[Scons-users] Using a Decider to prevent shared lib relinks

Bill Deegan bill at baddogconsulting.com
Tue May 26 12:06:00 EDT 2020


Also:

            dep.attributes.csig_is_lib_symbols_md5 = True # rather than object()



On Tue, May 26, 2020 at 8:24 AM Bill Deegan <bill at baddogconsulting.com>
wrote:

> Also I wouldn't do the whereis inside the function.
>
> def compute_symbols_md5(lib_filename):*    dumpbin = WhereIs('dumpbin', env['ENV']['PATH'])*
>     s = subprocess.check_output([dumpbin, '/exports', lib_filename])
>     return SCons.Util.MD5signature(s)
>
> And calling subprocess inside a decider also not really encouraged.
>
>
> On Tue, May 26, 2020 at 7:30 AM Andrew C. Morrow <
> andrew.c.morrow at gmail.com> wrote:
>
>>
>> Hi Olivier -
>>
>> You might like to read the blog post describing my investigations on this
>> topic:
>> https://engineering.mongodb.com/post/pruning-dynamic-rebuilds-with-libabigail
>>
>> Thanks,
>> Andrew
>>
>>
>> On Mon, May 25, 2020 at 9:56 PM <orenaud at coventor.com> wrote:
>>
>>> Thanks, your message and the recent message from Andrew C. Morrow ("My
>>> own Decider function") put me on the right track. The missing piece for me
>>> was to understand that it is possible, and even necessary, to assign the
>>> value of csig in this decider function. I now realize that it is more or
>>> less mentioned in the doc, in this excerpt:
>>>
>>> Note how the signature information for the dependency file has to get
>>> initialized via get_csig during each function call (this is mandatory!)
>>>
>>> Here is the solution I came up with. From my limited testing, it seems
>>> to do what I want:
>>>
>>> def compute_symbols_md5(lib_filename):
>>>     dumpbin = WhereIs('dumpbin', env['ENV']['PATH'])
>>>     s = subprocess.check_output([dumpbin, '/exports', lib_filename])
>>>     return SCons.Util.MD5signature(s)
>>>
>>> def is_lib(filename, env):
>>>     return filename.startswith(env.subst('$LIBPREFIX')) and filename.endswith(env.subst('$LIBSUFFIX'))
>>>
>>> def gen_my_decider(env):
>>>     default_decider = env.decide_target
>>>
>>>     def my_decider(dep, tgt, prev_ni, repo_node=None):
>>>         if not is_lib(str(dep), env):
>>>             return default_decider(dep, tgt, prev_ni, repo_node)
>>>
>>>         dep_node_info = dep.get_ninfo()
>>>
>>>         if not hasattr(dep.attributes, 'csig_is_lib_symbols_md5'):
>>>             dep_node_info.csig = compute_symbols_md5(str(dep))
>>>             dep.attributes.csig_is_lib_symbols_md5 = object()
>>>
>>>         if prev_ni is None or not hasattr(prev_ni, 'csig'):
>>>             return True
>>>
>>>         return prev_ni.csig != dep_node_info.csig
>>>
>>>     return my_decider
>>>
>>> my_env.Decider(gen_my_decider(my_env))
>>>
>>>
>>> I use `csig_is_lib_symbols_md5` to mark the lib nodes that contain a md5
>>> of the dumped symbols in `csig`, as opposed to a md5 of the file content
>>> (it works because the lib file is always represented by the same instance
>>> of a Node during a build session). I introduced `gen_my_decider` so that I
>>> have a place to store the "default" decider to use for non-lib files. I
>>> still need to distinguish between static lib file and dynamic lib file on
>>> windows.
>>>
>>> On 5/25/2020 11:26 PM, Bill Deegan wrote:
>>>
>>> Olivier,
>>>
>>> You're not the first to head down this path as it would be useful to
>>> (configurably) avoid such unnecessary relinks.
>>> I think MongoDB has implemented something like this.
>>>
>>> Please keep in mind that dep and tgt are not strings, the are File()
>>> nodes, which have access to information stored in .sconsign with regards to
>>> the signature of the file from the previous build.
>>> This is (currently) a md5sum of the full contents of the file.
>>> So to adequately do what you're thinking you'd need a way to store the
>>> ABI info (or a hash thereof) as the content signature (also known as csig)
>>> for the node(s) in question.
>>>
>>> So it's going to be a pretty deep dive into the code.
>>>
>>> Perhaps someone from MongoDB will chime in here with a pointer to their
>>> code/experience.
>>>
>>> -Bill
>>>
>>> On Mon, May 25, 2020 at 10:11 AM <orenaud at coventor.com> wrote:
>>>
>>>> Hi,
>>>>
>>>> I want to prevent unnecessary re-links against my shared libraries. By
>>>> "unnecessary", I mean the re-links that happen even when the list of
>>>> exported symbols did not change.
>>>>
>>>> In practice, I would like to use the output of `nm --extern-only` on
>>>> Linux or `dumpbin /exports` on Windows as the content to use for the
>>>> signature of the shared library.
>>>>
>>>> I was able to approximate this on a simple example, by explicitly
>>>> specifying the dependency of a target (myExe) that links against a specific
>>>> shared library (myLib.lib):
>>>>
>>>> def dumpLibSymbols(target, source, env):
>>>>     with open(str(target[0]), 'w') as dumpFile:
>>>>         nm = WhereIs('nm', env['ENV']['PATH'])
>>>>         subprocess.call([nm, '--extern-only', str(source[0])], stdout=dumpFile)
>>>>
>>>> dumpLibSymbolsAction = Action(dumpLibSymbols,
>>>>                               "Dumping Symbols for $SOURCE into $TARGET")
>>>> dumpLibSymbolsBuilder = Builder(action=dumpLibSymbolsAction,
>>>>                                 src_suffix="$LIBSUFFIX",
>>>>                                 suffix=".lib.exports")
>>>> env.Append(BUILDERS= {'DumpLibSymbols': dumpLibSymbolsBuilder})
>>>>
>>>> Ignore(myExe, 'myLib.lib')
>>>> Depends(myExe, 'myLib.lib.exports')
>>>>
>>>> Basically, it changes the dependency chain [myExe -> myLib.lib] to
>>>> [myExe -> myLib.lib.exports -> myLib.lib], without changing the link
>>>> command itself.
>>>>
>>>> While this example works, I am unable to generalize it: the calls to
>>>> Ignore and Depends must be explicit for all the pairs exe/lib (or lib/lib).
>>>> What I want is for it to be automatic for anything that depends on a shared
>>>> library.
>>>>
>>>> It seems to me that the Decider function (
>>>> https://scons.org/doc/production/HTML/scons-user/ch06.html#idm962
>>>> <https://urldefense.proofpoint.com/v2/url?u=https-3A__scons.org_doc_production_HTML_scons-2Duser_ch06.html-23idm962&d=DwMFaQ&c=RWI7EqL8K9lqtga8KxgfzvOYoob76EZWE0yAO85PVMQ&r=4kZkTHayCjeGAQcCy0E295XGr9xJGf-CHs_MqTnCu2I&m=BObPKvTudrvwwNERpOKI_nAnPLa7214U5QnkFJLchJs&s=w6CW5DrsZgcPu8gkgq_bP5jjjWlISlfBgzcaWyEoSLM&e=>)
>>>> is exactly what I need. Indeed, I want to decide a target is out of date
>>>> based on the output of nm/dumpbin. The problem is that I don't understand
>>>> how I can do that in practice. The example in the docs uses an unspecified
>>>> `specific_part_of_file_has_changed(dep, tgt)` function, but I don't see how
>>>> this function can do its job given its inputs. To me, this function needs
>>>> an additional information: the previous content of the dependency file (or
>>>> a hash of only the part it is interested in). Similarly, my custom Decider
>>>> would also need to know the previous content of nm/dumpbin in addition to
>>>> its current content.
>>>>
>>>> Can someone points me to a real world usage of the Decider function? Is
>>>> it the right tool I need, or did I overlook a better way to achieve my goal?
>>>>
>>>> Thanks,
>>>>
>>>> Olivier Renaud
>>>> _______________________________________________
>>>> Scons-users mailing list
>>>> Scons-users at scons.org
>>>> https://pairlist4.pair.net/mailman/listinfo/scons-users
>>>> <https://urldefense.proofpoint.com/v2/url?u=https-3A__pairlist4.pair.net_mailman_listinfo_scons-2Dusers&d=DwMFaQ&c=RWI7EqL8K9lqtga8KxgfzvOYoob76EZWE0yAO85PVMQ&r=4kZkTHayCjeGAQcCy0E295XGr9xJGf-CHs_MqTnCu2I&m=BObPKvTudrvwwNERpOKI_nAnPLa7214U5QnkFJLchJs&s=fHlnPOUCfVYT_y1hB7DzmO1UwCCHOf89CzXmmg0i1gk&e=>
>>>>
>>>
>>> _______________________________________________
>>> Scons-users mailing list
>>> Scons-users at scons.org
>>> https://pairlist4.pair.net/mailman/listinfo/scons-users
>>>
>> _______________________________________________
>> Scons-users mailing list
>> Scons-users at scons.org
>> https://pairlist4.pair.net/mailman/listinfo/scons-users
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://pairlist4.pair.net/pipermail/scons-users/attachments/20200526/a7e932c5/attachment.html>


More information about the Scons-users mailing list