YJIT: Compress BranchGenFn and BranchShape

[k0kubun]

We have only a few variants of BranchGenFn. Using this property, we could use an enum to represent each function and set it instead of a function pointer, which is the core idea of this PR.

I also made another change, making BranchShape part of the BranchGenFn enum, which I believe contributes to its readability since it's made obvious that some functions don't use BranchShapes other than Default. (I hoped it would make an enum of an enum 1 byte like #7368, but this time it didn't seem to work. But I want to keep this change anyway)

This reduces yjit_alloc_size by 0.28MB (3.1%) on 25 itrs of railsbench with arm64-darwin22 stats build.

Before

yjit_alloc_size:           9,026,109

After

yjit_alloc_size:           8,747,009

[XrXr]

I love the idea here. I think it also allows us to derive Debug on Branch now because the function pointer is gone. But, I think this conflicts with the raw pointer change I'm trying a bit. In particular, the gen_fun never changes for a branch, but its shape can, so in my diff I have { gen_fun: Fun, shape: Cell<BranchShape>, .. } to show this. I mean, it's not a huge problem and I can just make a new copy with the new branch shape and just have Cell<Fun>. I think I'd prefer that the mutability of the shape is in contrast though. I think both representation should give the Branch struct the same layout in the end (it recursively pulls out and aggregate enum variant bits at the top level). Change this if you want, it's not a big deal.

EDIT: actually, having the enum reflect the fact that some gen funs should/can only have the default shape (it's implicit), the representation here is probably better than being able to see the mutability of shape for some of the variants.

[jimmyhmiller]

Just wanted to say I love this change. Makes the code much more readable. Thanks for the great change :)

[k0kubun]

I think it also allows us to derive Debug on Branch now because the function pointer is gone.

Nice. I did so in 6b9dda5.

having the enum reflect the fact that some gen funs should/can only have the default shape (it's implicit), the representation here is probably better than being able to see the mutability of shape for some of the variants.

Yeah, I want to keep that property.

I can just make a new copy with the new branch shape and just have Cell

I wonder if you can make it BranchGenFn<Cell<BranchShape>> for some variants. If Rust doesn't allow that, then maybe I would live with Cell<BranchGenFn> in your branch, preferring the benefit of having the constraint between BranchGenFn and BranchShape.

[XrXr]

By the way, the fact that we couldn't derive(Debug) when there was a function pointer seems to be some gnarly higher rank type bound issue WRT lifetime of references in the function pointer parameters: rust-lang/rust#54508 rust-lang/rust#46989

[maximecb]

Really minor comments, othewise ready to merge.