[Nvidia][TritonGPU] Add phases to RemoveLayoutConversions

Author: zeroherolin

include/triton/Dialect/TritonGPU/Transforms/Passes.td

@@ -273,6 +273,26 @@ def TritonGPURemoveLayoutConversions : Pass<"tritongpu-remove-layout-conversions
   let dependentDialects = ["mlir::triton::gpu::TritonGPUDialect",
                            "mlir::triton::TritonDialect"];
 
+  // These enhancement phases are opt-in (default off) so the pass keeps its
+  // original behavior for every caller; the NVIDIA backend enables them
+  // explicitly (see third_party/nvidia/backend/compiler.py).
+  let options = [
+    Option<"enableCostBasedResolution", "enable-cost-based-resolution",
+           "bool", /*default=*/"false",
+           "Use cost-based conflict resolution instead of heuristic rules">,
+    Option<"enableBackwardPropagation", "enable-backward-propagation",
+           "bool", /*default=*/"false",
+           "Enable backward layout propagation from convert_layout ops">,
+    Option<"enableSmallComponentSolving", "enable-small-component-solving",
+           "bool", /*default=*/"false",
+           "Enable small local layout component solving">,
+    Option<"enableStoreLayoutRematerialization",
+           "enable-store-layout-rematerialization",
+           "bool", /*default=*/"false",
+           "Drop store/atomic writeback convert_layout ops by rematerializing "
+           "their address/mask chains in the value's layout">
+  ];
+
 }
 
 def TritonGPUOptimizeThreadLocality : Pass<"tritongpu-optimize-thread-locality", "mlir::ModuleOp"> {