llvm dev - Aug 2016 - How to make the optimizer eliminate `fadd zeroinitializer`?

I have the module:

target triple = "x86_64-pc-linux-gnu"

define void @round(<8 x float>*) {
_L1:
   %1 = load <8 x float>, <8 x float>* %0
   %2 = fadd <8 x float> %1, zeroinitializer
   store <8 x float> %2, <8 x float>* %0
   ret void
}


Optimizing it with -O3 does not remove the add with zero. Also options 
like -enable-no-infs-fp-math -enable-no-nans-fp-math 
-enable-unsafe-fp-math -enable-fp-mad do not eliminate it. I understand 
that floating point arithmetic exposes strange behavior that disables many 
obvious simplifications. I cannot see the corner case here. If there exist 
one, how can I convince 'opt' to remove the add with zero?

Adding ‘fast’ to the fadd should simplify it. Note that adding zero is not a
no-op in floating point arithmetic; adding negative zero is.

—escha 
> On Aug 12, 2016, at 1:53 PM, Henning Thielemann via llvm-dev <llvm-dev
at lists.llvm.org> wrote:
> 
> 
> I have the module:
> 
> target triple = "x86_64-pc-linux-gnu"
> 
> define void @round(<8 x float>*) {
> _L1:
>  %1 = load <8 x float>, <8 x float>* %0
>  %2 = fadd <8 x float> %1, zeroinitializer
>  store <8 x float> %2, <8 x float>* %0
>  ret void
> }
> 
> 
> Optimizing it with -O3 does not remove the add with zero. Also options like
-enable-no-infs-fp-math -enable-no-nans-fp-math -enable-unsafe-fp-math
-enable-fp-mad do not eliminate it. I understand that floating point arithmetic
exposes strange behavior that disables many obvious simplifications. I cannot
see the corner case here. If there exist one, how can I convince 'opt'
to remove the add with zero?
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

On Fri, 12 Aug 2016, escha at apple.com wrote:
> Adding ‘fast’ to the fadd should simplify it. Note that adding zero is 
> not a no-op in floating point arithmetic; adding negative zero is.
Ah, that is adding zero to negative zero is the corner case!

It seems that
   opt -enable-no-infs-fp-math  adds  ninf  globally
   opt -enable-no-nans-fp-math  adds  nnan  globally

but I do not see a counterpart for the flags 'nsz', 'arcp' and
'fast'. :-(