[Mlir-commits] [mlir] [mlir][vector] Clarify the semantics of BroadcastOp (PR #101928)

[Hugo Trachino]

================
@@ -2391,12 +2391,28 @@ mlir::vector::isBroadcastableTo(Type srcType, VectorType dstVectorType,
   // (all leading dimensions are simply duplicated).
   int64_t lead = dstRank - srcRank;
   for (int64_t r = 0; r < srcRank; ++r) {
+    bool mismatch = false;
+
+    // Check fixed-width dims
     int64_t srcDim = srcVectorType.getDimSize(r);
     int64_t dstDim = dstVectorType.getDimSize(lead + r);
-    if (srcDim != 1 && srcDim != dstDim) {
+    if ((srcDim != 1 && srcDim != dstDim))
+      mismatch = true;
+
+    // Check scalable flags
+    bool srcDimScalableFlag = srcVectorType.getScalableDims()[r];
+    bool dstDimScalableFlag = dstVectorType.getScalableDims()[lead + r];
+    if ((srcDim == 1 && srcDimScalableFlag && dstDim != 1) ||
+        (srcDimScalableFlag && !dstDimScalableFlag))
----------------
nujaa wrote:

It got me thinking, what would be the expected behaviour of something like:
```
 %0 = vector.broadcast %arg0 : vector<nxf32> to vector<[n]xf32>
```
IMO it should not be supported as physically equivalent to a usecase
```
%1 = vector.broadcast %arg0 : vector<nxf32> to vector<vscale*nxf32>
```
Which is not invalid for fixed dimensions. Do you think this handles the cases ?
```suggestion
        (srcDimScalableFlag != dstDimScalableFlag))
```

https://github.com/llvm/llvm-project/pull/101928