Re: [avr-gcc-list] [WinAVR 20090313] Bug in split-Pattern for swap

[Georg-Johann Lay]

Weddington, Eric schrieb:
>  
> It probably comes from the attached patch. This was a fix for GCC bug #11259:
> <http://gcc.gnu.org/bugzilla/show_bug.cgi?id=11259>
> This was committed in 4.4.
>
> Eric Weddington

Yes, that patch introduces the bug because operand0 resp. operand1 may 
be hard reg before reload, which is the case here (#0=r24, #1=r25).

The patterns are using constraint "0" which is pretty much useless for 
the split part of the define_insn_and_split before reload, and therefore 
no one actually ensures that input and output operands are the same. It 
will work for pseudos, but not if a hard register is involved.

The simplest fix is to split only if reload_completed or just omit the 
patch because in newer versions the UNSPEC_SWAP will have gone anyway.

Georg-Johann