[llvm] [WebAssembly] load_zero to initialise build_vector (PR #100610)

[Sam Parker via llvm-commits]

sparker-arm wrote:

> I guess the load_zero is generally more efficient than the load_splat?

Yes, it should be more easily codegen'd to a scalar load, of some sort. I noticed on my Arm box that these were quite costly in numerical loops. These are some of the uplifts I saw on V8:

```
polybench-2mm                    8.325%
polybench-atax                   3.03%
polybench-correlation            2.88%
polybench-covariance             1.773%
polybench-doitgen                0.833%
polybench-gemver                 1.792%
polybench-ludcmp                 5.075%
polybench-mvt                    2.703%

```


https://github.com/llvm/llvm-project/pull/100610