[Mlir-commits] [mlir] [mlir][ArmSME] Lower vector.extract/insert on SME tiles to MOVA intrinsics (PR #67786)

[Cullen Rhodes]

================
@@ -550,6 +550,113 @@ struct VectorOuterProductToArmSMELowering
   }
 };
 
+/// Lower `vector.extract` using SME MOVA intrinsics.
+///
+/// Example:
+/// ```
+/// %el = vector.extract %tile[%y,%x]: i32 from vector<[4]x[4]xi32>
+/// ```
+/// Becomes:
+/// ```
+/// %slice = arm_sme.move_tile_slice_to_vector %tile[%y]
+///            : vector<[4]xi32> from vector<[4]x[4]xi32>
+/// %el = vector.extract %slice[%x] : i32 from vector<[4]xi32>
+/// ```
+struct VectorExtractToArmSMELowering
+    : public ConvertOpToLLVMPattern<vector::ExtractOp> {
+  using ConvertOpToLLVMPattern<vector::ExtractOp>::ConvertOpToLLVMPattern;
+
+  LogicalResult
+  matchAndRewrite(vector::ExtractOp extractOp, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    VectorType sourceType = extractOp.getSourceVectorType();
+    if (!isValidSMETileVectorType(sourceType))
+      return failure();
+
+    auto loc = extractOp.getLoc();
+    auto position = extractOp.getMixedPosition();
+
+    Value sourceVector = extractOp.getVector();
+
+    if (position.empty()) {
+      rewriter.replaceOp(extractOp, sourceVector);
+      return success();
+    }
----------------
c-rhodes wrote:

I would think canonicalization already does this, I guess there's no harm doing it here as well given it's minimal code. I see `VectorExtractOpConversion` also does this, could you copy the comment from there?

```// Extract entire vector. Should be handled by folder, but just to be safe.```

https://github.com/llvm/llvm-project/pull/67786