[llvm] Calculate KnownBits from Metadata correctly for vector loads (PR #128908)

[via llvm-commits]

================
@@ -4004,22 +4004,25 @@ KnownBits SelectionDAG::computeKnownBits(SDValue Op, const APInt &DemandedElts,
         }
       }
     } else if (Op.getResNo() == 0) {
-      KnownBits Known0(!LD->getMemoryVT().isScalableVT()
-                           ? LD->getMemoryVT().getFixedSizeInBits()
-                           : BitWidth);
+      unsigned MemorySize = !LD->getMemoryVT().isScalableVT()
+                                ? LD->getMemoryVT().getFixedSizeInBits()
----------------
LU-JOHN wrote:

> > MemorySize is only used for the initial size for Known0 and to extend/truncate KnownMetadata to the width of Known0.
> 
> In no context does it make sense to use the full bitwidth of the memory size here. The casting may happen to make it work out, but I don't care to reason about whether it does or not. It only makes sense to operate on the scalar type

Using MemorySize for the initial size of Known0 is important for scalar 8-bit loads that zero-extend to 32-bit.  In this case MemorySize is 8 and BitWidth is 32.  This is used in the later code that zero-extends the KnownBits to 32-bits.  I think this part of the code is hard to understand, but this PR is not trying to address the rationale for the size of Known0.  It was only intended to ensure that a KnownBits of the correct size was passed to computeKnownBitsFromRangeMetadata.

https://github.com/llvm/llvm-project/pull/128908