On 28/08/23 08:10, John Paul Adrian Glaubitz wrote:
On Sun, 2023-08-27 at 10:46 +1000, Finn Thain wrote:
Not only mold but also most notably the following projects:
- LLVM
- Firebird Database
- OpenJDK
- Various Qt packages
And potentially more in the future, which may be anticipated on the basis
that "those users don't need a stable ABI any more, so let's just ignore
the portability issues in our code and leave the problem to the distros
and toolchain developers".
It's reasonable to assume that a 32-bit architecture uses 32-bit alignment and
I understand every single upstream project that doesn't want to care about obscure
design the decisions of some ABI designers of the past.
That is the precedent you would set.
No, I wouldn't set such precedent. I would fix something that has been broken
for years and has caused endless headaches for people maintaining the m68k port
in Linux distributions.
And since we have to break the ABI anyway to be able to use 64-bit time_t, I don't
see any valid reason to stick to the problematic 16-bit alignment used by the current
ABI.
Moreover, why is it that only a few developers have a problem with making
explicit their decisions regarding alignment of shorts? What actual pain
does it cause them to accept a patch to make their struct layouts plain?
The problem aren't upstream projects but the lack of manpower to work on all these
issues. Talk is cheap when there is hardly anyone doing this work.
I have invested a ton of work to get the m68k port into better shape and with the
help of the community, we even managed to land m68k support in LLVM. It was a HUGE
disappointment to me when the 16-bit alignment again caused trouble for a relevant
upstream project on m68k meaning that LLVM can currently not be used natively on
m68k.
It goes against the traditional ABIs, but practically no m68k Linux
binaries are published outside of distributions, so this not a
concern.
It is of concern to some users (though not all, apparently).
If these users really cared, they would actually help address these issues. I haven't
seen any contributions trying to address these issues outside my efforts and the efforts
of the Gentoo developers.
We need to break the ABI anyway with time_t going 64-bit, so it makes
sense to do these two things at the same time.
Fully agreed.
If the kernel breaks the ABI, that's a bug, not an excuse. Either you're
okay with proliferation of incompatible binaries and tools or there are
some criteria (yet to be identified AFAIK) which permit this bug.
It's not difficult to foresee fragmentation because it follows from the
manpower shortage. There will always be sufficient manpower to produce a
break that pleases a few. There may never be enough manpower to produce a
stable ABI that pleases everyone for the foreseeable future.
Again, talk is cheap. Show me the code.
I think -gnu32 sounds very reasonable.
You do? I think 32 is misleading in the absence of 16-bit or 64-bit
variants, and -gnu is misleading if other tooling like LLVM already
supports malign-int. Moreover, it's impossible to align to a bit count in
general. Not that you'd want to -- it's actually the natural alignment of
shorts that is at issue, AIUI.
Yes, I do and that's just my personal opinion. But as I said, I am open to
other naming suggestions.
So, for naming purposes, the proposal might be described as either the ABI
du jour (leading to -abi23 for 2023) or the new ABI for ever (leading to
-abin as in -gnuabin32 on MIPS).
That's why I suggested we can look how the ARM developers will name their
triplet when switching to 64-bit time_t on 32-bit ARM systems.
If it's the former, perhaps you should not push it upstream. If it's the
latter, perhaps this redesign should seek to address real shortcomings
with the existing ABI, including problems which (for all I know) may have
entirely prevented some people from using it thus far. That is, it should
consider silicon beyond 680x0.
It's a historic architecture. We don't have to redesign everything. It's enough
to address the most pressing issues and these are 16-bit alignment and 32-bit
time_t.
If the idea is really to endeavor on a new ABI for m68k, it means a different
loader and the question: will it be interoperable with current m68k ABI in the
sense that i686 is interoperable with x86_64? It would allow to keep old binaries
running, similar to what old ABI did for 32 to 64 bits transition.
It would require take care that some possible shared data structures (such as
pthread_mutex_t and alike) have the same layout and alignment, add some support
to ldconfig to differentiated between DSO with different ABIs (either through
e_flags as ARM, PT_GNU_PROPERTY used by aarch64 or x86_64, or something else),
bump the required minimum kernel (for 64 bit time_t support), and check current
status of the port.
I am bringing the later because I fixed some recent m68k build issues [1], that
seems to be from gcc changes over the years (as hinted by Andreas Schwab) where
compiler changed some internal defined flags and it was not reflected on glibc
(for a short, it seems that -mcpu=680X0 does not already define __mc68020__).
The build fix is straightforward, but it raised question whether something
else is not broken and has not been caught yet.
Waldemar Brodkorb has posted his results on running glibc 2.38 on qemu and
it shows a lot of regression:
949 FAIL
3344 PASS
99 UNSUPPORTED
16 XFAIL
2 XPASS
I guess the math failures are from the extra rounding and exception testing, which
requires a fully compliant IEEE 754 fp unit (which I guess m68k does not provide).
The last m68k testsuite report where from 2.26 release [1] running under ARAnyM,
which shows the port is a better shape.
I also noted that gcc on mc68060 changed the __DEC_EVAL_METHOD__ to 2, which makes
glibc tests to fail to build (since it assumes __DEC_EVAL_METHOD__ equal 0). This
again raised questions on how the math library would behave depending of the target
chip.
All of this issues and potentially work required for a new ABI makes me wonder
if is really worth to keep *2* distinct ABIs for m68k. Yes, m68k can follow the
MIPS mess and have 28 different ABIs that fails to be fully interoperable; but
I think that if you really want to on this 'gnu32' journey, I think it will be
better to just deprecate the m68k current ABI, remove it from glibc; and move
everything to new ABI.
[1] https://sourceware.org/bugzilla/show_bug.cgi?id=30740
[2] https://sourceware.org/bugzilla/show_bug.cgi?id=30740#c16
[3] https://sourceware.org/glibc/wiki/Release/2.26#M68K
