[Opt] Improve strength reduction optimization for "pow"

[xumingkuan]

Related issue = #1141 #944

Related PR = #1065

WANT_LGTM = any

[Click here for the format server]

[k-ye]

Thanks! My only concern is the floating comparisons here. IMHO except for 0.5, we probably should do the value cast the other way around, i.e.

• int -> int
• float -> if super close to an int, return an int

Maybe the cast result needs to be std::optional<int> or something?

[xumingkuan]

I think for any int32 x, x == (float64)x. Or shall we optimize (say) a ** (2 + eps) to a * a when fast_math=True?

[k-ye]

I think for any int32 x, x == (float64)x.

I see, glad to learn about this..

Or shall we optimize (say) a ** (2 + eps) to a * a when fast_math=True?

I'll leave this up to you :)