Hi everyone
I'm new here. I read the mailing list policy, but please correct me if
I'm doing anything wrong.
In our C codebase, we're using a fair amount of static inline
__attribute__((always_inline)). This is arguably not a great decision,
because it forces inlining under questionable conditions, i.e. in cold
functions and unlikely branches. We're looking into switching some
non-trivial functions to extern inline instead.
When testing I noticed that GCC avoids inlining in
__attribute__((cold)) functions, but not in unlikely branches, as
signaled by __builtin_expect(cond, 0), even if the caller is hot. See
https://godbolt.org/z/sxnYh5PsE. The unlikely branch is moved to the
bottom, but it still seems beneficial to avoid inlining to reduce
instruction cache pressure for the rest of the hot functions. I wasn't
able to achieve this behavior with __builtin_expect or
__builtin_expect_with_probability with a very high probability.
Coincidentally, Clang does the exact opposite, inlining in cold
functions, but not in unlikely branches.
This is just one example, and I don't know anything about the
heuristics GCC uses to determine if some function is inlined. I wanted
to bring this up nonetheless to see if this situation may be improved.
I wasn't sure if this belongs in the bug tracker, let me know if I
should move it there instead.
Thank you.
Ilija
