Three bugs in the async version of UnpWriteBuf (used only in the async
decompression path) were causing data corruption when decompressing RAR5
archives containing executable filters (E8/E8E9/ARM) whose blocks spanned
the 4MB write boundary (UNPACK_MAX_WRITE):
1. WrittenFileSize was only incremented inside `if (OutMem != null)`, but
should always be incremented after ApplyFilter (matching sync behavior).
2. UnpSomeRead = true was missing from the filter processing path.
3. In the NotAllFiltersProcessed else branch (filter intersects write border):
- WrPtr was not updated to WrittenBorder, corrupting the window pointer
used to compute WriteBorder on the next flush cycle
- Remaining filters had NextWindow set to true (inverted logic), when
sync correctly sets them to false so they are processed next time
All three changes align the async path with the sync UnpWriteBuf logic.
When SeekableSharpCompressStream.StopRecording() was called (e.g. after the tar
probe in GZipFactory.TryOpenReader), it cleared _recordedPosition without seeking
back to it. This left the underlying FileStream at the advanced position reached
during the probe, so the subsequent GZipReader would start reading from the
wrong offset, producing Key=null and 0 extracted bytes.
Fix: seek back to _recordedPosition before clearing it, matching the behavior of
the non-seekable SharpCompressStream.StopRecording() which rewinds _logicalPosition.
Also enhance the existing GZip_ReaderFactory_FlatGZip test to actually verify the
extracted content (not just MoveToNextEntry returns true), and add an async variant.
Closes#1391
Two releases since the generator was adopted here, both fixing things this
repository ran into.
2.0.41 emits the using directives of the source file into the generated one.
Documentation is copied across verbatim and a cref in it resolves against the
file it lands in, so crefs which relied on a using went unresolved and the
compiler reported CS1574 for each.
2.0.42 spaces a rewritten argument list the way it was written. The generated
extension calls here were coming out as
IArchiveEntryExtensions.WriteTo(archiveEntry, streamToWriteTo,
Constants.BufferSize,
progress: progress) ;
and now come out as
IArchiveEntryExtensions.WriteTo(archiveEntry, streamToWriteTo,
Constants.BufferSize,
progress: progress);
No change to what is generated, only to how it is spaced and what its
documentation can refer to. Full test suite passes unchanged.
The five synchronous WriteTo/WriteToDirectory/WriteToFile overloads were
maintained by hand alongside their async counterparts, which had drifted:
the async ones report progress through ExtractionOptions.BufferSize while
the sync ones did not, and each fix has had to be made twice.
Zomp.SyncMethodGenerator produces the synchronous copy from the async
source at compile time. The async methods are the only ones written out;
the sync versions are emitted into the same partial class, keeping the
public API identical - same names, same overloads, same signatures.
The generator is a build-time only dependency (PrivateAssets="all"), so
it adds nothing to the shipped package and no runtime reference.
Documentation summaries lose the word "asynchronously" because a summary
is now shared by both copies of the method.
