Re: [arm] Fix invalid store instructions

[Richard Earnshaw <rearnsha at gcc dot gnu dot org>]

On Mon, 2004-06-21 at 13:38, Paul Brook wrote:
> > My feeling is that this is an MI bug not a target specific one, so I
> > think it should really be fixed in the mid-end.
> >
> > There is no way of representing an early-clobber type behaviour in the
> > constraints for a memory operand with pre/post operations so the
> > compiler really must assume that it does happen and prevent any further
> > use of the base register in the insn.
> 
> FWIW We also encounter similar issues with base+index load-double 
> instructions, even when base value writeback is not specified. These are much 
> easier to reproduce, eg.
> long long foo(char *p, int n){return *(long long *)(p+n);}
> Without hacks in the doubleword assembly output routines this would compile to
> 
> foo:
>   ldrd r0, [r0, r1] @illegal because destination {r0-r1} overlaps index {r1}
>   bx lr
> 

It should be possible to fix this with a simple '&' on the loaded value
(ie just treat it as an early-clobber).

> > It's unclear to me at this time whether this has always been a
> > theoretical problem with the compiler, or if it's something new that's
> > been introduced by the tree-ssa changes.  I'm not aware that we suffered
> > from this problem before those changes were merged.
> 
> The problem is still present, just harder to reproduce.
> The example I gave generates invalid code with compilers from csl-arm-branch 
> (3.4 based) and 3.3.2.

OK.  Then I guess it's always been a problem.  I still feel that it
really needs to be fixed in an MI way, or there needs to be a MI way of
representing the restriction (possibly a target hook).  We should really
be preventing these invalid instructions from being created by the
compiler, rather than allowing them and trying to fix up the mess
afterwards.

R.