Symbol versions in libc.a

["Joseph S. Myers" <joseph at codesourcery dot com>]

I'm looking at removing --disable-versioning, as proposed in 
<http://sourceware.org/ml/libc-alpha/2013-07/msg00174.html>.

Mostly this is straightforward.  The presumption that versioning is always 
enabled for shared-library builds means that $(versioning) makefile 
conditionals become $(build-shared) (the conditionals may not always be 
needed at all, but $(build-shared) is a safe conversion).  The 
corresponding preprocessor macro is DO_VERSIONING.

When DO_VERSIONING is used together with SHARED, it's a simple matter of 
removing the DO_VERSIONING conditional.  There are however places where 
it's used on its own.  A couple of those in testcases (that are only built 
when shared libraries are built) are unproblematic.  The problem comes in 
shlib-compat.h and libc-symbols.h.

In shlib-compat.h, we have

#if defined DO_VERSIONING
/* Since there is just one set of .d files generated, we need to
   include this unconditionally to have the dependency noticed properly.  */
#include <abi-versions.h>	/* header generated by abi-versions.awk */
#endif

which appears to assert a need for the #include even when building objects 
for static libraries, if (versioned) shared libraries are built at all.  I 
don't know what "just one set of .d files generated" is referring to - I 
certainly see separate .o.d and .os.d files - so I suspect this is in fact 
an obsolete comment and a conditional on SHARED would be fine.

In libc-symbols.h, DO_VERSIONING on its own controls the definitions of 
macros such as symbol_version and default_symbol_version.  The effect is 
that if those macros are used outside a conditional on SHARED, and the 
object is being built for libc.a, what happens depends on whether shared 
libraries are being built at all.  In particular, various objects in 
libc.a on powerpc (at least) contain symbol version information, and I 
suspect that they should not.  I'm inclined to say this is a bug in the 
relevant source files - they shouldn't be using symbol_version and 
default_symbol_version outside of conditionals on SHARED - but also that 
this information in libc.a objects is not useful and so it is safe to 
change the DO_VERSIONING conditional in libc-symbols.h to one on SHARED 
(so meaning libc.a no longer has this versioning information).

Comments?

-- 
Joseph S. Myers
joseph@codesourcery.com