Re: [PATCH v2] x86/fpu: use _Alignof to avoid UB in TYPE_ALIGN
From: Nick Desaulniers
Date: Wed Oct 05 2022 - 14:58:15 EST
On Wed, Oct 5, 2022 at 11:38 AM Nick Desaulniers
> On Wed, Oct 5, 2022 at 11:30 AM Nick Desaulniers
> <ndesaulniers@xxxxxxxxxx> wrote:
> > On Wed, Oct 5, 2022 at 12:29 AM YingChi Long <me@xxxxxxxxx> wrote:
> > >
> > > Kindly ping :)
> > Hi YingChi,
> > Sorry for the delay in review.
> > I think https://godbolt.org/z/sPs1GEhbT has convinced me that
> > TYPE_ALIGN is analogous to _Alignof and not __alignof__; so your patch
> > is correct to use _Alignof rather than __alignof__. I think that test
> > case demonstrates this clearer than the other links in the commit
> > message. Please consider replacing the existing godbolt links with
> > that one if you agree.
> > Please reword the paragraphs in the commit message from:
> > ```
> > In PATCH v1 "TYPE_ALIGN" was substituted with "__alignof__" which is a
> > GCC extension, which returns the *preferred alignment*, that is
> > different from C11 "_Alignof" returning *ABI alignment*. For example, on
> > i386 __alignof__(long long) evaluates to 8 but _Alignof(long long)
> > evaluates to 4. See godbolt links below.
> > In this patch, I'd like to use "__alignof__" to "_Alignof" to preserve
> > the behavior here.
> > ```
> > to:
> > ```
> > ISO C11 _Alignof is subtly different from the GNU C extension
> > __alignof__. _Alignof expressions evaluate to a multiple of the object
> > size, while __alignof__ expressions evaluate to the alignment dictated
> > by the target machine's ABI. In the case of long long on i386,
> > _Alignof (long long) is 8 while __alignof__ (long long) is 4.
> Oops, and I had that backwards.
> In the case of long long on i386, _Alignof (long long) is 4 while
> __alignof__ (long long) is 8.
> So I guess my commentary on "multiple of the object size" is
> wrong...hmm...this wording can probably be improved further still...
Section 6.2.8 "Alignment of objects" refers to "fundamental alignment"
and "extended alignment."
I wonder if it would be precise to say that "_Alignof evaluates to the
fundamental alignment while __alignof__ evaluates to the extended
alignment (which is implementation defined, typically by the machine
specific ABI)." Though even that seems imprecise since it sounds like
a fundamental alignment could be less than or equal to what alignof
Grepping for `alignment requirement` turns up perhaps relevant
portions of the spec.
> > The macro TYPE_ALIGN we're replacing has behavior that matches
> > _Alignof rather than __alignof__.
> > ```
> > In particular, I think it's best to avoid language like "returns" in
> > favor of "evaluates to" since these are expressions, not function
> > calls. I think it's also good to avoid the term "preferred alignment"
> > since that isn't meaningful; it looks like it was pulled from one of
> > the GCC bug reports rather than the GCC docs or latest ISO C standard
> > (https://www.open-std.org/jtc1/sc22/wg14/www/docs/n3054.pdf). I'm not
> > sure that the links to the GCC bug tracker add anything meaningful
> > here; I think those can get dropped, too. It's also perhaps confusing
> > to refer to earlier versions of the patch. One thing you can do is
> > include comments like that "below the fold" in a commit message as a
> > meta comment to reviewers. See
> > https://lore.kernel.org/llvm/20220512205545.992288-1-twd2.me@xxxxxxxxx/
> > as an example of commentary "below the fold" on differences between
> > patch versions. Text in that area is discarded by git when a patch is
> > applied.
> > With those changes to the commit message in a v3, I'd be happy to sign
> > off on the change. Thanks for your work on this!
> > --
> > Thanks,
> > ~Nick Desaulniers
> ~Nick Desaulniers