The idea is to allow codegen to exploit specific known constants. It
seems that LLVM by itself will not generate split loop blocks by an
input argument that is a loop constant. Whereas we know, from PNG, that
some constants yield much better code. In this case
coef[i] = coef[i] + coef[i - n]
This has a destructive dependency chain for `n = 1`, should never get an
argument of 0 and is almost embarrassingly parallel in SIMD if n >= 8
where we can increase the amount of data loaded at once for each
independent loop iteration.
By splitting the loop we make the compiler apply independent
optimization passes to each of case, then have a fallback for things we
did not cover with vectorized possibilities.
Prepares extensibility for the eventual case of supporting multi-planar
image layouts that are to be written into a multi-planar buffer as
indicated by the caller in some way (possibly not via read_*_bytes but a
method taking a wrapper around the output buffer instead). Through the
struct we can later add an indicator that can be used by the caller to
decide between methods of fetching the pixel data as appropriate.
Having the value reader itself take care of the bigtiff and limit
semantics is a little cleaner, with the TagReader being responsible
for the interface of grabbing specific types of IFD values.
This allows us to recombine the reader and other file data in a Decoder
with any IFD, whether that of an image or an extension IFD as indicated
by an IFD pointer such as Exif and GPS.
The previous strategy could not keep track of all cycles such as
reflexive ones at the very start. Additionally, it would fail to be able
to iterate after seeking since it was overly cautious and verified the
next IFD position before allowing the read of a current IFD.
The new implementation is furthermore proof against having multiple
independent chains of IFD which may occur when iterating SubIFDs for
thumbnails. (Note that we do not verify the tree-structure, only a
freedom from cycles).
A. Molzer