[llvm-dev] RFC: Adding llvm::ThinStream

[Zachary Turner via llvm-dev]

On Wed, Feb 22, 2017 at 11:29 AM David Blaikie <dblaikie at gmail.com> wrote:

> Haven't got to this but would like to take a look/review it before it goes
> in.
>
> *skimming over some of the description*
>
> Sounds like 'stream' might not be the right terminology - since they
> return pointers into data that (I think) remains valid for the life of the
> stream? (this also makes me wonder a bit about memory usage if the
> cross-block operation is used a lot (causing allocations) but the values
> are then discarded by the user - the memory can't be reused, it's
> effectively leaked)
>
> cross-block allocation is the exception though because PDB is weird, I
don't expect that to be heavily used in other places.  That said, in the
PDB's implementation of BinaryStream we take this into account and cache
allocation requests at a given offset.  The data isn't just random, it's
structured into records, and so we expect reads for the same data to always
be at the same offsets (if you've got a list of records, you're going to
read records 1, 2, and 3, not some random bytes in the middle).  Because of
that, we can cache allocation requests for a particular offset and size,
and if we see another request come in that was previously cached, we just
return that.  It still handles the "weird" case anyway just for the sake of
completeness, but I don't think it's actually needed.



> Also - the wrappers ThinStreamReader/Writer - do they benefit
> substantially from being thin-stream aware, rather than abstractions around
> any stream of bytes? (since they transform the data anyway)
>
This is actually a surprisingly handy abstraction.  Since a BinaryStream
(I've abandoned the term Thin btw, but we can discuss the name orthogonally
if you like) most likely represents some kind of structured format with
records, you often have quite a bit of data to read out.  And since
BinaryStream can be any arbitrary implementation, you don't always have a
single list of bytes representing everything.  I'll use the canonical PDB
case where you've got all these blocks spread around.  I might know that I
have 1000 records, but not where they come from.  Any given record might
cross a boundary and return me a pointer allocated from the stream's
internal BumpPtrAllocator.  But if I just want to read 1000 records, it's
convenient to *treat* the entire thing as one contiguous sequence.  So I
just make a BinaryStreamReader and iterate over it until it becomes empty,
reading out the records.  If I had to give it a sequence of bytes, this
would become very cumbersome.

That said, there is one implementation of BinaryStream that is just a
sequence of bytes.  So this is general enough to support that.


> Oh, reinterpret casting... hrm. That kind of file reading/writing scheme
> usually makes me a bit uncomfortable due to portability concerns (having to
> align, byte swap, etc, structs to match the on-disk format can make those
> structures problematic to work with - if you have to byte swap anyway,
> you'd need to copy the data out of the underlying buffer anyway, right?)
>
For byte swapping usually this is done by using llvm's endian aware types.
If you have a struct with a bunch of llvm::support::ulittle16_t's, then it
doesn't matter what platform you're on, you can reinterpret_cast the
pointer to that struct, and when you read the values they will be correct.

On the other hand, sometimes it's convenient to just say "give me an
integer", and in that case the read function lets you specify what
endianness you want (defaults to host) and it copies the value into an
output parameter rather than returning a pointer.  In either case, the user
of the API never has to deal with byte swapping.

I'll upload a patch shortly.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170222/54a13d9b/attachment.html>