Add polygeist NoopOp

Use a noop to preserve structure of gpu kernels through optimizations

Author: ivanradanov

include/polygeist/Passes/Passes.h

@@ -6,6 +6,12 @@
 #include <memory>
 
 enum PolygeistAlternativesMode { PAM_Static, PAM_PGO_Profile, PAM_PGO_Opt };
+enum PolygeistGPUStructureMode {
+  PGSM_Discard,
+  PGSM_BlockThreadWrappers,
+  PGSM_ThreadNoop,
+  PGSM_BlockThreadNoops
+};
 
 namespace mlir {
 class PatternRewriter;
@@ -25,9 +31,9 @@ std::unique_ptr<Pass> createCPUifyPass(StringRef method = "");
 std::unique_ptr<Pass> createBarrierRemovalContinuation();
 std::unique_ptr<Pass> detectReductionPass();
 std::unique_ptr<Pass> createRemoveTrivialUsePass();
-std::unique_ptr<Pass>
-createParallelLowerPass(bool wrapParallelOps = false,
-                        bool preserveGPUKernelStructure = false);
+std::unique_ptr<Pass> createParallelLowerPass(
+    bool wrapParallelOps = false,
+    PolygeistGPUStructureMode gpuKernelStructureMode = PGSM_Discard);
 std::unique_ptr<Pass> createCudaRTLowerPass();
 std::unique_ptr<Pass>
 createConvertPolygeistToLLVMPass(const LowerToLLVMOptions &options,
@@ -36,7 +42,8 @@ std::unique_ptr<Pass> createConvertPolygeistToLLVMPass();
 std::unique_ptr<Pass> createForBreakToWhilePass();
 std::unique_ptr<Pass>
 createConvertParallelToGPUPass1(bool useOriginalThreadNums = false);
-std::unique_ptr<Pass> createConvertParallelToGPUPass2(bool emitGPUKernelLaunchBounds = true);
+std::unique_ptr<Pass>
+createConvertParallelToGPUPass2(bool emitGPUKernelLaunchBounds = true);
 std::unique_ptr<Pass> createGpuSerializeToCubinPass(
     StringRef triple, StringRef arch, StringRef features, int llvmOptLevel,
     int ptxasOptLevel, std::string ptxasPath, std::string libDevicePath,

include/polygeist/PolygeistOps.td

@@ -18,6 +18,17 @@ include "mlir/IR/SymbolInterfaces.td"
 include "mlir/Dialect/LLVMIR/LLVMOpBase.td"
 include "mlir/Dialect/LLVMIR/LLVMOpsInterfaces.td"
 
+def NoopOp
+    : Polygeist_Op<"noop",
+                   [DeclareOpInterfaceMethods<MemoryEffectsOpInterface>]> {
+  let summary = "Noop for preventing folding or transformations";
+  let arguments = (ins Variadic<Index>:$blockDims);
+  let skipDefaultBuilders = 1;
+  let builders = [
+      OpBuilder<(ins "ValueRange":$indices)>];
+  let description = [{}];
+}
+
 def CacheLoad
     : Polygeist_Op<"cacheload"> {
 

lib/polygeist/Ops.cpp

@@ -42,6 +42,30 @@ using namespace mlir::arith;
 llvm::cl::opt<bool> BarrierOpt("barrier-opt", llvm::cl::init(true),
                                llvm::cl::desc("Optimize barriers"));
 
+//===----------------------------------------------------------------------===//
+// NoopOp
+//===----------------------------------------------------------------------===//
+
+struct NoopResource : public SideEffects::Resource::Base<NoopResource> {
+  MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID(NoopResource)
+
+  StringRef getName() final { return "<NoopResource>"; }
+};
+
+void NoopOp::build(OpBuilder &builder, OperationState &result,
+                   ValueRange indices) {
+  result.addOperands(indices);
+}
+
+void NoopOp::getEffects(
+    SmallVectorImpl<MemoryEffects::EffectInstance> &effects) {
+  // TODO CHECK is it okay to ::get() a new resource every time?
+  SideEffects::Resource *resource = NoopResource::get();
+  MemoryEffects::Effect *effect =
+      MemoryEffects::Effect::get<MemoryEffects::Write>();
+  effects.emplace_back(effect, resource);
+}
+
 //===----------------------------------------------------------------------===//
 // GPUErrorOp
 //===----------------------------------------------------------------------===//
@@ -652,6 +676,8 @@ static bool mayAlias(Value v, Value v2) {
 
 bool mayAlias(MemoryEffects::EffectInstance a,
               MemoryEffects::EffectInstance b) {
+  if (a.getResource()->getResourceID() != b.getResource()->getResourceID())
+    return false;
   if (Value v2 = b.getValue()) {
     return mayAlias(a, v2);
   } else if (Value v = a.getValue()) {