[lld] r316731 - De-template EhFrameSection. NFC.

[Rafael Avila de Espindola via llvm-commits]

Rui Ueyama <ruiu at google.com> writes:

> I did not benchmark these particular patches as it seems neutral to
> performance to me. Do you have any concerns? I think as we improved
> performance in the last few months, we actually have some room to aim for
> readability and simplicity even if it slows down the linker a little bit.

I have no problem with an informed decision about performance versus
readability.

The key is in the informed part. It is at least non obvious that
checking endianess for so many reads and writes will not have an impact.

I benchmarked this particular patch and it was fine, but we recently
regressed on linking both the linux kernel and firefox/scylla. Those
regressions were also not completely obvious from reading the patch.

For this reason I think setting up performance tracking bots should be a
priority for lld.

Galina, which machines do we still have available for setting up bots?
For this particular case it is critical that the machines be real
hardware (no vm) and not in use for anything else. Having many cores is
not important for now since we will be benchmarking single threaded
performance. Lots (32 GB) of ram would be nice, but not critical.

Cheers,
Rafael