[NVPTX] add an optional early copy of byval arguments (llvm#113384)

byval arguments in NVPTX are special. We're only allowed to read from
them using a special instruction, and if we ever need to write to them
or take an address, we must make a local copy and use it, instead.

The problem is that local copies are very expensive, and we create them
very late in the compilation pipeline, so LLVM does not have much of a
chance to eliminate them, if they turn out to be unnecessary.

One way around that is to create such copies early on, and let them
percolate through the optimizations. The copying itself will never
trigger creation of another copy later on, as the reads are allowed. If
LLVM can eliminate it, it's a win. It the full optimization pipeline
can't remove the copy, that's as good as it gets in terms of the effort
we could've done, and it's certainly a much better effort than what we
do now.

This early injection of the copies has potential to create undesireable
side-effects, so it's disabled by default, for now, until it sees more
testing.

llvm/lib/Target/NVPTX/NVPTX.h

@@ -70,6 +70,10 @@ struct GenericToNVVMPass : PassInfoMixin<GenericToNVVMPass> {
   PreservedAnalyses run(Module &M, ModuleAnalysisManager &AM);
 };
 
+struct NVPTXCopyByValArgsPass : PassInfoMixin<NVPTXCopyByValArgsPass> {
+  PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
+};
+
 namespace NVPTX {
 enum DrvInterface {
   NVCL,

llvm/lib/Target/NVPTX/NVPTXLowerArgs.cpp

@@ -543,6 +543,33 @@ struct ArgUseChecker : PtrUseVisitor<ArgUseChecker> {
       PI.setAborted(&II);
   }
 }; // struct ArgUseChecker
+
+void copyByValParam(Function &F, Argument &Arg) {
+  LLVM_DEBUG(dbgs() << "Creating a local copy of " << Arg << "\n");
+  // Otherwise we have to create a temporary copy.
+  BasicBlock::iterator FirstInst = F.getEntryBlock().begin();
+  Type *StructType = Arg.getParamByValType();
+  const DataLayout &DL = F.getDataLayout();
+  AllocaInst *AllocA = new AllocaInst(StructType, DL.getAllocaAddrSpace(),
+                                      Arg.getName(), FirstInst);
+  // Set the alignment to alignment of the byval parameter. This is because,
+  // later load/stores assume that alignment, and we are going to replace
+  // the use of the byval parameter with this alloca instruction.
+  AllocA->setAlignment(F.getParamAlign(Arg.getArgNo())
+                           .value_or(DL.getPrefTypeAlign(StructType)));
+  Arg.replaceAllUsesWith(AllocA);
+
+  Value *ArgInParam = new AddrSpaceCastInst(
+      &Arg, PointerType::get(Arg.getContext(), ADDRESS_SPACE_PARAM),
+      Arg.getName(), FirstInst);
+  // Be sure to propagate alignment to this load; LLVM doesn't know that NVPTX
+  // addrspacecast preserves alignment.  Since params are constant, this load
+  // is definitely not volatile.
+  const auto ArgSize = *AllocA->getAllocationSize(DL);
+  IRBuilder<> IRB(&*FirstInst);
+  IRB.CreateMemCpy(AllocA, AllocA->getAlign(), ArgInParam, AllocA->getAlign(),
+                   ArgSize);
+}
 } // namespace
 
 void NVPTXLowerArgs::handleByValParam(const NVPTXTargetMachine &TM,
@@ -558,7 +585,7 @@ void NVPTXLowerArgs::handleByValParam(const NVPTXTargetMachine &TM,
 
   ArgUseChecker AUC(DL, IsGridConstant);
   ArgUseChecker::PtrInfo PI = AUC.visitArgPtr(*Arg);
-  bool ArgUseIsReadOnly  = !(PI.isEscaped() || PI.isAborted());
+  bool ArgUseIsReadOnly = !(PI.isEscaped() || PI.isAborted());
   // Easy case, accessing parameter directly is fine.
   if (ArgUseIsReadOnly && AUC.Conditionals.empty()) {
     // Convert all loads and intermediate operations to use parameter AS and
@@ -587,7 +614,6 @@ void NVPTXLowerArgs::handleByValParam(const NVPTXTargetMachine &TM,
   // However, we're still not allowed to write to it. If the user specified
   // `__grid_constant__` for the argument, we'll consider escaped pointer as
   // read-only.
-  unsigned AS = DL.getAllocaAddrSpace();
   if (HasCvtaParam && (ArgUseIsReadOnly || IsGridConstant)) {
     LLVM_DEBUG(dbgs() << "Using non-copy pointer to " << *Arg << "\n");
     // Replace all argument pointer uses (which might include a device function
@@ -612,29 +638,8 @@ void NVPTXLowerArgs::handleByValParam(const NVPTXTargetMachine &TM,
 
     // Do not replace Arg in the cast to param space
     CastToParam->setOperand(0, Arg);
-  } else {
-    LLVM_DEBUG(dbgs() << "Creating a local copy of " << *Arg << "\n");
-    // Otherwise we have to create a temporary copy.
-    AllocaInst *AllocA =
-        new AllocaInst(StructType, AS, Arg->getName(), FirstInst);
-    // Set the alignment to alignment of the byval parameter. This is because,
-    // later load/stores assume that alignment, and we are going to replace
-    // the use of the byval parameter with this alloca instruction.
-    AllocA->setAlignment(Func->getParamAlign(Arg->getArgNo())
-                             .value_or(DL.getPrefTypeAlign(StructType)));
-    Arg->replaceAllUsesWith(AllocA);
-
-    Value *ArgInParam = new AddrSpaceCastInst(
-        Arg, PointerType::get(Arg->getContext(), ADDRESS_SPACE_PARAM),
-        Arg->getName(), FirstInst);
-    // Be sure to propagate alignment to this load; LLVM doesn't know that NVPTX
-    // addrspacecast preserves alignment.  Since params are constant, this load
-    // is definitely not volatile.
-    const auto ArgSize = *AllocA->getAllocationSize(DL);
-    IRBuilder<> IRB(&*FirstInst);
-    IRB.CreateMemCpy(AllocA, AllocA->getAlign(), ArgInParam, AllocA->getAlign(),
-                     ArgSize);
-  }
+  } else
+    copyByValParam(*Func, *Arg);
 }
 
 void NVPTXLowerArgs::markPointerAsGlobal(Value *Ptr) {
@@ -734,3 +739,22 @@ bool NVPTXLowerArgs::runOnFunction(Function &F) {
 }
 
 FunctionPass *llvm::createNVPTXLowerArgsPass() { return new NVPTXLowerArgs(); }
+
+static bool copyFunctionByValArgs(Function &F) {
+  LLVM_DEBUG(dbgs() << "Creating a copy of byval args of " << F.getName()
+                    << "\n");
+  bool Changed = false;
+  for (Argument &Arg : F.args())
+    if (Arg.getType()->isPointerTy() && Arg.hasByValAttr() &&
+        !(isParamGridConstant(Arg) && isKernelFunction(F))) {
+      copyByValParam(F, Arg);
+      Changed = true;
+    }
+  return Changed;
+}
+
+PreservedAnalyses NVPTXCopyByValArgsPass::run(Function &F,
+                                              FunctionAnalysisManager &AM) {
+  return copyFunctionByValArgs(F) ? PreservedAnalyses::none()
+                                  : PreservedAnalyses::all();
+}

llvm/lib/Target/NVPTX/NVPTXPassRegistry.def

@@ -37,4 +37,5 @@ FUNCTION_ALIAS_ANALYSIS("nvptx-aa", NVPTXAA())
 #endif
 FUNCTION_PASS("nvvm-intr-range", NVVMIntrRangePass())
 FUNCTION_PASS("nvvm-reflect", NVVMReflectPass())
+FUNCTION_PASS("nvptx-copy-byval-args", NVPTXCopyByValArgsPass())
 #undef FUNCTION_PASS

llvm/lib/Target/NVPTX/NVPTXTargetMachine.cpp

@@ -64,6 +64,29 @@ static cl::opt<bool> UseShortPointersOpt(
         "Use 32-bit pointers for accessing const/local/shared address spaces."),
     cl::init(false), cl::Hidden);
 
+// byval arguments in NVPTX are special. We're only allowed to read from them
+// using a special instruction, and if we ever need to write to them or take an
+// address, we must make a local copy and use it, instead.
+//
+// The problem is that local copies are very expensive, and we create them very
+// late in the compilation pipeline, so LLVM does not have much of a chance to
+// eliminate them, if they turn out to be unnecessary.
+//
+// One way around that is to create such copies early on, and let them percolate
+// through the optimizations. The copying itself will never trigger creation of
+// another copy later on, as the reads are allowed. If LLVM can eliminate it,
+// it's a win. It the full optimization pipeline can't remove the copy, that's
+// as good as it gets in terms of the effort we could've done, and it's
+// certainly a much better effort than what we do now.
+//
+// This early injection of the copies has potential to create undesireable
+// side-effects, so it's disabled by default, for now, until it sees more
+// testing.
+static cl::opt<bool> EarlyByValArgsCopy(
+    "nvptx-early-byval-copy",
+    cl::desc("Create a copy of byval function arguments early."),
+    cl::init(false), cl::Hidden);
+
 namespace llvm {
 
 void initializeGenericToNVVMLegacyPassPass(PassRegistry &);
@@ -236,6 +259,8 @@ void NVPTXTargetMachine::registerPassBuilderCallbacks(PassBuilder &PB) {
         // Note: NVVMIntrRangePass was causing numerical discrepancies at one
         // point, if issues crop up, consider disabling.
         FPM.addPass(NVVMIntrRangePass());
+        if (EarlyByValArgsCopy)
+          FPM.addPass(NVPTXCopyByValArgsPass());
         PM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
       });
 }