[llvm] [SROA] adjust struct layout requirements for struct to vector canonicalization (PR #201967)

[Antonio Frighetto via llvm-commits]

================
@@ -5125,9 +5126,20 @@ static FixedVectorType *tryCanonicalizeStructToVector(StructType *STy,
       !IsIntegralPointerTy)
     return nullptr;
 
+  // Ensure the struct is tightly packed so that the bit-layout is the same as
+  // the corresponding vector. For example, this prevents a miscompile for
+  // { i5, i5 }, which has padding after each i5 field, whereas <i5, i5> has
+  // tightly packed elements and trailing padding. So if you loaded an
+  // { i5, i5 } memory layout as <i5, i5> you wouldn't correctly load each
+  // element.
----------------
antoniofrighetto wrote:

```suggestion
  // Ensure the struct is tightly packed so that the bit-layout is the same as
  // the corresponding vector.
```

https://github.com/llvm/llvm-project/pull/201967