[Mlir-commits] [mlir] [MLIR][OpenMP] Add `omp.private` op (PR #80955)

[Kiran Chandramohan]

================
@@ -133,6 +134,84 @@ def DeclareTargetAttr : OpenMP_Attr<"DeclareTarget", "declaretarget"> {
   let assemblyFormat = "`<` struct(params) `>`";
 }
 
+//===----------------------------------------------------------------------===//
+// 2.19.4 Data-Sharing Attribute Clauses
+//===----------------------------------------------------------------------===//
+
+def PrivateClauseOp : OpenMP_Op<"private", [
+    IsolatedFromAbove, FunctionOpInterface
+  ]> {
+  let summary = "Outline [first]private logic in a separate op.";
+  let description = [{
+    Using this operation, the dialect can model the data-sharing attributes of
+    `private` and `firstprivate` variables on the IR level. This means that of
+    "eagerly" privatizing variables in the frontend, we can instead model which
+    variables should be privatized and only materialze the privatization when
+    necessary; e.g. directly before lowring to LLVM IR.
+
+    Examples:
+    ---------
+    * `private(x)` would be emitted as:
+    ```mlir
+    omp.private @x.privatizer : (!fir.ref<i32>) -> !fir.ref<i32> {
+    ^bb0(%arg0: !fir.ref<i32>):
+      %0 = fir.alloca i32
+      omp.yield(%0 : !fir.ref<i32>)
+    }
+    ```
+
+    * `firstprivate(x)` would be emitted as:
+    ```mlir
+    omp.private @x.privatizer : (!fir.ref<i32>) -> !fir.ref<i32> {
+    ^bb0(%arg0: !fir.ref<i32>):
+      %0 = fir.alloca i32
+      %1 = fir.load %arg0 : !fir.ref<i32>
+      fir.store %1 to %0 : !fir.ref<i32>
+      omp.yield(%0 : !fir.ref<i32>)
+    }
+    ```
----------------
kiranchandramohan wrote:

I would prefer if we split this into two regions for firstprivate.
1. That does the allocation
2. That does the firstprivate copy
In future we can have another one that does finalization/deallocation.

I think this will allow us to use similar representation for lastprivate. It will also permit us to reuse code from copyprivate for firstprivate.
Also, I believe this is what OpenACC does.

Alternatively, we can remove the firstprivate part from this patch and push on with  the private only.

https://github.com/llvm/llvm-project/pull/80955