On Tue, 17 Dec 2024 at 16:47, Alexei Starovoitov
<alexei.starovoitov@xxxxxxxxx> wrote:
>
> Since we're on this topic, Daniel is looking to reuse format_decode()
> in bpf_bprintf_prepare() to get rid of our manual format validation.
That was literally why I started looking into this - the many separate
type formats actually end up causing format_decode() (and the callers)
to have to generate multiple different cases, which then in turn
either cause a jump table, or - more commonly due to the CPU indirect
branch mitigations - a chain of conditionals that are fairly ugly.
Compressing the state table for the types from 11 down to 4 types
helps a bit, but then also dealing with the "smaller than int" things
as just 'int' (with the formatting flags that are separate) also ends
up avoiding some unnecessary and extra cases.
Because in the end, 'size_t' and 'long' are the same thing, even on
architectures like 32-bit x86 where 'size_t' really is 'unsigned int'
- simply because the only thing that matters for fetching the value is
the size, which is 32-bit.
(The whole "is it signed" and the truncation to smaller-than-int etc
is then something we have to handle anyway in by the 'printf_spec'
thing).
So I have a patch series to clean some of this up and avoid the extra
states. I'm not entirely happy with it, though, and I've been going
back and forth on some of the code, so I'm not ready to post it or
have anybody use it as a basis for some "real" cleanups.
I guess I could at least post the "turn 11 different types into 4"
part. I have other things in there, but that part seems fairly
unambiguously good.
Let me go separate that part out and maybe people can point out where
I've done something silly.
Linus
