[AMDGPU] Added MsgPack format PAL metadata

[tpr]

Summary

PAL metadata now supports both the old linear reg=val pairs format and
the new MsgPack format.

The MsgPack format uses YAML as its textual representation. On output to
YAML, a mnemonic name is provided for some hardware registers.

Change-Id: I2bbaabaaca4b3574f7e03b80fbef7c7a69d06a94

[scott.linder]

lib/Target/AMDGPU/Utils/AMDGPUPALMetadata.cpp

52–57

After thinking more about this and talking to @t-tye and others, I think it might be worth considering other ways to convey metadata through IR. We might also be able to avoid the issue of the front-end needing to make calls into the back-end if we just agree on a scheme and document it; if there is a contract between the front-end and back-end then you can safely put all of the code in the AMDGPU back-end without the private-headers issue.

I'm not sure I like the non-human-readable-blob approach. Since we have the YAML-Msgpack interop could this be a YAML string? I think this is something we would like to incorporate for HSA as well, and one requirement would be that it is readable and editable in IR.

[tpr]

lib/Target/AMDGPU/Utils/AMDGPUPALMetadata.cpp

52–57

Are you thinking we'd put the YAML string into an IR metadata string? I was thinking about that for a completely different application, but it looks to me like an IR string containing multiple lines just gets jammed into a single line with escapes for the newlines, so it is still not very editable.

Where we (in the PAL metadata world) have a requirement for editability is when the ELF is disassembled and reassembled, and that works fine with these changes because the PAL metadata is disassembled as a YAML block of text.

[tpr]

lib/Target/AMDGPU/Utils/AMDGPUPALMetadata.cpp

52–57

Just to add to this: What swayed me to use a binary blob of msgpack:

1. This is used in an online compiler in a graphics device driver, so that made me think that a textual format, or something more complicated, is not suitable.
2. The contract is that the frontend provides this msgpack document, and LLVM mostly does not have to understand it, except that LLVM adds to it, and in a few cases ORs a value into an existing int value in the document tree if there is already one there. So LLVM does not fully understand the metadata part of the PAL ABI, only the bits it needs to add to.

[tpr]

V2: Like the previous commit, no longer supports a method for LLPC to

call to write the PAL metadata into IR metadata. The plan now is
that LLPC will use MsgPackDocument and put the msgpack binary blob
into IR metadata itself.

[tpr]

V3: Rebased, and addressed "single char in single quotes" review comment from the other change.

[scott.linder]

include/llvm/BinaryFormat/ELF.h

1369 ↗

(On Diff #191256)

I thought the goal was to use NT_AMDGPU_METADATA = 32 for both HSA and PAL when we made the move to MsgPack; can we make this switch now?

[tpr]

V4: Generate vendor=AMDGPU, type=32 for the metadata note record, as now expected by PAL.

[scott.linder]

LGTM, thanks! One small change to a comment.

lib/Target/AMDGPU/Utils/AMDGPUPALMetadata.h

95

Can we drop ELF::NT_AMD_AMDGPU_PAL_METADATA_MSGPACK here?

include/llvm/Support/AMDGPUMetadata.h

 
 } // end namespace HSAMD
 
 //===----------------------------------------------------------------------===//
 // PAL metadata.
 //===----------------------------------------------------------------------===//
 namespace PALMD {
 
-/// PAL metadata assembler directive.
+/// PAL metadata (old linear format) assembler directive.
 constexpr char AssemblerDirective[] = ".amd_amdgpu_pal_metadata";
 
+/// PAL metadata (new MsgPack format) beginning assembler directive.
+constexpr char AssemblerDirectiveBegin[] = ".amdgpu_pal_metadata";
+
+/// PAL metadata (new MsgPack format) ending assembler directive.
+constexpr char AssemblerDirectiveEnd[] = ".end_amdgpu_pal_metadata";
+
 /// PAL metadata keys.
 enum Key : uint32_t {
   R_2E12_COMPUTE_PGM_RSRC1 = 0x2e12,
   R_2D4A_SPI_SHADER_PGM_RSRC1_LS = 0x2d4a,
   R_2D0A_SPI_SHADER_PGM_RSRC1_HS = 0x2d0a,
   R_2CCA_SPI_SHADER_PGM_RSRC1_ES = 0x2cca,
   R_2C8A_SPI_SHADER_PGM_RSRC1_GS = 0x2c8a,
   R_2C4A_SPI_SHADER_PGM_RSRC1_VS = 0x2c4a,

lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp

 // frontend as LLVM metadata. Once all functions are written, the PAL metadata
 // is then written as a single block in the .note section.
 void AMDGPUAsmPrinter::EmitPALMetadata(const MachineFunction &MF,
        const SIProgramInfo &CurrentProgramInfo) {
   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
   auto CC = MF.getFunction().getCallingConv();
   auto MD = getTargetStreamer()->getPALMetadata();
 
+  MD->setEntryPoint(CC, MF.getFunction().getName());
   MD->setNumUsedVgprs(CC, CurrentProgramInfo.NumVGPRsForWavesPerEU);
   MD->setNumUsedSgprs(CC, CurrentProgramInfo.NumSGPRsForWavesPerEU);
   if (AMDGPU::isCompute(MF.getFunction().getCallingConv())) {
     MD->setRsrc1(CC, CurrentProgramInfo.ComputePGMRSrc1);
     MD->setRsrc2(CC, CurrentProgramInfo.ComputePGMRSrc2);
   } else {
     MD->setRsrc1(CC, S_00B028_VGPRS(CurrentProgramInfo.VGPRBlocks) |
         S_00B028_SGPRS(CurrentProgramInfo.SGPRBlocks));

lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp

private:
   bool ParseDirectiveHSACodeObjectISA();
   bool ParseAMDKernelCodeTValue(StringRef ID, amd_kernel_code_t &Header);
   bool ParseDirectiveAMDKernelCodeT();
   bool subtargetHasRegister(const MCRegisterInfo &MRI, unsigned RegNo) const;
   bool ParseDirectiveAMDGPUHsaKernel();
 
   bool ParseDirectiveISAVersion();
   bool ParseDirectiveHSAMetadata();
+  bool ParseDirectivePALMetadataBegin();
   bool ParseDirectivePALMetadata();
 
+  /// Common code to parse out a block of text (typically YAML) between start and
+  /// end directives.
+  bool ParseToEndDirective(const char *AssemblerDirectiveBegin,
+                           const char *AssemblerDirectiveEnd,
+                           std::string &CollectString);
+
   bool AddNextRegisterToList(unsigned& Reg, unsigned& RegWidth,
                              RegisterKind RegKind, unsigned Reg1,
                              unsigned RegNum);
   bool ParseAMDGPURegister(RegisterKind& RegKind, unsigned& Reg,
                            unsigned& RegNum, unsigned& RegWidth,
                            unsigned *DwordRegIndex);
   Optional<StringRef> getGprCountSymbolName(RegisterKind RegKind);
   void initializeGprCountSymbol(RegisterKind RegKind);
bool AMDGPUAsmParser::ParseDirectiveHSAMetadata() {
 
   if (getSTI().getTargetTriple().getOS() != Triple::AMDHSA) {
     return Error(getParser().getTok().getLoc(),
                  (Twine(AssemblerDirectiveBegin) + Twine(" directive is "
                  "not available on non-amdhsa OSes")).str());
   }
 
   std::string HSAMetadataString;
-  raw_string_ostream YamlStream(HSAMetadataString);
+  if (ParseToEndDirective(AssemblerDirectiveBegin, AssemblerDirectiveEnd,
+                          HSAMetadataString))
+    return true;
+
+  if (IsaInfo::hasCodeObjectV3(&getSTI())) {
+    if (!getTargetStreamer().EmitHSAMetadataV3(HSAMetadataString))
+      return Error(getParser().getTok().getLoc(), "invalid HSA metadata");
+  } else {
+    if (!getTargetStreamer().EmitHSAMetadataV2(HSAMetadataString))
+      return Error(getParser().getTok().getLoc(), "invalid HSA metadata");
+  }
+
+  return false;
+}
+
+/// Common code to parse out a block of text (typically YAML) between start and
+/// end directives.
+bool AMDGPUAsmParser::ParseToEndDirective(const char *AssemblerDirectiveBegin,
+                                          const char *AssemblerDirectiveEnd,
+                                          std::string &CollectString) {
+
+  raw_string_ostream CollectStream(CollectString);
 
   getLexer().setSkipSpace(false);
 
   bool FoundEnd = false;
   while (!getLexer().is(AsmToken::Eof)) {
     while (getLexer().is(AsmToken::Space)) {
-      YamlStream << getLexer().getTok().getString();
+      CollectStream << getLexer().getTok().getString();
       Lex();
     }
 
     if (getLexer().is(AsmToken::Identifier)) {
       StringRef ID = getLexer().getTok().getIdentifier();
       if (ID == AssemblerDirectiveEnd) {
         Lex();
         FoundEnd = true;
         break;
       }
     }
 
-    YamlStream << Parser.parseStringToEndOfStatement()
+    CollectStream << Parser.parseStringToEndOfStatement()
                   << getContext().getAsmInfo()->getSeparatorString();
 
     Parser.eatToEndOfStatement();
   }
 
   getLexer().setSkipSpace(true);
 
   if (getLexer().is(AsmToken::Eof) && !FoundEnd) {
     return TokError(Twine("expected directive ") +
-                    Twine(HSAMD::AssemblerDirectiveEnd) + Twine(" not found"));
+                    Twine(AssemblerDirectiveEnd) + Twine(" not found"));
   }
 
-  YamlStream.flush();
-
-  if (IsaInfo::hasCodeObjectV3(&getSTI())) {
-    if (!getTargetStreamer().EmitHSAMetadataV3(HSAMetadataString))
-      return Error(getParser().getTok().getLoc(), "invalid HSA metadata");
-  } else {
-    if (!getTargetStreamer().EmitHSAMetadataV2(HSAMetadataString))
-      return Error(getParser().getTok().getLoc(), "invalid HSA metadata");
-  }
+  CollectStream.flush();
+  return false;
 
+/// Parse the assembler directive for new MsgPack-format PAL metadata.
+bool AMDGPUAsmParser::ParseDirectivePALMetadataBegin() {
+  std::string String;
+  if (ParseToEndDirective(AMDGPU::PALMD::AssemblerDirectiveBegin,
+                          AMDGPU::PALMD::AssemblerDirectiveEnd, String))
+    return true;
+
+  auto PALMetadata = getTargetStreamer().getPALMetadata();
+  if (!PALMetadata->setFromString(String))
+    return Error(getParser().getTok().getLoc(), "invalid PAL metadata");
   return false;
 }
 
+/// Parse the assembler directive for old linear-format PAL metadata.
 bool AMDGPUAsmParser::ParseDirectivePALMetadata() {
   if (getSTI().getTargetTriple().getOS() != Triple::AMDPAL) {
     return Error(getParser().getTok().getLoc(),
                  (Twine(PALMD::AssemblerDirective) + Twine(" directive is "
                  "not available on non-amdpal OSes")).str());
   }
 
   auto PALMetadata = getTargetStreamer().getPALMetadata();
if (AMDGPU::IsaInfo::hasCodeObjectV3(&getSTI())) {
 
     if (IDVal == ".amd_amdgpu_isa")
       return ParseDirectiveISAVersion();
 
     if (IDVal == AMDGPU::HSAMD::AssemblerDirectiveBegin)
       return ParseDirectiveHSAMetadata();
   }
 
+  if (IDVal == PALMD::AssemblerDirectiveBegin)
+    return ParseDirectivePALMetadataBegin();
+
   if (IDVal == PALMD::AssemblerDirective)
     return ParseDirectivePALMetadata();
 
   return true;
 }
 
 bool AMDGPUAsmParser::subtargetHasRegister(const MCRegisterInfo &MRI,
                                            unsigned RegNo) const {

lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp

 MCELFStreamer &AMDGPUTargetELFStreamer::getStreamer() {
   return static_cast<MCELFStreamer &>(Streamer);
 }
 
 // A hook for emitting stuff at the end.
 // We use it for emitting the accumulated PAL metadata as a .note record.
 void AMDGPUTargetELFStreamer::finish() {
   std::string Blob;
-  unsigned Type = ELF::NT_AMD_AMDGPU_PAL_METADATA;
+  const char *Vendor = getPALMetadata()->getVendor();
+  unsigned Type = getPALMetadata()->getType();
   getPALMetadata()->toBlob(Type, Blob);
   if (Blob.empty())
     return;
-  EmitNote(ElfNote::NoteNameV2,
-           MCConstantExpr::create(Blob.size(), getContext()), Type,
+  EmitNote(Vendor, MCConstantExpr::create(Blob.size(), getContext()), Type,
            [&](MCELFStreamer &OS) { OS.EmitBytes(Blob); });
 }
 
 void AMDGPUTargetELFStreamer::EmitNote(
     StringRef Name, const MCExpr *DescSZ, unsigned NoteType,
     function_ref<void(MCELFStreamer &)> EmitDesc) {
   auto &S = getStreamer();
   auto &Context = S.getContext();

lib/Target/AMDGPU/Utils/AMDGPUPALMetadata.h

 /// PAL metadata handling
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPUPALMETADATA_H
 #define LLVM_LIB_TARGET_AMDGPU_AMDGPUPALMETADATA_H
 
 #include "llvm/ADT/StringRef.h"
+#include "llvm/BinaryFormat/MsgPackDocument.h"
 #include <map>
 
 namespace llvm {
 
 class AMDGPUTargetStreamer;
 class formatted_raw_ostream;
 class MCStreamer;
 class Module;
 
 class AMDGPUPALMetadata {
-  std::map<uint32_t, uint32_t> Registers;
+  unsigned BlobType = 0;
+  msgpack::Document MsgPackDoc;
+  msgpack::DocNode Registers;
+  msgpack::DocNode HwStages;
 
 public:
   // Read the amdgpu.pal.metadata supplied by the frontend, ready for
   // per-function modification.
   void readFromIR(Module &M);
 
   // Set PAL metadata from a binary blob from the applicable .note record.
   // Returns false if bad format.  Blob must remain valid for the lifetime of
public:
 
   // Get a register from the metadata, or 0 if not currently set.
   unsigned getRegister(unsigned Reg);
 
   // Set a register in the metadata.
   // In fact this ORs the value into any previous setting of the register.
   void setRegister(unsigned Reg, unsigned Val);
 
+  // Set the entry point name for one shader.
+  void setEntryPoint(unsigned CC, StringRef Name);
+
   // Set the number of used vgprs in the metadata. This is an optional advisory
   // record for logging etc; wave dispatch actually uses the rsrc1 register for
   // the shader stage to determine the number of vgprs to allocate.
   void setNumUsedVgprs(unsigned CC, unsigned Val);
 
   // Set the number of used sgprs in the metadata. This is an optional advisory
   // record for logging etc; wave dispatch actually uses the rsrc1 register for
   // the shader stage to determine the number of sgprs to allocate.
   void setNumUsedSgprs(unsigned CC, unsigned Val);
 
   // Set the scratch size in the metadata.
   void setScratchSize(unsigned CC, unsigned Val);
 
-  // Emit the accumulated PAL metadata as an asm directive.
+  // Emit the accumulated PAL metadata as asm directives.
   // This is called from AMDGPUTargetAsmStreamer::Finish().
   void toString(std::string &S);
 
+  // Set PAL metadata from YAML text.
+  bool setFromString(StringRef S);
+
+  // Get .note record vendor name of metadata blob to be emitted.
+  const char *getVendor() const;
+
+  // Get .note record type of metadata blob to be emitted:
+  // ELF::NT_AMD_AMDGPU_PAL_METADATA (legacy key=val format), or
+  // ELF::NT_AMD_AMDGPU_PAL_METADATA_MSGPACK or ELF::NT_AMDGPU_METADATA
+  // (MsgPack format).
+  unsigned getType() const;
+
   // Emit the accumulated PAL metadata as a binary blob.
   // This is called from AMDGPUTargetELFStreamer::Finish().
   void toBlob(unsigned Type, std::string &S);
+
+  // Get the msgpack::Document for the PAL metadata.
+  msgpack::Document *getMsgPackDoc() { return &MsgPackDoc; }
+
+private:
+  // Return whether the blob type is legacy PAL metadata.
+  bool isLegacy() const;
+
+  // Reference (create if necessary) the node for the registers map.
+  msgpack::DocNode &refRegisters();
+
+  // Get (create if necessary) the registers map.
+  msgpack::MapDocNode getRegisters();
+
+  // Get (create if necessary) the .hardware_stages entry for the given calling
+  // convention.
+  msgpack::MapDocNode getHwStage(unsigned CC);
+
+  bool setFromLegacyBlob(StringRef Blob);
+  bool setFromMsgPackBlob(StringRef Blob);
+  void toLegacyBlob(std::string &Blob);
+  void toMsgPackBlob(std::string &Blob);
 };
 
 } // end namespace llvm
 
 #endif // LLVM_LIB_TARGET_AMDGPU_AMDGPUPALMETADATA_H

lib/Target/AMDGPU/Utils/AMDGPUPALMetadata.cpp

 #include "llvm/BinaryFormat/ELF.h"
 #include "llvm/IR/CallingConv.h"
 #include "llvm/Support/AMDGPUMetadata.h"
 #include "llvm/Support/EndianStream.h"
 
 using namespace llvm;
 using namespace llvm::AMDGPU;
 
-// Read the amdgpu.pal.metadata supplied by the
-// frontend into our Registers, ready for per-function modification.  It
-// is a NamedMD containing an MDTuple containing a number of MDNodes each of
-// which is an integer value, and each two integer values forms a key=value
-// pair that we store as Registers[key]=value in the map.
+// Read the PAL metadata from IR metadata, where it was put by the frontend.
 void AMDGPUPALMetadata::readFromIR(Module &M) {
-  auto NamedMD = M.getNamedMetadata("amdgpu.pal.metadata");
+  auto NamedMD = M.getNamedMetadata("amdgpu.pal.metadata.msgpack");
+  if (NamedMD && NamedMD->getNumOperands()) {
+    // This is the new msgpack format for metadata. It is a NamedMD containing
+    // an MDTuple containing an MDString containing the msgpack data.
+    BlobType = ELF::NT_AMDGPU_METADATA;
+    auto MDN = dyn_cast<MDTuple>(NamedMD->getOperand(0));
+    if (MDN && MDN->getNumOperands()) {
+      if (auto MDS = dyn_cast<MDString>(MDN->getOperand(0)))
+        setFromMsgPackBlob(MDS->getString());
+    }
+    return;
+  }
+  BlobType = ELF::NT_AMD_AMDGPU_PAL_METADATA;
+  NamedMD = M.getNamedMetadata("amdgpu.pal.metadata");
   if (!NamedMD || !NamedMD->getNumOperands())
     return;
+  // This is the old reg=value pair format for metadata. It is a NamedMD
+  // containing an MDTuple containing a number of MDNodes each of which is an
+  // integer value, and each two integer values forms a key=value pair that we
+  // store as Registers[key]=value in the map.
   auto Tuple = dyn_cast<MDTuple>(NamedMD->getOperand(0));
   if (!Tuple)
     return;
   for (unsigned I = 0, E = Tuple->getNumOperands() & -2; I != E; I += 2) {
     auto Key = mdconst::dyn_extract<ConstantInt>(Tuple->getOperand(I));
     auto Val = mdconst::dyn_extract<ConstantInt>(Tuple->getOperand(I + 1));
     if (!Key || !Val)
       continue;
-    Registers[Key->getZExtValue()] = Val->getZExtValue();
+    setRegister(Key->getZExtValue(), Val->getZExtValue());
   }
 }
 
 // Set PAL metadata from a binary blob from the applicable .note record.
 // Returns false if bad format.  Blob must remain valid for the lifetime of the
 // Metadata.
 bool AMDGPUPALMetadata::setFromBlob(unsigned Type, StringRef Blob) {
-  assert(Type == ELF::NT_AMD_AMDGPU_PAL_METADATA);
+  BlobType = Type;
+  if (Type == ELF::NT_AMD_AMDGPU_PAL_METADATA)
+    return setFromLegacyBlob(Blob);
+  return setFromMsgPackBlob(Blob);
+}
+
+// Set PAL metadata from legacy (array of key=value pairs) blob.
+bool AMDGPUPALMetadata::setFromLegacyBlob(StringRef Blob) {
   auto Data = reinterpret_cast<const uint32_t *>(Blob.data());
   for (unsigned I = 0; I != Blob.size() / sizeof(uint32_t) / 2; ++I)
     setRegister(Data[I * 2], Data[I * 2 + 1]);
   return true;
 }
 
+// Set PAL metadata from msgpack blob.
+bool AMDGPUPALMetadata::setFromMsgPackBlob(StringRef Blob) {
+  msgpack::Reader Reader(Blob);
+  return MsgPackDoc.readFromBlob(Blob, /*Multi=*/false);
+}
+
 // Given the calling convention, calculate the register number for rsrc1. In
 // principle the register number could change in future hardware, but we know
 // it is the same for gfx6-9 (except that LS and ES don't exist on gfx9), so
 // we can use fixed values.
 static unsigned getRsrc1Reg(CallingConv::ID CC) {
   switch (CC) {
   default:
     return PALMD::R_2E12_COMPUTE_PGM_RSRC1;

 
 // Set the SPI_PS_INPUT_ADDR register in the metadata.
 // In fact this ORs the value into any previous setting of the register.
 void AMDGPUPALMetadata::setSpiPsInputAddr(unsigned Val) {
   setRegister(PALMD::R_A1B4_SPI_PS_INPUT_ADDR, Val);
 }
 
 // Get a register from the metadata, or 0 if not currently set.
-unsigned AMDGPUPALMetadata::getRegister(unsigned Reg) { return Registers[Reg]; }
+unsigned AMDGPUPALMetadata::getRegister(unsigned Reg) {
+  auto Regs = getRegisters();
+  auto It = Regs.find(MsgPackDoc.getNode(Reg));
+  if (It == Regs.end())
+    return 0;
+  auto N = It->second;
+  if (N.getKind() != msgpack::Type::UInt)
+    return 0;
+  return N.getUInt();
+}
 
 // Set a register in the metadata.
 // In fact this ORs the value into any previous setting of the register.
 void AMDGPUPALMetadata::setRegister(unsigned Reg, unsigned Val) {
-  Registers[Reg] |= Val;
+  if (!isLegacy()) {
+    // In the new MsgPack format, ignore register numbered >= 0x10000000. It
+    // is a PAL ABI pseudo-register in the old non-MsgPack format.
+    if (Reg >= 0x10000000)
+      return;
+  }
+  auto &N = getRegisters()[MsgPackDoc.getNode(Reg)];
+  if (N.getKind() == msgpack::Type::UInt)
+    Val |= N.getUInt();
+  N = N.getDocument()->getNode(Val);
 }
 
-// Set the number of used vgprs in the metadata. This is an optional advisory
-// record for logging etc; wave dispatch actually uses the rsrc1 register for
-// the shader stage to determine the number of vgprs to allocate.
+// Set the entry point name for one shader.
+void AMDGPUPALMetadata::setEntryPoint(unsigned CC, StringRef Name) {
+  if (isLegacy())
+    return;
+  // Msgpack format.
+  getHwStage(CC)[".entry_point"] = MsgPackDoc.getNode(Name, /*Copy=*/true);
+}
+
+// Set the number of used vgprs in the metadata. This is an optional
+// advisory record for logging etc; wave dispatch actually uses the rsrc1
+// register for the shader stage to determine the number of vgprs to
+// allocate.
 void AMDGPUPALMetadata::setNumUsedVgprs(CallingConv::ID CC, unsigned Val) {
+  if (isLegacy()) {
+    // Old non-msgpack format.
     unsigned NumUsedVgprsKey = getScratchSizeKey(CC) +
                                PALMD::Key::VS_NUM_USED_VGPRS -
                                PALMD::Key::VS_SCRATCH_SIZE;
-  Registers[NumUsedVgprsKey] = Val;
+    setRegister(NumUsedVgprsKey, Val);
+    return;
+  }
+  // Msgpack format.
+  getHwStage(CC)[".vgpr_count"] = MsgPackDoc.getNode(Val);
 }
 
 // Set the number of used sgprs in the metadata. This is an optional advisory
 // record for logging etc; wave dispatch actually uses the rsrc1 register for
 // the shader stage to determine the number of sgprs to allocate.
 void AMDGPUPALMetadata::setNumUsedSgprs(CallingConv::ID CC, unsigned Val) {
+  if (isLegacy()) {
+    // Old non-msgpack format.
     unsigned NumUsedSgprsKey = getScratchSizeKey(CC) +
                                PALMD::Key::VS_NUM_USED_SGPRS -
                                PALMD::Key::VS_SCRATCH_SIZE;
-  Registers[NumUsedSgprsKey] = Val;
+    setRegister(NumUsedSgprsKey, Val);
+    return;
+  }
+  // Msgpack format.
+  getHwStage(CC)[".sgpr_count"] = MsgPackDoc.getNode(Val);
 }
 
 // Set the scratch size in the metadata.
 void AMDGPUPALMetadata::setScratchSize(CallingConv::ID CC, unsigned Val) {
-  Registers[getScratchSizeKey(CC)] = Val;
+  if (isLegacy()) {
+    // Old non-msgpack format.
+    setRegister(getScratchSizeKey(CC), Val);
+    return;
+  }
+  // Msgpack format.
+  getHwStage(CC)[".scratch_memory_size"] = MsgPackDoc.getNode(Val);
+}
+
+// Convert a register number to name, for display by toString().
+// Returns nullptr if none.
+static const char *getRegisterName(unsigned RegNum) {
+  // Table of registers.
+  static const struct RegInfo {
+    unsigned Num;
+    const char *Name;
+  } RegInfoTable[] = {
+      // Registers that code generation sets/modifies metadata for.
+      {PALMD::R_2C4A_SPI_SHADER_PGM_RSRC1_VS, "SPI_SHADER_PGM_RSRC1_VS"},
+      {PALMD::R_2C4A_SPI_SHADER_PGM_RSRC1_VS + 1, "SPI_SHADER_PGM_RSRC2_VS"},
+      {PALMD::R_2D4A_SPI_SHADER_PGM_RSRC1_LS, "SPI_SHADER_PGM_RSRC1_LS"},
+      {PALMD::R_2D4A_SPI_SHADER_PGM_RSRC1_LS + 1, "SPI_SHADER_PGM_RSRC2_LS"},
+      {PALMD::R_2D0A_SPI_SHADER_PGM_RSRC1_HS, "SPI_SHADER_PGM_RSRC1_HS"},
+      {PALMD::R_2D0A_SPI_SHADER_PGM_RSRC1_HS + 1, "SPI_SHADER_PGM_RSRC2_HS"},
+      {PALMD::R_2CCA_SPI_SHADER_PGM_RSRC1_ES, "SPI_SHADER_PGM_RSRC1_ES"},
+      {PALMD::R_2CCA_SPI_SHADER_PGM_RSRC1_ES + 1, "SPI_SHADER_PGM_RSRC2_ES"},
+      {PALMD::R_2C8A_SPI_SHADER_PGM_RSRC1_GS, "SPI_SHADER_PGM_RSRC1_GS"},
+      {PALMD::R_2C8A_SPI_SHADER_PGM_RSRC1_GS + 1, "SPI_SHADER_PGM_RSRC2_GS"},
+      {PALMD::R_2E12_COMPUTE_PGM_RSRC1, "COMPUTE_PGM_RSRC1"},
+      {PALMD::R_2E12_COMPUTE_PGM_RSRC1 + 1, "COMPUTE_PGM_RSRC2"},
+      {PALMD::R_2C0A_SPI_SHADER_PGM_RSRC1_PS, "SPI_SHADER_PGM_RSRC1_PS"},
+      {PALMD::R_2C0A_SPI_SHADER_PGM_RSRC1_PS + 1, "SPI_SHADER_PGM_RSRC2_PS"},
+      {PALMD::R_A1B3_SPI_PS_INPUT_ENA, "SPI_PS_INPUT_ENA"},
+      {PALMD::R_A1B4_SPI_PS_INPUT_ADDR, "SPI_PS_INPUT_ADDR"},
+
+      // Registers not known to code generation.
+      {0x2c07, "SPI_SHADER_PGM_RSRC3_PS"},
+      {0x2c46, "SPI_SHADER_PGM_RSRC3_VS"},
+      {0x2c87, "SPI_SHADER_PGM_RSRC3_GS"},
+      {0x2cc7, "SPI_SHADER_PGM_RSRC3_ES"},
+      {0x2d07, "SPI_SHADER_PGM_RSRC3_HS"},
+      {0x2d47, "SPI_SHADER_PGM_RSRC3_LS"},
+
+      {0xa1c3, "SPI_SHADER_POS_FORMAT"},
+      {0xa1b1, "SPI_VS_OUT_CONFIG"},
+      {0xa207, "PA_CL_VS_OUT_CNTL"},
+      {0xa204, "PA_CL_CLIP_CNTL"},
+      {0xa206, "PA_CL_VTE_CNTL"},
+      {0xa2f9, "PA_SU_VTX_CNTL"},
+      {0xa293, "PA_SC_MODE_CNTL_1"},
+      {0xa2a1, "VGT_PRIMITIVEID_EN"},
+      {0x2c81, "SPI_SHADER_PGM_RSRC4_GS"},
+      {0x2e18, "COMPUTE_TMPRING_SIZE"},
+      {0xa1b5, "SPI_INTERP_CONTROL_0"},
+      {0xa1ba, "SPI_TMPRING_SIZE"},
+      {0xa1c4, "SPI_SHADER_Z_FORMAT"},
+      {0xa1c5, "SPI_SHADER_COL_FORMAT"},
+      {0xa203, "DB_SHADER_CONTROL"},
+      {0xa08f, "CB_SHADER_MASK"},
+      {0xa1b6, "SPI_PS_IN_CONTROL"},
+      {0xa191, "SPI_PS_INPUT_CNTL_0"},
+      {0xa192, "SPI_PS_INPUT_CNTL_1"},
+      {0xa193, "SPI_PS_INPUT_CNTL_2"},
+      {0xa194, "SPI_PS_INPUT_CNTL_3"},
+      {0xa195, "SPI_PS_INPUT_CNTL_4"},
+      {0xa196, "SPI_PS_INPUT_CNTL_5"},
+      {0xa197, "SPI_PS_INPUT_CNTL_6"},
+      {0xa198, "SPI_PS_INPUT_CNTL_7"},
+      {0xa199, "SPI_PS_INPUT_CNTL_8"},
+      {0xa19a, "SPI_PS_INPUT_CNTL_9"},
+      {0xa19b, "SPI_PS_INPUT_CNTL_10"},
+      {0xa19c, "SPI_PS_INPUT_CNTL_11"},
+      {0xa19d, "SPI_PS_INPUT_CNTL_12"},
+      {0xa19e, "SPI_PS_INPUT_CNTL_13"},
+      {0xa19f, "SPI_PS_INPUT_CNTL_14"},
+      {0xa1a0, "SPI_PS_INPUT_CNTL_15"},
+      {0xa1a1, "SPI_PS_INPUT_CNTL_16"},
+      {0xa1a2, "SPI_PS_INPUT_CNTL_17"},
+      {0xa1a3, "SPI_PS_INPUT_CNTL_18"},
+      {0xa1a4, "SPI_PS_INPUT_CNTL_19"},
+      {0xa1a5, "SPI_PS_INPUT_CNTL_20"},
+      {0xa1a6, "SPI_PS_INPUT_CNTL_21"},
+      {0xa1a7, "SPI_PS_INPUT_CNTL_22"},
+      {0xa1a8, "SPI_PS_INPUT_CNTL_23"},
+      {0xa1a9, "SPI_PS_INPUT_CNTL_24"},
+      {0xa1aa, "SPI_PS_INPUT_CNTL_25"},
+      {0xa1ab, "SPI_PS_INPUT_CNTL_26"},
+      {0xa1ac, "SPI_PS_INPUT_CNTL_27"},
+      {0xa1ad, "SPI_PS_INPUT_CNTL_28"},
+      {0xa1ae, "SPI_PS_INPUT_CNTL_29"},
+      {0xa1af, "SPI_PS_INPUT_CNTL_30"},
+      {0xa1b0, "SPI_PS_INPUT_CNTL_31"},
+
+      {0xa2ce, "VGT_GS_MAX_VERT_OUT"},
+      {0xa2ab, "VGT_ESGS_RING_ITEMSIZE"},
+      {0xa290, "VGT_GS_MODE"},
+      {0xa291, "VGT_GS_ONCHIP_CNTL"},
+      {0xa2d7, "VGT_GS_VERT_ITEMSIZE"},
+      {0xa2d8, "VGT_GS_VERT_ITEMSIZE_1"},
+      {0xa2d9, "VGT_GS_VERT_ITEMSIZE_2"},
+      {0xa2da, "VGT_GS_VERT_ITEMSIZE_3"},
+      {0xa298, "VGT_GSVS_RING_OFFSET_1"},
+      {0xa299, "VGT_GSVS_RING_OFFSET_2"},
+      {0xa29a, "VGT_GSVS_RING_OFFSET_3"},
+
+      {0xa2e4, "VGT_GS_INSTANCE_CNT"},
+      {0xa297, "VGT_GS_PER_VS"},
+      {0xa29b, "VGT_GS_OUT_PRIM_TYPE"},
+      {0xa2ac, "VGT_GSVS_RING_ITEMSIZE"},
+
+      {0xa2d5, "VGT_SHADER_STAGES_EN"},
+      {0xa2ad, "VGT_REUSE_OFF"},
+      {0xa1b8, "SPI_BARYC_CNTL"},
+
+      {0x2c4c, "SPI_SHADER_USER_DATA_VS_0"},
+      {0x2c4d, "SPI_SHADER_USER_DATA_VS_1"},
+      {0x2c4e, "SPI_SHADER_USER_DATA_VS_2"},
+      {0x2c4f, "SPI_SHADER_USER_DATA_VS_3"},
+      {0x2c50, "SPI_SHADER_USER_DATA_VS_4"},
+      {0x2c51, "SPI_SHADER_USER_DATA_VS_5"},
+      {0x2c52, "SPI_SHADER_USER_DATA_VS_6"},
+      {0x2c53, "SPI_SHADER_USER_DATA_VS_7"},
+      {0x2c54, "SPI_SHADER_USER_DATA_VS_8"},
+      {0x2c55, "SPI_SHADER_USER_DATA_VS_9"},
+      {0x2c56, "SPI_SHADER_USER_DATA_VS_10"},
+      {0x2c57, "SPI_SHADER_USER_DATA_VS_11"},
+      {0x2c58, "SPI_SHADER_USER_DATA_VS_12"},
+      {0x2c59, "SPI_SHADER_USER_DATA_VS_13"},
+      {0x2c5a, "SPI_SHADER_USER_DATA_VS_14"},
+      {0x2c5b, "SPI_SHADER_USER_DATA_VS_15"},
+      {0x2c5c, "SPI_SHADER_USER_DATA_VS_16"},
+      {0x2c5d, "SPI_SHADER_USER_DATA_VS_17"},
+      {0x2c5e, "SPI_SHADER_USER_DATA_VS_18"},
+      {0x2c5f, "SPI_SHADER_USER_DATA_VS_19"},
+      {0x2c60, "SPI_SHADER_USER_DATA_VS_20"},
+      {0x2c61, "SPI_SHADER_USER_DATA_VS_21"},
+      {0x2c62, "SPI_SHADER_USER_DATA_VS_22"},
+      {0x2c63, "SPI_SHADER_USER_DATA_VS_23"},
+      {0x2c64, "SPI_SHADER_USER_DATA_VS_24"},
+      {0x2c65, "SPI_SHADER_USER_DATA_VS_25"},
+      {0x2c66, "SPI_SHADER_USER_DATA_VS_26"},
+      {0x2c67, "SPI_SHADER_USER_DATA_VS_27"},
+      {0x2c68, "SPI_SHADER_USER_DATA_VS_28"},
+      {0x2c69, "SPI_SHADER_USER_DATA_VS_29"},
+      {0x2c6a, "SPI_SHADER_USER_DATA_VS_30"},
+      {0x2c6b, "SPI_SHADER_USER_DATA_VS_31"},
+
+      {0x2ccc, "SPI_SHADER_USER_DATA_ES_0"},
+      {0x2ccd, "SPI_SHADER_USER_DATA_ES_1"},
+      {0x2cce, "SPI_SHADER_USER_DATA_ES_2"},
+      {0x2ccf, "SPI_SHADER_USER_DATA_ES_3"},
+      {0x2cd0, "SPI_SHADER_USER_DATA_ES_4"},
+      {0x2cd1, "SPI_SHADER_USER_DATA_ES_5"},
+      {0x2cd2, "SPI_SHADER_USER_DATA_ES_6"},
+      {0x2cd3, "SPI_SHADER_USER_DATA_ES_7"},
+      {0x2cd4, "SPI_SHADER_USER_DATA_ES_8"},
+      {0x2cd5, "SPI_SHADER_USER_DATA_ES_9"},
+      {0x2cd6, "SPI_SHADER_USER_DATA_ES_10"},
+      {0x2cd7, "SPI_SHADER_USER_DATA_ES_11"},
+      {0x2cd8, "SPI_SHADER_USER_DATA_ES_12"},
+      {0x2cd9, "SPI_SHADER_USER_DATA_ES_13"},
+      {0x2cda, "SPI_SHADER_USER_DATA_ES_14"},
+      {0x2cdb, "SPI_SHADER_USER_DATA_ES_15"},
+      {0x2cdc, "SPI_SHADER_USER_DATA_ES_16"},
+      {0x2cdd, "SPI_SHADER_USER_DATA_ES_17"},
+      {0x2cde, "SPI_SHADER_USER_DATA_ES_18"},
+      {0x2cdf, "SPI_SHADER_USER_DATA_ES_19"},
+      {0x2ce0, "SPI_SHADER_USER_DATA_ES_20"},
+      {0x2ce1, "SPI_SHADER_USER_DATA_ES_21"},
+      {0x2ce2, "SPI_SHADER_USER_DATA_ES_22"},
+      {0x2ce3, "SPI_SHADER_USER_DATA_ES_23"},
+      {0x2ce4, "SPI_SHADER_USER_DATA_ES_24"},
+      {0x2ce5, "SPI_SHADER_USER_DATA_ES_25"},
+      {0x2ce6, "SPI_SHADER_USER_DATA_ES_26"},
+      {0x2ce7, "SPI_SHADER_USER_DATA_ES_27"},
+      {0x2ce8, "SPI_SHADER_USER_DATA_ES_28"},
+      {0x2ce9, "SPI_SHADER_USER_DATA_ES_29"},
+      {0x2cea, "SPI_SHADER_USER_DATA_ES_30"},
+      {0x2ceb, "SPI_SHADER_USER_DATA_ES_31"},
+
+      {0x2c0c, "SPI_SHADER_USER_DATA_PS_0"},
+      {0x2c0d, "SPI_SHADER_USER_DATA_PS_1"},
+      {0x2c0e, "SPI_SHADER_USER_DATA_PS_2"},
+      {0x2c0f, "SPI_SHADER_USER_DATA_PS_3"},
+      {0x2c10, "SPI_SHADER_USER_DATA_PS_4"},
+      {0x2c11, "SPI_SHADER_USER_DATA_PS_5"},
+      {0x2c12, "SPI_SHADER_USER_DATA_PS_6"},
+      {0x2c13, "SPI_SHADER_USER_DATA_PS_7"},
+      {0x2c14, "SPI_SHADER_USER_DATA_PS_8"},
+      {0x2c15, "SPI_SHADER_USER_DATA_PS_9"},
+      {0x2c16, "SPI_SHADER_USER_DATA_PS_10"},
+      {0x2c17, "SPI_SHADER_USER_DATA_PS_11"},
+      {0x2c18, "SPI_SHADER_USER_DATA_PS_12"},
+      {0x2c19, "SPI_SHADER_USER_DATA_PS_13"},
+      {0x2c1a, "SPI_SHADER_USER_DATA_PS_14"},
+      {0x2c1b, "SPI_SHADER_USER_DATA_PS_15"},
+      {0x2c1c, "SPI_SHADER_USER_DATA_PS_16"},
+      {0x2c1d, "SPI_SHADER_USER_DATA_PS_17"},
+      {0x2c1e, "SPI_SHADER_USER_DATA_PS_18"},
+      {0x2c1f, "SPI_SHADER_USER_DATA_PS_19"},
+      {0x2c20, "SPI_SHADER_USER_DATA_PS_20"},
+      {0x2c21, "SPI_SHADER_USER_DATA_PS_21"},
+      {0x2c22, "SPI_SHADER_USER_DATA_PS_22"},
+      {0x2c23, "SPI_SHADER_USER_DATA_PS_23"},
+      {0x2c24, "SPI_SHADER_USER_DATA_PS_24"},
+      {0x2c25, "SPI_SHADER_USER_DATA_PS_25"},
+      {0x2c26, "SPI_SHADER_USER_DATA_PS_26"},
+      {0x2c27, "SPI_SHADER_USER_DATA_PS_27"},
+      {0x2c28, "SPI_SHADER_USER_DATA_PS_28"},
+      {0x2c29, "SPI_SHADER_USER_DATA_PS_29"},
+      {0x2c2a, "SPI_SHADER_USER_DATA_PS_30"},
+      {0x2c2b, "SPI_SHADER_USER_DATA_PS_31"},
+
+      {0x2e40, "COMPUTE_USER_DATA_0"},
+      {0x2e41, "COMPUTE_USER_DATA_1"},
+      {0x2e42, "COMPUTE_USER_DATA_2"},
+      {0x2e43, "COMPUTE_USER_DATA_3"},
+      {0x2e44, "COMPUTE_USER_DATA_4"},
+      {0x2e45, "COMPUTE_USER_DATA_5"},
+      {0x2e46, "COMPUTE_USER_DATA_6"},
+      {0x2e47, "COMPUTE_USER_DATA_7"},
+      {0x2e48, "COMPUTE_USER_DATA_8"},
+      {0x2e49, "COMPUTE_USER_DATA_9"},
+      {0x2e4a, "COMPUTE_USER_DATA_10"},
+      {0x2e4b, "COMPUTE_USER_DATA_11"},
+      {0x2e4c, "COMPUTE_USER_DATA_12"},
+      {0x2e4d, "COMPUTE_USER_DATA_13"},
+      {0x2e4e, "COMPUTE_USER_DATA_14"},
+      {0x2e4f, "COMPUTE_USER_DATA_15"},
+
+      {0x2e07, "COMPUTE_NUM_THREAD_X"},
+      {0x2e08, "COMPUTE_NUM_THREAD_Y"},
+      {0x2e09, "COMPUTE_NUM_THREAD_Z"},
+      {0xa2db, "VGT_TF_PARAM"},
+      {0xa2d6, "VGT_LS_HS_CONFIG"},
+      {0xa287, "VGT_HOS_MIN_TESS_LEVEL"},
+      {0xa286, "VGT_HOS_MAX_TESS_LEVEL"},
+      {0xa2f8, "PA_SC_AA_CONFIG"},
+      {0xa310, "PA_SC_SHADER_CONTROL"},
+      {0xa313, "PA_SC_CONSERVATIVE_RASTERIZATION_CNTL"},
+
+      {0x2d0c, "SPI_SHADER_USER_DATA_LS_0"},
+      {0x2d0d, "SPI_SHADER_USER_DATA_LS_1"},
+      {0x2d0e, "SPI_SHADER_USER_DATA_LS_2"},
+      {0x2d0f, "SPI_SHADER_USER_DATA_LS_3"},
+      {0x2d10, "SPI_SHADER_USER_DATA_LS_4"},
+      {0x2d11, "SPI_SHADER_USER_DATA_LS_5"},
+      {0x2d12, "SPI_SHADER_USER_DATA_LS_6"},
+      {0x2d13, "SPI_SHADER_USER_DATA_LS_7"},
+      {0x2d14, "SPI_SHADER_USER_DATA_LS_8"},
+      {0x2d15, "SPI_SHADER_USER_DATA_LS_9"},
+      {0x2d16, "SPI_SHADER_USER_DATA_LS_10"},
+      {0x2d17, "SPI_SHADER_USER_DATA_LS_11"},
+      {0x2d18, "SPI_SHADER_USER_DATA_LS_12"},
+      {0x2d19, "SPI_SHADER_USER_DATA_LS_13"},
+      {0x2d1a, "SPI_SHADER_USER_DATA_LS_14"},
+      {0x2d1b, "SPI_SHADER_USER_DATA_LS_15"},
+      {0x2d1c, "SPI_SHADER_USER_DATA_LS_16"},
+      {0x2d1d, "SPI_SHADER_USER_DATA_LS_17"},
+      {0x2d1e, "SPI_SHADER_USER_DATA_LS_18"},
+      {0x2d1f, "SPI_SHADER_USER_DATA_LS_19"},
+      {0x2d20, "SPI_SHADER_USER_DATA_LS_20"},
+      {0x2d21, "SPI_SHADER_USER_DATA_LS_21"},
+      {0x2d22, "SPI_SHADER_USER_DATA_LS_22"},
+      {0x2d23, "SPI_SHADER_USER_DATA_LS_23"},
+      {0x2d24, "SPI_SHADER_USER_DATA_LS_24"},
+      {0x2d25, "SPI_SHADER_USER_DATA_LS_25"},
+      {0x2d26, "SPI_SHADER_USER_DATA_LS_26"},
+      {0x2d27, "SPI_SHADER_USER_DATA_LS_27"},
+      {0x2d28, "SPI_SHADER_USER_DATA_LS_28"},
+      {0x2d29, "SPI_SHADER_USER_DATA_LS_29"},
+      {0x2d2a, "SPI_SHADER_USER_DATA_LS_30"},
+      {0x2d2b, "SPI_SHADER_USER_DATA_LS_31"},
+
+      {0xa2aa, "IA_MULTI_VGT_PARAM"},
+      {0xa2a5, "VGT_GS_MAX_PRIMS_PER_SUBGROUP"},
+      {0xa2e6, "VGT_STRMOUT_BUFFER_CONFIG"},
+      {0xa2e5, "VGT_STRMOUT_CONFIG"},
+      {0xa2b5, "VGT_STRMOUT_VTX_STRIDE_0"},
+      {0xa2b9, "VGT_STRMOUT_VTX_STRIDE_1"},
+      {0xa2bd, "VGT_STRMOUT_VTX_STRIDE_2"},
+      {0xa2c1, "VGT_STRMOUT_VTX_STRIDE_3"},
+      {0xa316, "VGT_VERTEX_REUSE_BLOCK_CNTL"},
+
+      {0, nullptr}};
+  auto Entry = RegInfoTable;
+  for (; Entry->Num && Entry->Num != RegNum; ++Entry)
+    ;
+  return Entry->Name;
 }
 
 // Convert the accumulated PAL metadata into an asm directive.
 void AMDGPUPALMetadata::toString(std::string &String) {
   String.clear();
-  if (Registers.empty())
-    return;
   raw_string_ostream Stream(String);
+  if (isLegacy()) {
+    if (MsgPackDoc.getRoot().getKind() == msgpack::Type::Nil)
+      return;
+    // Old linear reg=val format.
     Stream << '\t' << AMDGPU::PALMD::AssemblerDirective << ' ';
-  for (auto I = Registers.begin(), E = Registers.end(); I != E; ++I) {
-    if (I != Registers.begin())
+    auto Regs = getRegisters();
+    for (auto I = Regs.begin(), E = Regs.end(); I != E; ++I) {
+      if (I != Regs.begin())
         Stream << ',';
-    Stream << "0x" << Twine::utohexstr(I->first) << ",0x"
-           << Twine::utohexstr(I->second);
+      unsigned Reg = I->first.getUInt();
+      unsigned Val = I->second.getUInt();
+      Stream << "0x" << Twine::utohexstr(Reg) << ",0x" << Twine::utohexstr(Val);
     }
     Stream << '\n';
+    return;
+  }
+
+  // New msgpack-based format -- output as YAML (with unsigned numbers in hex),
+  // but first change the registers map to use names.
+  MsgPackDoc.setHexMode();
+  auto &RegsObj = refRegisters();
+  auto OrigRegs = RegsObj.getMap();
+  RegsObj = MsgPackDoc.getMapNode();
+  for (auto I : OrigRegs) {
+    auto Key = I.first;
+    if (const char *RegName = getRegisterName(Key.getUInt())) {
+      std::string KeyName = Key.toString();
+      KeyName += " (";
+      KeyName += RegName;
+      KeyName += ')';
+      Key = MsgPackDoc.getNode(KeyName, /*Copy=*/true);
+    }
+    RegsObj.getMap()[Key] = I.second;
+  }
+
+  // Output as YAML.
+  Stream << '\t' << AMDGPU::PALMD::AssemblerDirectiveBegin << '\n';
+  MsgPackDoc.toYAML(Stream);
+  Stream << '\t' << AMDGPU::PALMD::AssemblerDirectiveEnd << '\n';
+
+  // Restore original registers map.
+  RegsObj = OrigRegs;
 }
 
 // Convert the accumulated PAL metadata into a binary blob for writing as
 // a .note record of the specified AMD type.
 void AMDGPUPALMetadata::toBlob(unsigned Type, std::string &Blob) {
+  if (Type == ELF::NT_AMD_AMDGPU_PAL_METADATA)
+    toLegacyBlob(Blob);
+  else
+    toMsgPackBlob(Blob);
+}
+
+void AMDGPUPALMetadata::toLegacyBlob(std::string &Blob) {
   Blob.clear();
-  if (Type != ELF::NT_AMD_AMDGPU_PAL_METADATA)
-    return;
-  if (Registers.empty())
+  auto Registers = getRegisters();
+  if (Registers.getMap().empty())
     return;
   raw_string_ostream OS(Blob);
   support::endian::Writer EW(OS, support::endianness::little);
-  for (auto I : Registers) {
-    EW.write(uint32_t(I.first));
-    EW.write(uint32_t(I.second));
-  }
+  for (auto I : Registers.getMap()) {
+    EW.write(uint32_t(I.first.getUInt()));
+    EW.write(uint32_t(I.second.getUInt()));
+  }
+
+void AMDGPUPALMetadata::toMsgPackBlob(std::string &Blob) {
+  Blob.clear();
+  MsgPackDoc.writeToBlob(Blob);
+}
+
+// Set PAL metadata from YAML text. Returns false if failed.
+bool AMDGPUPALMetadata::setFromString(StringRef S) {
+  BlobType = ELF::NT_AMDGPU_METADATA;
+  if (!MsgPackDoc.fromYAML(S))
+    return false;
+
+  // In the registers map, some keys may be of the form "0xa191
+  // (SPI_PS_INPUT_CNTL_0)", in which case the YAML input code made it a
+  // string. We need to turn it into a number.
+  auto &RegsObj = refRegisters();
+  auto OrigRegs = RegsObj;
+  RegsObj = MsgPackDoc.getMapNode();
+  Registers = RegsObj.getMap();
+  bool Ok = true;
+  for (auto I : OrigRegs.getMap()) {
+    auto Key = I.first;
+    if (Key.getKind() == msgpack::Type::String) {
+      StringRef S = Key.getString();
+      uint64_t Val;
+      if (S.consumeInteger(0, Val)) {
+        Ok = false;
+        errs() << "Unrecognized PAL metadata register key '" << S << "'\n";
+        continue;
+      }
+      Key = MsgPackDoc.getNode(uint64_t(Val));
+    }
+    Registers.getMap()[Key] = I.second;
+  }
+  return Ok;
+}
+
+// Reference (create if necessary) the node for the registers map.
+msgpack::DocNode &AMDGPUPALMetadata::refRegisters() {
+  auto &N =
+      MsgPackDoc.getRoot()
+          .getMap(/*Convert=*/true)[MsgPackDoc.getNode("amdpal.pipelines")]
+          .getArray(/*Convert=*/true)[0]
+          .getMap(/*Convert=*/true)[MsgPackDoc.getNode(".registers")];
+  N.getMap(/*Convert=*/true);
+  return N;
+}
+
+// Get (create if necessary) the registers map.
+msgpack::MapDocNode AMDGPUPALMetadata::getRegisters() {
+  if (Registers.isEmpty())
+    Registers = refRegisters();
+  return Registers.getMap();
+}
+
+// Return the PAL metadata hardware shader stage name.
+static const char *getStageName(CallingConv::ID CC) {
+  switch (CC) {
+  case CallingConv::AMDGPU_PS:
+    return ".ps";
+  case CallingConv::AMDGPU_VS:
+    return ".vs";
+  case CallingConv::AMDGPU_GS:
+    return ".gs";
+  case CallingConv::AMDGPU_ES:
+    return ".es";
+  case CallingConv::AMDGPU_HS:
+    return ".hs";
+  case CallingConv::AMDGPU_LS:
+    return ".ls";
+  default:
+    return ".cs";
+  }
+}
+
+// Get (create if necessary) the .hardware_stages entry for the given calling
+// convention.
+msgpack::MapDocNode AMDGPUPALMetadata::getHwStage(unsigned CC) {
+  if (HwStages.isEmpty())
+    HwStages = MsgPackDoc.getRoot()
+                   .getMap(/*Convert=*/true)["amdpal.pipelines"]
+                   .getArray(/*Convert=*/true)[0]
+                   .getMap(/*Convert=*/true)[".hardware_stages"]
+                   .getMap(/*Convert=*/true);
+  return HwStages.getMap()[getStageName(CC)].getMap(/*Convert=*/true);
+}
+
+// Get .note record vendor name of metadata blob to be emitted.
+const char *AMDGPUPALMetadata::getVendor() const {
+  return isLegacy() ? ElfNote::NoteNameV2 : ElfNote::NoteNameV3;
+}
+
+// Get .note record type of metadata blob to be emitted:
+// ELF::NT_AMD_AMDGPU_PAL_METADATA (legacy key=val format), or
+// ELF::NT_AMDGPU_METADATA (MsgPack format).
+unsigned AMDGPUPALMetadata::getType() const {
+  return BlobType ? BlobType : unsigned(ELF::NT_AMDGPU_METADATA);
+}
+
+// Return whether the blob type is legacy PAL metadata.
+bool AMDGPUPALMetadata::isLegacy() const {
+  return BlobType == ELF::NT_AMD_AMDGPU_PAL_METADATA;
 }
 

test/CodeGen/AMDGPU/amdpal-msgpack-cs.ll

+; RUN: llc -mtriple=amdgcn--amdpal -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI -enable-var-scope %s
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI -enable-var-scope %s
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GFX9 -enable-var-scope %s
+
+; amdpal compute shader: check for 0x2e12 (COMPUTE_PGM_RSRC1) in pal metadata
+; GCN-LABEL: {{^}}cs_amdpal:
+; GCN: .amdgpu_pal_metadata
+; GCN: 0x2e12 (COMPUTE_PGM_RSRC1)
+define amdgpu_cs half @cs_amdpal(half %arg0) {
+  %add = fadd half %arg0, 1.0
+  ret half %add
+}
+
+; Force MsgPack format metadata
+!amdgpu.pal.metadata.msgpack = !{!0}
+!0 = !{!""}

test/CodeGen/AMDGPU/amdpal-msgpack-es.ll

+; RUN: llc -mtriple=amdgcn--amdpal -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
+
+; amdpal evaluation shader: check for 0x2cca (SPI_SHADER_PGM_RSRC1_ES) in pal metadata
+; GCN-LABEL: {{^}}es_amdpal:
+; GCN: .amdgpu_pal_metadata
+; GCN: 0x2cca (SPI_SHADER_PGM_RSRC1_ES)
+define amdgpu_es half @es_amdpal(half %arg0) {
+  %add = fadd half %arg0, 1.0
+  ret half %add
+}
+
+; Force MsgPack format metadata
+!amdgpu.pal.metadata.msgpack = !{!0}
+!0 = !{!""}

test/CodeGen/AMDGPU/amdpal-msgpack-gs.ll

+; RUN: llc -mtriple=amdgcn--amdpal -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GFX9 -enable-var-scope %s
+
+; amdpal geometry shader: check for 0x2c8a (SPI_SHADER_PGM_RSRC1_GS) in pal metadata
+; GCN-LABEL: {{^}}gs_amdpal:
+; GCN: .amdgpu_pal_metadata
+; GCN: 0x2c8a (SPI_SHADER_PGM_RSRC1_GS)
+define amdgpu_gs half @gs_amdpal(half %arg0) {
+  %add = fadd half %arg0, 1.0
+  ret half %add
+}
+
+; Force MsgPack format metadata
+!amdgpu.pal.metadata.msgpack = !{!0}
+!0 = !{!""}

test/CodeGen/AMDGPU/amdpal-msgpack-hs.ll

+; RUN: llc -mtriple=amdgcn--amdpal -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GFX9 -enable-var-scope %s
+
+; amdpal hull shader: check for 0x2d0a (SPI_SHADER_PGM_RSRC1_HS) in pal metadata
+; GCN-LABEL: {{^}}hs_amdpal:
+; GCN: .amdgpu_pal_metadata
+; GCN: 0x2d0a (SPI_SHADER_PGM_RSRC1_HS)
+define amdgpu_hs half @hs_amdpal(half %arg0) {
+  %add = fadd half %arg0, 1.0
+  ret half %add
+}
+
+; Force MsgPack format metadata
+!amdgpu.pal.metadata.msgpack = !{!0}
+!0 = !{!""}

test/CodeGen/AMDGPU/amdpal-msgpack-ls.ll

+; RUN: llc -mtriple=amdgcn--amdpal -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
+
+; amdpal load shader: check for 0x2d4a (SPI_SHADER_PGM_RSRC1_LS) in pal metadata
+; GCN-LABEL: {{^}}ls_amdpal:
+; GCN: .amdgpu_pal_metadata
+; GCN: 0x2d4a (SPI_SHADER_PGM_RSRC1_LS)
+define amdgpu_ls half @ls_amdpal(half %arg0) {
+  %add = fadd half %arg0, 1.0
+  ret half %add
+}
+
+; Force MsgPack format metadata
+!amdgpu.pal.metadata.msgpack = !{!0}
+!0 = !{!""}

test/CodeGen/AMDGPU/amdpal-msgpack-ps.ll

+; RUN: llc -mtriple=amdgcn--amdpal -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GFX9 -enable-var-scope %s
+
+; amdpal pixel shader: check for 0x2c0a (SPI_SHADER_PGM_RSRC1_PS) in pal
+; metadata. Check for 0x2c0b (SPI_SHADER_PGM_RSRC2_PS) in pal metadata, and
+; it has a value starting 0x42 as it is set to 0x42000000 in the metadata
+; below. Also check that .internal_pipeline_hash is propagated.
+; GCN-LABEL: {{^}}ps_amdpal:
+; GCN: .amdgpu_pal_metadata
+; GCN:         .internal_pipeline_hash:
+; GCN-NEXT:      - 0x123456789abcdef0
+; GCN-NEXT:      - 0xfedcba9876543210
+; GCN:         .registers:
+; GCN:           0x2c0a (SPI_SHADER_PGM_RSRC1_PS):
+; GCN:           0x2c0b (SPI_SHADER_PGM_RSRC2_PS): 0x42
+define amdgpu_ps half @ps_amdpal(half %arg0) {
+  %add = fadd half %arg0, 1.0
+  ret half %add
+}
+
+; amdgpu.pal.metadata.msgpack represents this:
+;
+;   .amdgpu_pal_metadata
+; ---
+; amdpal.pipelines: 
+;   - .internal_pipeline_hash:
+;       - 0x123456789abcdef0
+;       - 0xfedcba9876543210
+;     .registers:
+;       0x2c0b (SPI_SHADER_PGM_RSRC2_PS): 0x42000000
+; ...
+;   .end_amdgpu_pal_metadata
+
+!amdgpu.pal.metadata.msgpack = !{!0}
+!0 = !{!"\81\b0\61\6d\64\70\61\6c\2e\70\69\70\65\6c\69\6e\65\73\91\82\b7\2e\69\6e\74\65\72\6e\61\6c\5f\70\69\70\65\6c\69\6e\65\5f\68\61\73\68\92\cf\12\34\56\78\9a\bc\de\f0\cf\fe\dc\ba\98\76\54\32\10\aa\2e\72\65\67\69\73\74\65\72\73\81\cd\2c\0b\ce\42\00\00\00"};

test/CodeGen/AMDGPU/amdpal-msgpack-psenable.ll

+; RUN: llc -mtriple=amdgcn--amdpal -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GFX9 -enable-var-scope %s
+
+; This pixel shader does not use the result of its interpolation, so it would
+; end up with an interpolation mode set in PSAddr but not PSEnable. This test tests
+; the workaround that ensures that an interpolation mode is also set in PSEnable.
+; GCN-LABEL: {{^}}amdpal_psenable:
+; GCN: .amdgpu_pal_metadata
+; GCN: 0xa1b3 (SPI_PS_INPUT_ENA): 0x2
+; GCN: 0xa1b4 (SPI_PS_INPUT_ADDR): 0x2
+define amdgpu_ps void @amdpal_psenable(i32 inreg, i32 inreg, i32 inreg, i32 inreg %m0, <2 x float> %pos) #6 {
+  %inst23 = extractelement <2 x float> %pos, i32 0
+  %inst24 = extractelement <2 x float> %pos, i32 1
+  %inst25 = tail call float @llvm.amdgcn.interp.p1(float %inst23, i32 0, i32 0, i32 %m0)
+  %inst26 = tail call float @llvm.amdgcn.interp.p2(float %inst25, float %inst24, i32 0, i32 0, i32 %m0)
+  ret void
+}
+
+declare float @llvm.amdgcn.interp.p1(float, i32, i32, i32) #2
+declare float @llvm.amdgcn.interp.p2(float, float, i32, i32, i32) #2
+
+attributes #6 = { nounwind "InitialPSInputAddr"="2" }
+
+; Force MsgPack format metadata
+!amdgpu.pal.metadata.msgpack = !{!0}
+!0 = !{!""}

test/CodeGen/AMDGPU/amdpal-msgpack-vs.ll

+; RUN: llc -mtriple=amdgcn--amdpal -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GFX9 -enable-var-scope %s
+
+; amdpal vertex shader: check for 45352 (SPI_SHADER_PGM_RSRC1_VS) in pal metadata
+; GCN-LABEL: {{^}}vs_amdpal:
+; GCN: .amdgpu_pal_metadata
+; GCN: 0x2c4a (SPI_SHADER_PGM_RSRC1_VS)
+define amdgpu_vs half @vs_amdpal(half %arg0) {
+  %add = fadd half %arg0, 1.0
+  ret half %add
+}
+
+; Force MsgPack format metadata
+!amdgpu.pal.metadata.msgpack = !{!0}
+!0 = !{!""}

test/MC/AMDGPU/pal-msgpack.s

+// RUN: llvm-mc -triple amdgcn--amdpal -mcpu=kaveri -show-encoding %s | FileCheck %s --check-prefix=ASM
+// RUN: llvm-mc -filetype=obj -triple amdgcn--amdpal -mcpu=kaveri -show-encoding %s | llvm-readobj -symbols -s -sd | FileCheck %s --check-prefix=ELF
+
+  .amdgpu_pal_metadata
+---
+amdpal.pipelines: 
+  - .hardware_stages: 
+      .ps:             
+        .entry_point:    ps_amdpal
+        .scratch_memory_size: 0
+        .sgpr_count:     0x1
+        .vgpr_count:     0x1
+    .internal_pipeline_hash: 
+      - 0x123456789abcdef0
+      - 0xfedcba9876543210
+    .registers:      
+      0x2c0a (SPI_SHADER_PGM_RSRC1_PS): 0
+      0x2c0b (SPI_SHADER_PGM_RSRC2_PS): 0x42000000
+      0xa1b3 (SPI_PS_INPUT_ENA): 0x1
+      0xa1b4 (SPI_PS_INPUT_ADDR): 0x1
+...
+  .end_amdgpu_pal_metadata
+
+// ASM:   .amdgpu_pal_metadata
+// ASM: ---
+// ASM: amdpal.pipelines: 
+// ASM:   - .hardware_stages: 
+// ASM:       .ps:             
+// ASM:         .entry_point:    ps_amdpal
+// ASM:         .scratch_memory_size: 0
+// ASM:         .sgpr_count:     0x1
+// ASM:         .vgpr_count:     0x1
+// ASM:     .internal_pipeline_hash: 
+// ASM:       - 0x123456789abcdef0
+// ASM:       - 0xfedcba9876543210
+// ASM:     .registers:      
+// ASM:       0x2c0a (SPI_SHADER_PGM_RSRC1_PS): 0
+// ASM:       0x2c0b (SPI_SHADER_PGM_RSRC2_PS): 0x42000000
+// ASM:       0xa1b3 (SPI_PS_INPUT_ENA): 0x1
+// ASM:       0xa1b4 (SPI_PS_INPUT_ADDR): 0x1
+// ASM: ...
+// ASM:   .end_amdgpu_pal_metadata
+
+// ELF: SHT_NOTE
+// ELF:       0000: 04000000 BD000000 0D000000 414D4400  |............AMD.|
+// ELF:       0010: 81B0616D 6470616C 2E706970 656C696E  |..amdpal.pipelin|
+// ELF:       0020: 65739183 B02E6861 72647761 72655F73  |es....hardware_s|
+// ELF:       0030: 74616765 7381A32E 707384AC 2E656E74  |tages...ps...ent|
+// ELF:       0040: 72795F70 6F696E74 A970735F 616D6470  |ry_point.ps_amdp|
+// ELF:       0050: 616CB42E 73637261 7463685F 6D656D6F  |al..scratch_memo|
+// ELF:       0060: 72795F73 697A6500 AB2E7367 70725F63  |ry_size...sgpr_c|
+// ELF:       0070: 6F756E74 01AB2E76 6770725F 636F756E  |ount...vgpr_coun|
+// ELF:       0080: 7401B72E 696E7465 726E616C 5F706970  |t...internal_pip|
+// ELF:       0090: 656C696E 655F6861 736892CF 12345678  |eline_hash...4Vx|
+// ELF:       00A0: 9ABCDEF0 CFFEDCBA 98765432 10AA2E72  |.........vT2...r|
+// ELF:       00B0: 65676973 74657273 84CD2C0A 00CD2C0B  |egisters..,...,.|
+// ELF:       00C0: CE420000 00CDA1B3 01CDA1B4 01000000  |.B..............|