hg diff seems to traverse unnecessary dirs

[Jeff Squyres]

Greetings, and thanks for Mercurial!  A bunch of us use it in the Open  
MPI project in conjunction with the master SVN repository for side  
development (https://svn.open-mpi.org/trac/ompi/wiki/UsingMercurial,  
if you care).

I noticed a peculiar behavior this morning (using hg v1.0.1): it seems  
like "hg diff" traverses into directories that it shouldn't need to.   
I have not dived into the hg source code at all; this is mainly from a  
"least astonishment" user perspective.

I was testing Open MPI's "make dist" this morning and had made a test  
tarball.  I expanded that tarball in the same tree as my Mercurial  
repository+checkout and started running "configure" in that tree.  So  
it looked like this:

   # My hg repo and checkout tree
   $HOME/hg/romio-1.0.7/
   # Where I expanded my test tarball (not tracked by hg)
   $HOME/hg/romio-1.0.7/openmpi-1.4a1hg37ea0dd213fb/

I then was checking my work in the main tree and ran "hg diff" on a  
specific subdirectory.  I was surprised to see the following:

-----
[4:34] svbu-mpi:~/hg/romio-1.0.7 % hg di ompi/mca/io/romio/
abort: No such file or directory: /home/jsquyres/hg/romio-1.0.7/./ 
openmpi-1.4a1hg37ea0dd213fb/confdir.3997
[4:34] svbu-mpi:~/hg/romio-1.0.7 %
-----

I can guess what happened: configure runs lots of tests in short-lived  
subdirectories.  It creates a subdir, runs a test, and then whacks the  
subdir.  So if hg scans a directory, it might see a subdirectory but  
by the time it tries to traverse into that subdirectory, it could be  
gone.

But I asked hg to traverse into a specific subdirectory (ompi/mca/io/ 
romio) -- why did it look elsewhere?  Indeed, the openmpi-v1.4* tree  
is not tracked by Mercurial at all -- why did an "hg diff" traverse so  
deep into it at all?  The "least astonishment" factor here is that I  
thought it would be safe to run "hg diff" on the tree, even though I  
was actively changing files in a not-tracked-by-Mercurial  
subdirectory.  Put simply: it surprised me that activity in an area  
not tracked by Mercurial caused a Mercurial command to fail.

Open MPI is a goodly-sized code base (Mercurial tracks around 4500  
files); several Mercurial operations take a few seconds, even on a  
local/fast file system.  I always wondered why that was, especially  
when you run what should be a small operation (e.g., "hg diff" on ompi/ 
mca/io/romio is a pretty small subtree out of the overall code base;  
"svn diff" on the same tree is nearly instantaneous).  Could this  
"traversing unneeded directories" behavior be contributing to the  
"slow" operations?

-- 
Jeff Squyres
Cisco Systems