[TritonGPU] LICM outer loop before flattening (triton-lang#6010)

Ops in prologue/epilogue can't get hoisted by LICM after the loop is
flattened, so LICM the outer loop before. We still don't want to LICM
the inner loop because it can significantly increase liveranges.

Author: Mogball

lib/Dialect/TritonGPU/Transforms/FuseNestedLoops.cpp

@@ -2,6 +2,7 @@
 #include "mlir/Dialect/UB/IR/UBOps.h"
 #include "mlir/IR/Dominance.h"
 #include "mlir/IR/ImplicitLocOpBuilder.h"
+#include "mlir/Transforms/LoopInvariantCodeMotionUtils.h"
 #include "mlir/Transforms/RegionUtils.h"
 #include "triton/Dialect/Triton/IR/Dialect.h"
 #include "triton/Dialect/TritonGPU/Transforms/Passes.h"
@@ -1053,6 +1054,7 @@ static LogicalResult preprocessLoopNest(const LoopNest &nest,
   scf::ForOp &outerLoop = nest.root->loop;
   scf::ForOp &innerLoop = nest.root->children.front()->loop;
 
+  moveLoopInvariantCode(outerLoop);
   optimizeEpilogueDependencies(outerLoop, innerLoop, domInfo);
   return speculateInnerLoopLength(outerLoop, innerLoop, domInfo);
 }