On Tue, Feb 19, 2013 at 7:10 PM, Benjamin Kaduk <ka...@mit.edu> wrote:
> We've had a few commits over the years to help parallel builds along by
> splitting rules with two targets into two separate rules, one a pure
> dependency and the other with the actual contents of the rule (usually
> compile_et).
>
> A typical example is 147aeeb6c59b5f4a9f8a795a91e0c4**2ecf80278c
> (gerrit/257) (trimmed):
> -budb_errs.c budb_client.h: budb_errs.et budb_client.p.h
> +budb_client.h: budb_errs.c
> +
> +budb_errs.c: budb_errs.et budb_client.p.h
> $(RM) -f budb_client.h budb_errs.c; ${COMPILE_ET} -p ${srcdir}
> budb_errs -h budb_client
>
> However, these changes seem to cause problems for FreeBSD make, even for
> serial builds. The build log shows a failure to install budb_client.h , No
> such file or directory, but then continues on to compile linktest before
> bailing out (sorry for the poorly wrapped copy/paste):
>
> install -o root -g wheel -m 444 JUAFS/libjuafs.a
> /usr/ports/net/openafs/work/**openafs-1.6.2/lib/libjuafs.a
> install -o root -g wheel -m 444 UAFS/libuafs.a
> /usr/ports/net/openafs/work/**openafs-1.6.2/lib/libuafs.a
> case amd64_fbsd_91 in alpha_dux*|sgi_*|sun4x_*|**
> sunx86_*|rs_aix*|*linux*|hp_**ux11*|ia64_hpux*|*_darwin_1*|***nbsd*|*obsd*|*fbsd*)
> cd src && cd shlibafsauthent && make all ;; *) echo Not building shared
> libafsauthent for amd64_fbsd_91 ;; esac
> install -o root -g wheel -m 444 budb_client.h /usr/ports/net/openafs/work/
> **openafs-1.6.2/include/afs/**budb_client.h
> install: budb_client.h: No such file or directory
> *** [/usr/ports/net/openafs/work/**openafs-1.6.2/include/afs/**budb_client.h]
> Error code 71
> cc -g -O -I/usr/ports/net/openafs/work/**openafs-1.6.2/src/config
> -I/usr/ports/net/openafs/work/**openafs-1.6.2/include -I. -I. -O2 -pipe
> -fPIC -c budb_errs.c
> [ yes != "" ] || case amd64_fbsd_91 in alpha_dux*|sgi_*|sun*_5*|rs_**
> aix*|*linux*|hp_ux11*|ia64_**hpux*|*[nof]bsd*) cd src && cd tptserver &&
> make all ;; *_darwin_[1-6][0-9]) echo Not building MT ptserver for
> amd64_fbsd_91 ;; *_darwin_*) cd src && cd tptserver && make all ;; *)
> echo Not building MT ptserver for amd64_fbsd_91 ;; esac
>
>
>
> The make debugging output is not immediately enlightening, but provides
> some hints.
>
> budb_server (???) depends on ${TOP_INCDIR}/afs/budb_client.**h depends on
> budb_client.h depends on budb_errs.c depends on budb_errs.et. Yet
> budb_errs.c has been generated from the rule for budb_errs.h (${COMPILE_ET}
> -p ${srcdir} budb_errs) and is up-to-date, so make's internal bookkeeping
> sees the empty rule to make budb_client.h from budb_errs.c, claims succes
> (I don't see this in the debugging log, though?!) and tries to install the
> header.
>
> The problem seems to stem from there being two calls to compile_et using
> budb_errs.et, one of which makes budb_errs.h and the other which makes
> budb_client.h. Swapping the order in which the multiple targets are split
> out, so that budb_errs.c depends on budb_client.h which is made via
> compile_et, lets my build finish here, but it's probably not the best
> solution.
>
> Is there reason to not make a budb_client.h that is the contents of
> budb_client.p.h with #include <budb_errs.h> added? Alternately we could
> process the template manually with the shell.
>
> [tbudb has the same issue]
>
> -Ben
>
Sounds like we want commit 0d0b5048a5b68d4079b13609775f9234183d1947 from
master?
Marc
