[gentoo-dev] .LIBPATTERNS harmful?

[gentoo-dev] .LIBPATTERNS harmful?

[Steven J Long]

Hi,

I've been working with GNU make quite a lot recently, and I came across the 
.LIBPATTERNS variable. This variable means that make expands all -lname 
prerequisites via a library path search of /lib and /usr/lib *before* any 
command sees it. (It searches local paths set in the makefile first, which 
is useful for linking to built libs, though imo the build-system is better 
off using -L parameters in LDFLAGS for those.)

You can read about it in 'info make' Section 4.5.6 (just hit / and type 
LIBPATTERN<Enter> to find it.)

The default setting is active in make as installed, as it should be, which 
you can verify with: make -p -f /dev/null|grep -F LIBPATTERN
  
I can find nothing overriding it in portage, which makes sense, since in 
general one cannot know if the package in question uses gmake .LIBPATTERNS 
to link to locally-built libs. However I can't help thinking of it as 
harmful for a package manager, since a command like ld would be given a 
parameter of say, /usr/lib/libfoo.so, not -lfoo, meaning LDFLAGS would be 
irrelevant for its lookup.

My feeling is that build-systems reliant on the default gmake behaviour for 
locally-built libs (ie not setting any -L params and also having to link 
locally) would be rare, but it's just that: a gut-feeling with no data. 
Preferably they'd be marked as such so that the package manager could deal 
with that corner-case, while patches to supply local -L params could be 
worked on, in advance of submission upstream.

I'd hope upstream would accept them, since it makes cross-development 
easier. (One definitely does not want make expanding -lname to a library in 
/lib or /usr/lib in that case, and it's better to error out if the library 
can't be found than link to host libs.)

The reason I bring it up is because we have been discussing library linkage 
issues wrt initramfs. I also seem to recall quite a few blog posts and 
discussions of arbitrary linkage to libs in /usr. .LIBPATTERNS not being 
empty would certainly explain that.

Regards,
Steve.
-- 
#friendly-coders -- We're friendly, but we're not /that/ friendly ;-)

Re: [gentoo-dev] .LIBPATTERNS harmful?

[Mike Frysinger]

[-- Attachment #1: Type: Text/Plain, Size: 1441 bytes --]

On Sunday 22 April 2012 00:44:11 Steven J Long wrote:
> I can find nothing overriding it in portage, which makes sense, since in
> general one cannot know if the package in question uses gmake .LIBPATTERNS
> to link to locally-built libs. However I can't help thinking of it as
> harmful for a package manager, since a command like ld would be given a
> parameter of say, /usr/lib/libfoo.so, not -lfoo, meaning LDFLAGS would be
> irrelevant for its lookup.

.LIBPATTERNS only matters if you specify the -lfoo in the dependency, and then 
link in via an automatic make variable.

e.g. this:
$ cat Makefile
all: test
test: -lm
$ echo 'main(){}' > test.c
$ make
cc     test.c /usr/lib/libm.so   -o test

so the easy answer is: don't add -lfoo flags as dependencies to make targets.  
if you want to have something link in a library, do:
$ cat Makefile
all: test
test: LDLIBS += -lm
$ make
cc     test.c  -lm -o test

> I'd hope upstream would accept them, since it makes cross-development
> easier. (One definitely does not want make expanding -lname to a library in
> /lib or /usr/lib in that case, and it's better to error out if the library
> can't be found than link to host libs.)

i've seen this usage in only one or two packages before.  and when i notified 
the respective upstream, they weren't really doing it on purpose, so a simple 
patch (like i showed above) they were fine with merging.
-mike

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 836 bytes --]

[gentoo-dev] Re: .LIBPATTERNS harmful?

[Steven J Long]

Mike Frysinger wrote:

> On Sunday 22 April 2012 00:44:11 Steven J Long wrote:
>> I can find nothing overriding it in portage, which makes sense, since in
>> general one cannot know if the package in question uses gmake
>> .LIBPATTERNS to link to locally-built libs. However I can't help thinking
>> of it as harmful for a package manager, since a command like ld would be
>> given a parameter of say, /usr/lib/libfoo.so, not -lfoo, meaning LDFLAGS
>> would be irrelevant for its lookup.
> 
> .LIBPATTERNS only matters if you specify the -lfoo in the dependency, and
> then link in via an automatic make variable.
>
Indeed. But that is accepted, conventional usage of make: it's why $+ 
exists, for example.
 
> e.g. this:
> $ cat Makefile
> all: test
> test: -lm
> $ echo 'main(){}' > test.c
> $ make
> cc     test.c /usr/lib/libm.so   -o test
> 
> so the easy answer is: don't add -lfoo flags as dependencies to make
> targets. if you want to have something link in a library, do:
> $ cat Makefile
> all: test
> test: LDLIBS += -lm
> $ make
> cc     test.c  -lm -o test
>
The problem with target-specific variables, is that they also apply to any 
prerequisites (and their prerequisites..) that get updated. While LDLIBS is 
only used for linking final executables, it's not a general solution, since 
it's possible for an object file to depend on a built executable (eg if the 
binary is used to build an input file, or is a test program, or happens to 
be a prerequisite of another target in the chain):

$ echo 'main(){}' > bar.c
$ echo 'const char * foo(void) { return "fubar"; }' > foo.c
$ printf '%s\n' '#include <stdio.h>' 'const char * foo(void);' 'int main(){' 
' printf ("Result: %s\n", foo()); }' > test.c

$ cat Makefile
test: foo.o -lm
foo.o: bar
$ make
cc     bar.c   -o bar
cc    -c -o foo.o foo.c
cc     test.c foo.o /usr/lib/libm.so   -o test

vs:
$ cat Makefile
test: foo.o
test: LDLIBS += -lm
foo.o: bar
$ make
cc     bar.c  -lm -o bar
cc    -c -o foo.o foo.c
cc     test.c foo.o  -lm -o test

Furthermore, if we apply the method iteratively (let's say bar uses 
pthreads) and add:
bar: LDLIBS += -lpthread
..then we get:
cc     bar.c  -lm -lpthread -o bar
..meaning that test's lib/s will be searched before those specified by bar.
For the gallery, this is all a no-no in general: just because a lib should 
be linked into one object, does not mean it should link into another, unless 
it's explicitly been specified. Libs are searched in the order they appear 
on the command-line, so test's libs would interpose symbols in bar's link.

All this makes the method package-specific, so it has to be done (and 
tested) on a case-by-case basis, and imo liable to random breakage due to 
interposition on prerequisite linkages. Worse, the value is unpredictable, 
since it varies according to which targets are being updated during the make 
run (and have thus added to LDLIBS for their prerequisites.) While this 
might not matter so much for distro-builds, it can mess up development 
builds.

In testing how to fix this, I found running:
make .LIBPATTERNS=
..gave me:
make: *** No rule to make target `-lpthread', needed by `bar'.  Stop.
So I just added a .PHONY line:
$ cat Makefile
test: foo.o -lm
foo.o: bar
bar: -lpthread
.PHONY: -lm -lpthread

..and oddly enough, I found this alone, was enough to disable the expansion 
of those libs:
$ make
cc     bar.c -lpthread   -o bar
cc    -c -o foo.o foo.c
cc     test.c foo.o -lm   -o test

I got the same result with: make .LIBPATTERNS:= so using both seems like the 
best general solution, since we're then guaranteed make will do no libname 
substitution, and we can use conventional -lfoo deps for external libs.

>> I'd hope upstream would accept them, since it makes cross-development
>> easier. (One definitely does not want make expanding -lname to a library
>> in /lib or /usr/lib in that case, and it's better to error out if the
>> library can't be found than link to host libs.)
> 
> i've seen this usage in only one or two packages before.  and when i
> notified the respective upstream, they weren't really doing it on purpose,
> so a simple patch (like i showed above) they were fine with merging.

Thanks, that's exactly the kind of knowledge I don't have, and it's good to 
know that no-one really wants -lfoo looked up and substituted by make.

Adding -lname specified prereqs to .PHONY is a simple fix, although if 
they're internal there won't be any lookup done by make. Personally I think 
that's a good thing, but I don't know how it'd affect things; for instance a 
package building okay now with a dep on an internal -lfoo which is expanded 
to a target which can be built.

But as you state, you've not seen it done on purpose (and it would mess up 
cross-compiles) so I guess the fix there is for the internal lib to be 
specified as a filename, libfoo.so or .a (and perhaps -L .)?

In any event, we seem to agree that we don't want .LIBPATTERNS expansions 
happening.

Personally, I feel setting .LIBPATTERNS empty by default, either via make 
command-line (or setenv unless the package is marked as needing the 
expansions) would be 'correct', in that it would pick up potential problems 
straightaway, but I don't have the knowledge to assess the consequences. It 
wouldn't affect packages using the LDLIBS+= target-specific setting, at 
least, but would break packages using -lfoo prerequisites via automatic 
variables, which haven't been patched to use LDLIBS+= or add .PHONY deps.

Might be something to consider for ebuild-developer mode, so new ebuilds 
don't come in with the potential for borked linkage.

A case-by-case fix would be to add .LIBPATTERNS:=  to the makefile when 
adding .PHONY deps.

Anyhow, thanks for discussion and sharing your know-how; it means I now know 
how to handle it in our builds at least (.PHONY for external libs specified 
as -lfoo, .LIBPATTERNS set empty, filenames for internal libs and -L params 
for their directories.)

Regards,
Steve.
-- 
#friendly-coders -- Where everybody knows your nickname ;-)