[llvm] [AArch64] Fix type mismatch in bitconvert + vec_extract patterns (PR #183549)

[David Green via llvm-commits]

================
@@ -8212,10 +8212,8 @@ defm : Neon_INS_elt_ext_pattern<v8i16, v4i16, v2i32, INSvi16lane, DUPi16, hsub,
 
 // bitcast of an extract
 // f32 bitcast(vector_extract(v4i32 src, 0)) -> EXTRACT_SUBREG(src)
-def : Pat<(f32 (bitconvert (i32 (vector_extract v16i8:$src, (i64 0))))),
----------------
davemgreen wrote:

You have removed this one because it doesn't come up anywhere? I guess my preference would be to keep it, to keep symmetry with the other i16 types. I can see the argument for not having untested patterns, but it leaves a bit of a wonky backend that has so many missing holes that we end up missing optimizations again and again. I didn't see it come up anywhere at the though, so I guess it isn't doing much good at the moment.

https://github.com/llvm/llvm-project/pull/183549