Re: [PATCH 0/4] -ffreestanding/-fno-builtin-* patches

[Nick Desaulniers <ndesaulniers@google.com>]

On Tue, Aug 18, 2020 at 1:27 PM Nick Desaulniers
<ndesaulniers@google.com> wrote:
>
> On Tue, Aug 18, 2020 at 1:24 PM Arvind Sankar <nivedita@alum.mit.edu> wrote:
> >
> > On Tue, Aug 18, 2020 at 12:13:22PM -0700, Linus Torvalds wrote:
> > > On Tue, Aug 18, 2020 at 12:03 PM H. Peter Anvin <hpa@zytor.com> wrote:
> > > >
> > > > I'm not saying "change the semantics", nor am I saying that playing
> > > > whack-a-mole *for a limited time* is unreasonable. But I would like to go back
> > > > to the compiler authors and get them to implement such a #pragma: "this
> > > > freestanding implementation *does* support *this specific library function*,
> > > > and you are free to call it."
> > >
> > > I'd much rather just see the library functions as builtins that always
> > > do the right thing (with the fallback being "just call the standard
> > > function").
> > >
> > > IOW, there's nothing wrong with -ffreestanding if you then also have
> > > __builtin_memcpy() etc, and they do the sane compiler optimizations
> > > for memcpy().
> > >
> > > What we want to avoid is the compiler making *assumptions* based on
> > > standard names, because we may implement some of those things
> > > differently.
> > >
> >
> > -ffreestanding as it stands today does have __builtin_memcpy and
> > friends. But you need to then use #define memcpy __builtin_memcpy etc,
> > which is messy and also doesn't fully express what you want. #pragma, or
> > even just allowing -fbuiltin-foo options would be useful.

I do really like the idea of -fbuiltin-foo.  For example, you'd specify:

-ffreestanding -fbuiltin-bcmp

as an example. `-ffreestanding` would opt you out of ALL libcall
optimizations, `-fbuiltin-bcmp` would then opt you back in to
transforms that produce bcmp.  That way you're informing the compiler
more precisely about the environment you'd be targeting.  It feels
symmetric to existing `-fno-` flags (clang makes -f vs -fno- pretty
easy when there is such symmetry).  And it's already convention that
if you specify multiple conflicting compiler flags, then the latter
one specified "wins."  In that sense, turning back on specific
libcalls after disabling the rest looks more ergonomic to me.

Maybe Eli or David have thoughts on why that may or may not be as
ergonomic or possible to implement as I imagine?

> >
> > The two compilers have some peculiarities, which means you really can't
> > have functions with the same name that do something else if you want to
> > use builtins at all, and can also lead to missed optimizations.
> >
> > For eg, __builtin_strchr(s,'\0') can be optimized to strlen. gcc will
> > optimize it that way even if -ffreestanding is used (so strlen has to
> > mean strlen), while clang won't, so it misses a potential optimization.
> > This is admittedly a silly example, but you could imagine something like
> > strncpy being optimized to memcpy+memset if the source length was
> > previously computed.
> >
> > PS: clang optimizes sprintf, but doesn't provide __builtin_sprintf?
>
> https://bugs.llvm.org/show_bug.cgi?id=47224
> --
> Thanks,
> ~Nick Desaulniers



-- 
Thanks,
~Nick Desaulniers