Should we have a "don't simplify floating point at all" flag?

[Duncan Sands]

Hi David,

>> What would the community feel about the addition of a flag which led to
> the
>> behaviour of not applying any simplification to
>> floating point instructions (but only affecting those simplifications)? I
>> should stress that we're not committing to implementing this,
>> more getting a feel for if it's worth investigating feasibility some more.
>
> | I think it would be better to add more information to the IR so that the
> | optimizers can make correct decisions.  For example, LLVM has been lacking
> | proper support for rounding modes forever, maybe now is a good time to fix
> that?
>
> That may be the better long-term solution. One of the issues is that it's
> often
> a combination of the "set rounding mode" with fixed features of the hardware
> (eg, the hardware only offers flush-denormals-to-zero behaviour) that
> affects
> things. I haven't (at least yet) really looked in any detail at just how
> complex
> all this gets.

I have no idea what would be the best approach for representing all this in the
IR.  If it isn't feasible to have a system that covers every case, we can
always have a "funky" mode (for example) which causes the optimizers to not make
any assumptions about how floating point works, in effect disabling most
floating point optimizations.

Ciao, Duncan.