMap
}
void deallocateBuckets() {
- if (Small)
+ if (Small || getLargeRep()->NumBuckets == 0)
return;
deallocate_buffer(getLargeRep()->Buckets,
@@ -1096,46 +1108,22 @@ class SmallDenseMap
if (AtLeast > InlineBuckets)
AtLeast = std::max<unsigned>(64, NextPowerOf2(AtLeast - 1));
- if (Small) {
- // First move the inline buckets into a temporary storage.
- AlignedCharArrayUnion<BucketT[InlineBuckets]> TmpStorage;
- BucketT *TmpBegin = reinterpret_cast<BucketT *>(&TmpStorage);
- BucketT *TmpEnd = TmpBegin;
+ SmallDenseMap Tmp(AtLeast, ExactBucketCount{});
+ Tmp.moveFrom(*this);
- // Loop over the buckets, moving non-empty, non-tombstones into the
- // temporary storage. Have the loop move the TmpEnd forward as it goes.
- const KeyT EmptyKey = KeyInfoT::getEmptyKey();
- const KeyT TombstoneKey = KeyInfoT::getTombstoneKey();
- for (BucketT &B : inlineBuckets()) {
- if (!KeyInfoT::isEqual(B.getFirst(), EmptyKey) &&
- !KeyInfoT::isEqual(B.getFirst(), TombstoneKey)) {
- assert(size_t(TmpEnd - TmpBegin) < InlineBuckets &&
- "Too many inline buckets!");
- ::new (&TmpEnd->getFirst()) KeyT(std::move(B.getFirst()));
- ::new (&TmpEnd->getSecond()) ValueT(std::move(B.getSecond()));
- ++TmpEnd;
- B.getSecond().~ValueT();
- }
- B.getFirst().~KeyT();
- }
-
- // AtLeast == InlineBuckets can happen if there are many tombstones,
- // and grow() is used to remove them. Usually we always switch to the
- // large rep here.
- allocateBuckets(AtLeast);
- this->moveFromOldBuckets(llvm::make_range(TmpBegin, TmpEnd));
- return;
+ if (Tmp.Small) {
+ // Use moveFrom in those rare cases where we stay in the small mode. This
+ // can happen when we have many tombstones.
+ this->BaseT::initEmpty();
+ this->moveFrom(Tmp);
+ Tmp.Small = false;
+ Tmp.getLargeRep()->NumBuckets = 0;
+ } else {
+ Small = false;
+ NumTombstones = 0;
+ *getLargeRep() = std::move(*Tmp.getLargeRep());
+ Tmp.getLargeRep()->NumBuckets = 0;
}
-
- LargeRep OldRep = std::move(*getLargeRep());
- getLargeRep()->~LargeRep();
- allocateBuckets(AtLeast);
-
- this->moveFromOldBuckets(OldRep.buckets());
-
- // Free the old table.
- deallocate_buffer(OldRep.Buckets, sizeof(BucketT) * OldRep.NumBuckets,
- alignof(BucketT));
}
// Plan how to shrink the bucket table. Return:
|
kazutakahirata
deleted the
cleanup_20251112_ADT_DenseMap_SmallDenseMap_grow
branch
|
LSan reports that this change caused memory leaks: https://lab.llvm.org/buildbot/#/builders/24/builds/14677/steps/12/logs/stdio Can you please take a look? |
also hit lsan problem with mlir tools. @kazutakahirata can you please take a look ? |
Thanks! I'll post a patch soon. |
I've posted #168011. |
6 participants
Without this patch, SmallDenseMap::grow has two separate code paths to
grow the bucket array. The code path to handle the small mode has its
own traversal over the bucket array. This patch simplifies this logic
as follows:
Remarks:
This patch adds moveFromImpl to move key/value pairs.
moveFromOldBuckets is updated to use the new helper function.
This patch adds a private constructor to SmallDenseMap that takes an
exact number of buckets, accompanied by tag ExactBucketCount.
This patch adds a fast path to deallocateBuckets in case
getLargeRep()->NumBuckets == 0, just like destroyAll. This path is
used to destruct zombie instances after moves.
In somewhat rare cases, we "grow" from the small mode to the small
mode when there are many tombstones in the inline storage. This is
handled with another call to moveFrom.