ping @r-devulap

Wouldn't preventing the store from writing to these elements be a simpler and efficient fix? That is, change
_mm512_mask_storeu_pd(op, load_mask, res); to _mm512_mask_storeu_pd(op, or(glibc_mask, load_mask) , res); I think that might fix it or am I missing something?

@r-devulap,

I think that might fix it or am I missing something?

yes, sure it will fix it but I just tried to take advantage and slightly improve performance.

I removed unlikely hint to avoid flush the pipe and I added a constant hint
which leads to increase the performance in between 10 to 20% for mixed and out-ranged values.

actually using unlikely leads to a roughly 10% increase in performance for in-ranged values but it has a bad impact on
out-ranged values. I will update #17587 after this pr getting merged.

Thanks Sayed. I expect we will be tweaking the SIMD code going forward once it is all working.

actually using unlikely leads to a roughly 10% increase in performance for in-ranged values but it has a bad impact on
out-ranged values.

Very surprising to see any measurable difference with just the use of __builtin_expect .