[pacman-dev] stripping binaries in makepkg

[Allan McRae]

On 23/09/10 15:08, Dan McGee wrote:
> On Wed, Sep 22, 2010 at 11:55 PM, Sven-Hendrik Haase<sh at lutzhaase.com>  wrote:
>>   I'd also like an answer to this. Why not do away with the individual
>> directories? Locating the binaries even in a fairly large package will
>> be fairly quick in any case.
>>
>> -- Sven-Hendrik
>>
>>
>
> Commit 721ceee1e2c9b18425d84cf39f6541b2f04072b3 might offer some
> insight, and then there is the original code from the git import:
>
> +# strip binaries
> +if [ "$NOSTRIP" = "0" ]; then
> +       msg "Stripping debugging symbols from libraries..."
> +       find pkg/{,usr,usr/local,opt/*}/lib -type f -not -name "*.dll"
> -not -name "*.exe" \
> +               -exec /usr/bin/strip --strip-debug '{}' \; 2>&1 \
> +               | grep -v "No such file" | grep -v "format not recognized"
> +       msg "Stripping symbols from binaries..."
> +       find pkg/{,usr,usr/local,opt/*}/{bin,sbin} -type f -not -name
> "*.dll" -not -name "*.exe" \
> +               -exec /usr/bin/strip '{}' \; 2>&1 \
> +               | grep -v "No such file" | grep -v "format not recognized"
> +fi
>
> This leads me to believe stripping *.exe or *.dll files is bad news,
> and since we don't hardcode that anymore, we also limit where we
> traverse in the package filesystem.
>

This should not be an issue now as we do more testing on file types 
before stripping.  makepkg will not touch .dll files and .exe ones are 
probably cross-compiled and should have !strip in their PKGBUILD anyway.


So...  with that in mind I did some testing on this.

For the "gcc-snapshot" package in the AUR, the time taken by the current 
stripping is 9 sec.  Adjusting the strip to look at all files takes that 
to 9.3 sec.  So even with the addition of running "file" on all headers 
etc, the majority of the time seems to be taken with the strip command.

Another example, python 3 takes ~28 sec with the current method and ~29 
when testing all files.

Lets take a very extreme example - glest-data, a 110M arch=any package. 
  That takes 0.01 sec with the current method and 9 sec when checking 
every file.  That is about as extreme as possible...


My conclusion?  For most binary packages, the time taken in the 
stripping step is mainly used for the actual stripping of files and not 
testing which files to strip.  I think this justifies removing the 
STRIP_DIRS variable from makepkg.  For large architecture independent 
packages where the difference from this change is noticeable, the 
packager should add !strip to their PKGBUILD.

Sound OK?

Allan