Before this commit, the DLAStep MergePointeesOfPointerUnion wasn't
really equipped for dealing with LayoutTypeSystemNodes representing
types imported from Model.
There was some code that tried to deal with them but it was mostly an
afterthought and it wasn't robust.
This commit fixes that, and is able to handle a strict superset of the
scenarios envisioned before, while preserving the property that nodes
representing types imported from Model should be preserved (in
particular their size).
This commit extends the CompactCompatibleArrays DLAStep to also consider
non-strided accesses.
Strided accesses are still always considered first, and they are still
considered the only real source of information on arrays, but after
having tried to compact all the compatible arrays,
CompactCompatibleArrays now also considers non-strided instance edges to
see if they can be compacted with the rest of the inferred array.
This has shown to handle gracefully a number of real-world examples and
reduce unions.
Before this commit, the CompactCompatibleArray DLAStep had that caused
the compaction to be sensible to the order of the pairs of edges that
were compacted, and that also caused the compaction operation to overrun
the end of the containing struct in some corner cases.
This commit fixes both bugs.
Now the compaction routine is not ordering-sensitive anymore.
For each pair of edges <A, B> to compact it tries to compact them in 2
possible ways: by aligning A to B shifting A to lower offsets, and by
aligning B to A shifting B to lower offsets.
If both succeed, it picks the best result among the two, which doesn't
depend on the ordering of A and B.
This fixes the order-sensitiveness.
While reworking this logic, the logic was fixed so that if any of the
two compaction attempts causes to overrun the end of the containing
struct, the attempt is not considered successful anymore, and it's
discarded altogether.
This commit adds caching for mutual reachability among children of a
node, in SimplifyInstanceAtOffset0.
This makes the DLAStep 10% to 50% faster on real-world benchmarks we
have measured, such as `updatedb.plocate` and `df` for Ubuntu 22.04
x86_64.
llvm::EquivalenceClasses is an efficient data structure from LLVM, to
compute equivalence classes among objects with Tarjan's union-find.
This commit uses that to avoid an hand-crafted very inefficient
algorithm.
Before this commit, the dla::Step CompactCompatibleArrays was
occasionally generating artificial nodes that were larger than the inner
data, forcing their size to Stride - AvailableSlack.
This commit fixes the problem keeping track of the actual size that is
consumed in the array element, and forcing the size of the artificial
nodes to that.
The dla::Step ArrangeAccessesHierarchically looks throught the DLA graph
and for each node with many outgoing instance edges tries to see if some
of them are hierachically contained within each other.
In order to do this, before this commit, we were building an auxiliary
graph representing this hierarchy, and in order to build it we had to
perform a number of comparisons between edges that was quadratic with
the number of edges.
Moreover, once all the comparisons were done, we had a deep graph
representing inclusion between edges, but we only cared about the
top-level of this graph, i.e. only the edges that contained other edges
hierarchically, but were not contained in other edges (we'll call them
the root edges).
So we were doing a quadratic number of comparisons but possibly many of
them were useless.
Finally, all the edges that were included in root edges, were pushed
down, but only for a single layer, because they needed to be re-compared
later with the children of the root edges they were being pushed
through. This latter part was responsible for a lot of wasted
computation that just needed to be done over and over at all the layers.
Overall this algorithm was doing a lot of wasted computation.
This commit replaces this logic with a new algorithm.
Now we keep track only of the root edges, and we compare only root edges
with other root edges.
Initially all edges are root edges.
Then we start comparing them.
If a root edge A is included in another one B, then A is not a root edge
anymore, and all the edges that were previously found to be included in
A are not included in B.
This algorithm still does a worst case of quadratic number of
comparisons, but drastically reduces the amount of useless computation
that is redone later. In particular:
- in cases where there are a lot of root edges (meaning that only a few,
or no edge can be included in others) we do a number of comparison
close to quadratic, but we're only pushing non-root edges down, so
we'll never have to redo comparisons in deeper layers
- in cases where there is only a small number of root edges, we're doing
a number of comparisons close to linear, and we never compare non-root
edges with each other, so we're saving a lot of computation that would
be wasted (because it would need to be redone in deeper layers).
Before this commit we were manually building a topological ordering,
that resulted in worse overall performance in some scenarios.
This commit changes the dla::Step to temporarily introduce a fake root
node, and computes a RPOT from there, which overall yields better
results.
Before this commit, a logic mistake made the AvailableSlack too large
for arrays that started near the beginning of their parent.
This in turn allowed a compacted array to "underflow" on lower addresses
w.r.t. the parent.
This commit fixes the problem by properly computing the AvailableSlack,
restraining it in cases where the array under analysis is close to the
beginning of the parent.
This commit enables ArrangeAccessesHierachically to push down pointer
edges when rearranging accesses hierarchically.
In particular, if an edge N is being pushed down another edge M, and M
is at offset 0, and N reaches a pointer edge PE that points to a
grandparent G of M at offset 0, then we want to push down the pointer
edge PE so that it points to the target of M, unless there is another
edge M' different from M, such that M' has the same properties of M but
target(M) != target(M').
Before this commit, ArrangeAccesseHierachically was removing nodes from
the graph, instead of merging them.
This degraded the information collected on the graph useful for debug.
This commit drops the use of removeNode method, and switches to using
mergeNodes, that preserves the useful debug information when merging.
Before this commit postProcessMerge was making some strong assumptions
about dla::CollapseSingleChild::collapseSingle that no longer hold.
Namely the (now) wrong assumption is that collapseSingle would not break
post_order iteration, which is now false because collapseSingle can
change the incoming edges of the node it's called on, and those incoming
edges are on the visit stack of the post_order iteration itself.
This required making a copy of the post order, and resulted in dropping
the postProcessMerge function alltogether.
Before this commit we were using int64_t for Offset, Strides, and
TripCounts.
Originally, this choice was intended because for some time we envisioned
actually having a use for negative values, but in the end we decided
there's no use for those.
This commit switches all to unsigned integers, allowing to remove some
static_cast across the codebase, and overall easing typicall computation
we have to perform on those fields.
A logic bug caused the wrong detection of nodes with many fields.
As a result, the field deduplication was only executed on a subset of
the real candidates.
This caused the following two problems.
- Degradation of the quality of the results, i.e. unions that could be
deduplicated were not deduplicated.
- Failing assertions in rare cases.
The algorithm assumes that if it's looking at the children
(C1, ..., Cn) of node A to be deduplicated, then all (C1, ..., Cn)
have beed already visited, hence they don't have children that should
be duplicated because they would have been deduplicated already, when
looking at (C1, ..., Cn). The bug possibly caused to miss the
deduplication of children of Ci (for some i) causing assertions to
fail when looking at A to deduplicate (C1, ..., Cn).
Before this commit, the dla::Step depended on InterferingInfo, but this
was superfluous.
This commit drops the dependency and enables the Step to run before
ComputeNonInterferingComponents.
This also avoids the need to run ComputeNonInterferingComponents twice
(one before and one after DeduplicateUnionFields), so it can only run
once.
Before this commit, CollapseSingleChild Step was thinkering around
with InterferingInfo.
Now CollapseSingleChild does not care about InterferingInfo anymore, so
the code handling it can just be dropped.
Alessandro Di Federico