[Mlir-commits] [mlir] [mlir][amdgpu] Introduce `assume_subgroup_uniform` op (PR #152740)

[Jakub Kuderski]

================
@@ -635,6 +636,37 @@ def AMDGPU_DPPOp : AMDGPU_Op<"dpp",
   let hasVerifier = 1;
 }
 
+def AMDGPU_AssumeSubgroupUniformOp : AMDGPU_Op<"assume_subgroup_uniform",
+    [NoMemoryEffect, AllTypesMatch<["result", "src"]>,
+    DeclareOpInterfaceMethods<InferIntRangeInterface, ["inferResultRanges"]>,
+    DeclareOpInterfaceMethods<ConditionallySpeculatable, ["getSpeculatability"]>] #
+    ElementwiseMappable.traits>,
+  Arguments<(ins AnyType:$src,
+                 DefaultValuedAttr<UnitAttr, "false">:$all_lanes)> {
+  let summary = "Assumes value is unform across the lanes in subgroup";
+  let description = [{
+      This op is a compiler hint to help backend put values into scalar registers.
+
+      If `src` value is uniform across all the active subgroup lanes it is
+      returned unchanged, otherwise result is poison.
+
+      If `all_lanes` is set, the value is assumed to be uniform across all the
+      subgroup lanes, this can allow to speculate it out of control flow, which
+      may change the current active lanes, i.e:
+      ```
----------------
kuhar wrote:

```suggestion
      ```mlir
```

https://github.com/llvm/llvm-project/pull/152740