diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.h b/llvm/lib/Target/AMDGPU/AMDGPU.h
--- a/llvm/lib/Target/AMDGPU/AMDGPU.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPU.h
@@ -115,6 +115,11 @@
 void initializeAMDGPULowerKernelAttributesPass(PassRegistry &);
 extern char &AMDGPULowerKernelAttributesID;
 
+struct AMDGPULowerKernelAttributesPass
+    : PassInfoMixin<AMDGPULowerKernelAttributesPass> {
+  PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
+};
+
 void initializeAMDGPUPropagateAttributesEarlyPass(PassRegistry &);
 extern char &AMDGPUPropagateAttributesEarlyID;
 
diff --git a/llvm/lib/Target/AMDGPU/AMDGPULowerKernelAttributes.cpp b/llvm/lib/Target/AMDGPU/AMDGPULowerKernelAttributes.cpp
--- a/llvm/lib/Target/AMDGPU/AMDGPULowerKernelAttributes.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULowerKernelAttributes.cpp
@@ -19,7 +19,9 @@
 #include "llvm/CodeGen/TargetPassConfig.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/Function.h"
+#include "llvm/IR/InstIterator.h"
 #include "llvm/IR/Instructions.h"
+#include "llvm/IR/PassManager.h"
 #include "llvm/IR/PatternMatch.h"
 #include "llvm/Pass.h"
 
@@ -41,16 +43,11 @@
 };
 
 class AMDGPULowerKernelAttributes : public ModulePass {
-  Module *Mod = nullptr;
-
 public:
   static char ID;
 
   AMDGPULowerKernelAttributes() : ModulePass(ID) {}
 
-  bool processUse(CallInst *CI);
-
-  bool doInitialization(Module &M) override;
   bool runOnModule(Module &M) override;
 
   StringRef getPassName() const override {
@@ -64,12 +61,7 @@
 
 } // end anonymous namespace
 
-bool AMDGPULowerKernelAttributes::doInitialization(Module &M) {
-  Mod = &M;
-  return false;
-}
-
-bool AMDGPULowerKernelAttributes::processUse(CallInst *CI) {
+static bool processUse(CallInst *CI) {
   Function *F = CI->getParent()->getParent();
 
   auto MD = F->getMetadata("reqd_work_group_size");
@@ -89,7 +81,7 @@
   Value *GridSizeY = nullptr;
   Value *GridSizeZ = nullptr;
 
-  const DataLayout &DL = Mod->getDataLayout();
+  const DataLayout &DL = F->getParent()->getDataLayout();
 
   // We expect to see several GEP users, casted to the appropriate type and
   // loaded.
@@ -239,7 +231,7 @@
   StringRef DispatchPtrName
     = Intrinsic::getName(Intrinsic::amdgcn_dispatch_ptr);
 
-  Function *DispatchPtr = Mod->getFunction(DispatchPtrName);
+  Function *DispatchPtr = M.getFunction(DispatchPtrName);
   if (!DispatchPtr) // Dispatch ptr not used.
     return false;
 
@@ -267,3 +259,22 @@
 ModulePass *llvm::createAMDGPULowerKernelAttributesPass() {
   return new AMDGPULowerKernelAttributes();
 }
+
+PreservedAnalyses
+AMDGPULowerKernelAttributesPass::run(Function &F, FunctionAnalysisManager &AM) {
+  StringRef DispatchPtrName =
+      Intrinsic::getName(Intrinsic::amdgcn_dispatch_ptr);
+
+  Function *DispatchPtr = F.getParent()->getFunction(DispatchPtrName);
+  if (!DispatchPtr) // Dispatch ptr not used.
+    return PreservedAnalyses::all();
+
+  for (Instruction &I : instructions(F)) {
+    if (CallInst *CI = dyn_cast<CallInst>(&I)) {
+      if (CI->getCalledFunction() == DispatchPtr)
+        processUse(CI);
+    }
+  }
+
+  return PreservedAnalyses::all();
+}
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
@@ -508,6 +508,10 @@
           PM.addPass(AMDGPUPromoteAllocaToVectorPass(*this));
           return true;
         }
+        if (PassName == "amdgpu-lower-kernel-attributes") {
+          PM.addPass(AMDGPULowerKernelAttributesPass());
+          return true;
+        }
         return false;
       });
 
@@ -530,6 +534,10 @@
           // but before SROA to increase SROA opportunities.
           FPM.addPass(InferAddressSpacesPass());
 
+          // This should run after inlining to have any chance of doing
+          // anything, and before other cleanup optimizations.
+          FPM.addPass(AMDGPULowerKernelAttributesPass());
+
           if (Level != PassBuilder::OptimizationLevel::O0) {
             // Promote alloca to vector before SROA and loop unroll. If we
             // manage to eliminate allocas before unroll we may choose to unroll
diff --git a/llvm/test/CodeGen/AMDGPU/reqd-work-group-size.ll b/llvm/test/CodeGen/AMDGPU/reqd-work-group-size.ll
--- a/llvm/test/CodeGen/AMDGPU/reqd-work-group-size.ll
+++ b/llvm/test/CodeGen/AMDGPU/reqd-work-group-size.ll
@@ -1,4 +1,5 @@
 ; RUN: opt -mtriple=amdgcn-amd-amdhsa -S -amdgpu-lower-kernel-attributes -instcombine %s | FileCheck -enable-var-scope %s
+; RUN: opt -mtriple=amdgcn-amd-amdhsa -S -passes=amdgpu-lower-kernel-attributes,instcombine %s | FileCheck -enable-var-scope %s
 
 target datalayout = "n32"
 
diff --git a/llvm/tools/opt/opt.cpp b/llvm/tools/opt/opt.cpp
--- a/llvm/tools/opt/opt.cpp
+++ b/llvm/tools/opt/opt.cpp
@@ -464,7 +464,7 @@
 static bool shouldPinPassToLegacyPM(StringRef Pass) {
   std::vector<StringRef> PassNameExactToIgnore = {
       "amdgpu-simplifylib", "amdgpu-usenative", "amdgpu-promote-alloca",
-      "amdgpu-promote-alloca-to-vector"};
+      "amdgpu-promote-alloca-to-vector", "amdgpu-lower-kernel-attributes"};
   for (const auto &P : PassNameExactToIgnore)
     if (Pass == P)
       return false;