Re: [RFC] PowerPC select_section / unique_section

[Alan Modra <amodra at bigpond dot net dot au>]

On Fri, Aug 30, 2002 at 11:27:25AM -0400, David Edelsohn wrote:
> >>>>> Alan Modra writes:
> 
> Alan> Uh, oh.  If I understand binds_local_p and mark_constant_function
> Alan> correctly, setting flag_pic is actually a bug-fix when compiling
> Alan> powerpc64-linux shared libs.  We have the standard ELF binding of
> Alan> global syms.  That is, global functions may be overridden by functions
> Alan> in another shared library or by the main application.
> 
> Alan> So powerpc64-linux-gcc should allow -fpic/PIC to twiddle flag_pic for
> Alan> binds_local_p, and users should set -fPIC when compiling shared libs
> Alan> as is common on other ELF targets.  We could use another flag, because
> Alan> like that annoying rs6000.c warning says "all code is position
> Alan> independent" on ppc64, but that would make powerpc64-linux just that
> Alan> more odd.  Lots of packages set -fPIC to mean "compile me code for a
> Alan> shared library".
> 
> 	This is what the patch that I applied to both gcc-3.2 and the
> trunk already does, without utilizing the generic infrastructure.  The
> PowerPC port currently does not use the targetm.binds_local_p.  We can
> discuss evolving to the generic infrastructure for GCC 3.4.

The PowerPC back-end code doesn't use binds_local_p, but
mark_constant_function does.

I think I'm correct in claiming that powerpc64-linux-gcc lacks a way
to say "I want this code compiled for a shared library;  Don't try to
analyze global functions for pure/const as the function in this file
may not be the one called at runtime."

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre