[Bug default/27995] New: abidiff performance regression

[Bug default/27995] New: abidiff performance regression

[gprocida at google dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=27995

            Bug ID: 27995
           Summary: abidiff performance regression
           Product: libabigail
           Version: unspecified
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: default
          Assignee: dodji at redhat dot com
          Reporter: gprocida at google dot com
                CC: libabigail at sourceware dot org
  Target Milestone: ---

Created attachment 13501
  --> https://sourceware.org/bugzilla/attachment.cgi?id=13501&action=edit
XML file for self-comparison

abidiff comparing XML with identical XML has become a lot slower, at least for
certain Android kernel ABI XML files.

Run time changes from 3-5 seconds to more than 5 minutes. This does cause
trouble with test infrastructure etc.

I've bisected this to the following recent commit:

7a9fa3fe5a823d1627e83571dd862b3f07196ae0 "abixml reader: Fix recursive type
definition handling"

I'll try to do some profiling today. A sample XML file that triggers this is
attached.

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

[Bug default/27995] abidiff performance regression

[gprocida at google dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=27995

--- Comment #1 from gprocida at google dot com ---
Profiling is suggesting this is (has become?) very hot:

abigail::comparison::diff::traverse(abigail::comparison::diff_node_visitor&)

with 25203 non-recursive and 116915966 recursive calls.

With 7a9fa3fe5a823d1627e83571dd862b3f07196ae0~1, this function isn't even in
the gprof call graph.

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

[Bug default/27995] abidiff performance regression

[dodji at redhat dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=27995

dodji at redhat dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2021-06-21

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

[Bug default/27995] abidiff performance regression

[dodji at redhat dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=27995

--- Comment #2 from dodji at redhat dot com ---
Okay, I am still looking into this, but I have a patch set 
(still not fully documented) for it at on the temporary working branch named
'PR27995'.  It can be browsed at
https://sourceware.org/git/?p=libabigail.git;a=shortlog;h=refs/heads/PR27995.

You might want to test it in your environment to see how it behaves.  I am
still looking into some fallouts of this patch.  It seems to be working good
enough from my testing so far, though.

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

[Bug default/27995] abidiff performance regression

[gprocida at google dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=27995

--- Comment #3 from gprocida at google dot com ---
This does indeed fix the issue.

It may also speed up other cases where abidiff takes quite a while. Checking...

A kernel ABI XML diff I'm using for performance measurement at origin/PR27995:

real    1m25.813s
user    1m25.361s
sys     0m0.452s

And at current origin/master:

real    2m47.301s
user    2m46.856s
sys     0m0.442s

This is a real improvement!

I even diff'd the diffs and the new one looks better. There was an odd line,
mentioning indirect sub-type changes but not followed by any diff text, in the
old one.

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

[Bug default/27995] abidiff performance regression

[gprocida at google dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=27995

--- Comment #4 from gprocida at google dot com ---
Verified as (still) fixed at current master.

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

[Bug default/27995] abidiff performance regression

[gprocida at google dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=27995

gprocida at google dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
             Status|ASSIGNED                    |RESOLVED

--- Comment #5 from gprocida at google dot com ---
This was fixed in commit 46b1ab08b0d1201599c474646514ae72978bd120 "Bug 27995 -
Self comparison error from abixml file".

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