Fix tbaa annotation on union selector bytes inside of structs

[gbaraldi]

We currently cause a alias analysis contradiction by saying that the unionselbytes are on the stack, even if they are on a struct. LLVM is then able to figure out that we giving it a impossible alias situation (the object doesn't alias itself) and triggers UB.

https://godbolt.org/z/ssEKMzsPf

We may want to do a benchmarks run on this to see if anything too critical hasn't regressed.

Fixes #54599

[fingolfin]

"it's" -> "its" ?

[KristofferC]

Is it possible to minimise this a bit or is that tricky? It's a lot of code under the hood here and with changes to that code it's possible a regression would not be found?

[topolarity]

Uh oh, I found this example that fails LLVM assertions even with this PR:

bar54599() = Base.inferencebarrier(true) ? (Base.PkgId(Test),1) : nothing
function foo54599()
    pkginfo = @noinline bar54599()
    pkgid = pkginfo !== nothing ? pkginfo[1] : nothing
    pkgid.uuid !== nothing ? pkgid.uuid : false
end

@test foo54599() !== nothing

julia: /home/topolarity/repos/julia/src/llvm-alloc-opt.cpp:408: void {anonymous}::Optimizer::insertLifetime(llvm::Value*, llvm::Constant*, llvm::Instruction*): Assertion `cur.p_cur != cur.p_end' failed.

[gbaraldi]

@topolarity that is a separate crash 😄, that is fixed in #54690.

[oscardssmith]

in that case, should we merge this?

[gbaraldi]

I wasn't able to minimize that test to I gues so

[topolarity]

The crash I pointed out was in search of a more minimal test case.

Here's one that works:

bar54599() = Base.inferencebarrier(true) ? (Base.PkgId(Main),1) : nothing
function foo54599()
    pkginfo = @noinline bar54599()
    pkgid = pkginfo !== nothing ? pkginfo[1] : nothing
    @noinline println(devnull, pkgid)
    pkgid.uuid !== nothing ? pkgid.uuid : false
end
foo54599() # segfaults