[Mlir-commits] [mlir] [mlir][Vector] Improve vector.transferx store-to-load-forwarding (PR #171840)

[Jakub Kuderski]

================
@@ -5118,6 +5118,23 @@ Speculation::Speculatability TransferReadOp::getSpeculatability() {
   return Speculation::NotSpeculatable;
 }
 
+/// Given a projected permutation, inverse an affine map, making the unused dims
+/// 0 in the result.
+static AffineMap inverseWithUnusedDims(AffineMap map) {
+  assert(map.isProjectedPermutation() &&
+         "expected a projected permutation map");
+  SmallVector<AffineExpr> results(map.getNumInputs(),
+                                  getAffineConstantExpr(0, map.getContext()));
+  for (auto [idx, result] : llvm::enumerate(map.getResults())) {
+    assert(isa<AffineDimExpr>(result) &&
+           "expected only dim exprs in a projected permutation map");
----------------
kuhar wrote:

nit: I'd turn this into a comment -- cast already asserts internally

https://github.com/llvm/llvm-project/pull/171840