[llvm] Add case to TargetLowering::getValueType() to fix AMDGPU <N x ptr(7)> crash (PR #126642)

[Matt Arsenault via llvm-commits]

arsenm wrote:

Looking through the backtrace, this doesn't look like the first kludge inserted to handle irregular cases. The LoopVectorizer itself it pre-filtering tokens, invalid vector elements, and scalable vectors. Computing the number of registers used shouldn't really require figuring out the exact set of EVTs, and the users shouldn't need to worry about the edge cases either.

I really hate this chain of interfaces, but I suppose you could override getNumRegisters (which is already overridable) for this case. 

https://github.com/llvm/llvm-project/pull/126642