On Mon, Sep 30, 2019 at 3:08 PM Miguel Ojeda
[off-list ref] wrote:
On Mon, Sep 30, 2019 at 11:50 PM Nick Desaulniers
[off-list ref] wrote:
quoted
So __attribute__((always_inline)) doesn't guarantee that code will be
inlined. [...] inline and __attribute__((always_inline))
are a heuristic laden mess and should not be relied upon.
Small note: in GCC, __attribute__((always_inline)) is documented as
actually guaranteeing to either inline or error otherwise (although
see the remark for indirect calls):
"Failure to inline such a function is diagnosed as an error. Note
Not an error, but a warning at least: https://godbolt.org/z/_V5im1.
That's interesting, so it has multiple semantics, because it's also
documented to inline even when no optimizations are specified. So
when someone uses __attribute__((always_inline)) without a comment,
it's not clear whether they mean for there to be a warning when this
is not inlined, or for it to be inlined at -O0 (guess not for the
kernel), or both. If the kernel wants to enforce the former, why not
set `-Werror=attributes`? Maybe that warning is too broad? Seems
like a recipe for subtly broken code found at runtime, when we'd
rather have stronger compile time guarantees.
that if such a function is called indirectly the compiler may or may
not inline it depending on optimization level and a failure to inline
an indirect call may or may not be diagnosed."
As for LLVM/Clang, no idea, since it does not say anything about it in
the docs -- but from what you say, it is a weaker guarantee.
Filed https://bugs.llvm.org/show_bug.cgi?id=43517
--
Thanks,
~Nick Desaulniers