[Mlir-commits] [mlir] [mlir][vector] Document `ConvertVectorStore` + unify var names (nfc) (PR #126422)

[Andrzej Warzyński]

================
@@ -480,7 +559,8 @@ struct ConvertVectorStore final : OpConversionPattern<vector::StoreOp> {
     // vector<4xi8>
 
     auto origElements = valueToStore.getType().getNumElements();
-    bool isAlignedEmulation = origElements % numSrcElemsPerDest == 0;
+    // Note, per-element-alignment was already verified above.
----------------
banach-space wrote:

AFAIK, all patterns require per-element alignment, yes. We should extract that condition somewhere to avoid repeating it.

On a related note, given how quickly this logic is growing, I feel that we should try to avoid emulating: `i3`, `i5`, `i7`. Unless that's really required. I am just worried that the potential cost of supporting that would be relatively high.

https://github.com/llvm/llvm-project/pull/126422