Index: llvm/lib/Target/NVPTX/NVPTX.h
===================================================================
--- llvm/lib/Target/NVPTX/NVPTX.h
+++ llvm/lib/Target/NVPTX/NVPTX.h
@@ -46,7 +46,7 @@
 ModulePass *createNVPTXAssignValidGlobalNamesPass();
 ModulePass *createGenericToNVVMPass();
 FunctionPass *createNVVMIntrRangePass(unsigned int SmVersion);
-FunctionPass *createNVVMReflectPass();
+FunctionPass *createNVVMReflectPass(unsigned int SmVersion);
 MachineFunctionPass *createNVPTXPrologEpilogPass();
 MachineFunctionPass *createNVPTXReplaceImageHandlesPass();
 FunctionPass *createNVPTXImageOptimizerPass();
Index: llvm/lib/Target/NVPTX/NVPTXTargetMachine.cpp
===================================================================
--- llvm/lib/Target/NVPTX/NVPTXTargetMachine.cpp
+++ llvm/lib/Target/NVPTX/NVPTXTargetMachine.cpp
@@ -195,7 +195,7 @@
   Builder.addExtension(
     PassManagerBuilder::EP_EarlyAsPossible,
     [&](const PassManagerBuilder &, legacy::PassManagerBase &PM) {
-      PM.add(createNVVMReflectPass());
+      PM.add(createNVVMReflectPass(Subtarget.getSmVersion()));
       PM.add(createNVVMIntrRangePass(Subtarget.getSmVersion()));
     });
 }
@@ -258,7 +258,8 @@
   // it here does nothing.  But since we need it for correctness when lowering
   // to NVPTX, run it here too, in case whoever built our pass pipeline didn't
   // call addEarlyAsPossiblePasses.
-  addPass(createNVVMReflectPass());
+  const NVPTXSubtarget &ST = *getTM<NVPTXTargetMachine>().getSubtargetImpl();
+  addPass(createNVVMReflectPass(ST.getSmVersion()));
 
   if (getOptLevel() != CodeGenOpt::None)
     addPass(createNVPTXImageOptimizerPass());
Index: llvm/lib/Target/NVPTX/NVVMReflect.cpp
===================================================================
--- llvm/lib/Target/NVPTX/NVVMReflect.cpp
+++ llvm/lib/Target/NVPTX/NVVMReflect.cpp
@@ -50,7 +50,9 @@
 class NVVMReflect : public FunctionPass {
 public:
   static char ID;
-  NVVMReflect() : FunctionPass(ID) {
+  unsigned int SmVersion;
+  NVVMReflect() : NVVMReflect(0) {}
+  explicit NVVMReflect(unsigned int Sm) : FunctionPass(ID), SmVersion(Sm) {
     initializeNVVMReflectPass(*PassRegistry::getPassRegistry());
   }
 
@@ -58,7 +60,9 @@
 };
 }
 
-FunctionPass *llvm::createNVVMReflectPass() { return new NVVMReflect(); }
+FunctionPass *llvm::createNVVMReflectPass(unsigned int SmVersion) {
+  return new NVVMReflect(SmVersion);
+}
 
 static cl::opt<bool>
 NVVMReflectEnabled("nvvm-reflect-enable", cl::init(true), cl::Hidden,
@@ -163,6 +167,8 @@
       if (auto *Flag = mdconst::extract_or_null<ConstantInt>(
               F.getParent()->getModuleFlag("nvvm-reflect-ftz")))
         ReflectVal = Flag->getSExtValue();
+    } else if (ReflectArg == "__CUDA_ARCH") {
+      ReflectVal = SmVersion * 10;
     }
     Call->replaceAllUsesWith(ConstantInt::get(Call->getType(), ReflectVal));
     ToRemove.push_back(Call);
Index: llvm/test/CodeGen/NVPTX/nvvm-reflect-arch.ll
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/NVPTX/nvvm-reflect-arch.ll
@@ -0,0 +1,21 @@
+; Libdevice in recent CUDA versions relies on __CUDA_ARCH reflecting GPU type.
+; Verify that __nvvm_reflect() is replaced with an appropriate value.
+;
+; RUN: opt %s -S -nvvm-reflect -O2 -mtriple=nvptx64 \
+; RUN:   | FileCheck %s --check-prefixes=COMMON,SM20
+; RUN: opt %s -S -nvvm-reflect -O2 -mtriple=nvptx64 -mcpu=sm_35 \
+; RUN:   | FileCheck %s --check-prefixes=COMMON,SM35
+
+@"$str" = private addrspace(1) constant [12 x i8] c"__CUDA_ARCH\00"
+
+declare i32 @__nvvm_reflect(i8*)
+
+; COMMON-LABEL: @foo
+define i32 @foo(float %a, float %b) {
+; COMMON-NOT: call i32 @__nvvm_reflect
+  %reflect = call i32 @__nvvm_reflect(i8* addrspacecast (i8 addrspace(1)* getelementptr inbounds ([12 x i8], [12 x i8] addrspace(1)* @"$str", i32 0, i32 0) to i8*))
+; SM20: ret i32 200  
+; SM35: ret i32 350  
+  ret i32 %reflect
+}
+