[Mlir-commits] [mlir] [mlir][scf] Uplift `scf.while` to `scf.for` (PR #76108)

[Ivan Butygin]

================
@@ -154,4 +154,20 @@ def SCFForToWhileLoop : Pass<"scf-for-to-while"> {
   }];
 }
 
+def SCFUpliftWhileToFor : Pass<"scf-uplift-while-to-for"> {
+  let summary = "Uplift scf.while ops to scf.for";
+  let description = [{
+    This pass tries to uplift `scf.while` ops to `scf.for` if they have a
+    compatible form. `scf.while` are left unchanged if uplifting is not
+    possible.
+  }];
+
+  let options = [
+    Option<"indexBitWidth", "index-bitwidth", "unsigned",
+           /*default=*/"64",
+           "Bitwidth of index type.">,
+  ];
----------------
Hardcode84 wrote:

Non-index type support in `scf.for` was a late addition, and not all parts in pipeline support it (e.g. `scf.parallel` only support index), and we never really needed non-index `scf.for`s internally. Also, using index type can be actually beneficially as it reduces amount of  `index_cast`s needed when working with tensors/memrefs.

Anyway, I'll change pattern to support any integer type and will probably do maximizing index type as separate transform.

https://github.com/llvm/llvm-project/pull/76108