[llvm] r344172 - [LV] Add a new reduction pattern match

[Renato Golin via llvm-commits]

On Tue, 23 Oct 2018 at 15:02, Benjamin Kramer <benny.kra at gmail.com> wrote:
>> +  Value *Op1, *Op2;
>> +  if (m_FAdd(m_Value(Op1), m_Value(Op2)).match(I1) ||
>> +      m_FSub(m_Value(Op1), m_Value(Op2)).match(I1))
>> +    return InstDesc(Kind == RK_FloatAdd, SI);
>
> Are these patterns safe without checking for associative math?

Hi Benjamin,

I don't think so, but the instructions are all "fast" and the
vector-width is forced. I may be wrong, but I thought that reduction
patterns already checked for that.

>  Seeing some numerical weirdness after this change, but haven't looked closely.

I'll have a look, too.

--renato