[AMDGPU] Emit kernel code properties as code object metadata

[kzhuravl]

Summary

These are not required for low level runtime

[kzhuravl]

I need to add the test.

[arsenm]

LGTM except for the lack of test

[yaxunl]

LGTM with test added

[kzhuravl]

Rebase + add a test case.

[yaxunl]

LGTM. Thanks!

lib/Target/AMDGPU/AMDGPUAsmPrinter.h

 /// \file
 /// \brief AMDGPU Assembly printer class.
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPUASMPRINTER_H
 #define LLVM_LIB_TARGET_AMDGPU_AMDGPUASMPRINTER_H
 
+#include "AMDKernelCodeT.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/CodeGen/AsmPrinter.h"
 #include <cstddef>
 #include <cstdint>
 #include <limits>
 #include <memory>
 #include <string>
 #include <vector>
struct SIProgramInfo {
     // Bonus information for debugging.
     bool VCCUsed = false;
     uint64_t CodeLen = 0;
 
     SIProgramInfo() = default;
   };
 
   void getSIProgramInfo(SIProgramInfo &Out, const MachineFunction &MF) const;
+  void getAmdKernelCode(amd_kernel_code_t &Out, const SIProgramInfo &KernelInfo,
+                        const MachineFunction &MF) const;
   void findNumUsedRegistersSI(const MachineFunction &MF,
                               unsigned &NumSGPR,
                               unsigned &NumVGPR) const;
 
   /// \brief Emit register usage information so that the GPU driver
   /// can correctly setup the GPU state.
   void EmitProgramInfoR600(const MachineFunction &MF);
   void EmitProgramInfoSI(const MachineFunction &MF, const SIProgramInfo &KernelInfo);
-  void EmitAmdKernelCodeT(const MachineFunction &MF,
-                          const SIProgramInfo &KernelInfo) const;
 
 public:
   explicit AMDGPUAsmPrinter(TargetMachine &TM,
                             std::unique_ptr<MCStreamer> Streamer);
 
   StringRef getPassName() const override;
 
   const MCSubtargetInfo* getSTI() const;

lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp

 //===----------------------------------------------------------------------===//
 //
 
 #include "AMDGPUAsmPrinter.h"
 #include "MCTargetDesc/AMDGPUTargetStreamer.h"
 #include "InstPrinter/AMDGPUInstPrinter.h"
 #include "Utils/AMDGPUBaseInfo.h"
 #include "AMDGPU.h"
-#include "AMDKernelCodeT.h"
 #include "AMDGPUSubtarget.h"
 #include "R600Defines.h"
 #include "R600MachineFunctionInfo.h"
 #include "R600RegisterInfo.h"
 #include "SIDefines.h"
 #include "SIMachineFunctionInfo.h"
 #include "SIInstrInfo.h"
 #include "SIRegisterInfo.h"
bool AMDGPUAsmPrinter::isBlockOnlyReachableByFallthrough(
   // the start of the block is needed.  to the start of the block.
   // XXX - Is there a smarter way to check this?
   return (MBB->back().getOpcode() != AMDGPU::S_SETPC_B64);
 }
 
 void AMDGPUAsmPrinter::EmitFunctionBodyStart() {
   const AMDGPUSubtarget &STM = MF->getSubtarget<AMDGPUSubtarget>();
   SIProgramInfo KernelInfo;
+  amd_kernel_code_t KernelCode;
   if (STM.isAmdCodeObjectV2(*MF)) {
     getSIProgramInfo(KernelInfo, *MF);
-    EmitAmdKernelCodeT(*MF, KernelInfo);
+    getAmdKernelCode(KernelCode, KernelInfo, *MF);
+
+    OutStreamer->SwitchSection(getObjFileLowering().getTextSection());
+    getTargetStreamer().EmitAMDKernelCodeT(KernelCode);
   }
 
   if (TM.getTargetTriple().getOS() != Triple::AMDHSA)
     return;
-  getTargetStreamer().EmitKernelCodeObjectMetadata(*MF->getFunction());
+  getTargetStreamer().EmitKernelCodeObjectMetadata(*MF->getFunction(),
+                                                   KernelCode);
 }
 
 void AMDGPUAsmPrinter::EmitFunctionEntryLabel() {
   const SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
   const AMDGPUSubtarget &STM = MF->getSubtarget<AMDGPUSubtarget>();
   if (MFI->isKernel() && STM.isAmdCodeObjectV2(*MF)) {
     SmallString<128> SymbolName;
     getNameWithPrefix(SymbolName, MF->getFunction()),
case 8:
     return AMD_ELEMENT_8_BYTES;
   case 16:
     return AMD_ELEMENT_16_BYTES;
   default:
     llvm_unreachable("invalid private_element_size");
   }
 }
 
-void AMDGPUAsmPrinter::EmitAmdKernelCodeT(const MachineFunction &MF,
-                                         const SIProgramInfo &KernelInfo) const {
+void AMDGPUAsmPrinter::getAmdKernelCode(amd_kernel_code_t &Out,
+                                        const SIProgramInfo &KernelInfo,
+                                        const MachineFunction &MF) const {
   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
   const SISubtarget &STM = MF.getSubtarget<SISubtarget>();
-  amd_kernel_code_t header;
 
-  AMDGPU::initDefaultAMDKernelCodeT(header, STM.getFeatureBits());
+  AMDGPU::initDefaultAMDKernelCodeT(Out, STM.getFeatureBits());
 
-  header.compute_pgm_resource_registers =
+  Out.compute_pgm_resource_registers =
       KernelInfo.ComputePGMRSrc1 |
       (KernelInfo.ComputePGMRSrc2 << 32);
-  header.code_properties = AMD_CODE_PROPERTY_IS_PTR64;
-
+  Out.code_properties = AMD_CODE_PROPERTY_IS_PTR64;
 
-  AMD_HSA_BITS_SET(header.code_properties,
+  AMD_HSA_BITS_SET(Out.code_properties,
                    AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE,
                    getElementByteSizeValue(STM.getMaxPrivateElementSize()));
 
   if (MFI->hasPrivateSegmentBuffer()) {
-    header.code_properties |=
+    Out.code_properties |=
       AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER;
   }
 
   if (MFI->hasDispatchPtr())
-    header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR;
+    Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR;
 
   if (MFI->hasQueuePtr())
-    header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR;
+    Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR;
 
   if (MFI->hasKernargSegmentPtr())
-    header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR;
+    Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR;
 
   if (MFI->hasDispatchID())
-    header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID;
+    Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID;
 
   if (MFI->hasFlatScratchInit())
-    header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT;
-
-  // TODO: Private segment size
+    Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT;
 
   if (MFI->hasGridWorkgroupCountX()) {
-    header.code_properties |=
+    Out.code_properties |=
       AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_X;
   }
 
   if (MFI->hasGridWorkgroupCountY()) {
-    header.code_properties |=
+    Out.code_properties |=
       AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Y;
   }
 
   if (MFI->hasGridWorkgroupCountZ()) {
-    header.code_properties |=
+    Out.code_properties |=
       AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Z;
   }
 
   if (MFI->hasDispatchPtr())
-    header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR;
+    Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR;
 
   if (STM.debuggerSupported())
-    header.code_properties |= AMD_CODE_PROPERTY_IS_DEBUG_SUPPORTED;
+    Out.code_properties |= AMD_CODE_PROPERTY_IS_DEBUG_SUPPORTED;
 
   if (STM.isXNACKEnabled())
-    header.code_properties |= AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED;
+    Out.code_properties |= AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED;
 
   // FIXME: Should use getKernArgSize
-  header.kernarg_segment_byte_size =
+  Out.kernarg_segment_byte_size =
     STM.getKernArgSegmentSize(MF, MFI->getABIArgOffset());
-  header.wavefront_sgpr_count = KernelInfo.NumSGPR;
-  header.workitem_vgpr_count = KernelInfo.NumVGPR;
-  header.workitem_private_segment_byte_size = KernelInfo.ScratchSize;
-  header.workgroup_group_segment_byte_size = KernelInfo.LDSSize;
-  header.reserved_vgpr_first = KernelInfo.ReservedVGPRFirst;
-  header.reserved_vgpr_count = KernelInfo.ReservedVGPRCount;
+  Out.wavefront_sgpr_count = KernelInfo.NumSGPR;
+  Out.workitem_vgpr_count = KernelInfo.NumVGPR;
+  Out.workitem_private_segment_byte_size = KernelInfo.ScratchSize;
+  Out.workgroup_group_segment_byte_size = KernelInfo.LDSSize;
+  Out.reserved_vgpr_first = KernelInfo.ReservedVGPRFirst;
+  Out.reserved_vgpr_count = KernelInfo.ReservedVGPRCount;
 
   // These alignment values are specified in powers of two, so alignment =
   // 2^n.  The minimum alignment is 2^4 = 16.
-  header.kernarg_segment_alignment = std::max((size_t)4,
+  Out.kernarg_segment_alignment = std::max((size_t)4,
       countTrailingZeros(MFI->getMaxKernArgAlign()));
 
   if (STM.debuggerEmitPrologue()) {
-    header.debug_wavefront_private_segment_offset_sgpr =
+    Out.debug_wavefront_private_segment_offset_sgpr =
       KernelInfo.DebuggerWavefrontPrivateSegmentOffsetSGPR;
-    header.debug_private_segment_buffer_sgpr =
+    Out.debug_private_segment_buffer_sgpr =
       KernelInfo.DebuggerPrivateSegmentBufferSGPR;
   }
-
-  OutStreamer->SwitchSection(getObjFileLowering().getTextSection());
-  getTargetStreamer().EmitAMDKernelCodeT(header);
 }
 
 bool AMDGPUAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
                                        unsigned AsmVariant,
                                        const char *ExtraCode, raw_ostream &O) {
   if (ExtraCode && ExtraCode[0]) {
     if (ExtraCode[1] != 0)
       return true; // Unknown modifier.

lib/Target/AMDGPU/MCTargetDesc/AMDGPUCodeObjectMetadata.h

struct Metadata final {
   std::string mTypeName = std::string();
 
   /// \brief Default constructor.
   Metadata() = default;
 };
 
 } // end namespace Arg
 
+//===----------------------------------------------------------------------===//
+// Kernel Code Properties Metadata.
+//===----------------------------------------------------------------------===//
+namespace CodeProps {
+
+namespace Key {
+/// \brief Key for Kernel::CodeProps::mKernargSegmentSize.
+constexpr char KernargSegmentSize[] = "KernargSegmentSize";
+/// \brief Key for Kernel::CodeProps::mWorkgroupGroupSegmentSize.
+constexpr char WorkgroupGroupSegmentSize[] = "WorkgroupGroupSegmentSize";
+/// \brief Key for Kernel::CodeProps::mWorkitemPrivateSegmentSize.
+constexpr char WorkitemPrivateSegmentSize[] = "WorkitemPrivateSegmentSize";
+/// \brief Key for Kernel::CodeProps::mWavefrontNumSGPRs.
+constexpr char WavefrontNumSGPRs[] = "WavefrontNumSGPRs";
+/// \brief Key for Kernel::CodeProps::mWorkitemNumVGPRs.
+constexpr char WorkitemNumVGPRs[] = "WorkitemNumVGPRs";
+/// \brief Key for Kernel::CodeProps::mKernargSegmentAlign.
+constexpr char KernargSegmentAlign[] = "KernargSegmentAlign";
+/// \brief Key for Kernel::CodeProps::mGroupSegmentAlign.
+constexpr char GroupSegmentAlign[] = "GroupSegmentAlign";
+/// \brief Key for Kernel::CodeProps::mPrivateSegmentAlign.
+constexpr char PrivateSegmentAlign[] = "PrivateSegmentAlign";
+/// \brief Key for Kernel::CodeProps::mWavefrontSize.
+constexpr char WavefrontSize[] = "WavefrontSize";
+} // end namespace Key
+
+/// \brief In-memory representation of kernel code properties metadata.
+struct Metadata final {
+  /// \brief Size in bytes of the kernarg segment memory. Kernarg segment memory
+  /// holds the values of the arguments to the kernel. Optional.
+  uint64_t mKernargSegmentSize = 0;
+  /// \brief Size in bytes of the group segment memory required by a workgroup.
+  /// This value does not include any dynamically allocated group segment memory
+  /// that may be added when the kernel is dispatched. Optional.
+  uint32_t mWorkgroupGroupSegmentSize = 0;
+  /// \brief Size in bytes of the private segment memory required by a workitem.
+  /// Private segment memory includes arg, spill and private segments. Optional.
+  uint32_t mWorkitemPrivateSegmentSize = 0;
+  /// \brief Total number of SGPRs used by a wavefront. Optional.
+  uint16_t mWavefrontNumSGPRs = 0;
+  /// \brief Total number of VGPRs used by a workitem. Optional.
+  uint16_t mWorkitemNumVGPRs = 0;
+  /// \brief Maximum byte alignment of variables used by the kernel in the
+  /// kernarg memory segment. Expressed as a power of two. Optional.
+  uint8_t mKernargSegmentAlign = 0;
+  /// \brief Maximum byte alignment of variables used by the kernel in the
+  /// group memory segment. Expressed as a power of two. Optional.
+  uint8_t mGroupSegmentAlign = 0;
+  /// \brief Maximum byte alignment of variables used by the kernel in the
+  /// private memory segment. Expressed as a power of two. Optional.
+  uint8_t mPrivateSegmentAlign = 0;
+  /// \brief Wavefront size. Expressed as a power of two. Optional.
+  uint8_t mWavefrontSize = 0;
+
+  /// \brief Default constructor.
+  Metadata() = default;
+
+  /// \returns True if kernel code properties metadata is empty, false
+  /// otherwise.
+  bool empty() const {
+    return !notEmpty();
+  }
+
+  /// \returns True if kernel code properties metadata is not empty, false
+  /// otherwise.
+  bool notEmpty() const {
+    return mKernargSegmentSize || mWorkgroupGroupSegmentSize ||
+           mWorkitemPrivateSegmentSize || mWavefrontNumSGPRs ||
+           mWorkitemNumVGPRs || mKernargSegmentAlign || mGroupSegmentAlign ||
+           mPrivateSegmentAlign || mWavefrontSize;
+  }
+};
+
+} // end namespace CodeProps
+
 namespace Key {
 /// \brief Key for Kernel::Metadata::mName.
 constexpr char Name[] = "Name";
 /// \brief Key for Kernel::Metadata::mLanguage.
 constexpr char Language[] = "Language";
 /// \brief Key for Kernel::Metadata::mLanguageVersion.
 constexpr char LanguageVersion[] = "LanguageVersion";
 /// \brief Key for Kernel::Metadata::mAttrs.
 constexpr char Attrs[] = "Attrs";
 /// \brief Key for Kernel::Metadata::mArgs.
 constexpr char Args[] = "Args";
+/// \brief Key for Kernel::Metadata::mCodeProps.
+constexpr char CodeProps[] = "CodeProps";
 } // end namespace Key
 
 /// \brief In-memory representation of kernel metadata.
 struct Metadata final {
   /// \brief Name. Required.
   std::string mName = std::string();
   /// \brief Language. Optional.
   std::string mLanguage = std::string();
   /// \brief Language version. Optional.
   std::vector<uint8_t> mLanguageVersion = std::vector<uint8_t>();
   /// \brief Attributes metadata. Optional.
   Attrs::Metadata mAttrs = Attrs::Metadata();
   /// \brief Arguments metadata. Optional.
   std::vector<Arg::Metadata> mArgs = std::vector<Arg::Metadata>();
+  /// \brief Code properties metadata. Optional.
+  CodeProps::Metadata mCodeProps = CodeProps::Metadata();
 
   /// \brief Default constructor.
   Metadata() = default;
 };
 
 } // end namespace Kernel
 
 namespace Key {

lib/Target/AMDGPU/MCTargetDesc/AMDGPUCodeObjectMetadataStreamer.h

 ///
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_LIB_TARGET_AMDGPU_MCTARGETDESC_AMDGPUCODEOBJECTMETADATASTREAMER_H
 #define LLVM_LIB_TARGET_AMDGPU_MCTARGETDESC_AMDGPUCODEOBJECTMETADATASTREAMER_H
 
 #include "AMDGPUCodeObjectMetadata.h"
+#include "AMDKernelCodeT.h"
 #include "llvm/ADT/StringRef.h"
 
 namespace llvm {
 
 class Argument;
 class DataLayout;
 class FeatureBitset;
 class Function;
private:
   void emitKernelArgs(const Function &Func);
 
   void emitKernelArg(const Argument &Arg);
 
   void emitKernelArg(const DataLayout &DL, Type *Ty, ValueKind ValueKind,
                      StringRef TypeQual = "", StringRef BaseTypeName = "",
                      StringRef AccQual = "", StringRef Name = "",
                      StringRef TypeName = "");
+
+  void emitKernelCodeProps(const amd_kernel_code_t &KernelCode);
+
 public:
   MetadataStreamer() = default;
   ~MetadataStreamer() = default;
 
   void begin(const FeatureBitset &Features, const Module &Mod);
 
   void end() {}
 
-  void emitKernel(const Function &Func);
+  void emitKernel(const Function &Func, const amd_kernel_code_t &KernelCode);
 
   ErrorOr<std::string> toYamlString();
 
   ErrorOr<std::string> toYamlString(const FeatureBitset &Features,
                                     StringRef YamlString);
 };
 
 } // end namespace CodeObject
 } // end namespace AMDGPU
 } // end namespace llvm
 
 #endif // LLVM_LIB_TARGET_AMDGPU_MCTARGETDESC_AMDGPUCODEOBJECTMETADATASTREAMER_H

lib/Target/AMDGPU/MCTargetDesc/AMDGPUCodeObjectMetadataStreamer.cpp

static void mapping(IO &YIO, Kernel::Arg::Metadata &MD) {
     YIO.mapOptional(Kernel::Arg::Key::IsRestrict, MD.mIsRestrict, false);
     YIO.mapOptional(Kernel::Arg::Key::IsVolatile, MD.mIsVolatile, false);
     YIO.mapOptional(Kernel::Arg::Key::Name, MD.mName, std::string());
     YIO.mapOptional(Kernel::Arg::Key::TypeName, MD.mTypeName, std::string());
   }
 };
 
 template <>
+struct MappingTraits<Kernel::CodeProps::Metadata> {
+  static void mapping(IO &YIO, Kernel::CodeProps::Metadata &MD) {
+    YIO.mapOptional(Kernel::CodeProps::Key::KernargSegmentSize,
+                    MD.mKernargSegmentSize, uint64_t(0));
+    YIO.mapOptional(Kernel::CodeProps::Key::WorkgroupGroupSegmentSize,
+                    MD.mWorkgroupGroupSegmentSize, uint32_t(0));
+    YIO.mapOptional(Kernel::CodeProps::Key::WorkitemPrivateSegmentSize,
+                    MD.mWorkitemPrivateSegmentSize, uint32_t(0));
+    YIO.mapOptional(Kernel::CodeProps::Key::WavefrontNumSGPRs,
+                    MD.mWavefrontNumSGPRs, uint16_t(0));
+    YIO.mapOptional(Kernel::CodeProps::Key::WorkitemNumVGPRs,
+                    MD.mWorkitemNumVGPRs, uint16_t(0));
+    YIO.mapOptional(Kernel::CodeProps::Key::KernargSegmentAlign,
+                    MD.mKernargSegmentAlign, uint8_t(0));
+    YIO.mapOptional(Kernel::CodeProps::Key::GroupSegmentAlign,
+                    MD.mGroupSegmentAlign, uint8_t(0));
+    YIO.mapOptional(Kernel::CodeProps::Key::PrivateSegmentAlign,
+                    MD.mPrivateSegmentAlign, uint8_t(0));
+    YIO.mapOptional(Kernel::CodeProps::Key::WavefrontSize,
+                    MD.mWavefrontSize, uint8_t(0));
+  }
+};
+
+template <>
 struct MappingTraits<Kernel::Metadata> {
   static void mapping(IO &YIO, Kernel::Metadata &MD) {
     YIO.mapRequired(Kernel::Key::Name, MD.mName);
     YIO.mapOptional(Kernel::Key::Language, MD.mLanguage, std::string());
     YIO.mapOptional(Kernel::Key::LanguageVersion, MD.mLanguageVersion,
                     std::vector<uint8_t>());
     if (!MD.mAttrs.empty() || !YIO.outputting())
       YIO.mapOptional(Kernel::Key::Attrs, MD.mAttrs);
     if (!MD.mArgs.empty() || !YIO.outputting())
       YIO.mapOptional(Kernel::Key::Args, MD.mArgs);
+    if (!MD.mCodeProps.empty() || !YIO.outputting())
+      YIO.mapOptional(Kernel::Key::CodeProps, MD.mCodeProps);
   }
 };
 
 template <>
 struct MappingTraits<CodeObject::Metadata> {
   static void mapping(IO &YIO, CodeObject::Metadata &MD) {
     YIO.mapRequired(Key::Version, MD.mVersion);
     YIO.mapOptional(Key::Isa, MD.mIsa);
for (StringRef Key : SplitTypeQuals) {
     if (P)
       *P = true;
   }
 
   Arg.mName = Name;
   Arg.mTypeName = TypeName;
 }
 
+void MetadataStreamer::emitKernelCodeProps(
+    const amd_kernel_code_t &KernelCode) {
+  auto &CodeProps = CodeObjectMetadata.mKernels.back().mCodeProps;
+
+  CodeProps.mKernargSegmentSize = KernelCode.kernarg_segment_byte_size;
+  CodeProps.mWorkgroupGroupSegmentSize =
+      KernelCode.workgroup_group_segment_byte_size;
+  CodeProps.mWorkitemPrivateSegmentSize =
+      KernelCode.workitem_private_segment_byte_size;
+  CodeProps.mWavefrontNumSGPRs = KernelCode.wavefront_sgpr_count;
+  CodeProps.mWorkitemNumVGPRs = KernelCode.workitem_vgpr_count;
+  CodeProps.mKernargSegmentAlign = KernelCode.kernarg_segment_alignment;
+  CodeProps.mGroupSegmentAlign = KernelCode.group_segment_alignment;
+  CodeProps.mPrivateSegmentAlign = KernelCode.private_segment_alignment;
+  CodeProps.mWavefrontSize = KernelCode.wavefront_size;
+}
+
 void MetadataStreamer::begin(const FeatureBitset &Features, const Module &Mod) {
   emitVersion();
   emitIsa(Features);
   emitPrintf(Mod);
 }
 
-void MetadataStreamer::emitKernel(const Function &Func) {
+void MetadataStreamer::emitKernel(const Function &Func,
+                                  const amd_kernel_code_t &KernelCode) {
   if (Func.getCallingConv() != CallingConv::AMDGPU_KERNEL)
     return;
 
   CodeObjectMetadata.mKernels.push_back(Kernel::Metadata());
   auto &Kernel = CodeObjectMetadata.mKernels.back();
 
   Kernel.mName = Func.getName();
   emitKernelLanguage(Func);
   emitKernelAttrs(Func);
   emitKernelArgs(Func);
+  emitKernelCodeProps(KernelCode);
 }
 
 ErrorOr<std::string> MetadataStreamer::toYamlString() {
   auto YamlString = Metadata::toYamlString(CodeObjectMetadata);
   if (!YamlString)
     return YamlString;
 
   if (DumpCodeObjectMetadata)

lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.h

public:
 
   virtual void EmitAMDGPUHsaModuleScopeGlobal(StringRef GlobalName) = 0;
 
   virtual void EmitAMDGPUHsaProgramScopeGlobal(StringRef GlobalName) = 0;
 
   virtual void EmitStartOfCodeObjectMetadata(const FeatureBitset &Features,
                                              const Module &Mod);
 
-  virtual void EmitKernelCodeObjectMetadata(const Function &Func);
+  virtual void EmitKernelCodeObjectMetadata(
+      const Function &Func, const amd_kernel_code_t &KernelCode);
 
   virtual void EmitEndOfCodeObjectMetadata(const FeatureBitset &Features);
 
   /// \returns True on success, false on failure.
   virtual bool EmitCodeObjectMetadata(const FeatureBitset &Features,
                                       StringRef YamlString) = 0;
 };
 

lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp

 AMDGPUTargetStreamer::AMDGPUTargetStreamer(MCStreamer &S)
     : MCTargetStreamer(S) {}
 
 void AMDGPUTargetStreamer::EmitStartOfCodeObjectMetadata(
     const FeatureBitset &Features, const Module &Mod) {
   CodeObjectMetadataStreamer.begin(Features, Mod);
 }
 
-void AMDGPUTargetStreamer::EmitKernelCodeObjectMetadata(const Function &Func) {
-  CodeObjectMetadataStreamer.emitKernel(Func);
+void AMDGPUTargetStreamer::EmitKernelCodeObjectMetadata(
+    const Function &Func, const amd_kernel_code_t &KernelCode) {
+  CodeObjectMetadataStreamer.emitKernel(Func, KernelCode);
 }
 
 void AMDGPUTargetStreamer::EmitEndOfCodeObjectMetadata(
     const FeatureBitset &Features) {
   CodeObjectMetadataStreamer.end();
   EmitCodeObjectMetadata(Features,
                          CodeObjectMetadataStreamer.toYamlString().get());
 }

test/CodeGen/AMDGPU/code-object-metadata-from-llvm-ir-full.ll

 !94 = !{!"", !"", !"", !"", !"", !"", !""}
 !100 = !{!"1:1:4:%d\5Cn"}
 !101 = !{!"2:1:8:%g\5Cn"}
 
 ; NOTES: Displaying notes found at file offset 0x{{[0-9]+}}
 ; NOTES-NEXT: Owner    Data size    Description
 ; NOTES-NEXT: AMD      0x00000008   Unknown note type: (0x00000001)
 ; NOTES-NEXT: AMD      0x0000001b   Unknown note type: (0x00000003)
-; GFX700:     AMD      0x00007834   Unknown note type: (0x0000000a)
-; GFX800:     AMD      0x00007834   Unknown note type: (0x0000000a)
-; GFX900:     AMD      0x00007835   Unknown note type: (0x0000000a)
+; GFX700:     AMD      0x0000920c   Unknown note type: (0x0000000a)
+; GFX800:     AMD      0x0000922b   Unknown note type: (0x0000000a)
+; GFX900:     AMD      0x0000920d   Unknown note type: (0x0000000a)
 
 ; PARSER: AMDGPU Code Object Metadata Parser Test: PASS

test/CodeGen/AMDGPU/code-object-metadata-kernel-code-props.ll

+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -filetype=obj -o - < %s | llvm-readobj -amdgpu-code-object-metadata -elf-output-style=GNU -notes | FileCheck --check-prefix=CHECK --check-prefix=GFX700 --check-prefix=NOTES %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx800 -filetype=obj -o - < %s | llvm-readobj -amdgpu-code-object-metadata -elf-output-style=GNU -notes | FileCheck --check-prefix=CHECK --check-prefix=GFX800 --check-prefix=NOTES %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -filetype=obj -o - < %s | llvm-readobj -amdgpu-code-object-metadata -elf-output-style=GNU -notes | FileCheck --check-prefix=CHECK --check-prefix=GFX900 --check-prefix=NOTES %s
+
+; CHECK: ---
+; CHECK:  Version: [ 1, 0 ]
+
+; CHECK:  Kernels:
+; CHECK:    - Name: test
+; CHECK:      CodeProps:
+; CHECK:        KernargSegmentSize:  24
+; GFX700:       WavefrontNumSGPRs:   6
+; GFX800:       WavefrontNumSGPRs:   96
+; GFX900:       WavefrontNumSGPRs:   6
+; GFX700:       WorkitemNumVGPRs:    4
+; GFX800:       WorkitemNumVGPRs:    6
+; GFX900:       WorkitemNumVGPRs:    6
+; CHECK:        KernargSegmentAlign: 4
+; CHECK:        GroupSegmentAlign:   4
+; CHECK:        PrivateSegmentAlign: 4
+; CHECK:        WavefrontSize:       6
+define amdgpu_kernel void @test(
+    half addrspace(1)* %r,
+    half addrspace(1)* %a,
+    half addrspace(1)* %b) {
+entry:
+  %a.val = load half, half addrspace(1)* %a
+  %b.val = load half, half addrspace(1)* %b
+  %r.val = fadd half %a.val, %b.val
+  store half %r.val, half addrspace(1)* %r
+  ret void
+}

test/MC/AMDGPU/code-object-metadata-kernel-code-props.s

+// RUN: llvm-mc -triple=amdgcn-amd-amdhsa -mcpu=gfx700 -show-encoding %s | FileCheck --check-prefix=CHECK --check-prefix=GFX700 %s
+// RUN: llvm-mc -triple=amdgcn-amd-amdhsa -mcpu=gfx800 -show-encoding %s | FileCheck --check-prefix=CHECK --check-prefix=GFX800 %s
+// RUN: llvm-mc -triple=amdgcn-amd-amdhsa -mcpu=gfx900 -show-encoding %s | FileCheck --check-prefix=CHECK --check-prefix=GFX900 %s
+
+// CHECK:  .amdgpu_code_object_metadata
+// CHECK:    Version: [ 1, 0 ]
+// CHECK:    Kernels:
+// CHECK:      - Name: test_kernel
+// CHECK:        CodeProps:
+// CHECK:        KernargSegmentSize:         24
+// CHECK:        WorkitemPrivateSegmentSize: 16
+// CHECK:        WavefrontNumSGPRs:          6
+// CHECK:        WorkitemNumVGPRs:           12
+.amdgpu_code_object_metadata
+  Version: [ 1, 0 ]
+  Printf: [ '1:1:4:%d\n', '2:1:8:%g\n' ]
+  Kernels:
+    - Name:            test_kernel
+      CodeProps:
+        KernargSegmentSize:         24
+        WorkitemPrivateSegmentSize: 16
+        WavefrontNumSGPRs:          6
+        WorkitemNumVGPRs:           12
+.end_amdgpu_code_object_metadata