Re: [PATCH] x86: support i386 with Clang

[Nick Desaulniers]

On Mon, May 11, 2020 at 1:03 PM David Woodhouse <dwmw2@xxxxxxxxxxxxx> wrote:
>
> On Mon, 2020-05-11 at 13:01 -0700, Linus Torvalds wrote:
> > On Mon, May 11, 2020 at 12:52 PM Nick Desaulniers
> > <ndesaulniers@xxxxxxxxxx> wrote:
> > >
> > > Interesting approach.  Researching __builtin_choose_expr, it looks
> > > like it was cited as prior art for C11's _Generic keyword.
> >
> > Well, the thing that made me think that __builtin_choose_expr() would
> > work is that unlike the switch statement, you absolutely _have_ to do
> > the choice in the front end. You can't leave it as some kind of
> > optimization for later phases, because the choice od expression ends
> > up also determining the type of the result, so it isn't just a local
> > choice - it affects everything around that expression.
> >
> > But clang still doesn't like that "qi" constraint with a (non-chosen)
> > expression that has a "u64" type.
> >
> > I guess we can take the stupid extra cast, but I think it would at
> > least need a comment (maybe through a helper function) about why "qi"
> > needs it, but "ri" does not, and why the cast to "unsigned long" is
> > needed, even though "clearly" the type is already just 8 bits.
> >
> > Otherwise somebody will just remove that "obviously pointless" cast,
> > and gcc will eat the result happily, and clang will fail.
>
> I'm also mildly concerned that LLVM will start to whine about the 'ri'
> case too. It's odd that it doesn't, even when GCC does.

Ah, sorry, it took me a while to understand what case you meant by
this.  I recall you pointing this out previously in
https://bugs.llvm.org/show_bug.cgi?id=33587#c16, but at the time I
simply wasn't well versed enough in inline asm to understand.  The
case you're referring to is 64b operands, "r" constraint, -m32 target.
Yes, if I fix that: https://reviews.llvm.org/D79804, then this whole
patch needs to be reworked.  Back to the drawing board.
-- 
Thanks,
~Nick Desaulniers