[Haskell-cafe] A small Darcs anomoly

[Nick Bowler]

On 2011-04-28 15:23 +0000, malcolm.wallace wrote:
> Then I suggest that your build tools are broken.  Rebuilding should
> not depend on an _ordering_ between modification times of source and
> object, merely on whether the timestamp of the source file is
> different to its timestamp the last time we looked.  (This requires
> your build tools to keep a journal/log, yes, but it is the only safe
> way to do it.)

Right.  The /order/ of the timestamps is wrong when a build directory is
shared between repositories (isn't that what I said?).  Try it yourself
with cabal: it will fail.

Consider two repos, A and B, each with different versions of foo.x,
that (when compiled) produces the output foo.y.  We store the build in
the directory "C".

Initially, say A/foo.x has a mtime of 1, and B/foo.x has an mtime of 2.

We do a build of A, producing the output file C/foo.y.  say C/foo.y now
has a mtime of 3.

Now we do a build in B.  The build system sees that C/foo.y has a
mtime of 3, which is newer than B/foo.x's mtime of 2.  The build
system therefore does not rebuild C/foo.y.

> It is relatively common to change source files to have an older
> timestamp rather than a newer one.  This should not cause your build
> system to ignore them.  It can happen for any number of reasons:
> restoring from backup, switching repository, bisecting the history of
> a repo, clock skew on different machines, ....

All of these operations update the mtimes on the files...

-- 
Nick Bowler, Elliptic Technologies (http://www.elliptictech.com/)