[PATCH] D142556: [DebugInfo] Merge partially matching chains of inlined locations

Fri Feb 10 10:17:43 PST 2023

ayermolo added a comment.

In D142556#4118632 <https://reviews.llvm.org/D142556#4118632>, @dstenb wrote:

> @dblaikie, @ayermolo: I have now ran measurements on an otherwise unused machine to get rid of the noise.
>
> The measurements are from 21 number of opt binary builds per each build type (RelWithDebInfo with/without ThinLTO), using Clang built with the Release build type as before. The time is user+sys (I had missed the system time last round) taken from bash's `time`.
>
>   # With patch:
>   mean: 81.391325 (81m23.479476s)
>   median: 81.387383 (81m23.243000s)
>   std: 0.052490
>   
>   # Without patch:
>   mean: 81.216625 (81m12.997476s)
>   median: 81.216417 (81m12.985000s)
>   std: 0.058820
>   
>   # With patch (ThinLTO build):
>   mean: 90.751952 (90m45.117095s)
>   median: 90.761333 (90m45.680000s)
>   std: 0.045923
>   
>   # Without patch (ThinLTO build):
>   mean: 90.619793 (90m37.187571s)
>   median: 90.600950 (90m36.057000s)
>   std: 0.085509
>
> So, it seems like this patch adds ~10 seconds of CPU time for  these ~80-90 minute builds.
>
> Is this a sufficient measurement?
>
> @ayermolo: And in case you meant the effect on the section and file sizes for a ThinLTO build, those are available in my previous comment!

Ah sorry I am blind.
Another data point. I used this patch on some internal build with all the bells and whistles and increase was about 28MB total. Which is negligible.

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D142556/new/

https://reviews.llvm.org/D142556