[llvm-dev] Vectorization with fast-math on irregular ISA sub-sets

[Stephen Canon via llvm-dev]

> On Feb 8, 2016, at 2:25 PM, James Molloy via llvm-dev <llvm-dev at lists.llvm.org> wrote:
> 
> Sorry, on phone so cherry picking what I reply to :
> 
>> On 8 Feb 2016, at 19:15, Renato Golin <renato.golin at linaro.org> wrote:
>> 
>> For 16275, the fix is to disable loop vect. for no-fast-math + hasUnsafeAlgebra.
> 
> Do you think there is a set of people that care about IEEE accuracy in so far that they don't want FTZ, but *are* happy to reassociate FP operations? That seems fairly niche to me?

I agree.  FZ is usually relatively benign (it only causes major problems when programs expect x != y to imply that x - y != 0, an axiom of floating-point that’s broken in FZ).  Re-association more frequently causes significant instability.

I think it’s reasonable for unsafeAlgebra to imply "FZ is an allowed mode”.

– Steve