[llvm] [WebAssembly] Add support for avgr_u in loops (PR #153252)

[Jasmine Tang via llvm-commits]

================

----------------
badumbatish wrote:

Even at precommit, the compiler can turn the following this to avgr_u already by the pattern in the tablegen
```llvm
  %a = add nuw <16 x i8> %x, %y
  %b = add nuw <16 x i8> %a, <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1,
                              i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
  ; replacing udiv with lshr in @avgr_u_v16i8 of llvm/test/CodeGen/WebAssembly/simd-arith.ll
  %c = lshr <16 x i8> %b, <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1,
                           i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
  ret <16 x i8> %c
```

I think the point of the having the loop vectorizer test case is that the loop vectorizer will vectorize the loop into AVGCEIL_U, where we can match to it to produce avgr_u.

I'm not sure how to replicate this in raw llvm ir without going through the loop vectorizer?

https://github.com/llvm/llvm-project/pull/153252