[llvm] [AMDGPU] Split the true16 fcopysign pattern by uniformity (PR #226898)

[Jay Foad via llvm-commits]

jayfoad wrote:

> A uniform 16-bit value already lives in the low half of an SGPR,

That is not a guarantee. A uniform value can also live in a VGPR. I am concerned that this patch might just be another workaround for the underlying problem.

Where do the illegal REG_SEQUENCE nodes actually get introduced? Can we get MachineVerifier to reject them?

> confused DetectDeadLanes.

Just FYI #225809 might help DetectDeadLanes to not get confused by these illegal REG_SEQUENCE nodes. But that is just another workaround, not a proper fix.

https://github.com/llvm/llvm-project/pull/226898