[Mlir-commits] [mlir] [mlir][vector] Improve shape_cast lowering (PR #140800)

[James Newling]

================
@@ -157,41 +160,54 @@ class ShapeCastOpRewritePattern : public OpRewritePattern<vector::ShapeCastOp> {
   LogicalResult matchAndRewrite(vector::ShapeCastOp op,
                                 PatternRewriter &rewriter) const override {
     Location loc = op.getLoc();
-    auto sourceVectorType = op.getSourceVectorType();
-    auto resultVectorType = op.getResultVectorType();
+    VectorType sourceType = op.getSourceVectorType();
+    VectorType resultType = op.getResultVectorType();
 
-    if (sourceVectorType.isScalable() || resultVectorType.isScalable())
+    if (sourceType.isScalable() || resultType.isScalable())
       return failure();
 
-    // Special case for n-D / 1-D lowerings with better implementations.
-    int64_t srcRank = sourceVectorType.getRank();
-    int64_t resRank = resultVectorType.getRank();
-    if ((srcRank > 1 && resRank == 1) || (srcRank == 1 && resRank > 1))
+    // Special case for n-D / 1-D lowerings with implementations that use
+    // extract_strided_slice / insert_strided_slice.
----------------
newling wrote:

I didn't much like the logic spread over the 3 patterns (N->N, 1->N, N->1) as there isn't really anything special about then 1->N and N->1 cases. So I've done a fairly major update to the N->N pattern, so that now it handles then 1->N and N->1 cases. As the new change is quite significant, if you'd prefer it to be done in a separate PR I'm happy to postpone this 'unification', backtrack, and just make the minor suggestions to this PR that you suggested. 

I also unified the tests across the test file. The behavior for the 1->N and N->1 cases is unchanged by this PR though. 

https://github.com/llvm/llvm-project/pull/140800