DemoteFloat16 pass ignores vectors of Float16

[vchuravy]

As noticed by @giordano in #40308 (comment), when LLVM decided to vectorize, the demote Float16 pass doesn't trigger. Presumably this is due to

julia/src/llvm-demote-float16.cpp

Line 82 in fec6951

if (Op->getType() == T_float16) {

only checking if it is a T_float16 and not a vector of Float16.

While it is nice that it currently "accidentally" works, since we want on A64FX to trigger the emission of native Float16 operations,
we should fix this bug and simultaneously address #40216

[vtjnash]

This was fixed in #45249, until we decided to say that performance was more important than correctness and reverted it

[vchuravy]

Ah great that we have the code already :) Would have been great if the fixed for the conversion routine and the other correctness fixes were not co-mingled.

[vtjnash]

They aren't co-mingled really, you just have to pick a name for the function (or intrinsic) to call here https://github.com/JuliaLang/julia/pull/45249/files#diff-d6a6d037fdf95b0875b7d800187abc42b7ef9abf7004e36ad3ef392655144631R140

[gbaraldi]

It also ignores @fastmath, not sure if #45249 fixes that however.

[vtjnash]

The point of fastmath is to return less correct answers, so this pass isn't needed

[gbaraldi]

I thought generating float16 instructions when not available was bad, but it seems llvm just adds the conversions if necessary when generating native code

[giordano]

but it seems llvm just adds the conversions if necessary when generating native code

Then why we have the demotion pass in Julia again? 😬

[gbaraldi]

I don't know :)