[clang] [llvm] [AMDGPU] Match bitsin(typeof(x)) - ctpop(x) to s_bcnt0_i32 (PR #164847)

[Jay Foad via cfe-commits]

jayfoad wrote:

That;s just demonstrating that the InstCombine optimization "C - ctpop(X) => ctpop(~X) if C is bitwidth" has a oneuse restriction but your isel pattern does not. It's not clear to me which is better in this case. Generally oneuse restrictions lead to codegen with longer overall latency but lower register pressure, so it's a tradeoff.

https://github.com/llvm/llvm-project/pull/164847