Re: long double on powerpc64

[Rich Felker <dalias@libc.org>]

On Fri, Mar 11, 2016 at 02:04:35PM +0100, Szabolcs Nagy wrote:
> * Justin Cormack <justin@specialbusservice.com> [2016-03-11 11:19:22 +0000]:
> > On 11 March 2016 at 04:17, Rich Felker <dalias@libc.org> wrote:
> > > On Thu, Mar 10, 2016 at 09:16:36PM -0600, Bobby Bingham wrote:
> > >> I've been working on a PPC64 port of musl lately.  I've made some good
> > >> progress, and it's time to decide what to do about the long double type.
> > >>
> > >> The PPC64 ELFv2 ABI [1] calls for a 128 bit long double.  It allows an
> > >> implementation to choose to use either IEEE quad, or IBM double double,
> > >> with IEEE quad being preferred.
> > >>
> > >> On the compiler side, it looks like things are a bit of a mess.
> > >>
> > >> Clang only supports IBM double double on PPC64, AFAICS, and therefore
> > >> won't work for us currently.
> > >>
> > >> GCC support is more complicated.  It supports both 128 bit variants, as
> > >> well as supporting (and defaulting to) a 64 bit long double.  To get a
> > >> 128 bit long double, you must build gcc with --with-long-double-128 or
> > >> pass -mlong-double-128, and even then you get IBM double double.  To get
> > >> IEEE quad, you must additionally pass -mlong-double-128, though there
> > >> are whispers that the default may change in gcc 7 [2].
> > >>
> > >> The final piece of bad news is that gcc can't successfully build musl on
> > >> PPC64 with IEEE quad long double.  It chokes on even trivial code using
> > >> long double complex [3].  So only 64 bit long double is usable for now.
> > >>
> > >> The good news is that gcc's predefined macros are sufficient to detect
> > >> which long double variant is in use.  My current thinking is that we can
> > >> support both 64 bit long and IEEE quad as two powerpc64 subarchs, even
> > >> if we can only implement 64 bit for now.  Because it looks like the
> > >> future direction is for IEEE quad to become the default, I think that
> > >> should be the suffix-less subarch, and the 64 bit long double subarch
> > >> should have a -ld64 suffix or similar.
> > >
> > > My leaning would be to just go with ld64 if nobody has their act
> > > together for quad support, but let's see what people who want to use
> > > powerpc64 think about it. The only option that's not on the table is
> > > IBM double-double (because it's incompatible with musl's assumption of
> > > IEEE semantics; math-savvy people in the musl community already know
> > > this of course but I'm repeating it for the sake of possible
> > > newcomers).
> > 
> > I think it would be a mistake to only support ld64, I think Bobby's approach
> > of two architectures is probably better, and maybe look to retire ld64
> > eventually.
> 
> if long double is 64bit then the 128bit hw floats cannot be used
> with musl, because we don't want library support for __float128.

I don't follow. Are you saying we would not want quad if it requires
soft float? I think the idea here is that the baseline binaries that
work on all models would need to use soft float operations for long
double, but higher -march could use the hardware directly in the
future (and the soft support should correctly use the fenv from
hardware). That's the same as the situation for aarch64, no?

> note that the glibc position is that for __float128 support the
> minimum required gcc version has to change to gcc-7 because it is
> abi and libc needs complex support (which will not be in gcc-6 yet).
> https://gcc.gnu.org/ml/gcc-patches/2015-12/msg02222.html
> https://sourceware.org/ml/libc-alpha/2016-03/msg00193.html

Support for __float128 is separate from support for targets where long
double is IEEE quad, I think. The former is not interesting IMO.

> the problem with -ld64 suffix is that gcc-6 already has hardcoded
> dynamic linker names (i don't think we can change that now, before
> the gcc-6 release).
> 
> so i'd leave the dynlinker name as is, use 64bit ld for now and
> rediscuss the issue when ieee128 long double works in gcc-7

Any idea why IEEE quad support requires bleeding-edge gcc for some
targets when mips64 had it way back in gcc 4.2 or earlier?

Rich