[llvm] [AMDGPU] Add VOP1 DPP8 pseudo infrastructure (PR #196736)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> We want to avoid using real instructions in pattern matching because that could potentially blow the matcher table's 16-bit opcode.

How does that help? The opcode number goes *up*. Does it just happen that the real instructions get assigned higher numbers? 

https://github.com/llvm/llvm-project/pull/196736