Avoid race condition

[bmwiedemann]

There seems to be some race-condition in build.rs
that causes random build results to be produced for /usr/bin/difft.

This change seems to have negligible influence on overall build time.

If you make a better patch, I can test it.

[Wilfred]

Thanks for the PR!

Can you share more information? Are you getting invalid binaries, or valid but not bit-for-bit identical binaries? What platform are you building on?

I'd also be surprised if building all the C libraries in parallel wasn't still a big improvement for build time. Try removing .flag_if_supported("-Wno-unused-parameter") from build.rs and see if the build time is comparable with and without your change.

[bmwiedemann]

It produced different binary results. Probably all working, but did not test. I think, some elements were ordered differently.

You can use https://github.com/bmwiedemann/reproducibleopensuse/blob/master/filterdiff with 'strings' and difft binaries as params. (and the diff is probably more readable with lto = "off" )

I measured build times on a 2x16 core 2.6GHz host with build VMs:
without patch on 1-core-VM: 657s
with patch on 1-core-VM: 654s
without patch on 4-core-VM: 142s
with patch on 4-core-VM: 175s

So the patch does slightly speed up single-core builds but slow down multi-core builds. If we can avoid the variations in other ways, that would be good.

-Wno-unused-parameter seems to not make a difference.