[Mlir-commits] [mlir] [mlir][linalg] convert arith ops to destination-passing-style. (PR #157854)

[Andrzej Warzyński]

================
@@ -603,6 +610,94 @@ Value linalg::bufferizeToAllocation(
 }
 
 namespace {
+/// Rewrites an arith op operating on tensors, e.g.
+///  `%z = arith.addf %x, %y : tensor<5xf32>`
+/// into an equivalent linalg.generic in destination-passing-style.
+/// ```mlir
+/// %0 = tensor.empty() : tensor<5xf32>
+/// %1 = linalg.generic ...
+///        ins(%x, %y : tensor<5xf32>, tensor<5xf32>)
+///        outs(%0 : tensor<5xf32>) {
+///      ^bb0(%in: f32, %in_0: f32, %out: f32):
+///         %2 = arith.addf %in, %in_0 : f32
+///         linalg.yield %2 : f32
+///     } -> tensor<5xf32>
+template <typename OpTy>
+FailureOr<Operation *>
+rewriteArithInDestinationPassingStyle(RewriterBase &rewriter, OpTy op) {
+  // Reject ops such as `arith.constant` and `arith.select`.
+  // constants don't need dps conversion and select is a a `todo`.
----------------
banach-space wrote:

```suggestion
  // constants don't need dps conversion and select is a TODO.
```

https://github.com/llvm/llvm-project/pull/157854