Re: On overriding make variables from the environment...

2018-10-18 Thread Junio C Hamano
SZEDER Gábor  writes:

> So, then it's either 'config.mak', or passing a 'CC=$CC' argument to
> _all_ make commands, including those that are not supposed to build
> anything, but only run the tests.  I find the latter aesthetically not
> particularly pleasing.

The config.mak file is available for individual builder-testers to
customize their build, and in the context of this discussion, I
think the CI builder is just one particular individual who happens
to be non human.  If it is easy to throw suitable settings from within
the CI configuration .yaml files into config.mak, I'd think that is
exactly how the mechanism was invented to be used, so...


Re: On overriding make variables from the environment...

2018-10-18 Thread Johannes Schindelin
Hi Gábor,

On Wed, 17 Oct 2018, SZEDER Gábor wrote:

> On Tue, Oct 16, 2018 at 03:40:01PM -0700, Jonathan Nieder wrote:
> > SZEDER Gábor wrote:
> > > On Tue, Oct 16, 2018 at 02:54:56PM -0700, Jonathan Nieder wrote:
> > >> SZEDER Gábor wrote:
> > 
> > >>> Our Makefile has lines like these:
> > >>>
> > >>>   CFLAGS = -g -O2 -Wall
> > >>>   CC = cc
> > >>>   AR = ar
> > >>>   SPATCH = spatch
> > [...]
> > >>> I'm not sure what to do.  I'm fine with updating our 'ci/' scripts to
> > >>> explicitly respect CC in the environment (either by running 'make
> > >>> CC=$CC' or by writing $CC into 'config.mak').  Or I could update our
> > >>> Makefile to use '?=' for specific variables, but:
> > >>
> > >> That's a good question.  I don't have a strong opinion myself, so I
> > >> tend to trust larger projects like Linux to have thought this through
> > >> more, and they use 'CC = cc' as well.
> > >
> > > I don't think Linux is a good example to follow in this case, see e.g.
> > > 6d62c983f7 (Makefile: Change the default compiler from "gcc" to "cc",
> > > 2011-12-20) (in short: Git is supposed to be buildable with compilers
> > > other than GCC as well, while Linux not really, so they can hardcode
> > > 'CC = gcc').
> > 
> > Nowadays Linux builds with clang as well.  People also have other
> > reasons to override the CC setting (cross-compiling, etc) and to
> > override other settings like CFLAGS.
> > 
> > > Also, the projects I have on hand usually have 'CC = gcc' hardcoded in
> > > their Makefiles, too, but those Makefiles were generated by their
> > > ./configure script (which in turn by ./autogen.sh...), and those tend
> > > to write CC from the environment into the generated Makefiles.
> > 
> > Yes, I think that's what makes travis's setup normally work.  It makes
> > sense to me since ./configure is expected to have more implicit or
> > automatic behavior.  For "make", I kind of like that it doesn't depend
> > on environment variables that I am not thinking about when debugging a
> > reported build problem.
> > 
> > When building Git without autoconf, the corresponding place for
> > settings is config.mak.  Would it make sense for the ci scripts to
> > write a config.mak file?
> 
> A first I though it doesn't, but it turns out it acually does.
> 
> 'ci/run-build-and-tests.sh' basically runs:
> 
>   make
>   make test
> 
> I naively put a 'CC=$CC' argument at the end of the first command,
> thinking it should Just Work...  but then that second 'make test' got
> all clever on me, said "* new build flags", and then proceeded to
> rebuild everything with the default 'cc'.  (With the additional
> complication, that on Travis CI we actually run 'make --quiet test',
> which rebuilds everything, well, quietly...  so the rebuild itself is
> not even visible in the build logs.)
> 
> So, then it's either 'config.mak', or passing a 'CC=$CC' argument to
> _all_ make commands, including those that are not supposed to build
> anything, but only run the tests.  I find the latter aesthetically not
> particularly pleasing.

How about using `MAKEFLAGS`? I ran a quick test:

MAKEFLAGS='CC=blub' make -C .. git.o
make: Entering directory '/usr/src/git/wip'
* new build flags
CC git.o
/bin/sh: blub: command not found

In other words, you could add something like this to the ci/ script:

MAKEFLAGS="${MAKEFLAGS:+$MAKEFLAGS }CC=$CC"
export MAKEFLAGS

Ciao,
Dscho

Re: On overriding make variables from the environment...

2018-10-17 Thread SZEDER Gábor
On Tue, Oct 16, 2018 at 03:40:01PM -0700, Jonathan Nieder wrote:
> SZEDER Gábor wrote:
> > On Tue, Oct 16, 2018 at 02:54:56PM -0700, Jonathan Nieder wrote:
> >> SZEDER Gábor wrote:
> 
> >>> Our Makefile has lines like these:
> >>>
> >>>   CFLAGS = -g -O2 -Wall
> >>>   CC = cc
> >>>   AR = ar
> >>>   SPATCH = spatch
> [...]
> >>> I'm not sure what to do.  I'm fine with updating our 'ci/' scripts to
> >>> explicitly respect CC in the environment (either by running 'make
> >>> CC=$CC' or by writing $CC into 'config.mak').  Or I could update our
> >>> Makefile to use '?=' for specific variables, but:
> >>
> >> That's a good question.  I don't have a strong opinion myself, so I
> >> tend to trust larger projects like Linux to have thought this through
> >> more, and they use 'CC = cc' as well.
> >
> > I don't think Linux is a good example to follow in this case, see e.g.
> > 6d62c983f7 (Makefile: Change the default compiler from "gcc" to "cc",
> > 2011-12-20) (in short: Git is supposed to be buildable with compilers
> > other than GCC as well, while Linux not really, so they can hardcode
> > 'CC = gcc').
> 
> Nowadays Linux builds with clang as well.  People also have other
> reasons to override the CC setting (cross-compiling, etc) and to
> override other settings like CFLAGS.
> 
> > Also, the projects I have on hand usually have 'CC = gcc' hardcoded in
> > their Makefiles, too, but those Makefiles were generated by their
> > ./configure script (which in turn by ./autogen.sh...), and those tend
> > to write CC from the environment into the generated Makefiles.
> 
> Yes, I think that's what makes travis's setup normally work.  It makes
> sense to me since ./configure is expected to have more implicit or
> automatic behavior.  For "make", I kind of like that it doesn't depend
> on environment variables that I am not thinking about when debugging a
> reported build problem.
> 
> When building Git without autoconf, the corresponding place for
> settings is config.mak.  Would it make sense for the ci scripts to
> write a config.mak file?

A first I though it doesn't, but it turns out it acually does.

'ci/run-build-and-tests.sh' basically runs:

  make
  make test

I naively put a 'CC=$CC' argument at the end of the first command,
thinking it should Just Work...  but then that second 'make test' got
all clever on me, said "* new build flags", and then proceeded to
rebuild everything with the default 'cc'.  (With the additional
complication, that on Travis CI we actually run 'make --quiet test',
which rebuilds everything, well, quietly...  so the rebuild itself is
not even visible in the build logs.)

So, then it's either 'config.mak', or passing a 'CC=$CC' argument to
_all_ make commands, including those that are not supposed to build
anything, but only run the tests.  I find the latter aesthetically not
particularly pleasing.



Re: On overriding make variables from the environment...

2018-10-16 Thread Jonathan Nieder
SZEDER Gábor wrote:
> On Tue, Oct 16, 2018 at 02:54:56PM -0700, Jonathan Nieder wrote:
>> SZEDER Gábor wrote:

>>> Our Makefile has lines like these:
>>>
>>>   CFLAGS = -g -O2 -Wall
>>>   CC = cc
>>>   AR = ar
>>>   SPATCH = spatch
[...]
>>> I'm not sure what to do.  I'm fine with updating our 'ci/' scripts to
>>> explicitly respect CC in the environment (either by running 'make
>>> CC=$CC' or by writing $CC into 'config.mak').  Or I could update our
>>> Makefile to use '?=' for specific variables, but:
>>
>> That's a good question.  I don't have a strong opinion myself, so I
>> tend to trust larger projects like Linux to have thought this through
>> more, and they use 'CC = cc' as well.
>
> I don't think Linux is a good example to follow in this case, see e.g.
> 6d62c983f7 (Makefile: Change the default compiler from "gcc" to "cc",
> 2011-12-20) (in short: Git is supposed to be buildable with compilers
> other than GCC as well, while Linux not really, so they can hardcode
> 'CC = gcc').

Nowadays Linux builds with clang as well.  People also have other
reasons to override the CC setting (cross-compiling, etc) and to
override other settings like CFLAGS.

> Also, the projects I have on hand usually have 'CC = gcc' hardcoded in
> their Makefiles, too, but those Makefiles were generated by their
> ./configure script (which in turn by ./autogen.sh...), and those tend
> to write CC from the environment into the generated Makefiles.

Yes, I think that's what makes travis's setup normally work.  It makes
sense to me since ./configure is expected to have more implicit or
automatic behavior.  For "make", I kind of like that it doesn't depend
on environment variables that I am not thinking about when debugging a
reported build problem.

When building Git without autoconf, the corresponding place for
settings is config.mak.  Would it make sense for the ci scripts to
write a config.mak file?

Thanks,
Jonathan


Re: On overriding make variables from the environment...

2018-10-16 Thread SZEDER Gábor
On Tue, Oct 16, 2018 at 02:54:56PM -0700, Jonathan Nieder wrote:
> SZEDER Gábor wrote:
> > Our Makefile has lines like these:
> >
> >   CFLAGS = -g -O2 -Wall
> >   CC = cc
> >   AR = ar
> >   SPATCH = spatch
> >
> > Note the use of '=', not '?='.
> [...]
> > I'm not sure what to do.  I'm fine with updating our 'ci/' scripts to
> > explicitly respect CC in the environment (either by running 'make
> > CC=$CC' or by writing $CC into 'config.mak').  Or I could update our
> > Makefile to use '?=' for specific variables, but:
> 
> That's a good question.  I don't have a strong opinion myself, so I
> tend to trust larger projects like Linux to have thought this through
> more, and they use 'CC = cc' as well.

I don't think Linux is a good example to follow in this case, see e.g.
6d62c983f7 (Makefile: Change the default compiler from "gcc" to "cc",
2011-12-20) (in short: Git is supposed to be buildable with compilers
other than GCC as well, while Linux not really, so they can hardcode
'CC = gcc').

Also, the projects I have on hand usually have 'CC = gcc' hardcoded in
their Makefiles, too, but those Makefiles were generated by their
./configure script (which in turn by ./autogen.sh...), and those tend
to write CC from the environment into the generated Makefiles.



Re: On overriding make variables from the environment...

2018-10-16 Thread Jonathan Nieder
Hi,

SZEDER Gábor wrote:

> Our Makefile has lines like these:
>
>   CFLAGS = -g -O2 -Wall
>   CC = cc
>   AR = ar
>   SPATCH = spatch
>
> Note the use of '=', not '?='.
[...]
> I'm not sure what to do.  I'm fine with updating our 'ci/' scripts to
> explicitly respect CC in the environment (either by running 'make
> CC=$CC' or by writing $CC into 'config.mak').  Or I could update our
> Makefile to use '?=' for specific variables, but:

That's a good question.  I don't have a strong opinion myself, so I
tend to trust larger projects like Linux to have thought this through
more, and they use 'CC = cc' as well.  So I'd lean toward the updating
'ci/' scripts approach, to do something like

make ${CC:+"CC=$CC"} ...

(or the equivalent multi-line construction).

That also has the bonus of being explicit.

Just my two cents,
Jonathan


On overriding make variables from the environment...

2018-10-16 Thread SZEDER Gábor
Our Makefile has lines like these:

  CFLAGS = -g -O2 -Wall
  CC = cc
  AR = ar
  SPATCH = spatch

Note the use of '=', not '?='.  This means that these variables can be
overridden from the command line, i.e. 'make CC=gcc-X.Y' will build
with that particular GCC version, but not from the environment, i.e.
'CC=gcc-X.Y make' will still build with 'cc'.

This can be confusing for developers who come from other projects
where they used to run 'CC=whatever make'.

And our build jobs on Travis CI are badly affected by this.  We have
dedicated build jobs to build Git with GCC and Clang both on Linux and
OSX from the very beginning (522354d70f (Add Travis CI support,
2015-11-27)).  But guess how Travis CI specifies which compiler to
use!  With 'export CC=gcc' and 'export CC=clang', respectively.
Consequently, our Clang Linux build job has always used gcc, because
that's where 'cc' points at on Linux by default, while the GCC OSX
build job has always used Clang.  Oh, well.  Furthermore, see
37fa4b3c78 (travis-ci: run gcc-8 on linux-gcc jobs, 2018-05-19), where
Duy added an 'export CC=gcc-8' in an attempt to use a more modern
compiler, but this had no effect either.

I'm not sure what to do.  I'm fine with updating our 'ci/' scripts to
explicitly respect CC in the environment (either by running 'make
CC=$CC' or by writing $CC into 'config.mak').  Or I could update our
Makefile to use '?=' for specific variables, but:

  - I'm afraid to break somebody's setup relying on the current
behavior and CC having different values in the environment and in
'config.mak'.

  - Where to stop, IOW which variables should be set with '?='?
CFLAGS, LDFLAGS, CC, AR, ..., SPATCH, SPATCH_FLAGS?  Dunno.  We
already have 'STRIP ?= strip' and there are variables that are
checked explicitly (e.g. 'DEVELOPER=y make' works).

Note also that prior to b05701c5b4 (Make CFLAGS overridable from
make command line., 2005-08-06) our Makefile used 'CC?=gcc' as
well.