Re: bug#13928: [PATCH] compile: don't place built object files in $(srcdir), ever ...

On 01/03/2015 12:23 AM, Stefano Lattarini wrote:
> Hi Eric.
>
> On 01/03/2015 12:14 AM, Eric Blake wrote:
>> On 01/02/2015 11:49 AM, Stefano Lattarini wrote:
>>> ... even when a source file is specified as '$(srdir)/foo.c' or
>>> '$(top_srcdir)/bar.c'. And ditto for dependency-tracking makefile
>>> fragments (those under '.deps' directories).
>>
>>> +++ b/NEWS
>>
>>> @@ -88,6 +88,21 @@ New in 1.16:
>>>       using a private target that is only meant to bootstrap the required
>>>       makefile fragments.
>>>
>>> +  - The 'subdir-object' option no longer causes object files corresponding
>>> +    to source files specified with an explicit '$(srcdir)' component to be
>>> +    placed in the source tree rather than in the build tree.
>>> +
>>> +    For example, if Makefile.am contains:
>>> +
>>> +        AUTOMAKE_OPTIONS = subdir-objects
>>> +        foo_SOURCES = $(srcdir)/foo.c $(srcdir)/s/bar.c $(top_srcdir)/baz.c
>>> +
>>> +    then "make all" will create 'foo.o' and 's/bar.o' $(builddir) rather
>>
>> s|'s/bar.o'|'s/bar.o' in|
>>
> Thanks, will fix before merging this in a non-rewindable branch (that
> can't happen before Automake 1.16 is released anyway).
>
>>> +    than in $(srcdir), and 'baz.o' in $(top_builddir) rather than in
>>> +    $(top_srcdir).
>>> +
>>> +    This was the second part of automake bug#13928.
>>> +
>>>   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>
>> And thanks for tackling this!
>>
> Wait to thank me, I've found another pre-existing bug in this area, affecting
> non-GNU makes :-/  And by that I mean non-borked ones, like BSD make and
> Solaris 10 CCS make.
>
> On the plus side, the bug only affects "make distclean" (causing spurious
> failures), and only for packages using a recursive setup and referencing
> source files in a parent directory from a subdir Make; so it's a minor one.
> On the negative side, I probably introduced it myself in some 1.12.x
> release...
>
> Hopefully I'll have a fix in a week or so (I'll be AFK for most time in the
> coming days).
 >
Also, it appears this is causing another new testsuite failure with BSD make.
The following squash-in diff seems to take care of that:

-*-*-*-

diff --git a/bin/automake.in b/bin/automake.in
index f4327fc..0c44703 100644
--- a/bin/automake.in
+++ b/bin/automake.in
@@ -1677,12 +1677,15 @@ sub handle_single_transform
                {
                  $this_obj_ext = $obj;
                  $derived_source = 0;
-                # Don't ever place built object files in $(srcdir), even when
-                # sources are specified explicitly as (say) '$(srcdir)/foo.c'
-                # or '$(top_srcdir)/foo.c'.  See automake bug#13928.
-                my @d = split '/', $directory;
-                if (@d)
+                do
                    {
+                    last unless option 'subdir-objects';
+                    # Don't ever place built object files in $(srcdir),
+                    # even when sources are specified explicitly as (say)
+                    # '$(srcdir)/foo.c' or '$(top_srcdir)/foo.c'.
+                    # See automake bug#13928.
+                    my @d = split '/', $directory;
+                    last unless @d;
                      my $d = $d[0];
                      if ($d eq '$(srcdir)' or $d eq '${srcdir}')
                        {
@@ -1690,7 +1693,7 @@ sub handle_single_transform
                        }
                      elsif ($d eq '$(top_srcdir)' or $d eq '${top_srcdir}')
                        {
-                       $d[0] = '$(top_builddir)';
+                        $d[0] = '$(top_builddir)';
                        }
                      $directory = join '/', @d;
                    }

-*-*-*-

I will re-run the whole testsuite with all the affected make implementations
in the coming days, to make sure this change doesn't introduce new failures.

Sorry for the noise,
   Stefano