codegen optimizations for unions

[vtjnash]

This seems to generate code that llvm is better at optimizing. Will definitely need to check nanosoldier though to see if this seems to trigger any regressions elsewhere.

[Keno]

Instcombine is quite expensive. Can you get away with instsimplify for your purpose?

[vtjnash]

Ah, I was assuming it was cheap since we call it very often. I don't actually need it.

[ararslan]

Looks like this broke the cfunction round-trip test on x86-64

[vtjnash]

@nanosoldier runbenchmarks(ALL, vs=":master")

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @jrevels

[martinholters]

Are the "simd" regressions real? I don't remember those as particularly noisy.

[ararslan]

We can always run again to see if the results are the same.

@nanosoldier runbenchmarks(ALL, vs=":master")

[vtjnash]

They seem to be "quasi-real", but the reason is hilarious. With the extra passes, LLVM is able to notice that the loop is computing the number 0 and can elid all of the SIMD computational work and replace it with a simple memset. So we end up profiling the quality of the system memset function 😆

[ararslan]

That is pretty amusing. Why would that cause a regression though? The system's memset could actually be slower than the full SIMD computations?

[vtjnash]

The system's memset could actually be slower than the full SIMD computations?

Yes. It is entirely reliant on the glibc version supporting the max vector width for the machine. The bottleneck is waiting for memory, making the SIMD calculations essentially "free".

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @jrevels