Re: [PATCH v2 09/18] RISC-V: The ABI implementation for 32-bit

[Alistair Francis <alistair23@gmail.com>]

On Fri, Jul 10, 2020 at 6:24 PM Maciej W. Rozycki <macro@wdc.com> wrote:
>
> On Fri, 10 Jul 2020, Alistair Francis wrote:
>
> > >  I note there is an extensive discussion on the way to move forward here:
> > > <https://sourceware.org/glibc/wiki/Y2038ProofnessDesign#utmp_types_and_APIs>
> > > We might as well try to implement it right away, so as to avoid being
> > > limited to 32-bit time records here.
> >
> > Is there an advantage of doing it now or can we put this off for the
> > next release?
>
>  The change is major enough it'll have to wait for the next development
> cycle anyway.  It shouldn't matter that much for RV32 glibc deployments
> though, given the amount of suitable hardware available.
>
> > >  NB some existing ports do have __WORDSIZE_TIME64_COMPAT32 set and cleared
> > > for their 64-bit and 32-bit ABIs respectively, as per the note in our
> > > top-level bits/wordsize.h, however this reflects the state as before we
> > > introduced the possibility for `__time_t' to be a 64-bit type with
> > > `__WORDSIZE == 32' ABIs.  Given the turn of events I think the note ought
> > > to be updated accordingly; I gather it was missed with commit 07fe93cd9850
> > > ("generic/typesizes.h: Add support for 32-bit arches with 64-bit types").
>
>  Will you be able to look into it?
>
>  The context here is before Y2038 changes __WORDSIZE_TIME64_COMPAT32 would
> only be clear for 64-bit ABIs with those 64-bit systems that do not have
> any 32-bit ABI (compatibility mode) to support, such as the DEC Alpha.
> And it would always be clear for 32-bit ABIs, so as to use the proper
> `__time_t' type without changing the width of actual data held there in
> the structure.
>
>  I'm not sure what the story is behind the S/390 port though; perhaps it
> does not support ABI coexistence in a single run-time environment.

I was just about to start this but Adhemerval has very kindly sent a
patch series already. I'm going to leave this as is (defining it as 1
for RV32 and RV64).


>
> > > >    int __cur_writer;
> > > >    int __shared;
> > > >    unsigned long int __pad1;
> > > >    unsigned long int __pad2;
> > > >    unsigned int __flags;
> > > > +#else
> > > > +# if __BYTE_ORDER == __BIG_ENDIAN
> > > > +  unsigned char __pad1;
> > > > +  unsigned char __pad2;
> > > > +  unsigned char __shared;
> > > > +  unsigned char __flags;
> > > > +# else
> > > > +  unsigned char __flags;
> > > > +  unsigned char __shared;
> > > > +  unsigned char __pad1;
> > > > +  unsigned char __pad2;
> > > > +# endif
> > > > +  int __cur_writer;
> > > > +#endif
> > > >  };
> > >
> > >  I note with this change the RV32 structure will use the generic layout as
> > > per sysdeps/nptl/bits/struct_rwlock.h, however regrettably RV64 does not.
> > > Would it make sense to instead have the layout the same between RV64 and
> > > RV32, perhaps by redefining `__pad1' and `__pad2' in terms of `unsigned
> > > long long' (which would have alignment implications though) or otherwise?
> >
> > I'm not sure which one is better. On one hand it seems better to be
> > more generic and therefore RV32 should use the generic interface. On
> > the other hand the more similar they are the better. I'm still leaning
> > towards we should be generic where possible.
>
>  It would be good to get a second opinion here.

The RV64 __pthread_rwlock_arch_t is exactly the same as the AArch64
version. My guess is that it was just copied to be the same.

The generic version also has this comment

/* Generic struct for both POSIX read-write lock.  New ports are expected
   to use the default layout, however archictetures can redefine it to add
   arch-specific extensions (such as lock-elision).  The struct have a size
   of 32 bytes on both LP32 and LP64 architectures.  */

As we aren't adding any arch-specific extensions I think we should
keep the generic one. It's unfortunate we didn't do that for RV64, but
too late now.

The only downside I see in using the generic version is what happens
if the value of _shared or flags overflows a char. There are currently
only 4 flags though, so I think it's ok.

>
> > >  Is there any benefit from having `__flags' and `__shared' (and the
> > > reserve) grouped within a single 32-bit word?  I gather there is, given
> > > the lengths gone to to match the bit lanes across the word regardless of
> > > the endianness.  But what is it?
> >
> > I have no idea.
>
>  Especially given this.
>
> > > > +# define _FP_DIV_MEAT_S(R, X, Y)     _FP_DIV_MEAT_1_udiv_norm (S, R, X, Y)
> > > > +# define _FP_DIV_MEAT_D(R, X, Y)     _FP_DIV_MEAT_2_udiv (D, R, X, Y)
> > > > +# define _FP_DIV_MEAT_Q(R, X, Y)     _FP_DIV_MEAT_4_udiv (Q, R, X, Y)
> > > > +
> > > > +# define _FP_NANFRAC_S               _FP_QNANBIT_S
> > > > +# define _FP_NANFRAC_D               _FP_QNANBIT_D, 0
> > > > +# define _FP_NANFRAC_Q               _FP_QNANBIT_Q, 0, 0, 0
> > >
> > >  Likewise.  There seems to be an established practice for this header
> > > across architectures to have no space between macro arguments or before
> > > the opening parenthesis.  This might help with the alignment.
> >
> > I still think it makes sense to follow the glibc style though, even if
> > other archs don't.
> >
> > Let me know if it should be a different way and I'll update it.
>
>  There is the issue of the discrepancy compared to the libgcc version, and
> while `diff -l' and `patch -l' solve that for manual processing, more
> sophisticated tools may not cope and require manual intervention.
>
>  Again, I would suggest getting a second opinion.

The current option matches the rest of the file, so I think it makes
sense to leave as is.

If this is a problem we can just change the spacing for the entire
file in a future patch.

Alistair

>
>   Maciej