[NVPTX][NVPTXLowerArgs] Add NewPM interface for NVPTXLowerArgs

[akshayrdeodhar]

Add a NewPM interface for NVPTXLowerArgs

[llvmbot]

@llvm/pr-subscribers-backend-nvptx

Author: Akshay Deodhar (akshayrdeodhar)

Changes

Add a NewPM interface for NVPTXLowerArgs

---

Full diff: https://github.com/llvm/llvm-project/pull/128960.diff

2 Files Affected:

• (modified) llvm/lib/Target/NVPTX/NVPTX.h (+10)
• (modified) llvm/lib/Target/NVPTX/NVPTXLowerArgs.cpp (+18-20)

diff --git a/llvm/lib/Target/NVPTX/NVPTX.h b/llvm/lib/Target/NVPTX/NVPTX.h
index ca915cd3f3732..44be26b12f0c4 100644
--- a/llvm/lib/Target/NVPTX/NVPTX.h
+++ b/llvm/lib/Target/NVPTX/NVPTX.h
@@ -18,6 +18,7 @@
 #include "llvm/Pass.h"
 #include "llvm/Support/AtomicOrdering.h"
 #include "llvm/Support/CodeGen.h"
+#include "llvm/Target/TargetMachine.h"
 
 namespace llvm {
 class FunctionPass;
@@ -74,6 +75,15 @@ struct NVPTXCopyByValArgsPass : PassInfoMixin<NVPTXCopyByValArgsPass> {
   PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
 };
 
+struct NVPTXLowerArgsPass : PassInfoMixin<NVPTXLowerArgsPass> {
+  private:
+    TargetMachine &TM;
+
+  public:
+    NVPTXLowerArgsPass(TargetMachine& TM) : TM(TM) {};
+    PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
+};
+
 namespace NVPTX {
 enum DrvInterface {
   NVCL,
diff --git a/llvm/lib/Target/NVPTX/NVPTXLowerArgs.cpp b/llvm/lib/Target/NVPTX/NVPTXLowerArgs.cpp
index c763b54c8dbfe..0a65a0df8af3c 100644
--- a/llvm/lib/Target/NVPTX/NVPTXLowerArgs.cpp
+++ b/llvm/lib/Target/NVPTX/NVPTXLowerArgs.cpp
@@ -168,17 +168,6 @@ namespace {
 class NVPTXLowerArgs : public FunctionPass {
   bool runOnFunction(Function &F) override;
 
-  bool runOnKernelFunction(const NVPTXTargetMachine &TM, Function &F);
-  bool runOnDeviceFunction(const NVPTXTargetMachine &TM, Function &F);
-
-  // handle byval parameters
-  void handleByValParam(const NVPTXTargetMachine &TM, Argument *Arg);
-  // Knowing Ptr must point to the global address space, this function
-  // addrspacecasts Ptr to global and then back to generic. This allows
-  // NVPTXInferAddressSpaces to fold the global-to-generic cast into
-  // loads/stores that appear later.
-  void markPointerAsGlobal(Value *Ptr);
-
 public:
   static char ID; // Pass identification, replacement for typeid
   NVPTXLowerArgs() : FunctionPass(ID) {}
@@ -571,8 +560,7 @@ void copyByValParam(Function &F, Argument &Arg) {
 }
 } // namespace
 
-void NVPTXLowerArgs::handleByValParam(const NVPTXTargetMachine &TM,
-                                      Argument *Arg) {
+static void handleByValParam(const NVPTXTargetMachine &TM, Argument *Arg) {
   Function *Func = Arg->getParent();
   bool HasCvtaParam =
       TM.getSubtargetImpl(*Func)->hasCvtaParam() && isKernelFunction(*Func);
@@ -641,7 +629,11 @@ void NVPTXLowerArgs::handleByValParam(const NVPTXTargetMachine &TM,
     copyByValParam(*Func, *Arg);
 }
 
-void NVPTXLowerArgs::markPointerAsGlobal(Value *Ptr) {
+// Knowing Ptr must point to the global address space, this function
+// addrspacecasts Ptr to global and then back to generic. This allows
+// NVPTXInferAddressSpaces to fold the global-to-generic cast into
+// loads/stores that appear later.
+static void markPointerAsGlobal(Value *Ptr) {
   if (Ptr->getType()->getPointerAddressSpace() != ADDRESS_SPACE_GENERIC)
     return;
 
@@ -670,13 +662,12 @@ void NVPTXLowerArgs::markPointerAsGlobal(Value *Ptr) {
 // =============================================================================
 // Main function for this pass.
 // =============================================================================
-bool NVPTXLowerArgs::runOnKernelFunction(const NVPTXTargetMachine &TM,
-                                         Function &F) {
+static bool runOnKernelFunction(const NVPTXTargetMachine &TM,Function &F) {
   // Copying of byval aggregates + SROA may result in pointers being loaded as
   // integers, followed by intotoptr. We may want to mark those as global, too,
   // but only if the loaded integer is used exclusively for conversion to a
   // pointer with inttoptr.
-  auto HandleIntToPtr = [this](Value &V) {
+  auto HandleIntToPtr = [](Value &V) {
     if (llvm::all_of(V.users(), [](User *U) { return isa<IntToPtrInst>(U); })) {
       SmallVector<User *, 16> UsersToUpdate(V.users());
       for (User *U : UsersToUpdate)
@@ -721,8 +712,7 @@ bool NVPTXLowerArgs::runOnKernelFunction(const NVPTXTargetMachine &TM,
 }
 
 // Device functions only need to copy byval args into local memory.
-bool NVPTXLowerArgs::runOnDeviceFunction(const NVPTXTargetMachine &TM,
-                                         Function &F) {
+static bool runOnDeviceFunction(const NVPTXTargetMachine &TM, Function &F) {
   LLVM_DEBUG(dbgs() << "Lowering function args of " << F.getName() << "\n");
   for (Argument &Arg : F.args())
     if (Arg.getType()->isPointerTy() && Arg.hasByValAttr())
@@ -736,7 +726,6 @@ bool NVPTXLowerArgs::runOnFunction(Function &F) {
   return isKernelFunction(F) ? runOnKernelFunction(TM, F)
                              : runOnDeviceFunction(TM, F);
 }
-
 FunctionPass *llvm::createNVPTXLowerArgsPass() { return new NVPTXLowerArgs(); }
 
 static bool copyFunctionByValArgs(Function &F) {
@@ -757,3 +746,12 @@ PreservedAnalyses NVPTXCopyByValArgsPass::run(Function &F,
   return copyFunctionByValArgs(F) ? PreservedAnalyses::none()
                                   : PreservedAnalyses::all();
 }
+
+PreservedAnalyses NVPTXLowerArgsPass::run(Function &F,
+                                          FunctionAnalysisManager &AM) {
+  auto &NTM = static_cast<NVPTXTargetMachine &>(TM);
+  bool Changed = isKernelFunction(F) ? runOnKernelFunction(NTM, F)
+                                     : runOnDeviceFunction(NTM, F);
+  return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
+
+}
\ No newline at end of file

[AlexMaclean]

Could you please add this new pass to NVPTXPassRegistry.def and write a simple lit test for it as well? Right now this looks like it is going to be dead code.

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[Artem-B]

LGTM with a few nits.

[AlexMaclean]

LGTM