[llvm] [LowerTypeTests] Skip declarations when determining Thumb support (PR #129599)

[Simon Tatham via llvm-commits]

statham-arm wrote:

> clang generally rewrites the triple in a way that makes querying the subarch from the triple reliable. Not sure if we expect non-clang frontends to do the same thing.

I don't know either :slightly_smiling_face: 

I can imagine in some cases even that might not be enough. For this particular purpose the subarch is all that's needed, but if we needed to query a particular SubtargetFeature to decide whether a certain instruction was safe to put in jump tables, surely the target triple wouldn't include _that_ much detail.

I don't think that's come up _yet_. In this same module there is the question of whether to put the Arm `bti` instruction in jump tables, but those are safe, because they turn into NOPs on architectures that don't support them. But I wouldn't rule out something like that in future.

I must say that if I were considering cleaning this code up, I'd start by trying to move the generation of detailed machine code jump tables into the backend, where it surely belongs, and where _all_ the necessary information is already conveniently available! Having to do it by emitting `asm` into the IR at a much earlier stage in the pipeline is surely not ideal, and if I remember rightly, the only reason it has to be done this way is because the offsets between jump table entries have to end up as constants in the IR that refers to the jump table. My instinct would be to look for a solution to _that_, before trying to find ways to make the bodgy `asm` generation itself more reliable.

But someone else's closed PR is also not the best place to discuss it :slightly_smiling_face: 

https://github.com/llvm/llvm-project/pull/129599