functions with `align $N` can get merged without `align $N`

[workingjubilee]

Godbolt of opt -passes="module(mergefunc)": https://godbolt.org/z/cPTc9j1b6

LLVMIR diff:

-define internal i64 @aligned_to_2p8() unnamed_addr align 256 {
-entry:
-  ret i64 4
-}
 
 define i64 @aligned_to_default() unnamed_addr {
 entry:
   ret i64 4
 }

"What's the problem? You said to merge the functions, and it did!"

Well, in Rust, generally function pointers aren't expected to point to unique blocks of code. For this reason, we specify unnamed_addr on... essentially every function, ish? However, for some cases we would like to be able to force the alignment of function entry symbols while still running function merging where applicable. In a more complex case where the code would produce a result that depends on the address of the code in question, the correct semantic for Rust would be that the value does depend on this alignment but may not be unique. Thus, we would prefer that merging functions is allowed to work in this case, but the resulting merged function takes the highest alignment of all merges.

[workingjubilee]

Note that I think unnamed_addr might be a red herring and this is almost entirely about visibility, I have only looked at this from the surface tinkering-around-with-LLVMIR level and haven't dug deep into LLVM's logic.

Oh no, this is because I over-reduced it. unnamed_addr does matter in some cases: https://godbolt.org/z/Wa14bd9Gc