Re: What's cooking in git.git (Feb 2011, #05; Wed, 23)

From: Johannes Schindelin <johannes.schindelin@gmx.de>
Date: Feb 27, 2011, 14:28 UTC
Message-ID: <alpine.DEB.1.00.1102271509540.1561@bonsai2>
In-Reply-To: <20110226110740.GA3096@tugrik.mns.mnsspb.ru>

Hi Kirill,

On Sat, 26 Feb 2011, Kirill Smelkov wrote:

Show 17 quoted lines

> On Fri, Feb 25, 2011 at 10:54:40PM +0100, Johannes Schindelin wrote:
> > 
> > On Fri, 25 Feb 2011, Kirill Smelkov wrote:
> > 
> > > On Fri, Feb 25, 2011 at 12:11:59PM +0100, Johannes Schindelin wrote:
> > > 
> > > > Further, I think that my beloved Shift+Insert will no longer work 
> > > > with your [2].
> > > 
> > > Probably yes,
> > 
> > In my experiment after rebuilding msys-1.0.dll, it still works.
> 
> xser32.dll has nothing to do with msys - it's just a fake stub for
> sh.exe. Before testing, have you "patched" sh.exe the way it is done in
> my ks/xser32.dll
> (http://repo.or.cz/w/msysgit/kirr.git/commit/9d952c74a52f577b2d16d4e4a489541a8fa7fbbd)

Ah. I had to add -Wl,-kill-at to the LDFLAGS so it works.

See the ks/xser32 branch (and the ks/msys branch) of msysgit.git.

Show 17 quoted lines

> > The problem for now is that when I time /share/msysGit/run-tests.sh, 
> > there is hardly any gain from your patches:
> > 
> > Old:
> > 
> > 
> > real    18m1.031s
> > user    6m17.861s
> > sys     19m25.257s
> > 
> > New:
> > 
> > real    17m54.500s
> > user    6m12.319s
> > sys     19m28.567s
> 
> Did you patch sh.exe to link to xser32.dll instead of user32.dll?

Now I did, and the difference is quite noticable:

real 15m37.281s user 5m6.934s sys 15m53.911s

(Note that I did not spend time to increase the N, so there is no point in putting a percentage on the difference.)

I wonder whether we could patch sh.exe so it loads user32.dll _lazily_ rather than using the xser32.dll hack?

(This would also fix the issue that I can no longer use Shift+Insert to paste the clipboard into the Git Bash...)

> Also I can't say for sure (hope yet) how sh-intensitive git tests are, 
> but at least running configure for say gettext or whatever should be 
> visibly faster, at least on wine.

They are pretty intensive, as they are shell scripts all over the place in their own right. I fear, however, that the real problem (maybe not on WINE, but with real Windows) is that sh.exe needs to provide too many POSIX-like things when starting new processes.

I fear, further, that the only way to make things more efficient on Windows is to get rid of the many exec() calls and do more things in-process. (If this is true, JGit should kick msysGit's ass.) This would not apply to the test suite, though, which needs to stay a shell script suite.

But maybe I am wrong, and the performance is lost through memory management and filesystem interaction.

Ciao, Dscho