[llvm] [BasicAA] Add Vscale GEP decomposition on variable index (PR #69152)

[Nikita Popov via llvm-commits]

nikic wrote:

> I've added a VScaleSentinel. It means that the Value cannot be used for the typesize any more though, so it currently treats vscale values as 64bit, possibly truncated to a smaller number of bits through the CastedValue. Let me know if you think a different way would be better.

Can we explicitly store the bit width of the original value, similar to how we store the sext/zext/trunc widths?

https://github.com/llvm/llvm-project/pull/69152