[llvm] [AMDGCN][SIWholeQuadMode] Handle case when SI_KILL_I1_TERMINATOR -1,0 is not the only terminator (PR #122922)

[Juan Manuel Martinez Caamaño via llvm-commits]

jmmartinez wrote:

> > or even at IR level.
> 
> We could do this if we started using / requiring using callbr for this intrinsic (which is something we should do). But I would expect this to have folded out before this point

That sound good. It would help us get rid of both pseudo instructions `SI_KILL_I1_PSEUDO` and `SI_KILL_F32_COND_IMM_PSEUDO` and retain only the terminator variants. I'll create a new ticket for this since it sounds like a bigger change.

The calls to `@llvm.amdgcn.kill(true)` are already optimized by InstCombine. We could move this somewhere where it's also executed with `-O0`, maybe inside SIAnnotateControlFlow ?

https://github.com/llvm/llvm-project/pull/122922