diff --git a/llvm/docs/AMDGPUUsage.rst b/llvm/docs/AMDGPUUsage.rst
index 5c6034753eb4a..40b393224f15d 100644
--- a/llvm/docs/AMDGPUUsage.rst
+++ b/llvm/docs/AMDGPUUsage.rst
@@ -5914,10 +5914,7 @@ additional 256 bytes to the kernel_code_entry_byte_offset. This addition
 facilitates the incorporation of a prologue to the kernel entry to handle cases
 where code designed for kernarg preloading is executed on hardware equipped with
 incompatible firmware. If hardware has compatible firmware the 256 bytes at the
-start of the kernel entry will be skipped. Additionally, the compiler backend
-may insert a trap instruction at the start of the kernel prologue to manage
-situations where kernarg preloading is attempted on hardware with incompatible
-firmware.
+start of the kernel entry will be skipped.
 
 With code object V5 and later, hidden kernel arguments that are normally
 accessed through the Implicit Argument Ptr, may be preloaded into User SGPRs.
diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.h b/llvm/lib/Target/AMDGPU/AMDGPU.h
index b9769a1baf4d1..ad5ee75f0c5d1 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPU.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPU.h
@@ -64,6 +64,7 @@ createAMDGPULowerModuleLDSLegacyPass(const AMDGPUTargetMachine *TM = nullptr);
 ModulePass *createAMDGPULowerBufferFatPointersPass();
 FunctionPass *createSIModeRegisterPass();
 FunctionPass *createGCNPreRAOptimizationsPass();
+FunctionPass *createAMDGPUPreloadKernArgPrologLegacyPass();
 
 struct AMDGPUSimplifyLibCallsPass : PassInfoMixin<AMDGPUSimplifyLibCallsPass> {
   AMDGPUSimplifyLibCallsPass() {}
@@ -230,6 +231,9 @@ extern char &AMDGPUPerfHintAnalysisLegacyID;
 void initializeGCNRegPressurePrinterPass(PassRegistry &);
 extern char &GCNRegPressurePrinterID;
 
+void initializeAMDGPUPreloadKernArgPrologLegacyPass(PassRegistry &);
+extern char &AMDGPUPreloadKernArgPrologLegacyID;
+
 // Passes common to R600 and SI
 FunctionPass *createAMDGPUPromoteAlloca();
 void initializeAMDGPUPromoteAllocaPass(PassRegistry&);
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUArgumentUsageInfo.h b/llvm/lib/Target/AMDGPU/AMDGPUArgumentUsageInfo.h
index 06b2f181c276c..e07d47381ecca 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUArgumentUsageInfo.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUArgumentUsageInfo.h
@@ -9,6 +9,7 @@
 #ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPUARGUMENTUSAGEINFO_H
 #define LLVM_LIB_TARGET_AMDGPU_AMDGPUARGUMENTUSAGEINFO_H
 
+#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/CodeGen/Register.h"
 #include "llvm/Pass.h"
@@ -161,6 +162,8 @@ struct AMDGPUFunctionArgInfo {
 
   // Map the index of preloaded kernel arguments to its descriptor.
   SmallDenseMap<int, KernArgPreloadDescriptor> PreloadKernArgs{};
+  // The first user SGPR allocated for kernarg preloading.
+  Register FirstKernArgPreloadReg;
 
   std::tuple<const ArgDescriptor *, const TargetRegisterClass *, LLT>
   getPreloadedValue(PreloadedValue Value) const;
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp b/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
index 90c341ac0819c..737b2f740d6f7 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
@@ -207,12 +207,6 @@ void AMDGPUAsmPrinter::emitFunctionBodyStart() {
 
   if (STM.isAmdHsaOS())
     HSAMetadataStream->emitKernel(*MF, CurrentProgramInfo);
-
-  if (MFI.getNumKernargPreloadedSGPRs() > 0) {
-    assert(AMDGPU::hasKernargPreload(STM));
-    getTargetStreamer()->EmitKernargPreloadHeader(*getGlobalSTI(),
-                                                  STM.isAmdHsaOS());
-  }
 }
 
 void AMDGPUAsmPrinter::emitFunctionBodyEnd() {
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUPreloadKernArgProlog.cpp b/llvm/lib/Target/AMDGPU/AMDGPUPreloadKernArgProlog.cpp
new file mode 100644
index 0000000000000..b3a2139dfd24e
--- /dev/null
+++ b/llvm/lib/Target/AMDGPU/AMDGPUPreloadKernArgProlog.cpp
@@ -0,0 +1,211 @@
+//===- AMDGPUPreloadKernArgProlog.cpp - Preload KernArg Prolog ------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file This pass creates a backward compatibility layer for kernel argument
+/// preloading in situations where code is compiled with kernel argument
+/// preloading enabled but executed on hardware without firmware support for it.
+///
+/// To avoid recompilation, the pass inserts a block at the beginning of the
+/// program that loads the kernel arguments into SGPRs using s_load
+/// instructions. This sets up the registers exactly as they would be on systems
+/// with compatible firmware.
+///
+/// This effectively creates two entry points for the kernel. Firmware that
+/// supports the feature will automatically jump past the first 256 bytes of the
+/// program, skipping the compatibility layer and directly starting execution on
+/// the optimized code path.
+///
+/// This pass should be run as late as possible to prevent any optimizations
+/// that might assume the padding is dead code or that the added prologue is a
+/// true predecessor of the kernel entry block.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPUPreloadKernArgProlog.h"
+#include "AMDGPU.h"
+#include "GCNSubtarget.h"
+#include "SIMachineFunctionInfo.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/TargetParser/TargetParser.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "amdgpu-preload-kern-arg-prolog"
+
+namespace {
+
+// Used to build s_loads maping user SGPRs to kernel arguments
+struct LoadConfig {
+  unsigned Size;
+  const TargetRegisterClass *RegClass;
+  unsigned Opcode;
+  Register LoadReg = Register();
+};
+
+class AMDGPUPreloadKernArgProlog {
+public:
+  AMDGPUPreloadKernArgProlog(MachineFunction &MF);
+
+  bool run();
+
+private:
+  MachineFunction &MF;
+  const GCNSubtarget &ST;
+  const SIMachineFunctionInfo &MFI;
+  const SIInstrInfo &TII;
+  const TargetRegisterInfo &TRI;
+
+  // Create a new block before the entry point to the kernel. Firmware that
+  // supports preloading kernel arguments will automatically jump past this
+  // block to the alternative kernel entry point.
+  void createBackCompatBlock(unsigned NumKernArgPreloadSGPRs);
+
+  // Add instructions to load kernel arguments into SGPRs.
+  void addBackCompatLoads(MachineBasicBlock *BackCompatMBB,
+                          Register KernArgSegmentPtr,
+                          unsigned NumKernArgPreloadSGPRs);
+};
+
+class AMDGPUPreloadKernArgPrologLegacy : public MachineFunctionPass {
+public:
+  static char ID;
+
+  AMDGPUPreloadKernArgPrologLegacy() : MachineFunctionPass(ID) {}
+
+  StringRef getPassName() const override {
+    return "AMDGPU Preload Kernel Arguments Prolog";
+  }
+
+  bool runOnMachineFunction(MachineFunction &MF) override;
+};
+
+} // end anonymous namespace
+
+char AMDGPUPreloadKernArgPrologLegacy::ID = 0;
+
+INITIALIZE_PASS(AMDGPUPreloadKernArgPrologLegacy, DEBUG_TYPE,
+                "AMDGPU Preload Kernel Arguments Prolog", false, false)
+
+char &llvm::AMDGPUPreloadKernArgPrologLegacyID =
+    AMDGPUPreloadKernArgPrologLegacy::ID;
+
+FunctionPass *llvm::createAMDGPUPreloadKernArgPrologLegacyPass() {
+  return new AMDGPUPreloadKernArgPrologLegacy();
+}
+
+bool AMDGPUPreloadKernArgPrologLegacy::runOnMachineFunction(
+    MachineFunction &MF) {
+  return AMDGPUPreloadKernArgProlog(MF).run();
+}
+
+AMDGPUPreloadKernArgProlog::AMDGPUPreloadKernArgProlog(MachineFunction &MF)
+    : MF(MF), ST(MF.getSubtarget<GCNSubtarget>()),
+      MFI(*MF.getInfo<SIMachineFunctionInfo>()), TII(*ST.getInstrInfo()),
+      TRI(*ST.getRegisterInfo()) {}
+
+bool AMDGPUPreloadKernArgProlog::run() {
+  if (!ST.hasKernargPreload())
+    return false;
+
+  unsigned NumKernArgPreloadSGPRs = MFI.getNumKernargPreloadedSGPRs();
+  if (!NumKernArgPreloadSGPRs)
+    return false;
+
+  createBackCompatBlock(NumKernArgPreloadSGPRs);
+  return true;
+}
+
+void AMDGPUPreloadKernArgProlog::createBackCompatBlock(
+    unsigned NumKernArgPreloadSGPRs) {
+  auto KernelEntryMBB = MF.begin();
+  MachineBasicBlock *BackCompatMBB = MF.CreateMachineBasicBlock();
+  MF.insert(KernelEntryMBB, BackCompatMBB);
+
+  assert(MFI.getUserSGPRInfo().hasKernargSegmentPtr() &&
+         "Kernel argument segment pointer register not set.");
+  Register KernArgSegmentPtr = MFI.getArgInfo().KernargSegmentPtr.getRegister();
+  BackCompatMBB->addLiveIn(KernArgSegmentPtr);
+
+  // Load kernel arguments to SGPRs
+  addBackCompatLoads(BackCompatMBB, KernArgSegmentPtr, NumKernArgPreloadSGPRs);
+
+  // Wait for loads to complete
+  AMDGPU::IsaVersion IV = AMDGPU::getIsaVersion(ST.getCPU());
+  unsigned Waitcnt =
+      AMDGPU::encodeWaitcnt(IV, getVmcntBitMask(IV), getExpcntBitMask(IV), 0);
+  BuildMI(BackCompatMBB, DebugLoc(), TII.get(AMDGPU::S_WAITCNT))
+      .addImm(Waitcnt);
+
+  // Branch to kernel start
+  BuildMI(BackCompatMBB, DebugLoc(), TII.get(AMDGPU::S_BRANCH))
+      .addMBB(&*KernelEntryMBB);
+  BackCompatMBB->addSuccessor(&*KernelEntryMBB);
+
+  // Create a new basic block for padding to 256 bytes
+  MachineBasicBlock *PadMBB = MF.CreateMachineBasicBlock();
+  MF.insert(++BackCompatMBB->getIterator(), PadMBB);
+  PadMBB->setAlignment(Align(256));
+  PadMBB->addSuccessor(&*KernelEntryMBB);
+}
+
+/// Find the largest possible load size that fits with SGPR alignment
+static LoadConfig getLoadParameters(const TargetRegisterInfo &TRI,
+                                    Register KernArgPreloadSGPR,
+                                    unsigned NumKernArgPreloadSGPRs) {
+  static constexpr LoadConfig Configs[] = {
+      {8, &AMDGPU::SReg_256RegClass, AMDGPU::S_LOAD_DWORDX8_IMM},
+      {4, &AMDGPU::SReg_128RegClass, AMDGPU::S_LOAD_DWORDX4_IMM},
+      {2, &AMDGPU::SReg_64RegClass, AMDGPU::S_LOAD_DWORDX2_IMM}};
+
+  for (const auto &Config : Configs) {
+    if (NumKernArgPreloadSGPRs >= Config.Size) {
+      Register LoadReg = TRI.getMatchingSuperReg(KernArgPreloadSGPR,
+                                                 AMDGPU::sub0, Config.RegClass);
+      if (LoadReg) {
+        LoadConfig C(Config);
+        C.LoadReg = LoadReg;
+        return C;
+      }
+    }
+  }
+
+  // Fallback to a single register
+  return LoadConfig{1, &AMDGPU::SReg_32RegClass, AMDGPU::S_LOAD_DWORD_IMM,
+                    KernArgPreloadSGPR};
+}
+
+void AMDGPUPreloadKernArgProlog::addBackCompatLoads(
+    MachineBasicBlock *BackCompatMBB, Register KernArgSegmentPtr,
+    unsigned NumKernArgPreloadSGPRs) {
+  Register KernArgPreloadSGPR = MFI.getArgInfo().FirstKernArgPreloadReg;
+  unsigned Offset = 0;
+  // Fill all user SGPRs used for kernarg preloading with sequential data from
+  // the kernarg segment
+  while (NumKernArgPreloadSGPRs > 0) {
+    LoadConfig Config =
+        getLoadParameters(TRI, KernArgPreloadSGPR, NumKernArgPreloadSGPRs);
+
+    BuildMI(BackCompatMBB, DebugLoc(), TII.get(Config.Opcode), Config.LoadReg)
+        .addReg(KernArgSegmentPtr)
+        .addImm(Offset)
+        .addImm(0);
+
+    Offset += 4 * Config.Size;
+    KernArgPreloadSGPR = KernArgPreloadSGPR.asMCReg() + Config.Size;
+    NumKernArgPreloadSGPRs -= Config.Size;
+  }
+}
+
+PreservedAnalyses
+AMDGPUPreloadKernArgPrologPass::run(MachineFunction &MF,
+                                    MachineFunctionAnalysisManager &) {
+  if (!AMDGPUPreloadKernArgProlog(MF).run())
+    return PreservedAnalyses::all();
+
+  return PreservedAnalyses::none();
+}
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUPreloadKernArgProlog.h b/llvm/lib/Target/AMDGPU/AMDGPUPreloadKernArgProlog.h
new file mode 100644
index 0000000000000..1216132923ade
--- /dev/null
+++ b/llvm/lib/Target/AMDGPU/AMDGPUPreloadKernArgProlog.h
@@ -0,0 +1,25 @@
+//===- AMDGPUPreloadKernargProlog.h ----------------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_LIB_TARGET_AMDGPU_PRELOAD_KERNARG_PROLOG_H
+#define LLVM_LIB_TARGET_AMDGPU_PRELOAD_KERNARG_PROLOG_H
+
+#include "llvm/CodeGen/MachinePassManager.h"
+
+namespace llvm {
+
+class AMDGPUPreloadKernArgPrologPass
+    : public PassInfoMixin<AMDGPUPreloadKernArgPrologPass> {
+public:
+  PreservedAnalyses run(MachineFunction &MF,
+                        MachineFunctionAnalysisManager &AM);
+};
+
+} // end namespace llvm
+
+#endif // LLVM_LIB_TARGET_AMDGPU_PRELOAD_KERNARG_PROLOG_H
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
index 7256eec89008a..0c9d7d00a8a4a 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
@@ -540,6 +540,7 @@ extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAMDGPUTarget() {
   initializeGCNPreRALongBranchRegPass(*PR);
   initializeGCNRewritePartialRegUsesPass(*PR);
   initializeGCNRegPressurePrinterPass(*PR);
+  initializeAMDGPUPreloadKernArgPrologLegacyPass(*PR);
 }
 
 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
@@ -1669,6 +1670,7 @@ void GCNPassConfig::addPreEmitPass() {
     addPass(&AMDGPUInsertDelayAluID);
 
   addPass(&BranchRelaxationPassID);
+  addPass(createAMDGPUPreloadKernArgPrologLegacyPass());
 }
 
 TargetPassConfig *GCNTargetMachine::createPassConfig(PassManagerBase &PM) {
diff --git a/llvm/lib/Target/AMDGPU/CMakeLists.txt b/llvm/lib/Target/AMDGPU/CMakeLists.txt
index 03038caab521d..97a0d59cfeeda 100644
--- a/llvm/lib/Target/AMDGPU/CMakeLists.txt
+++ b/llvm/lib/Target/AMDGPU/CMakeLists.txt
@@ -88,6 +88,7 @@ add_llvm_target(AMDGPUCodeGen
   AMDGPUPerfHintAnalysis.cpp
   AMDGPUPostLegalizerCombiner.cpp
   AMDGPUPreLegalizerCombiner.cpp
+  AMDGPUPreloadKernArgProlog.cpp
   AMDGPUPrintfRuntimeBinding.cpp
   AMDGPUPromoteAlloca.cpp
   AMDGPUPromoteKernelArguments.cpp
diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp
index ffde4d33f1341..eccd77d6c00f0 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp
@@ -338,15 +338,6 @@ bool AMDGPUTargetAsmStreamer::EmitHSAMetadata(
   return true;
 }
 
-bool AMDGPUTargetAsmStreamer::EmitKernargPreloadHeader(
-    const MCSubtargetInfo &STI, bool TrapEnabled) {
-  OS << (TrapEnabled ? "\ts_trap 2" : "\ts_endpgm")
-     << " ; Kernarg preload header. Trap with incompatible firmware that "
-        "doesn't support preloading kernel arguments.\n";
-  OS << "\t.fill 63, 4, 0xbf800000 ; s_nop 0\n";
-  return true;
-}
-
 bool AMDGPUTargetAsmStreamer::EmitCodeEnd(const MCSubtargetInfo &STI) {
   const uint32_t Encoded_s_code_end = 0xbf9f0000;
   const uint32_t Encoded_s_nop = 0xbf800000;
@@ -935,20 +926,6 @@ bool AMDGPUTargetELFStreamer::EmitHSAMetadata(msgpack::Document &HSAMetadataDoc,
   return true;
 }
 
-bool AMDGPUTargetELFStreamer::EmitKernargPreloadHeader(
-    const MCSubtargetInfo &STI, bool TrapEnabled) {
-  const uint32_t Encoded_s_nop = 0xbf800000;
-  const uint32_t Encoded_s_trap = 0xbf920002;
-  const uint32_t Encoded_s_endpgm = 0xbf810000;
-  const uint32_t TrapInstr = TrapEnabled ? Encoded_s_trap : Encoded_s_endpgm;
-  MCStreamer &OS = getStreamer();
-  OS.emitInt32(TrapInstr);
-  for (int i = 0; i < 63; ++i) {
-    OS.emitInt32(Encoded_s_nop);
-  }
-  return true;
-}
-
 bool AMDGPUTargetELFStreamer::EmitCodeEnd(const MCSubtargetInfo &STI) {
   const uint32_t Encoded_s_code_end = 0xbf9f0000;
   const uint32_t Encoded_s_nop = 0xbf800000;
diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.h b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.h
index 6a91ad06de5d1..9c49020850584 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.h
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.h
@@ -96,12 +96,6 @@ class AMDGPUTargetStreamer : public MCTargetStreamer {
   /// \returns True on success, false on failure.
   virtual bool EmitCodeEnd(const MCSubtargetInfo &STI) { return true; }
 
-  /// \returns True on success, false on failure.
-  virtual bool EmitKernargPreloadHeader(const MCSubtargetInfo &STI,
-                                        bool TrapEnabled) {
-    return true;
-  }
-
   virtual void
   EmitAmdhsaKernelDescriptor(const MCSubtargetInfo &STI, StringRef KernelName,
                              const AMDGPU::MCKernelDescriptor &KernelDescriptor,
@@ -168,10 +162,6 @@ class AMDGPUTargetAsmStreamer final : public AMDGPUTargetStreamer {
   /// \returns True on success, false on failure.
   bool EmitCodeEnd(const MCSubtargetInfo &STI) override;
 
-  /// \returns True on success, false on failure.
-  bool EmitKernargPreloadHeader(const MCSubtargetInfo &STI,
-                                bool TrapEnabled) override;
-
   void
   EmitAmdhsaKernelDescriptor(const MCSubtargetInfo &STI, StringRef KernelName,
                              const AMDGPU::MCKernelDescriptor &KernelDescriptor,
@@ -225,10 +215,6 @@ class AMDGPUTargetELFStreamer final : public AMDGPUTargetStreamer {
   /// \returns True on success, false on failure.
   bool EmitCodeEnd(const MCSubtargetInfo &STI) override;
 
-  /// \returns True on success, false on failure.
-  bool EmitKernargPreloadHeader(const MCSubtargetInfo &STI,
-                                bool TrapEnabled) override;
-
   void
   EmitAmdhsaKernelDescriptor(const MCSubtargetInfo &STI, StringRef KernelName,
                              const AMDGPU::MCKernelDescriptor &KernelDescriptor,
diff --git a/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp b/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
index 8d7df73f3cee8..169f1369fb543 100644
--- a/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
@@ -262,6 +262,8 @@ SmallVectorImpl<MCRegister> *SIMachineFunctionInfo::addPreloadedKernArg(
   // If the available register tuples are aligned with the kernarg to be
   // preloaded use that register, otherwise we need to use a set of SGPRs and
   // merge them.
+  if (!ArgInfo.FirstKernArgPreloadReg)
+    ArgInfo.FirstKernArgPreloadReg = getNextUserSGPR();
   Register PreloadReg =
       TRI.getMatchingSuperReg(getNextUserSGPR(), AMDGPU::sub0, RC);
   if (PreloadReg &&
diff --git a/llvm/test/CodeGen/AMDGPU/llc-pipeline.ll b/llvm/test/CodeGen/AMDGPU/llc-pipeline.ll
index e77f4f69e265b..b2708cf13cbf3 100644
--- a/llvm/test/CodeGen/AMDGPU/llc-pipeline.ll
+++ b/llvm/test/CodeGen/AMDGPU/llc-pipeline.ll
@@ -144,6 +144,7 @@
 ; GCN-O0-NEXT:        SI Final Branch Preparation
 ; GCN-O0-NEXT:        Post RA hazard recognizer
 ; GCN-O0-NEXT:        Branch relaxation pass
+; GCN-O0-NEXT:        AMDGPU Preload Kernel Arguments Prolog
 ; GCN-O0-NEXT:        Register Usage Information Collector Pass
 ; GCN-O0-NEXT:        Remove Loads Into Fake Uses
 ; GCN-O0-NEXT:        Live DEBUG_VALUE analysis
@@ -424,6 +425,7 @@
 ; GCN-O1-NEXT:        Post RA hazard recognizer
 ; GCN-O1-NEXT:        AMDGPU Insert Delay ALU
 ; GCN-O1-NEXT:        Branch relaxation pass
+; GCN-O1-NEXT:        AMDGPU Preload Kernel Arguments Prolog
 ; GCN-O1-NEXT:        Register Usage Information Collector Pass
 ; GCN-O1-NEXT:        Remove Loads Into Fake Uses
 ; GCN-O1-NEXT:        Live DEBUG_VALUE analysis
@@ -732,6 +734,7 @@
 ; GCN-O1-OPTS-NEXT:        Post RA hazard recognizer
 ; GCN-O1-OPTS-NEXT:        AMDGPU Insert Delay ALU
 ; GCN-O1-OPTS-NEXT:        Branch relaxation pass
+; GCN-O1-OPTS-NEXT:        AMDGPU Preload Kernel Arguments Prolog
 ; GCN-O1-OPTS-NEXT:        Register Usage Information Collector Pass
 ; GCN-O1-OPTS-NEXT:        Remove Loads Into Fake Uses
 ; GCN-O1-OPTS-NEXT:        Live DEBUG_VALUE analysis
@@ -1046,6 +1049,7 @@
 ; GCN-O2-NEXT:        Post RA hazard recognizer
 ; GCN-O2-NEXT:        AMDGPU Insert Delay ALU
 ; GCN-O2-NEXT:        Branch relaxation pass
+; GCN-O2-NEXT:        AMDGPU Preload Kernel Arguments Prolog
 ; GCN-O2-NEXT:        Register Usage Information Collector Pass
 ; GCN-O2-NEXT:        Remove Loads Into Fake Uses
 ; GCN-O2-NEXT:        Live DEBUG_VALUE analysis
@@ -1373,6 +1377,7 @@
 ; GCN-O3-NEXT:        Post RA hazard recognizer
 ; GCN-O3-NEXT:        AMDGPU Insert Delay ALU
 ; GCN-O3-NEXT:        Branch relaxation pass
+; GCN-O3-NEXT:        AMDGPU Preload Kernel Arguments Prolog
 ; GCN-O3-NEXT:        Register Usage Information Collector Pass
 ; GCN-O3-NEXT:        Remove Loads Into Fake Uses
 ; GCN-O3-NEXT:        Live DEBUG_VALUE analysis
diff --git a/llvm/test/CodeGen/AMDGPU/preload-implicit-kernargs.ll b/llvm/test/CodeGen/AMDGPU/preload-implicit-kernargs.ll
index 0c6d8dce193da..31beb7a3cce24 100644
--- a/llvm/test/CodeGen/AMDGPU/preload-implicit-kernargs.ll
+++ b/llvm/test/CodeGen/AMDGPU/preload-implicit-kernargs.ll
@@ -4,18 +4,28 @@
 
 define amdgpu_kernel void @preload_block_count_x(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-LABEL: preload_block_count_x:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dword s4, s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB0_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB0_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s4
 ; GFX940-NEXT:    global_store_dword v0, v1, s[2:3] sc0 sc1
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: preload_block_count_x:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dword s8, s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB0_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB0_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s8
 ; GFX90a-NEXT:    global_store_dword v0, v1, s[6:7]
@@ -28,18 +38,30 @@ define amdgpu_kernel void @preload_block_count_x(ptr addrspace(1) inreg %out) #0
 
 define amdgpu_kernel void @preload_unused_arg_block_count_x(ptr addrspace(1) inreg %out, i32 inreg) #0 {
 ; GFX940-LABEL: preload_unused_arg_block_count_x:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x8
+; GFX940-NEXT:    s_load_dword s6, s[0:1], 0x10
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB1_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB1_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s6
 ; GFX940-NEXT:    global_store_dword v0, v1, s[2:3] sc0 sc1
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: preload_unused_arg_block_count_x:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; GFX90a-NEXT:    s_load_dword s10, s[4:5], 0x10
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB1_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB1_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s10
 ; GFX90a-NEXT:    global_store_dword v0, v1, s[6:7]
@@ -52,9 +74,13 @@ define amdgpu_kernel void @preload_unused_arg_block_count_x(ptr addrspace(1) inr
 
 define amdgpu_kernel void @no_free_sgprs_block_count_x(ptr addrspace(1) inreg %out, i256 inreg) {
 ; GFX940-LABEL: no_free_sgprs_block_count_x:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x0
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB2_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB2_0:
 ; GFX940-NEXT:    s_load_dword s0, s[4:5], 0x28
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
@@ -63,9 +89,13 @@ define amdgpu_kernel void @no_free_sgprs_block_count_x(ptr addrspace(1) inreg %o
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: no_free_sgprs_block_count_x:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx4 s[12:15], s[8:9], 0x0
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB2_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB2_0:
 ; GFX90a-NEXT:    s_load_dword s0, s[8:9], 0x28
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
@@ -135,9 +165,13 @@ define amdgpu_kernel void @mixed_inreg_block_count_x(ptr addrspace(1) %out, i32
 
 define amdgpu_kernel void @incorrect_type_i64_block_count_x(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-LABEL: incorrect_type_i64_block_count_x:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB5_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB5_0:
 ; GFX940-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x8
 ; GFX940-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
@@ -146,9 +180,13 @@ define amdgpu_kernel void @incorrect_type_i64_block_count_x(ptr addrspace(1) inr
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: incorrect_type_i64_block_count_x:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB5_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB5_0:
 ; GFX90a-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
 ; GFX90a-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
@@ -163,9 +201,13 @@ define amdgpu_kernel void @incorrect_type_i64_block_count_x(ptr addrspace(1) inr
 
 define amdgpu_kernel void @incorrect_type_i16_block_count_x(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-LABEL: incorrect_type_i16_block_count_x:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB6_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB6_0:
 ; GFX940-NEXT:    s_load_dword s0, s[0:1], 0x8
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
@@ -174,9 +216,13 @@ define amdgpu_kernel void @incorrect_type_i16_block_count_x(ptr addrspace(1) inr
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: incorrect_type_i16_block_count_x:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB6_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB6_0:
 ; GFX90a-NEXT:    s_load_dword s0, s[4:5], 0x8
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
@@ -191,18 +237,28 @@ define amdgpu_kernel void @incorrect_type_i16_block_count_x(ptr addrspace(1) inr
 
 define amdgpu_kernel void @preload_block_count_y(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-LABEL: preload_block_count_y:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB7_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB7_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s5
 ; GFX940-NEXT:    global_store_dword v0, v1, s[2:3] sc0 sc1
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: preload_block_count_y:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB7_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB7_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s9
 ; GFX90a-NEXT:    global_store_dword v0, v1, s[6:7]
@@ -216,9 +272,13 @@ define amdgpu_kernel void @preload_block_count_y(ptr addrspace(1) inreg %out) #0
 
 define amdgpu_kernel void @random_incorrect_offset(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-LABEL: random_incorrect_offset:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB8_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB8_0:
 ; GFX940-NEXT:    s_mov_b32 s4, 8
 ; GFX940-NEXT:    s_load_dword s0, s[0:1], s4 offset:0x2
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
@@ -228,9 +288,13 @@ define amdgpu_kernel void @random_incorrect_offset(ptr addrspace(1) inreg %out)
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: random_incorrect_offset:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB8_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB8_0:
 ; GFX90a-NEXT:    s_mov_b32 s0, 8
 ; GFX90a-NEXT:    s_load_dword s0, s[4:5], s0 offset:0x2
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
@@ -247,18 +311,30 @@ define amdgpu_kernel void @random_incorrect_offset(ptr addrspace(1) inreg %out)
 
 define amdgpu_kernel void @preload_block_count_z(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-LABEL: preload_block_count_z:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x8
+; GFX940-NEXT:    s_load_dword s6, s[0:1], 0x10
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB9_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB9_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s6
 ; GFX940-NEXT:    global_store_dword v0, v1, s[2:3] sc0 sc1
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: preload_block_count_z:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; GFX90a-NEXT:    s_load_dword s10, s[4:5], 0x10
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB9_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB9_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s10
 ; GFX90a-NEXT:    global_store_dword v0, v1, s[6:7]
@@ -272,9 +348,15 @@ define amdgpu_kernel void @preload_block_count_z(ptr addrspace(1) inreg %out) #0
 
 define amdgpu_kernel void @preload_block_count_x_imparg_align_ptr_i8(ptr addrspace(1) inreg %out, i8 inreg %val) #0 {
 ; GFX940-LABEL: preload_block_count_x_imparg_align_ptr_i8:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x8
+; GFX940-NEXT:    s_load_dword s6, s[0:1], 0x10
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB10_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB10_0:
 ; GFX940-NEXT:    s_and_b32 s0, s4, 0xff
 ; GFX940-NEXT:    s_add_i32 s0, s6, s0
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
@@ -283,9 +365,15 @@ define amdgpu_kernel void @preload_block_count_x_imparg_align_ptr_i8(ptr addrspa
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: preload_block_count_x_imparg_align_ptr_i8:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; GFX90a-NEXT:    s_load_dword s10, s[4:5], 0x10
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB10_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB10_0:
 ; GFX90a-NEXT:    s_and_b32 s0, s8, 0xff
 ; GFX90a-NEXT:    s_add_i32 s0, s10, s0
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
@@ -302,9 +390,15 @@ define amdgpu_kernel void @preload_block_count_x_imparg_align_ptr_i8(ptr addrspa
 
 define amdgpu_kernel void @preload_block_count_xyz(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-LABEL: preload_block_count_xyz:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x8
+; GFX940-NEXT:    s_load_dword s6, s[0:1], 0x10
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB11_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB11_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v3, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v0, s4
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s5
@@ -313,9 +407,15 @@ define amdgpu_kernel void @preload_block_count_xyz(ptr addrspace(1) inreg %out)
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: preload_block_count_xyz:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; GFX90a-NEXT:    s_load_dword s10, s[4:5], 0x10
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB11_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB11_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v3, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, s8
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s9
@@ -338,9 +438,14 @@ define amdgpu_kernel void @preload_block_count_xyz(ptr addrspace(1) inreg %out)
 
 define amdgpu_kernel void @preload_workgroup_size_x(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-LABEL: preload_workgroup_size_x:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB12_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB12_0:
 ; GFX940-NEXT:    s_and_b32 s0, s7, 0xffff
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s0
@@ -348,9 +453,14 @@ define amdgpu_kernel void @preload_workgroup_size_x(ptr addrspace(1) inreg %out)
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: preload_workgroup_size_x:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB12_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB12_0:
 ; GFX90a-NEXT:    s_and_b32 s0, s11, 0xffff
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s0
@@ -366,9 +476,14 @@ define amdgpu_kernel void @preload_workgroup_size_x(ptr addrspace(1) inreg %out)
 
 define amdgpu_kernel void @preload_workgroup_size_y(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-LABEL: preload_workgroup_size_y:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB13_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB13_0:
 ; GFX940-NEXT:    s_lshr_b32 s0, s7, 16
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s0
@@ -376,9 +491,14 @@ define amdgpu_kernel void @preload_workgroup_size_y(ptr addrspace(1) inreg %out)
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: preload_workgroup_size_y:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB13_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB13_0:
 ; GFX90a-NEXT:    s_lshr_b32 s0, s11, 16
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s0
@@ -394,9 +514,15 @@ define amdgpu_kernel void @preload_workgroup_size_y(ptr addrspace(1) inreg %out)
 
 define amdgpu_kernel void @preload_workgroup_size_z(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-LABEL: preload_workgroup_size_z:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x8
+; GFX940-NEXT:    s_load_dword s8, s[0:1], 0x18
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB14_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB14_0:
 ; GFX940-NEXT:    s_and_b32 s0, s8, 0xffff
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s0
@@ -404,9 +530,15 @@ define amdgpu_kernel void @preload_workgroup_size_z(ptr addrspace(1) inreg %out)
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: preload_workgroup_size_z:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x8
+; GFX90a-NEXT:    s_load_dword s12, s[4:5], 0x18
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB14_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB14_0:
 ; GFX90a-NEXT:    s_and_b32 s0, s12, 0xffff
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s0
@@ -422,9 +554,15 @@ define amdgpu_kernel void @preload_workgroup_size_z(ptr addrspace(1) inreg %out)
 
 define amdgpu_kernel void @preload_workgroup_size_xyz(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-LABEL: preload_workgroup_size_xyz:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x8
+; GFX940-NEXT:    s_load_dword s8, s[0:1], 0x18
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB15_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB15_0:
 ; GFX940-NEXT:    s_lshr_b32 s0, s7, 16
 ; GFX940-NEXT:    s_and_b32 s1, s7, 0xffff
 ; GFX940-NEXT:    s_and_b32 s4, s8, 0xffff
@@ -436,9 +574,15 @@ define amdgpu_kernel void @preload_workgroup_size_xyz(ptr addrspace(1) inreg %ou
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: preload_workgroup_size_xyz:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x8
+; GFX90a-NEXT:    s_load_dword s12, s[4:5], 0x18
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB15_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB15_0:
 ; GFX90a-NEXT:    s_lshr_b32 s0, s11, 16
 ; GFX90a-NEXT:    s_and_b32 s1, s11, 0xffff
 ; GFX90a-NEXT:    s_and_b32 s2, s12, 0xffff
@@ -467,9 +611,15 @@ define amdgpu_kernel void @preload_workgroup_size_xyz(ptr addrspace(1) inreg %ou
 
 define amdgpu_kernel void @preload_remainder_x(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-LABEL: preload_remainder_x:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x8
+; GFX940-NEXT:    s_load_dword s8, s[0:1], 0x18
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB16_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB16_0:
 ; GFX940-NEXT:    s_lshr_b32 s0, s8, 16
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s0
@@ -477,9 +627,15 @@ define amdgpu_kernel void @preload_remainder_x(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: preload_remainder_x:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x8
+; GFX90a-NEXT:    s_load_dword s12, s[4:5], 0x18
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB16_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB16_0:
 ; GFX90a-NEXT:    s_lshr_b32 s0, s12, 16
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s0
@@ -495,9 +651,15 @@ define amdgpu_kernel void @preload_remainder_x(ptr addrspace(1) inreg %out) #0 {
 
 define amdgpu_kernel void @preloadremainder_y(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-LABEL: preloadremainder_y:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x8
+; GFX940-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x18
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB17_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB17_0:
 ; GFX940-NEXT:    s_and_b32 s0, s9, 0xffff
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s0
@@ -505,9 +667,15 @@ define amdgpu_kernel void @preloadremainder_y(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: preloadremainder_y:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x8
+; GFX90a-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x18
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB17_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB17_0:
 ; GFX90a-NEXT:    s_and_b32 s0, s13, 0xffff
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s0
@@ -523,9 +691,15 @@ define amdgpu_kernel void @preloadremainder_y(ptr addrspace(1) inreg %out) #0 {
 
 define amdgpu_kernel void @preloadremainder_z(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-LABEL: preloadremainder_z:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x8
+; GFX940-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x18
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB18_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB18_0:
 ; GFX940-NEXT:    s_lshr_b32 s0, s9, 16
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s0
@@ -533,9 +707,15 @@ define amdgpu_kernel void @preloadremainder_z(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: preloadremainder_z:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x8
+; GFX90a-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x18
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB18_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB18_0:
 ; GFX90a-NEXT:    s_lshr_b32 s0, s13, 16
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s0
@@ -551,9 +731,15 @@ define amdgpu_kernel void @preloadremainder_z(ptr addrspace(1) inreg %out) #0 {
 
 define amdgpu_kernel void @preloadremainder_xyz(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-LABEL: preloadremainder_xyz:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x8
+; GFX940-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x18
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB19_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB19_0:
 ; GFX940-NEXT:    s_lshr_b32 s0, s9, 16
 ; GFX940-NEXT:    s_lshr_b32 s1, s8, 16
 ; GFX940-NEXT:    s_and_b32 s4, s9, 0xffff
@@ -565,9 +751,15 @@ define amdgpu_kernel void @preloadremainder_xyz(ptr addrspace(1) inreg %out) #0
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: preloadremainder_xyz:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x8
+; GFX90a-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x18
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB19_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB19_0:
 ; GFX90a-NEXT:    s_lshr_b32 s0, s13, 16
 ; GFX90a-NEXT:    s_lshr_b32 s1, s12, 16
 ; GFX90a-NEXT:    s_and_b32 s2, s13, 0xffff
@@ -596,9 +788,13 @@ define amdgpu_kernel void @preloadremainder_xyz(ptr addrspace(1) inreg %out) #0
 
 define amdgpu_kernel void @no_free_sgprs_preloadremainder_z(ptr addrspace(1) inreg %out) {
 ; GFX940-LABEL: no_free_sgprs_preloadremainder_z:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x0
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB20_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB20_0:
 ; GFX940-NEXT:    s_lshr_b32 s0, s15, 16
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s0
@@ -606,9 +802,13 @@ define amdgpu_kernel void @no_free_sgprs_preloadremainder_z(ptr addrspace(1) inr
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: no_free_sgprs_preloadremainder_z:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[12:13], s[8:9], 0x0
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB20_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB20_0:
 ; GFX90a-NEXT:    s_load_dword s0, s[8:9], 0x1c
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
@@ -628,18 +828,31 @@ define amdgpu_kernel void @no_free_sgprs_preloadremainder_z(ptr addrspace(1) inr
 
 define amdgpu_kernel void @preload_block_max_user_sgprs(ptr addrspace(1) inreg %out, i192 inreg %t0, i32 inreg %t1) #0 {
 ; GFX940-LABEL: preload_block_max_user_sgprs:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x8
+; GFX940-NEXT:    s_load_dword s12, s[0:1], 0x28
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB21_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB21_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s12
 ; GFX940-NEXT:    global_store_dword v0, v1, s[2:3] sc0 sc1
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: preload_block_max_user_sgprs:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x8
+; GFX90a-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x18
+; GFX90a-NEXT:    s_load_dword s14, s[4:5], 0x20
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB21_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB21_0:
 ; GFX90a-NEXT:    s_load_dword s0, s[4:5], 0x28
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
@@ -654,9 +867,15 @@ define amdgpu_kernel void @preload_block_max_user_sgprs(ptr addrspace(1) inreg %
 
 define amdgpu_kernel void @preload_block_count_z_workgroup_size_z_remainder_z(ptr addrspace(1) inreg %out) #0 {
 ; GFX940-LABEL: preload_block_count_z_workgroup_size_z_remainder_z:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x8
+; GFX940-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x18
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB22_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB22_0:
 ; GFX940-NEXT:    s_lshr_b32 s0, s9, 16
 ; GFX940-NEXT:    s_and_b32 s1, s8, 0xffff
 ; GFX940-NEXT:    v_mov_b32_e32 v3, 0
@@ -667,9 +886,15 @@ define amdgpu_kernel void @preload_block_count_z_workgroup_size_z_remainder_z(pt
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: preload_block_count_z_workgroup_size_z_remainder_z:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x8
+; GFX90a-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x18
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB22_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB22_0:
 ; GFX90a-NEXT:    s_lshr_b32 s0, s13, 16
 ; GFX90a-NEXT:    s_and_b32 s1, s12, 0xffff
 ; GFX90a-NEXT:    v_mov_b32_e32 v3, 0
diff --git a/llvm/test/CodeGen/AMDGPU/preload-kernarg-header.ll b/llvm/test/CodeGen/AMDGPU/preload-kernarg-header.ll
index 5a03381447d0e..58f0b9657476c 100644
--- a/llvm/test/CodeGen/AMDGPU/preload-kernarg-header.ll
+++ b/llvm/test/CodeGen/AMDGPU/preload-kernarg-header.ll
@@ -1,23 +1,52 @@
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx940 -amdgpu-kernarg-preload-count=1 -asm-verbose=0 < %s | FileCheck -check-prefixes=GCN,HSA,ASM %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx940 -amdgpu-kernarg-preload-count=1 -filetype=obj < %s | llvm-objdump --arch=amdgcn --mcpu=gfx940 --disassemble - | FileCheck -check-prefixes=GCN,HSA,OBJ %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx940 -amdgpu-kernarg-preload-count=1 -filetype=obj < %s | llvm-objdump --arch=amdgcn --mcpu=gfx940 --disassemble - | FileCheck -check-prefixes=GCN,NON-HSA,OBJ %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx940 -amdgpu-kernarg-preload-count=1 -asm-verbose=0 < %s | llvm-mc -triple amdgcn-amd-amdhsa -mcpu=gfx940 -filetype=obj | llvm-objdump --arch=amdgcn --mcpu=gfx940 --disassemble - | FileCheck -check-prefixes=GCN,HSA,OBJ %s
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 -asm-verbose=0 < %s | FileCheck -check-prefixes=ASM %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 -filetype=obj < %s | llvm-objdump --arch=amdgcn --mcpu=gfx942 --disassemble - | FileCheck -check-prefixes=OBJ %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 -amdgpu-kernarg-preload-count=1 -asm-verbose=0 < %s | llvm-mc -triple amdgcn-amd-amdhsa -mcpu=gfx942 -filetype=obj | llvm-objdump --arch=amdgcn --mcpu=gfx942 --disassemble - | FileCheck -check-prefixes=OBJ %s
 
-; GCN: preload_kernarg_header
-; HSA: s_trap 2
-; NON-HSA: s_endpgm
-; ASM: .fill 63, 4, 0xbf800000 ; s_nop 0
-; OBJ-COUNT-63: s_nop 0
-define amdgpu_kernel void @preload_kernarg_header(ptr inreg %arg) {
+; OBJ: preload_ptr_kernarg_header
+; OBJ-COUNT-60: s_nop 0
+define amdgpu_kernel void @preload_ptr_kernarg_header(ptr inreg %arg) {
+; ASM-LABEL: preload_ptr_kernarg_header:
+; ASM:         s_load_dwordx2 s[8:9], s[4:5], 0x0
+; ASM-NEXT:    s_waitcnt lgkmcnt(0)
+; ASM-NEXT:    s_branch .LBB0_0
+; ASM-NEXT:    .p2align 8
+; ASM-NEXT:  .LBB0_0:
+; ASM-NEXT:    v_mov_b64_e32 v[0:1], s[8:9]
+; ASM-NEXT:    flat_store_dwordx2 v[0:1], v[0:1]
+; ASM-NEXT:    s_endpgm
     store ptr %arg, ptr %arg
     ret void
 }
 
-; GCN: non_kernel_function
-; GCN-NOT: s_trap 2
-; GCN-NOT: s_nop 0
-; GCN: flat_store
+; OBJ: preload_i32_kernarg_header
+; OBJ-COUNT-58: s_nop 0
+define amdgpu_kernel void @preload_i32_kernarg_header(ptr inreg %arg, i32 inreg %arg1) {
+; ASM-LABEL: preload_i32_kernarg_header:
+; ASM:         s_load_dwordx2 s[8:9], s[4:5], 0x0
+; ASM-NEXT:    s_load_dword s10, s[4:5], 0x8
+; ASM-NEXT:    s_waitcnt lgkmcnt(0)
+; ASM-NEXT:    s_branch .LBB1_0
+; ASM-NEXT:    .p2align 8
+; ASM-NEXT:  .LBB1_0:
+; ASM-NEXT:    v_mov_b64_e32 v[0:1], s[8:9]
+; ASM-NEXT:    v_mov_b32_e32 v2, s10
+; ASM-NEXT:    flat_store_dword v[0:1], v2
+; ASM-NEXT:    s_endpgm
+    store i32 %arg1, ptr %arg
+    ret void
+}
+
+; OBJ: non_kernel_function
+; ASM: non_kernel_function
+; OBJ-NOT: s_branch
+; ASM-NOT: s_branch
 define void @non_kernel_function(ptr %arg) {
+; ASM-LABEL: non_kernel_function:
+; ASM:         s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; ASM-NEXT:    flat_store_dwordx2 v[0:1], v[0:1]
+; ASM-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; ASM-NEXT:    s_setpc_b64 s[30:31]
     store ptr %arg, ptr %arg
     ret void
 }
diff --git a/llvm/test/CodeGen/AMDGPU/preload-kernargs.ll b/llvm/test/CodeGen/AMDGPU/preload-kernargs.ll
index 2aaffd7121ae9..0f60888bcb2f5 100644
--- a/llvm/test/CodeGen/AMDGPU/preload-kernargs.ll
+++ b/llvm/test/CodeGen/AMDGPU/preload-kernargs.ll
@@ -5,9 +5,14 @@
 
 define amdgpu_kernel void @ptr1_i8(ptr addrspace(1) inreg %out, i8 inreg %arg0) #0 {
 ; GFX940-LABEL: ptr1_i8:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dword s4, s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB0_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB0_0:
 ; GFX940-NEXT:    s_and_b32 s0, s4, 0xff
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s0
@@ -15,9 +20,14 @@ define amdgpu_kernel void @ptr1_i8(ptr addrspace(1) inreg %out, i8 inreg %arg0)
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: ptr1_i8:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dword s8, s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB0_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB0_0:
 ; GFX90a-NEXT:    s_and_b32 s0, s8, 0xff
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s0
@@ -30,9 +40,14 @@ define amdgpu_kernel void @ptr1_i8(ptr addrspace(1) inreg %out, i8 inreg %arg0)
 
 define amdgpu_kernel void @ptr1_i8_zext_arg(ptr addrspace(1) inreg %out, i8 zeroext inreg %arg0) #0 {
 ; GFX940-LABEL: ptr1_i8_zext_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dword s4, s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB1_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB1_0:
 ; GFX940-NEXT:    s_and_b32 s0, s4, 0xff
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s0
@@ -40,9 +55,14 @@ define amdgpu_kernel void @ptr1_i8_zext_arg(ptr addrspace(1) inreg %out, i8 zero
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: ptr1_i8_zext_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dword s8, s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB1_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB1_0:
 ; GFX90a-NEXT:    s_and_b32 s0, s8, 0xff
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s0
@@ -55,9 +75,14 @@ define amdgpu_kernel void @ptr1_i8_zext_arg(ptr addrspace(1) inreg %out, i8 zero
 
 define amdgpu_kernel void @ptr1_i16_preload_arg(ptr addrspace(1) inreg %out, i16 inreg %arg0) #0 {
 ; GFX940-LABEL: ptr1_i16_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dword s4, s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB2_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB2_0:
 ; GFX940-NEXT:    s_and_b32 s0, s4, 0xffff
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s0
@@ -65,9 +90,14 @@ define amdgpu_kernel void @ptr1_i16_preload_arg(ptr addrspace(1) inreg %out, i16
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: ptr1_i16_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dword s8, s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB2_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB2_0:
 ; GFX90a-NEXT:    s_and_b32 s0, s8, 0xffff
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s0
@@ -80,18 +110,28 @@ define amdgpu_kernel void @ptr1_i16_preload_arg(ptr addrspace(1) inreg %out, i16
 
 define amdgpu_kernel void @ptr1_i32_preload_arg(ptr addrspace(1) inreg %out, i32 inreg %arg0) #0 {
 ; GFX940-LABEL: ptr1_i32_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dword s4, s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB3_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB3_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s4
 ; GFX940-NEXT:    global_store_dword v0, v1, s[2:3] sc0 sc1
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: ptr1_i32_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dword s8, s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB3_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB3_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s8
 ; GFX90a-NEXT:    global_store_dword v0, v1, s[6:7]
@@ -103,9 +143,15 @@ define amdgpu_kernel void @ptr1_i32_preload_arg(ptr addrspace(1) inreg %out, i32
 
 define amdgpu_kernel void @i32_ptr1_i32_preload_arg(i32 inreg %arg0, ptr addrspace(1) inreg %out, i32 inreg %arg1) #0 {
 ; GFX940-LABEL: i32_ptr1_i32_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x8
+; GFX940-NEXT:    s_load_dword s6, s[0:1], 0x10
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB4_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB4_0:
 ; GFX940-NEXT:    s_add_i32 s0, s2, s6
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s0
@@ -113,9 +159,15 @@ define amdgpu_kernel void @i32_ptr1_i32_preload_arg(i32 inreg %arg0, ptr addrspa
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: i32_ptr1_i32_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; GFX90a-NEXT:    s_load_dword s10, s[4:5], 0x10
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB4_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB4_0:
 ; GFX90a-NEXT:    s_add_i32 s0, s6, s10
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s0
@@ -128,9 +180,14 @@ define amdgpu_kernel void @i32_ptr1_i32_preload_arg(i32 inreg %arg0, ptr addrspa
 
 define amdgpu_kernel void @ptr1_i16_i16_preload_arg(ptr addrspace(1) inreg %out, i16 inreg %arg0, i16 inreg %arg1) #0 {
 ; GFX940-LABEL: ptr1_i16_i16_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dword s4, s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB5_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB5_0:
 ; GFX940-NEXT:    s_lshr_b32 s0, s4, 16
 ; GFX940-NEXT:    s_and_b32 s1, s4, 0xffff
 ; GFX940-NEXT:    s_add_i32 s0, s1, s0
@@ -140,9 +197,14 @@ define amdgpu_kernel void @ptr1_i16_i16_preload_arg(ptr addrspace(1) inreg %out,
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: ptr1_i16_i16_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dword s8, s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB5_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB5_0:
 ; GFX90a-NEXT:    s_lshr_b32 s0, s8, 16
 ; GFX90a-NEXT:    s_and_b32 s1, s8, 0xffff
 ; GFX90a-NEXT:    s_add_i32 s0, s1, s0
@@ -159,18 +221,28 @@ define amdgpu_kernel void @ptr1_i16_i16_preload_arg(ptr addrspace(1) inreg %out,
 
 define amdgpu_kernel void @ptr1_v2i8_preload_arg(ptr addrspace(1) inreg %out, <2 x i8> inreg %in) #0 {
 ; GFX940-LABEL: ptr1_v2i8_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dword s4, s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB6_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB6_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s4
 ; GFX940-NEXT:    global_store_short v0, v1, s[2:3] sc0 sc1
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: ptr1_v2i8_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dword s8, s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB6_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB6_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s8
 ; GFX90a-NEXT:    global_store_short v0, v1, s[6:7]
@@ -182,9 +254,13 @@ define amdgpu_kernel void @ptr1_v2i8_preload_arg(ptr addrspace(1) inreg %out, <2
 
 define amdgpu_kernel void @byref_preload_arg(ptr addrspace(1) inreg %out, ptr addrspace(4) byref(i32) align(256) %in.byref, i32 %after.offset) #0 {
 ; GFX940-LABEL: byref_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB7_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB7_0:
 ; GFX940-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x100
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
@@ -197,9 +273,13 @@ define amdgpu_kernel void @byref_preload_arg(ptr addrspace(1) inreg %out, ptr ad
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: byref_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB7_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB7_0:
 ; GFX90a-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x100
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
@@ -220,9 +300,13 @@ define amdgpu_kernel void @byref_preload_arg(ptr addrspace(1) inreg %out, ptr ad
 
 define amdgpu_kernel void @byref_staggered_preload_arg(ptr addrspace(1) inreg %out, ptr addrspace(4) byref(i32) align(256) %in.byref, i32 inreg %after.offset) #0 {
 ; GFX940-LABEL: byref_staggered_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB8_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB8_0:
 ; GFX940-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x100
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
@@ -235,9 +319,13 @@ define amdgpu_kernel void @byref_staggered_preload_arg(ptr addrspace(1) inreg %o
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: byref_staggered_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB8_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB8_0:
 ; GFX90a-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x100
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
@@ -257,9 +345,13 @@ define amdgpu_kernel void @byref_staggered_preload_arg(ptr addrspace(1) inreg %o
 
 define amdgpu_kernel void @v8i32_arg(ptr addrspace(1) nocapture inreg %out, <8 x i32> inreg %in) #0 {
 ; GFX940-LABEL: v8i32_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB9_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB9_0:
 ; GFX940-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x20
 ; GFX940-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
@@ -277,9 +369,13 @@ define amdgpu_kernel void @v8i32_arg(ptr addrspace(1) nocapture inreg %out, <8 x
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: v8i32_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB9_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB9_0:
 ; GFX90a-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x20
 ; GFX90a-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
@@ -301,9 +397,14 @@ define amdgpu_kernel void @v8i32_arg(ptr addrspace(1) nocapture inreg %out, <8 x
 
 define amdgpu_kernel void @v3i16_preload_arg(ptr addrspace(1) nocapture inreg %out, <3 x i16> inreg %in) #0 {
 ; GFX940-LABEL: v3i16_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB10_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB10_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s5
 ; GFX940-NEXT:    global_store_short v0, v1, s[2:3] offset:4 sc0 sc1
@@ -312,9 +413,14 @@ define amdgpu_kernel void @v3i16_preload_arg(ptr addrspace(1) nocapture inreg %o
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: v3i16_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB10_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB10_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s9
 ; GFX90a-NEXT:    global_store_short v0, v1, s[6:7] offset:4
@@ -327,9 +433,15 @@ define amdgpu_kernel void @v3i16_preload_arg(ptr addrspace(1) nocapture inreg %o
 
 define amdgpu_kernel void @v3i32_preload_arg(ptr addrspace(1) nocapture inreg %out, <3 x i32> inreg %in) #0 {
 ; GFX940-LABEL: v3i32_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x8
+; GFX940-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x18
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB11_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB11_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v0, s6
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX940-NEXT:    v_mov_b32_e32 v2, s8
@@ -338,9 +450,15 @@ define amdgpu_kernel void @v3i32_preload_arg(ptr addrspace(1) nocapture inreg %o
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: v3i32_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x8
+; GFX90a-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x18
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB11_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB11_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, s10
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s11
 ; GFX90a-NEXT:    v_mov_b32_e32 v2, s12
@@ -353,9 +471,15 @@ define amdgpu_kernel void @v3i32_preload_arg(ptr addrspace(1) nocapture inreg %o
 
 define amdgpu_kernel void @v3f32_preload_arg(ptr addrspace(1) nocapture inreg %out, <3 x float> inreg %in) #0 {
 ; GFX940-LABEL: v3f32_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x8
+; GFX940-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x18
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB12_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB12_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v3, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v0, s6
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s7
@@ -364,9 +488,15 @@ define amdgpu_kernel void @v3f32_preload_arg(ptr addrspace(1) nocapture inreg %o
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: v3f32_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x8
+; GFX90a-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x18
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB12_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB12_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v3, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, s10
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s11
@@ -379,9 +509,14 @@ define amdgpu_kernel void @v3f32_preload_arg(ptr addrspace(1) nocapture inreg %o
 
 define amdgpu_kernel void @v5i8_preload_arg(ptr addrspace(1) nocapture inreg %out, <5 x i8> inreg %in) #0 {
 ; GFX940-LABEL: v5i8_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB13_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB13_0:
 ; GFX940-NEXT:    s_lshr_b32 s1, s4, 24
 ; GFX940-NEXT:    s_and_b32 s0, s4, 0xffff
 ; GFX940-NEXT:    s_lshl_b32 s1, s1, 8
@@ -397,9 +532,14 @@ define amdgpu_kernel void @v5i8_preload_arg(ptr addrspace(1) nocapture inreg %ou
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: v5i8_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB13_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB13_0:
 ; GFX90a-NEXT:    s_lshr_b32 s1, s8, 24
 ; GFX90a-NEXT:    s_lshl_b32 s1, s1, 8
 ; GFX90a-NEXT:    s_bfe_u32 s2, s8, 0x80010
@@ -419,9 +559,13 @@ define amdgpu_kernel void @v5i8_preload_arg(ptr addrspace(1) nocapture inreg %ou
 
 define amdgpu_kernel void @v5f64_arg(ptr addrspace(1) nocapture inreg %out, <5 x double> inreg %in) #0 {
 ; GFX940-LABEL: v5f64_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB14_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB14_0:
 ; GFX940-NEXT:    s_load_dwordx2 s[12:13], s[0:1], 0x60
 ; GFX940-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x40
 ; GFX940-NEXT:    v_mov_b32_e32 v4, 0
@@ -442,9 +586,13 @@ define amdgpu_kernel void @v5f64_arg(ptr addrspace(1) nocapture inreg %out, <5 x
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: v5f64_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB14_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB14_0:
 ; GFX90a-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x60
 ; GFX90a-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x40
 ; GFX90a-NEXT:    v_mov_b32_e32 v4, 0
@@ -469,9 +617,14 @@ define amdgpu_kernel void @v5f64_arg(ptr addrspace(1) nocapture inreg %out, <5 x
 
 define amdgpu_kernel void @v8i8_preload_arg(ptr addrspace(1) inreg %out, <8 x i8> inreg %in) #0 {
 ; GFX940-LABEL: v8i8_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB15_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB15_0:
 ; GFX940-NEXT:    s_lshr_b32 s1, s5, 24
 ; GFX940-NEXT:    s_and_b32 s0, s5, 0xffff
 ; GFX940-NEXT:    s_lshl_b32 s1, s1, 8
@@ -493,9 +646,14 @@ define amdgpu_kernel void @v8i8_preload_arg(ptr addrspace(1) inreg %out, <8 x i8
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: v8i8_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB15_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB15_0:
 ; GFX90a-NEXT:    s_lshr_b32 s1, s9, 24
 ; GFX90a-NEXT:    s_lshl_b32 s1, s1, 8
 ; GFX90a-NEXT:    s_bfe_u32 s2, s9, 0x80010
@@ -521,18 +679,28 @@ define amdgpu_kernel void @v8i8_preload_arg(ptr addrspace(1) inreg %out, <8 x i8
 
 define amdgpu_kernel void @i64_kernel_preload_arg(ptr addrspace(1) inreg %out, i64 inreg %a) #0 {
 ; GFX940-LABEL: i64_kernel_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB16_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB16_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX940-NEXT:    v_mov_b64_e32 v[0:1], s[4:5]
 ; GFX940-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3] sc0 sc1
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: i64_kernel_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB16_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB16_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX90a-NEXT:    v_pk_mov_b32 v[0:1], s[8:9], s[8:9] op_sel:[0,1]
 ; GFX90a-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
@@ -543,18 +711,28 @@ define amdgpu_kernel void @i64_kernel_preload_arg(ptr addrspace(1) inreg %out, i
 
 define amdgpu_kernel void @f64_kernel_preload_arg(ptr addrspace(1) inreg %out, double inreg %in) #0 {
 ; GFX940-LABEL: f64_kernel_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB17_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB17_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX940-NEXT:    v_mov_b64_e32 v[0:1], s[4:5]
 ; GFX940-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3] sc0 sc1
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: f64_kernel_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB17_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB17_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX90a-NEXT:    v_pk_mov_b32 v[0:1], s[8:9], s[8:9] op_sel:[0,1]
 ; GFX90a-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
@@ -565,18 +743,28 @@ define amdgpu_kernel void @f64_kernel_preload_arg(ptr addrspace(1) inreg %out, d
 
 define amdgpu_kernel void @half_kernel_preload_arg(ptr addrspace(1) inreg %out, half inreg %in) #0 {
 ; GFX940-LABEL: half_kernel_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dword s4, s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB18_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB18_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s4
 ; GFX940-NEXT:    global_store_short v0, v1, s[2:3] sc0 sc1
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: half_kernel_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dword s8, s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB18_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB18_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s8
 ; GFX90a-NEXT:    global_store_short v0, v1, s[6:7]
@@ -587,18 +775,28 @@ define amdgpu_kernel void @half_kernel_preload_arg(ptr addrspace(1) inreg %out,
 
 define amdgpu_kernel void @bfloat_kernel_preload_arg(ptr addrspace(1) inreg %out, bfloat inreg %in) #0 {
 ; GFX940-LABEL: bfloat_kernel_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dword s4, s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB19_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB19_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s4
 ; GFX940-NEXT:    global_store_short v0, v1, s[2:3] sc0 sc1
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: bfloat_kernel_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dword s8, s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB19_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB19_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s8
 ; GFX90a-NEXT:    global_store_short v0, v1, s[6:7]
@@ -609,18 +807,28 @@ define amdgpu_kernel void @bfloat_kernel_preload_arg(ptr addrspace(1) inreg %out
 
 define amdgpu_kernel void @v2bfloat_kernel_preload_arg(ptr addrspace(1) inreg %out, <2 x bfloat> inreg %in) #0 {
 ; GFX940-LABEL: v2bfloat_kernel_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dword s4, s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB20_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB20_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s4
 ; GFX940-NEXT:    global_store_dword v0, v1, s[2:3] sc0 sc1
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: v2bfloat_kernel_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dword s8, s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB20_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB20_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s8
 ; GFX90a-NEXT:    global_store_dword v0, v1, s[6:7]
@@ -631,9 +839,14 @@ define amdgpu_kernel void @v2bfloat_kernel_preload_arg(ptr addrspace(1) inreg %o
 
 define amdgpu_kernel void @v3bfloat_kernel_preload_arg(ptr addrspace(1) inreg %out, <3 x bfloat> inreg %in) #0 {
 ; GFX940-LABEL: v3bfloat_kernel_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB21_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB21_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s5
 ; GFX940-NEXT:    global_store_short v0, v1, s[2:3] offset:4 sc0 sc1
@@ -642,9 +855,14 @@ define amdgpu_kernel void @v3bfloat_kernel_preload_arg(ptr addrspace(1) inreg %o
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: v3bfloat_kernel_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB21_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB21_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s9
 ; GFX90a-NEXT:    global_store_short v0, v1, s[6:7] offset:4
@@ -657,9 +875,15 @@ define amdgpu_kernel void @v3bfloat_kernel_preload_arg(ptr addrspace(1) inreg %o
 
 define amdgpu_kernel void @v6bfloat_kernel_preload_arg(ptr addrspace(1) inreg %out, <6 x bfloat> inreg %in) #0 {
 ; GFX940-LABEL: v6bfloat_kernel_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x8
+; GFX940-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x18
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB22_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB22_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v0, s6
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX940-NEXT:    v_mov_b32_e32 v2, s8
@@ -668,9 +892,15 @@ define amdgpu_kernel void @v6bfloat_kernel_preload_arg(ptr addrspace(1) inreg %o
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: v6bfloat_kernel_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x8
+; GFX90a-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x18
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB22_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB22_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, s10
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s11
 ; GFX90a-NEXT:    v_mov_b32_e32 v2, s12
@@ -683,9 +913,14 @@ define amdgpu_kernel void @v6bfloat_kernel_preload_arg(ptr addrspace(1) inreg %o
 
 define amdgpu_kernel void @half_v7bfloat_kernel_preload_arg(ptr addrspace(1) inreg %out, half inreg %in, <7 x bfloat> inreg %in2, ptr addrspace(1) inreg %out2) #0 {
 ; GFX940-LABEL: half_v7bfloat_kernel_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB23_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB23_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v3, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v0, s4
 ; GFX940-NEXT:    global_store_short v3, v0, s[2:3] sc0 sc1
@@ -698,9 +933,14 @@ define amdgpu_kernel void @half_v7bfloat_kernel_preload_arg(ptr addrspace(1) inr
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: half_v7bfloat_kernel_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB23_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB23_0:
 ; GFX90a-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x20
 ; GFX90a-NEXT:    v_mov_b32_e32 v3, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, s8
@@ -720,9 +960,14 @@ define amdgpu_kernel void @half_v7bfloat_kernel_preload_arg(ptr addrspace(1) inr
 
 define amdgpu_kernel void @i1_kernel_preload_arg(ptr addrspace(1) inreg %out, i1 inreg %in) #0 {
 ; GFX940-LABEL: i1_kernel_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dword s4, s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB24_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB24_0:
 ; GFX940-NEXT:    s_and_b32 s0, s4, 1
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s0
@@ -730,9 +975,14 @@ define amdgpu_kernel void @i1_kernel_preload_arg(ptr addrspace(1) inreg %out, i1
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: i1_kernel_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dword s8, s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB24_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB24_0:
 ; GFX90a-NEXT:    s_and_b32 s0, s8, 1
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s0
@@ -744,9 +994,15 @@ define amdgpu_kernel void @i1_kernel_preload_arg(ptr addrspace(1) inreg %out, i1
 
 define amdgpu_kernel void @fp128_kernel_preload_arg(ptr addrspace(1) inreg %out, fp128 inreg %in) #0 {
 ; GFX940-LABEL: fp128_kernel_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x8
+; GFX940-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x18
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB25_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB25_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v0, s6
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s7
@@ -756,9 +1012,15 @@ define amdgpu_kernel void @fp128_kernel_preload_arg(ptr addrspace(1) inreg %out,
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: fp128_kernel_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x8
+; GFX90a-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x18
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB25_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB25_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, s10
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s11
@@ -772,9 +1034,14 @@ define amdgpu_kernel void @fp128_kernel_preload_arg(ptr addrspace(1) inreg %out,
 
 define amdgpu_kernel void @v7i8_kernel_preload_arg(ptr addrspace(1) inreg %out, <7 x i8> inreg %in) #0 {
 ; GFX940-LABEL: v7i8_kernel_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB26_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB26_0:
 ; GFX940-NEXT:    s_lshr_b32 s1, s4, 24
 ; GFX940-NEXT:    s_and_b32 s0, s4, 0xffff
 ; GFX940-NEXT:    s_lshl_b32 s1, s1, 8
@@ -791,9 +1058,14 @@ define amdgpu_kernel void @v7i8_kernel_preload_arg(ptr addrspace(1) inreg %out,
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: v7i8_kernel_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB26_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB26_0:
 ; GFX90a-NEXT:    s_lshr_b32 s1, s8, 24
 ; GFX90a-NEXT:    s_lshl_b32 s1, s1, 8
 ; GFX90a-NEXT:    s_bfe_u32 s2, s8, 0x80010
@@ -814,9 +1086,15 @@ define amdgpu_kernel void @v7i8_kernel_preload_arg(ptr addrspace(1) inreg %out,
 
 define amdgpu_kernel void @v7half_kernel_preload_arg(ptr addrspace(1) inreg %out, <7 x half> inreg %in) #0 {
 ; GFX940-LABEL: v7half_kernel_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x8
+; GFX940-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x18
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB27_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB27_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v3, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v0, s9
 ; GFX940-NEXT:    global_store_short v3, v0, s[2:3] offset:12 sc0 sc1
@@ -827,9 +1105,15 @@ define amdgpu_kernel void @v7half_kernel_preload_arg(ptr addrspace(1) inreg %out
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: v7half_kernel_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x8
+; GFX90a-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x18
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB27_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB27_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v3, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, s13
 ; GFX90a-NEXT:    global_store_short v3, v0, s[6:7] offset:12
@@ -842,29 +1126,37 @@ define amdgpu_kernel void @v7half_kernel_preload_arg(ptr addrspace(1) inreg %out
   ret void
 }
 
-define amdgpu_kernel void @i16_i32_kernel_preload_arg(ptr addrspace(1) %out, i16 inreg %in, i32 inreg %in2, ptr addrspace(1) inreg %out2) #0 {
+define amdgpu_kernel void @i16_i32_kernel_preload_arg(ptr addrspace(1) inreg %out, i16 inreg %in, i32 inreg %in2, ptr addrspace(1) inreg %out2) #0 {
 ; GFX940-LABEL: i16_i32_kernel_preload_arg:
-; GFX940:       ; %bb.0:
-; GFX940-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x0
-; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x10
-; GFX940-NEXT:    v_mov_b32_e32 v0, 0
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x8
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_mov_b32_e32 v1, s6
-; GFX940-NEXT:    v_mov_b32_e32 v2, s7
-; GFX940-NEXT:    global_store_short v0, v1, s[4:5] sc0 sc1
-; GFX940-NEXT:    global_store_dword v0, v2, s[2:3] sc0 sc1
+; GFX940-NEXT:    s_branch .LBB28_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB28_0:
+; GFX940-NEXT:    v_mov_b32_e32 v0, 0
+; GFX940-NEXT:    v_mov_b32_e32 v1, s4
+; GFX940-NEXT:    global_store_short v0, v1, s[2:3] sc0 sc1
+; GFX940-NEXT:    v_mov_b32_e32 v1, s5
+; GFX940-NEXT:    global_store_dword v0, v1, s[6:7] sc0 sc1
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: i16_i32_kernel_preload_arg:
-; GFX90a:       ; %bb.0:
-; GFX90a-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
-; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x10
-; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x8
 ; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX90a-NEXT:    v_mov_b32_e32 v1, s2
-; GFX90a-NEXT:    v_mov_b32_e32 v2, s3
-; GFX90a-NEXT:    global_store_short v0, v1, s[0:1]
-; GFX90a-NEXT:    global_store_dword v0, v2, s[6:7]
+; GFX90a-NEXT:    s_branch .LBB28_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB28_0:
+; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
+; GFX90a-NEXT:    v_mov_b32_e32 v1, s8
+; GFX90a-NEXT:    global_store_short v0, v1, s[6:7]
+; GFX90a-NEXT:    v_mov_b32_e32 v1, s9
+; GFX90a-NEXT:    global_store_dword v0, v1, s[10:11]
 ; GFX90a-NEXT:    s_endpgm
   store i16 %in, ptr addrspace(1) %out
   store i32 %in2, ptr addrspace(1) %out2
@@ -873,9 +1165,14 @@ define amdgpu_kernel void @i16_i32_kernel_preload_arg(ptr addrspace(1) %out, i16
 
 define amdgpu_kernel void @i16_v3i32_kernel_preload_arg(ptr addrspace(1) inreg %out, i16 inreg %in, <3 x i32> inreg %in2, ptr addrspace(1) inreg %out2) #0 {
 ; GFX940-LABEL: i16_v3i32_kernel_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB29_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB29_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v3, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v4, s4
 ; GFX940-NEXT:    v_mov_b32_e32 v0, s6
@@ -886,9 +1183,14 @@ define amdgpu_kernel void @i16_v3i32_kernel_preload_arg(ptr addrspace(1) inreg %
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: i16_v3i32_kernel_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx8 s[8:15], s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB29_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB29_0:
 ; GFX90a-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x20
 ; GFX90a-NEXT:    v_mov_b32_e32 v3, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v4, s8
@@ -906,9 +1208,14 @@ define amdgpu_kernel void @i16_v3i32_kernel_preload_arg(ptr addrspace(1) inreg %
 
 define amdgpu_kernel void @i16_i16_kernel_preload_arg(ptr addrspace(1) inreg %out, i16 inreg %in, i16 inreg %in2, ptr addrspace(1) inreg %out2) #0 {
 ; GFX940-LABEL: i16_i16_kernel_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB30_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB30_0:
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s4
 ; GFX940-NEXT:    global_store_short v0, v1, s[2:3] sc0 sc1
@@ -916,9 +1223,14 @@ define amdgpu_kernel void @i16_i16_kernel_preload_arg(ptr addrspace(1) inreg %ou
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: i16_i16_kernel_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB30_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB30_0:
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s8
 ; GFX90a-NEXT:    global_store_short v0, v1, s[6:7]
@@ -931,9 +1243,14 @@ define amdgpu_kernel void @i16_i16_kernel_preload_arg(ptr addrspace(1) inreg %ou
 
 define amdgpu_kernel void @i16_v2i8_kernel_preload_arg(ptr addrspace(1) inreg %out, i16 inreg %in, <2 x i8> inreg %in2, ptr addrspace(1) inreg %out2) #0 {
 ; GFX940-LABEL: i16_v2i8_kernel_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB31_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB31_0:
 ; GFX940-NEXT:    s_lshr_b32 s0, s4, 24
 ; GFX940-NEXT:    s_lshl_b32 s0, s0, 8
 ; GFX940-NEXT:    s_bfe_u32 s1, s4, 0x80010
@@ -946,9 +1263,14 @@ define amdgpu_kernel void @i16_v2i8_kernel_preload_arg(ptr addrspace(1) inreg %o
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: i16_v2i8_kernel_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB31_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB31_0:
 ; GFX90a-NEXT:    s_lshr_b32 s0, s8, 24
 ; GFX90a-NEXT:    s_lshl_b32 s0, s0, 8
 ; GFX90a-NEXT:    s_bfe_u32 s1, s8, 0x80010
@@ -968,9 +1290,13 @@ define amdgpu_kernel void @i16_v2i8_kernel_preload_arg(ptr addrspace(1) inreg %o
 
 define amdgpu_kernel void @i32_ptr1_i32_staggered_preload_arg(i32 inreg %arg0, ptr addrspace(1) %out, i32 inreg %arg1) #0 {
 ; GFX940-LABEL: i32_ptr1_i32_staggered_preload_arg:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB32_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB32_0:
 ; GFX940-NEXT:    s_load_dword s3, s[0:1], 0x10
 ; GFX940-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x8
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
@@ -981,9 +1307,13 @@ define amdgpu_kernel void @i32_ptr1_i32_staggered_preload_arg(i32 inreg %arg0, p
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: i32_ptr1_i32_staggered_preload_arg:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dword s6, s[4:5], 0x0
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB32_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB32_0:
 ; GFX90a-NEXT:    s_load_dword s2, s[4:5], 0x10
 ; GFX90a-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
@@ -999,9 +1329,14 @@ define amdgpu_kernel void @i32_ptr1_i32_staggered_preload_arg(i32 inreg %arg0, p
 
 define amdgpu_kernel void @ptr1_i8_trailing_unused(ptr addrspace(1) inreg %out, i8 inreg %arg0, i32 inreg %unused) #0 {
 ; GFX940-LABEL: ptr1_i8_trailing_unused:
-; GFX940:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX940-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX940-NEXT:  ; %bb.0:
+; GFX940:       ; %bb.1:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX940-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x8
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    s_branch .LBB33_0
+; GFX940-NEXT:    .p2align 8
+; GFX940-NEXT:  ; %bb.2:
+; GFX940-NEXT:  .LBB33_0:
 ; GFX940-NEXT:    s_and_b32 s0, s4, 0xff
 ; GFX940-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s0
@@ -1009,9 +1344,14 @@ define amdgpu_kernel void @ptr1_i8_trailing_unused(ptr addrspace(1) inreg %out,
 ; GFX940-NEXT:    s_endpgm
 ;
 ; GFX90a-LABEL: ptr1_i8_trailing_unused:
-; GFX90a:         s_trap 2 ; Kernarg preload header. Trap with incompatible firmware that doesn't support preloading kernel arguments.
-; GFX90a-NEXT:    .fill 63, 4, 0xbf800000 ; s_nop 0
-; GFX90a-NEXT:  ; %bb.0:
+; GFX90a:       ; %bb.1:
+; GFX90a-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x0
+; GFX90a-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; GFX90a-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90a-NEXT:    s_branch .LBB33_0
+; GFX90a-NEXT:    .p2align 8
+; GFX90a-NEXT:  ; %bb.2:
+; GFX90a-NEXT:  .LBB33_0:
 ; GFX90a-NEXT:    s_and_b32 s0, s8, 0xff
 ; GFX90a-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX90a-NEXT:    v_mov_b32_e32 v1, s0