Required build directory is not made by fmake

[drj11]

[even later edit: it all works in pmake, which makes this more annoying than relevant]

Summary made after I understood the problem some more:

The machinery introduced in mkdir.mk to avoid making directories that already existed (introduced in 630694b827d489425ede498d4a21f0306c39eee8) appears to break an initial build if using FreeBSD make (fmake).

---

I'm on Ubuntu 16.04 using fmake (FreeBSD make) from aptpackagefreebsd-duildutils`:

$ CC=clang fmake -r
"share/mk/mkdir.mk", line 36: warning: duplicate script for target "build/lib" ignored
"share/mk/mkdir.mk", line 36: warning: duplicate script for target "build/lib" ignored
clang -o build/src/adt/alloc.o -I include -ansi -pedantic -O3 -DNDEBUG  -c src/adt/alloc.c
error: unable to open output file 'build/src/adt/alloc.o': 'No such file or
      directory'
1 error generated.
*** [build/src/adt/alloc.o] Error code 1

Stop in /home/ac1xdrj/prj/libfsm.

I'm going to fiddle around making directories until it works.

Incidentally, the same toolchain built kgt just fine

[drj11]

(mostly notes for me now)

The output of fmake -r -V DIR suggests that DIR is being set correctly:

$ CC=clang fmake -r -V DIR
"share/mk/mkdir.mk", line 36: warning: duplicate script for target "build/lib" ignored
"share/mk/mkdir.mk", line 36: warning: duplicate script for target "build/lib" ignored
build/include/fsm build/include/re build/include build/src/adt build/src/print build/src/libfsm/cost build/src/libfsm/pred build/src/libfsm/print build/src/libfsm/walk build/src/libfsm/vm build/src/libfsm build/src/libre/class build/src/libre/dialect/glob build/src/libre/dialect/like build/src/libre/dialect/literal build/src/libre/dialect/native build/src/libre/dialect/pcre build/src/libre/dialect/sql build/src/libre/dialect build/src/libre/print build/src/libre build/src/fsm build/src/re build/src/retest build/src/lx/print build/tests/dump-lx build/src/lx build/src build/tests/capture build/tests/complement build/tests/intersect-ccuc build/tests/intersect build/tests/ir build/tests/eclosure build/tests/subtract-cuc build/tests/subtract build/tests/determinise build/tests/glush build/tests/glob build/tests/like build/tests/literal build/tests/minimise-rdrd build/tests/minimise build/tests/native build/tests/pcre-pcregrep build/tests/pcre build/tests/pcre-classes build/tests/pcre-anchor build/tests/pcre-flags build/tests/pcre-repeat build/tests/pred build/tests/reverse build/tests/trim build/tests/union build/tests/set build/tests/stateset build/tests/sql build/tests/hashset build/tests/queue build/tests/aho_corasick build/tests build/pc build/lib build/lib build/lib build/bin

[drj11]

I see that my trees for libfsm (which does not) and kgt (which works) reference different kmkf SHAs:

For libfsm

ac1xdrj@owl:~/prj/libfsm$ git submodule status
 f1bea106c89d5e39fe2be3ef52f8140d00b0608a share/git/kmkf (f1bea10)

For kgt

ac1xdrj@owl:~/prj/kgt$ git submodule status
 024867d34c7d8741d5f481e525f81845fe153823 share/git/kmkf (024867d)

And the mkdir.mk contained in kmkf is different. In fact the diff is:

ac1xdrj@owl:~/prj$ diff {kgt,libfsm}/share/git/kmkf/share/mk/mkdir.mk
10c10,29
< mkdir:: ${DIR}
---
> mkdir::
> 
> # Double-colon targets are always considered out of date,
> # so even though these directories are up to date, they
> # would always be recreated. Note there's no need to
> # re-create an out of date directory anyway.
> #
> # So we avoid that situation here by only depending on
> # directories which do not currently exist.
> 
> .for dir in ${BUILD} ${DIR}
> .if !exists(${dir})
> mkdir:: ${dir}
> .endif
> .endfor
> 
> # mkdir -p is used for ${BUILD} because that's used-specified
> # and we don't know what directories might exist above that.
> # Whereas for directories beneath it, we ought to account for
> # them all and so purposefully do not use -p.

(that is, the broken one has this extra loop instead of just the plain mkdir:: ${DIR} rule)

That seems suspicious.

[drj11]

well, I can workaround this changing the commit of the submodule:

cd share/git/kmkf/ 
git checkout 0248 # short for 024867d34c7d8741d5f481e525f81845fe153823

Is it possible that exists is buggy in FreeBSD make?

(it breaks on the linker step now, but if that is a problem, it clearly belongs in a different issue)

[katef]

I believe this is addressed by: katef/kmkf@596fea3
(which is now present on libfsm's main)