Re: AMDGPU and 16B stack alignment

[Nick Desaulniers]

On Tue, Oct 15, 2019 at 1:26 PM Arvind Sankar <nivedita@xxxxxxxxxxxx> wrote:
>
> On Tue, Oct 15, 2019 at 11:05:56AM -0700, Nick Desaulniers wrote:
> > Hmmm...I would have liked to remove it outright, as it is an ABI
> > mismatch that is likely to result in instability and non-fun-to-debug
> > runtime issues in the future.  I suspect my patch does work for GCC
> > 7.1+.  The question is: Do we want to either:
> > 1. mark AMDGPU broken for GCC < 7.1, or
> > 2. continue supporting it via stack alignment mismatch?
> >
> > 2 is brittle, and may break at any point in the future, but if it's
> > working for someone it does make me feel bad to outright disable it.
> > What I'd image 2 looks like is (psuedo code in a Makefile):
> >
> > if CC_IS_GCC && GCC_VERSION < 7.1:
> >   set stack alignment to 16B and hope for the best
> >
> > So my diff would be amended to keep the stack alignment flags, but
> > only to support GCC < 7.1.  And that assumes my change compiles with
> > GCC 7.1+. (Looks like it does for me locally with GCC 8.3, but I would
> > feel even more confident if someone with hardware to test on and GCC
> > 7.1+ could boot test).
> > --
> > Thanks,
> > ~Nick Desaulniers
>
> If we do keep it, would adding -mstackrealign make it more robust?
> That's simple and will only add the alignment to functions that require
> 16-byte alignment (at least on gcc).

I think there's also `-mincoming-stack-boundary=`.
https://github.com/ClangBuiltLinux/linux/issues/735#issuecomment-540038017

>
> Alternative is to use
> __attribute__((force_align_arg_pointer)) on functions that might be
> called from 8-byte-aligned code.

Which is hard to automate and easy to forget.  Likely a large diff to fix today.

>
> It looks like -mstackrealign should work from gcc 5.3 onwards.

The kernel would generally like to support GCC 4.9+.

There's plenty of different ways to keep layering on duct tape and
bailing wire to support differing ABIs, but that's just adding
technical debt that will have to be repaid one day.  That's why the
cleanest solution IMO is mark the driver broken for old toolchains,
and use a code-base-consistent stack alignment.  Bending over
backwards to support old toolchains means accepting stack alignment
mismatches, which is in the "unspecified behavior" ring of the
"undefined behavior" Venn diagram.  I have the same opinion on relying
on explicitly undefined behavior.

I'll send patches for fixing up Clang, but please consider my strong
advice to generally avoid stack alignment mismatches, regardless of
compiler.
--
Thanks,
~Nick Desaulniers