[llvm] [AMDGPU] Don't apply gfx950 fetch-window loop align to the wrong block (PR #221821)

[Akash Dutta via llvm-commits]

akadutta wrote:

> Not for this PR, but should the GFX10/GFX11 cache-line path also use the block being aligned instead of ML->getHeader() for the already-processed check and the LoopSize estimate?

Seems so to me. After rotation those two checks still look at getHeader() while placement aligns ChainBB, so they can skip the wrong block and mis-count padding. I'll wait for reviews on this before attempting to address the GFX10 ones.

https://github.com/llvm/llvm-project/pull/221821