Re: Problems with the 2003-01-23 patch (was Re: [PATCH] Fix s390 as)

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

On Wed, Feb 12, 2003 at 12:23:10PM +0100, Jakub Jelinek wrote:
> On Thu, Jan 23, 2003 at 11:21:59PM +1030, Alan Modra wrote:
> > This should fix the bug shown by Jakub's s390 testcase in a way that
> > preserves old behaviour regarding subtraction expressions.
> 
> Unfortunately, it seems to break things on various platforms
> (those which define their own TC_FORCE_RELOCATION).

I'm understanding better why sane people are afraid to touch the gas
reloc code.

> Say on Alpha
>         .text
> 1:      lda $15,-128($15)
>         .text   1
> 2:      ldgp $29,2b-1b($27)
> causes
> Fatal error: internal error? cannot generate `BFD_RELOC_ALPHA_GPDISP' relocation
> (this is distilled from much bigger assembly which actually did not change
> subsections in between the two labels, just a fragment break happened to
> be in between the two instructions (only on Alpha, not on i686->Alpha cross)
> and thus expr did not optimize that into constant 4.

Yeah, that's nasty.  Turning on gas listings exposes these problems too.

> Dunno if it is better to handle subtraction of two local labels in the same
> section in the generic TC_FORCE_RELOCATION_SUB_SAME implementation (and in
> the unlikely case some machine wants to override that behaviour, it could
> define its own TC_FORCE_RELOCATION_SUB_SAME), or call some new function
> in each arch_force_relocation before checking relocation type.

I'm not going to try to come up with a patch, as it's getting late here,
but my thoughts are to go back to using

#ifndef TC_FORCE_RELOCATION_SUB_SAME
#define TC_FORCE_RELOCATION_SUB_SAME(FIX, SEG)	\
  (! SEG_NORMAL (SEG))
#endif

as a default in write.c, and something special in tc-s390.h.

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre