[libffi] Use datarel encoding in libffi .eh_frame on Solaris 2/x86

While trying to make the libgcc unwinder use dl_iterate_phdr on Solaris
11, I ran into a gld bug (PR ld/12253) that broke libffi unwind tests
and all libjava execution tests: while Solaris 2/x86 uses
DW_EH_PE_datarel encoding for .eh_frame in 32-bit PIC code,
src/x86/sysv.S uses DW_EH_PE_pcrel.  Due to the above bug, .eh_frame_hdr
isn't sorted correctly with gld before 2.22.  To work around this issue
for older gld versions, I'd like to use the following workaround to
avoid the issue, namely also use datarel encoding on Solaris 2/x86.

Bootstrapped on i386-pc-solaris2.11 with CVS gas and gld (before the
fix) and i386-pc-solaris2.10 with Sun as and ld.  It fixes all the
testsuite regressions with gld and works without issues with Sun ld.

Ok for mainline?

	Rainer


2010-01-18  Rainer Orth  <ro@CeBiTec.Uni-Bielefeld.DE>

	* src/x86/sysv.S (.eh_frame) [__PIC__ && __sun__ && __svr4__]: Use
	datarel encoding.
	(.LASFDE1) [__PIC__ && __sun__ && __svr4__]: Use @GOTOFF for FDE
	initial location.
	(.LASFDE2, LASFDE3): Likewise.

On 11/26/2010 10:45 AM, Rainer Orth wrote:
> 	* src/x86/sysv.S (.eh_frame) [__PIC__ && __sun__ && __svr4__]: Use
> 	datarel encoding.
> 	(.LASFDE1) [__PIC__ && __sun__ && __svr4__]: Use @GOTOFF for FDE
> 	initial location.
> 	(.LASFDE2, LASFDE3): Likewise.

This is getting complex enough that I think we should not 
replicate this logic 4 times.  Would you please pull this
out into a set of macros near the top of the file?  E.g.

#if defined __PIC__
# if defined __sun__ && defined __svr4__
#  define FDE_ENCODING		0x30	/* datarel */
#  define FDE_ENCODE(X)		X@GOTOFF
# else
#  define FDE_ENCODING		0x1b	/* pcrel sdata4 */
#  if defined HAVE_AS_X86_PCREL
#   define FDE_ENCODE(X)	X-.
#  else
#   define FDE_ENCODE(X)	X@rel
#  endif
# endif
#else
# define FDE_ENCODING		0	/* absolute */
# define FDE_ENCODE(X)		X
#endif

And, frankly, for gcc 4.7 we ought to consider requiring a
recent version of binutils and scrap all of this explicit
unwind info in favor of CFI directives.



r~

Richard Henderson <rth@redhat.com> writes:

> On 11/26/2010 10:45 AM, Rainer Orth wrote:
>> 	* src/x86/sysv.S (.eh_frame) [__PIC__ && __sun__ && __svr4__]: Use
>> 	datarel encoding.
>> 	(.LASFDE1) [__PIC__ && __sun__ && __svr4__]: Use @GOTOFF for FDE
>> 	initial location.
>> 	(.LASFDE2, LASFDE3): Likewise.
>
> This is getting complex enough that I think we should not 
> replicate this logic 4 times.  Would you please pull this
> out into a set of macros near the top of the file?  E.g.
>
> #if defined __PIC__
> # if defined __sun__ && defined __svr4__
> #  define FDE_ENCODING		0x30	/* datarel */
> #  define FDE_ENCODE(X)		X@GOTOFF
> # else
> #  define FDE_ENCODING		0x1b	/* pcrel sdata4 */
> #  if defined HAVE_AS_X86_PCREL
> #   define FDE_ENCODE(X)	X-.
> #  else
> #   define FDE_ENCODE(X)	X@rel
> #  endif
> # endif
> #else
> # define FDE_ENCODING		0	/* absolute */
> # define FDE_ENCODE(X)		X
> #endif

Sure, I'll give this a try.

> And, frankly, for gcc 4.7 we ought to consider requiring a
> recent version of binutils and scrap all of this explicit
> unwind info in favor of CFI directives.

We have to be very careful here: libffi is used by several other
projects and is supposed to be buildable with vendor compilers and
assemblers also.

	Rainer