[Mlir-commits] [mlir] [Mlir] decompose generic by unfolding projected permutation crash fix (PR #122449)

[Andrzej Warzyński]

================
@@ -69,3 +69,35 @@ func.func @broadcast_only(%x : tensor<2x16x32xf32>, %y:  tensor<2x32xf32>, %z :
 // CHECK: %[[X_bc:.+]] = linalg.broadcast ins(%[[Y]] : tensor<2x32xf32>) outs(%[[E0]] : tensor<2x16x32xf32>) dimensions = [1]
 // CHECK: {{.*}} = linalg.div ins(%[[X]], %[[X_bc]] : tensor<2x16x32xf32>, tensor<2x16x32xf32>) outs(%arg2 : tensor<2x16x32xf32>) -> tensor<2x16x32xf32>
 // CHECK-NOT: linalg.generic
+
+// -----
+
+// unsupported currently.
+
+func.func @no_decompose_on_scalar() -> tensor<2x2xi32> {
----------------
banach-space wrote:

> "Negative scalar" makes me think -1 rather than not matching case.

How do I un-see and recover from this? :) 

Note, it's "Negative - scalar input", rather than "Negative scalar". Also, "Negative scalar" could be "-10" (i.e. ambiguous), where's "negative" as in "negative case" is not ambiguous. At least that's how I look at it 😅 

That said, it's not perfect. Things would be easier if we just agreed on a convention and followed it (so that "negative" tests are easy to identify).

https://github.com/llvm/llvm-project/pull/122449