[WIP] Remove redundant casts in LLVMIR

mlir/include/mlir/Dialect/Rock/Passes.h

@@ -41,6 +41,7 @@ namespace rock {
 #define GEN_PASS_DECL_ROCKVIEWTOTRANSFORMPASS
 #define GEN_PASS_DECL_ROCKDETECTFLASHDECODINGPASS
 #define GEN_PASS_DECL_ROCKLOWERREDUCEPASS
+#define GEN_PASS_DECL_ROCKREMOVEREDUNDANTCASTSPASS
 #define GEN_PASS_DECL_ROCKPREPARELLVMPASS
 #define GEN_PASS_DECL_ROCKCHECKRESIDENCYPASS
 #define GEN_PASS_DECL_ROCKVECTORIZEFUSIONSPASS

mlir/include/mlir/Dialect/Rock/Passes.td

@@ -176,6 +176,20 @@ def RockLowerReducePass : Pass<"rock-lower-reduce", "::mlir::func::FuncOp"> {
   let dependentDialects = ["rock::RockDialect", "func::FuncDialect", "gpu::GPUDialect"];
 }
 
+def RockRemoveRedundantCastsPass
+    : Pass<"rock-remove-redundant-casts", "::mlir::LLVM::LLVMFuncOp"> {
+  let summary = "Remove redundant fptrunc/fpext pairs through buffers at LLVM "
+                "dialect level";
+  let description = [{
+    Detects patterns at the LLVM dialect level where wider float values are
+    truncated (llvm.fptrunc) to a narrower type, stored to a buffer, then loaded
+    and extended (llvm.fpext) back to the original wider type. This pass
+    redirects the loads to read the wide values directly, eliminating the
+    fpext and preserving precision.
+  }];
+  let dependentDialects = ["LLVM::LLVMDialect"];
+}
+
 def RockPrepareLLVMPass : Pass<"rock-prepare-llvm", "::mlir::LLVM::LLVMFuncOp"> {
   let summary = "prepare the generated code for llvm";
   let dependentDialects = ["ROCDL::ROCDLDialect"];

mlir/lib/Dialect/Rock/Pipelines/Pipelines.cpp

@@ -305,6 +305,7 @@ void rock::buildBackendPipeline(OpPassManager &pm,
   // descriptors. (Mainly we want the `extractvalue` fold).
   llvmFuncPm.addPass(createCanonicalizerPass());
   llvmFuncPm.addPass(createCSEPass());
+  llvmFuncPm.addPass(rock::createRockRemoveRedundantCastsPass());
   llvmFuncPm.addPass(rock::createRockPrepareLLVMPass());
   if (options.compile) {
     GpuROCDLAttachTargetOptions opts;

mlir/lib/Dialect/Rock/Transforms/CMakeLists.txt

@@ -37,6 +37,7 @@ add_rocmlir_dialect_library(MLIRRockTransforms
   BlockwiseLoadTileToThreadwise.cpp
   AnnotateLiveness.cpp
   AddAsyncWait.cpp
+  RemoveRedundantCasts.cpp
   LowerRockOpsToROCDLOps.cpp
 
   ADDITIONAL_HEADER_DIRS