[llvm] [AMDGPU] - Add constant folding for s_bitreplicate (PR #72366)

[Jessica Del via llvm-commits]

OutOfCache wrote:


> Right. The behaviour is supposed to be "the same" at a high level, i.e. it performs the same calculation. But low level details (like clobbering physical registers) simply don't make any sense in LLVM IR.
> 
> > Is this possible because at the IR level, we are not aware of SCC yet and the compiler handles SCC during the lowering process?
> 
> Yes, exactly.

Thank you so much for the explanation! That makes a lot of sense.

https://github.com/llvm/llvm-project/pull/72366