Released Clang 12 generates bad code for the original loop in here.
While this is a compiler bug plain and simple, we still have to deal
with it.
This is related to the SLP vectorizer, and in particular the two
reverse subtracts in the butterflies for the second half to avoid
unary negates.
Use the more regular dataflow that has the unary negates in it
(we can at least fold one of them into a constant, namely for A2)
and introduce a few temporaries that also make alias analysis (and
possible block-level vectorization) a whole let easier while I'm at
it.
This fixes the codegen issues on Clang 12, which now produces a
working decoder, and I expect the single unary negate that we
actually gain per iteration of this loop is not a significant
perf concern. (There are bigger fish to fry here regardless.)
Fixes issue #1152.
Some parameters do not get used, or only when certain config
defines are set. Explicitly mark them as unused to make compilers
happy.
Fixes issue #396.
This is definitely unnecessary, or at least I can't find anything
in the Vorbis spec that would indicate anything special happening
here.
Fixes issue #816.
When start_decoder() fails it may already have allocated memory
for .vendor and/or .comment_list. Call vorbis_deinit() to free
any allocated memory.
Fixes issue #1051.
Not an actual bug, it just looked wonky, but this code runs
with code lengths that are verified to be in range (<32) by
the length-reading code. Anyway.
Fixes issue #901.
Put the formats that start with a clear magic number first,
the dodgy ones that don't have much of a distinctive header
should be tested for later after we've ruled out the clearer
ones.
Fixes issue #787, hopefully. (Never got a clean repro.)
It's implementation-specified behavior. Writing this code and then
relying on compiler strength reduction to turn it back into shifts
feels extremely silly but it is what it is.
Fixes issue #1097.
Define lrot in a way that doesn't involve UB when n==0.
Also, the previous patch ensures that n <= 15 for all callers
of stbi__extend_receive, so can remove the (less restrictive)
bounds check for 0 <= n < 17 (the bounds of stbi__bmask)
entirely.
Fixes issue #1065.
extend_receive implicitly requires n <= 15 (code length);
the maximum that actually makes sense for 8-bit baseline JPEG is
11, but 15 is the natural limit for us because the AC coding path
stores the number of magnitude bits in a nibble.
Check that DC delta bits are in range before attempting to call
extend_receive.
Fixes issue #1108.
Keep STB_DXT_DITHER so as not to break existing code that tries
to enable it, but just leave it permanently off. I originally
introduced it somewhat superstitiously because of the RGB565
endpoint resolution but it never improved either perceptual quality
or objective quality metrics, and the code is appreciably simpler
without it.
Also fix a "potentially uninitialized variable" warning.
This is a modified version of Alan's original PR that keeps the
table generator in the file (in case there's interest) and also
replaces the expand[] tables with math, since it's trivial.
Fixes issue #1117.
Fabian Giesen