codegen: use direct value PHI for pointer-free aggregate types

[KristofferC]

Fixes #59906 from my testing, takes the runtime of the code in there from 4000 ns to 1000 ns on a M4 macbook.

Developed together with Claude 🤖

My understanding of the code here is rudimentary but I still put up this PR in case it is helpful.

[vtjnash]

It is correctly implemented, but has quite a few known catastrophic performance failures, so we should not do this.

[KristofferC]

It is correctly implemented, but has quite a few known catastrophic performance failures, so we should not do this.

Do you have an example (would nanosoldier show it)? Any other way to get back the performance lost in 25cbe00?

[vtjnash]

Yeah, looks like an LLVM bug (this particularly one has been a super common one, and was supposed to be fixed by opaque pointers). But we can add the hack back that usually does okay to often work around that

[vtjnash]

We just need to audit all calls to emit_static_alloca and make sure they use the old (pre-opaque pointer) GEP type instead of actually benefitting from LLVM's enormous amount of opaque pointer work. I suspect the SROA pass is still at fault here for the performance issues.

[oscardssmith]

wait, we want to use the old version?

[gbaraldi]

llvm/llvm-project#164308 I did some snooping around and opened that. I couldn't minimize it much further but the difference is LLVM thinks that an alloca of floats is meaningfully different than an alloca of ints. If anyone wants to take a stab the issue is probably in https://github.com/llvm/llvm-project/blob/e6b4a21849f0588b1c4fb39802a3999d7ac51dad/llvm/lib/Transforms/Scalar/SROA.cpp#L4885-L4966