Please help prep for a new libffi release

Please help prep for a new libffi release

[Anthony Green]

I'd like to start prepping for a new release.

Over the past couple of weeks I've made some improvements to the
Github Actions CI infrastructure, notably taking advantage of the GCC
Compile Farm to test a number of previously untested platforms,
including the new loongarch64 port, Solaris, AIX, Aarch64, including
an M1 Linux host, and several combinations of macos & compiler -- 21
platforms in all.

The CI tests all come up green today, so it should be easy to track
regressions as we go through any patch backlog.

The test results, however, are not all perfect.  Some ports are still
not passing struct args by value, the Cygwin results are a mess
(uninvestigated), and some of the qemu-based testing is broken (eg.
m68k, s390).  The complete list of problems is available here:
https://github.com/libffi/rlgl-policy/blob/master/XFAIL

If you aren't familiar with "red light green light", this file
contains pattern matching expressions we use to filter the dejagnu
results in order to decide whether or not a test run is "good enough".
It doesn't make sense to xfail tests in the dejagnu test source when
the problem is in the simulator we're testing with, so we use rlgl as
a higher-level tool to evaluate the test results with an extra layer
of "expected fail" policy.

Any additional eyes-on or help with known problems would be
appreciated.  Also, if you are sitting on any patches, please let me
know.

Thanks!

AG

Re: Please help prep for a new libffi release

[Anthony Green]

On Wed, Sep 7, 2022 at 9:44 AM Anthony Green <green@moxielogic.com> wrote:
>
>  the Cygwin results are a mess

I fixed this porting static trampolines to Cygwin.  Cygwin emulates
Linux enough (eg. /proc filesystem), that turning on static
trampolines was easy.  The Cygwin results are pretty clean now.

AG

Re: Please help prep for a new libffi release

[Anthony Green]

I spent some time trying to fix the 32-bit x86 results.  As it
happens, libffi is doing things mostly right.  It's GCC that can't
return small structs (<= 16 byte) properly for any of the MS ABIs.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105932
I'm just going to ignore this for now and force the CI tests to pass
with new rlgl policy.

I'll probably spin out the new release in the next day or two.  As far
as I can tell there are no regressions.

These are the highlights:

        All struct args are passed by value, regardless of size, as per ABIs.
        Enable static trampolines for Cygwin.
        Add support for Loongson's LoonArch64 architecture.
        Fix x32 static trampolines.
        Fix 32-bit x86 stdcall stack corruption.
        Fix ILP32 aarch64 support.

AG

On Mon, Sep 12, 2022 at 10:19 AM Anthony Green <green@moxielogic.com> wrote:
>
> On Wed, Sep 7, 2022 at 9:44 AM Anthony Green <green@moxielogic.com> wrote:
> >
> >  the Cygwin results are a mess
>
> I fixed this porting static trampolines to Cygwin.  Cygwin emulates
> Linux enough (eg. /proc filesystem), that turning on static
> trampolines was easy.  The Cygwin results are pretty clean now.
>
> AG