[configurator] more flexible #define parsing

[rgrinberg]

Allow duplicate values for the same key as long as they are the same.

@rwmjones see if this fixes the issue for you

[nojb]

LGTM

[rwmjones]

I'm going to test this right now.

[rwmjones]

Yes this does appear to work here, thanks. I have pushed this preliminary commit to the Fedora branch of dune where we'll use it to build the packages that failed last time (https://bugzilla.redhat.com/show_bug.cgi?id=1870368#c17).

[rwmjones]

We don't seem to be quite there. There is another dune error which may be the same:

Error: Duplicate values for CAIRO_VERSION_MAJOR:
- 00000000001
- 01

Full log here: https://kojipkgs.fedoraproject.org//work/tasks/6856/50036856/build.log

[rwmjones]

I grabbed the test program from the log and compiled it with LTO:

$ gcc -c `rpm --eval '%{optflags}'` `rpm --eval '%{__global_ldflags}'` test.c `pkg-config cairo --cflags`
$ strings test.o  | grep BEGIN
BEGIN-0-01-END
BEGIN-1-016-END
BEGIN-1-00000000016-END
BEGIN-0-00000000001-END

Note again that what you're seeing here is a GCC intermediate form. When using LTO the compiler does only a preliminary parse stage. Actual compilation happens in the linker.

[rgrinberg]

I suppose we could drop leading 0’s for integer valued probes. This is obviously quite fishy though. Could you show the configure script in question? I want to make sure they are integer probes.

…

On Aug 24, 2020, 1:08 AM -0700, rwmjones ***@***.***>, wrote: I grabbed the test program from the log and compiled it with LTO: $ gcc -c `rpm --eval '%{optflags}'` `rpm --eval '%{__global_ldflags}'` test.c `pkg-config cairo --cflags` $ strings test.o | grep BEGIN BEGIN-0-01-END BEGIN-1-016-END BEGIN-1-00000000016-END BEGIN-0-00000000001-END Note again that what you're seeing here is a GCC intermediate form. When using LTO the compiler does only a preliminary parse stage. Actual compilation happens in the linker. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[rwmjones]

I added a dummy main function to the end of the test function. Note we have to somehow have main depend on s0 and s1 otherwise they are optimized out.

main()
{
  return (int)&s0 + (int)&s1;
}

Compiling to a final program, this was the output:

$ strings a.out  | grep BEGIN
BEGIN-1-00000000016-END
BEGIN-0-00000000001-END

I still think this whole approach is suspect and studying what autoconf does is the way to go.

[rwmjones]

Could you show the configure script in question?

See the log: https://kojipkgs.fedoraproject.org//work/tasks/6856/50036856/build.log

[rgrinberg]

I still think this whole approach is suspect and studying what autoconf does is the way to go.

We'd happily accept a PR taking us in this direction.

[rwmjones]

It's unlikely to happen soon. Is there any workaround I can use to get this working so I can build cairo?

[rwmjones]

Are the CFLAGS embedded in dune when it is built, or do they come from the environment when dune is used?

[rgrinberg]

Neither. They come from the cflags found in with ocamlc -config.

…

On Aug 24, 2020, 1:55 AM -0700, rwmjones ***@***.***>, wrote: Are the CFLAGS embedded in dune when it is built, or do they come from the environment when dune is used? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[rwmjones]

Is it possible to adjust these when building dune? Our ocamlc -config includes LTO flags, and that's how we want it to be because distros are generally moving to enabling LTO everywhere. We just want dune to use a slightly different set of flags when running these tests as a workaround.

[rgrinberg]

I pushed another attempt at a workaround. See if that works.

[rgrinberg]

Is it possible to adjust these when building dune? Our ocamlc -config includes LTO flags, and that's how we want it to be because distros are generally moving to enabling LTO everywhere. We just want dune to use a slightly different set of flags when running these tests as a workaround.

It's not possible, and I'm not sure it's the right way to solve this problem. Are there any other build tools where you configure the flags it passes at build time? We could just make dune respect CFLAGS and CPPFLAGS as well.

Let's see if @avsm has an opinion on this. I recall he wrestled with similar issues.

[rwmjones]

Yes that fixes ocaml-cairo (and didn't break ocaml-ssl), thanks. I'll add this patch to Fedora as well and continue rebuilding packages with it.

[rgrinberg]

@nojb Could you give this another glance if you don't mind? The implementation is now different.

[nojb]

LGTM, modulo the suggested simplification.