[Mlir-commits] [mlir] [MLIR][XeGPU] Introduce `xegpu::uArch` usage in target-sensitive passes (PR #163801)

[Artem Kroviakov]

================
@@ -43,7 +43,12 @@ def XeGPUPropagateLayout : Pass<"xegpu-propagate-layout"> {
   let options = [Option<
     "printOnly", "print-analysis-only", "bool",
     /*default=*/"false",
-    "Print the result of layout propagation analysis and exit.">];
+    "Print the result of layout propagation analysis and exit.">,
+    Option<
+    "assumeUnrolled", "assume-unrolled", "bool",
----------------
akroviakov wrote:

Not an enum, but a string.
For subgroup, we must have a user layout on anchor ops to propagate? It's not like lane/inst fields, which are tightly coupled to hw subgroup size and/or instruction size.
Anyway, this is a topic for a different PR. For now, we can do `lane` and `inst`.

https://github.com/llvm/llvm-project/pull/163801