[lld] r308544 - Speed up gdb index creation.

[Sean Silva via llvm-commits]

On Fri, Jul 21, 2017 at 9:14 AM, George Rimar via llvm-commits <
llvm-commits at lists.llvm.org> wrote:

> >>Not sure. I would recommend doing another benchmark run. We have to
> >>compare the delta from lld to lld --gdb-index to the delta from gold to
> >>gold --gdb-index.
> >>
> >>Cheers,
> >>Rafael
> >
> >I can do these benchmarks tomorrow.
> >
> >Also  going to look in profiler to check where it is
> >slow downs now.
> >
> >George.
>
> I did these benchmarks today, used clang objects and reproduce from
> clang-gdb-index folder
> of lld-speed-test.tar.xz archive you posted earlier (thanks, btw !) and
> results are below:
>
> * ld.gold with --gdb-index: 26.54s
> * ld.lld     with --gdb-index: 15.33s
> * ld.gold w/o --gdb-index: 21.65s
> * ld.lld     w/o --gdb-index: 10.14s
>
> Both LLD and ld.gold spends about 5 seconds for building gdb-index.
>

This is awesome!

At least in these rough measurements, we still seem to be 5.2 vs 4.9
seconds, so LLD is about 6% slower, which suggests that there's still some
opportunities for single-thread speedup (unless gold's implementation is of
gdb-index is hyperoptimized, though I doubt it). We generally have found
that with decent profiling, we can generally be measurably faster than gold
in the single-threaded case. Or in other words, say we target being at
least 10% faster than gold than gdb-index as reasonable to achieve with
some profiling. Then we have 6+10 = 16% speedup on the table, which is
fairly significant.

-- Sean Silva

>
> George.
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170724/ef13a8cd/attachment.html>