[Bug default/29464] abidw performance regression on vmlinux

["gprocida at google dot com" <sourceware-bugzilla@sourceware.org>]

https://sourceware.org/bugzilla/show_bug.cgi?id=29464

--- Comment #3 from gprocida at google dot com ---
Hi Dodji.

I have done some tests by merging dodji-perf (e6f222c8) into AOSP and running
our tests. In more detail:

I noticed that there was a merge conflict of "dwarf-reader: Revamp the
canonical type DIE propagation algorithm" with my hack that limited comparisons
per DIE pair to 10000. Given it seemed that the changes were working in the
same area, I reverted my hack and remerged.

I ran into trouble as your development branch is not known to AOSP, so I
actually ended up with: revert, merge, cherry-pick x 2.

I need to retest without any AOSP things. However, here is some initial
feedback:

1. Testing (unintentionally) at the second last cherry-pick ("dwarf-reader:
Revamp the canonical type DIE propagation algorithm") showed some test timeouts
on our standard kernel tests.

This means I cannot push that series to AOSP. I will consider reinstating my
hack and trying again, after other tests.

The test suite run also shows other issues:
* name changes like "long long unsigned" to "unsigned long long int"
* linearly probed - and so unstable - hash-based type ids for types like
"unsigned int" and "unsigned short int" (the strings that get hashed are the
same for the different types)

The last of these issues will also need to be resolved - we may decide to do it
in post-processing, but I imagine it's only a small change in libabigail.

2. Testing at the last cherry-pick ("Allow restricting analyzed decls to
exported symbols") also had test timeouts. However, more seriously, the run
time (for a 4.19 kernel) is much worse than in 1.

I haven't yet done any ABI comparisons using this abidw's output as it still
running at over 30 minutes.

-- 
You are receiving this mail because:
You are on the CC list for the bug.