[LLVMGPU] Add loop invariant code motion before software pipelining (#12540)

This cleans up the IR and prevents making invariant op loop carried
dependency.

compiler/src/iree/compiler/Codegen/LLVMGPU/Passes.cpp

diff --git a/compiler/src/iree/compiler/Codegen/LLVMGPU/Passes.cpp b/compiler/src/iree/compiler/Codegen/LLVMGPU/Passes.cpp
index 44ec121..30186c6 100644
--- a/compiler/src/iree/compiler/Codegen/LLVMGPU/Passes.cpp
+++ b/compiler/src/iree/compiler/Codegen/LLVMGPU/Passes.cpp
@@ -215,6 +215,9 @@
   nestedModulePM.addNestedPass<func::FuncOp>(
       createOptimizeVectorTransferPass());
 
+  // Hoist loop invariant code to avoid pipelining it.
+  nestedModulePM.addNestedPass<func::FuncOp>(
+      createLoopInvariantCodeMotionPass());
   // Pipeline memory operations.
   nestedModulePM.addNestedPass<func::FuncOp>(createGPUPipeliningPass());
 }
@@ -270,6 +273,9 @@
   nestedModulePM.addPass(createCanonicalizerPass());
   nestedModulePM.addPass(createCSEPass());
 
+  // Hoist loop invariant code to avoid pipelining it.
+  nestedModulePM.addNestedPass<func::FuncOp>(
+      createLoopInvariantCodeMotionPass());
   PipeliningSchedulingStrategy schedule =
       llvmgpuUseMMASync ? PipeliningSchedulingStrategy::nvidiaTensorCore
                         : PipeliningSchedulingStrategy::loadGlobalStage0;