[Mlir-commits] [mlir] [mlir][transform] LISH: Add transform op (PR #70630)

[Matthias Springer]

================
@@ -691,6 +691,65 @@ def GetTypeOp : TransformDialectOp<"get_type",
                        "functional-type(operands, results)";
 }
 
+def HoistLoopInvariantSubsetsOp
+    : TransformDialectOp<"hoist_loop_invariant_subsets",
+        [TransformOpInterface, TransformEachOpTrait,
+         DeclareOpInterfaceMethods<MemoryEffectsOpInterface>,
+         ReportTrackingListenerFailuresOpTrait]> {
+  let summary = "Hoist loop invariant subset ops";
+  let description = [{
+    This transform hoist loop-invariant subset ops out of loop-like ops. It
+    looks for matching subset extraction/insertion op pairs and hoists them. The
+    loop body operates on a newly introduced region iter_arg.
+
+    Example:
+    ```
+    %r = scf.for ... iter_args(%t = %a) -> (tensor<?xf32>) {
+      %0 = tensor.extract_slice %t[0][5][1] : tensor<?xf32> to tensor<5xf32>
+      %1 = "test.foo"(%0) : (tensor<5xf32>) -> (tensor<5xf32>)
+      %2 = tensor.insert_slice %1 into %t[0][5][1]
+          : tensor<5xf32> into tensor<?xf32>
+      scf.yield %2 : tensor<?xf32>
+    }
+    ```
+    Is transformed to:
+    ```
+    %0 = tensor.extract_slice %a[0][5][1] : tensor<?xf32> to tensor<5xf32>
+    %new_loop:2 = scf.for ... iter_args(%t = %a, %h = %0) -> (tensor<?xf32>) {
+      %1 = "test.foo"(%h) : (tensor<5xf32>) -> (tensor<5xf32>)
+      scf.yield %t, %2 : tensor<?xf32>, tensor<5xf32>
+    }
+    %r = tensor.insert_slice %new_loop#1 into %new_loop#0
+        : tensor<5xf32> into tensor<?xf32>
+    ```
+
+    Subset ops are hoisted only if there are no conflicting subset ops. E.g.,
+    if there were a second overlapping extraction in the above example, no ops
+    could be hoisted safely.
+
+    This transform looks for `LoopLikeOpInterface` ops within the targeted op,
+    including the target op itself. It attempts hoisting on all found loop-like
+    ops.
+
+    This transform reads the target handle and modifies the payload.
+
+    TODO: Make this op more targeted if needed. I.e., apply the transformation
+    only to the targeted `LoopLikeOpInterface` op.
----------------
matthias-springer wrote:

This transform op replaces `transform.structured.hoist_redundant_vector_transfer` and I tried to keep the same API for easier migration (in downstream projects). I'm going to change it to the more targeted variant. The only thing that needs to be added is an additional `structured.match` in test cases. And then it will also be consistent with `transform.apply_licm`.

https://github.com/llvm/llvm-project/pull/70630