[llvm] r212443 - CodeGen: it turns out that NAND is not the same thing as BIC. At all.

[Hal Finkel]

----- Original Message -----
> From: "Duncan Sands" <duncan.sands at gmail.com>
> To: "Hal Finkel" <hfinkel at anl.gov>
> Cc: llvm-commits at cs.uiuc.edu
> Sent: Monday, July 7, 2014 9:03:40 AM
> Subject: Re: [llvm] r212443 - CodeGen: it turns out that NAND is not the same thing as BIC. At all.
> 
> Hi Hal,
> 
> >>> We've been performing the wrong operation on ARM for "atomicrmw
> >>> nand" for
> >>> years, since "a NAND b" is "~(a & b)" rather than ARM's very
> >>> tempting "a & ~b".
> >>> This bled over into the generic expansion pass.
> >>>
> >>> So I assume no-one has ever actually tried to do an atomic nand
> >>> in
> >>> the real
> >>> world. Oh well.
> >>
> >> this may have been a feature, not a mistake: it might simply have
> >> been trying to
> >> be compatible with GCC which also did a & ~b for atomic nand for
> >> years (it was
> >> even documented to be this way).  The semantics were changed to
> >> ~(a &
> >> b) in GCC
> >> 4.4.  See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=37908
> >
> > Is this a problem, then, because we claim compatibility with gcc
> > 4.2.1? I wonder if this breaks anything in the wild.
> 
> in my opinion this was a gcc misfeature that is better off dead, so I
> suggest
> updating the docs to note that we deviate from gcc 4.2.1 (and follow
> gcc-4.4+)
> in defining atomic nand to be ~(a & b).

Sounds good to me.

 -Hal

> 
> Ciao, Duncan.
> 
> 

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory