[RemoveDIs] Add bitcode format

If --write-experimental-debuginfo-iterators-to-bitcode is true (default false)
and --expermental-debuginfo-iterators is also true then the new debug info
format (non-instruction records) is written to bitcode directly.

Added the following records:

    FUNC_CODE_DEBUG_RECORD_LABEL
    FUNC_CODE_DEBUG_RECORD_VALUE
    FUNC_CODE_DEBUG_RECORD_DECLARE
    FUNC_CODE_DEBUG_RECORD_ASSIGN
    FUNC_CODE_DEBUG_RECORD_VALUE_SIMPLE

The last one has an abbrev in FUNCTION_BLOCK BLOCK_INFO. Incidentally, this uses
the last value available without widening the code-length for FUNCTION_BLOCK
from 4 to 5 bits.

Records are formatted as follows:

    All DbgRecord start with:
      1. DILocation

      FUNC_CODE_DEBUG_RECORD_LABEL
        2. DILabel

      DPValues then share common fields:
        2. DILocalVariable
        3. DIExpression

        FUNC_CODE_DEBUG_RECORD_VALUE
          4. Location Metadata

        FUNC_CODE_DEBUG_RECORD_DECLARE
          4. Location Metadata

        FUNC_CODE_DEBUG_RECORD_VALUE_SIMPLE
	  4. Location Value (single)

        FUNC_CODE_DEBUG_RECORD_ASSIGN
	  4. Location Metadata
	  5. DIAssignID
	  6. DIExpression (address)
	  7. Location Metadata (address)

Encoding the DILocation metadata reference directly appeared to yield smaller
bitcode files than encoding the operands seperately (as is done with instruction
DILocations).

FUNC_CODE_DEBUG_RECORD_VALUE_SIMPLE is by far the most common DbgRecord record
in optimized code (order of 5x-10x over other kinds). Unoptimized code should
only contain FUNC_CODE_DEBUG_RECORD_DECLARE.

Author: OCHyams

llvm/include/llvm/Bitcode/LLVMBitCodes.h

@@ -624,6 +624,17 @@ enum FunctionCodes {
                                   //             operation, align, vol,
                                   //             ordering, synchscope]
   FUNC_CODE_BLOCKADDR_USERS = 60, // BLOCKADDR_USERS: [value...]
+
+  FUNC_CODE_DEBUG_RECORD_VALUE =
+      61, // [DILocation, DILocalVariable, DIExpression, ValueAsMetadata]
+  FUNC_CODE_DEBUG_RECORD_DECLARE =
+      62, // [DILocation, DILocalVariable, DIExpression, ValueAsMetadata]
+  FUNC_CODE_DEBUG_RECORD_ASSIGN =
+      63, // [DILocation, DILocalVariable, DIExpression, ValueAsMetadata,
+          //  DIAssignID, DIExpression (addr), ValueAsMetadata (addr)]
+  FUNC_CODE_DEBUG_RECORD_VALUE_SIMPLE =
+      64, // [DILocation, DILocalVariable, DIExpression, Value]
+  FUNC_CODE_DEBUG_RECORD_LABEL = 65, // DPVALUE: [DILocation, DILabel]
 };
 
 enum UseListCodes {

llvm/lib/Bitcode/Reader/BitcodeAnalyzer.cpp

@@ -270,6 +270,11 @@ GetCodeName(unsigned CodeID, unsigned BlockID,
       STRINGIFY_CODE(FUNC_CODE, INST_CMPXCHG)
       STRINGIFY_CODE(FUNC_CODE, INST_CALLBR)
       STRINGIFY_CODE(FUNC_CODE, BLOCKADDR_USERS)
+      STRINGIFY_CODE(FUNC_CODE, DEBUG_RECORD_DECLARE)
+      STRINGIFY_CODE(FUNC_CODE, DEBUG_RECORD_VALUE)
+      STRINGIFY_CODE(FUNC_CODE, DEBUG_RECORD_ASSIGN)
+      STRINGIFY_CODE(FUNC_CODE, DEBUG_RECORD_VALUE_SIMPLE)
+      STRINGIFY_CODE(FUNC_CODE, DEBUG_RECORD_LABEL)
     }
   case bitc::VALUE_SYMTAB_BLOCK_ID:
     switch (CodeID) {

llvm/lib/Bitcode/Reader/BitcodeReader.cpp

@@ -100,9 +100,6 @@ static cl::opt<bool> ExpandConstantExprs(
     cl::desc(
         "Expand constant expressions to instructions for testing purposes"));
 
-// Declare external flag for whether we're using the new debug-info format.
-extern llvm::cl::opt<bool> UseNewDbgInfoFormat;
-
 namespace {
 
 enum {
@@ -4279,6 +4276,10 @@ Error BitcodeReader::parseGlobalIndirectSymbolRecord(
 Error BitcodeReader::parseModule(uint64_t ResumeBit,
                                  bool ShouldLazyLoadMetadata,
                                  ParserCallbacks Callbacks) {
+  // Force the debug-info mode into the old format for now.
+  // FIXME: Remove this once all tools support RemoveDIs.
+  TheModule->IsNewDbgInfoFormat = false;
+
   this->ValueTypeCallback = std::move(Callbacks.ValueType);
   if (ResumeBit) {
     if (Error JumpFailed = Stream.JumpToBit(ResumeBit))
@@ -6398,6 +6399,89 @@ Error BitcodeReader::parseFunctionBody(Function *F) {
       InstructionList.push_back(I);
       break;
     }
+    case bitc::FUNC_CODE_DEBUG_RECORD_LABEL: {
+      // DPLabels are placed after the Instructions that they are attached to.
+      Instruction *Inst = getLastInstruction();
+      if (!Inst)
+        return error("Invalid dbg record: missing instruction");
+      DILocation *DIL = cast<DILocation>(getFnMetadataByID(Record[0]));
+      DILabel *Label = cast<DILabel>(getFnMetadataByID(Record[1]));
+      Inst->getParent()->insertDPValueBefore(new DPLabel(Label, DIL),
+                                             Inst->getIterator());
+      continue; // This isn't an instruction.
+    }
+    case bitc::FUNC_CODE_DEBUG_RECORD_VALUE_SIMPLE:
+    case bitc::FUNC_CODE_DEBUG_RECORD_VALUE:
+    case bitc::FUNC_CODE_DEBUG_RECORD_DECLARE:
+    case bitc::FUNC_CODE_DEBUG_RECORD_ASSIGN: {
+      // DPValues are placed after the Instructions that they are attached to.
+      Instruction *Inst = getLastInstruction();
+      if (!Inst)
+        return error("Invalid dbg record: missing instruction");
+
+      // First 3 fields are common to all kinds:
+      //   DILocation, DILocalVariable, DIExpression
+      // dbg_value (FUNC_CODE_DEBUG_RECORD_VALUE)
+      //   ..., LocationMetadata
+      // dbg_value (FUNC_CODE_DEBUG_RECORD_VALUE_SIMPLE - abbrev'd)
+      //   ..., Value
+      // dbg_declare (FUNC_CODE_DEBUG_RECORD_DECLARE)
+      //   ..., LocationMetadata
+      // dbg_assign (FUNC_CODE_DEBUG_RECORD_ASSIGN)
+      //   ..., LocationMetadata, DIAssignID, DIExpression, LocationMetadata
+      unsigned Slot = 0;
+      // Common fields (0-2).
+      DILocation *DIL = cast<DILocation>(getFnMetadataByID(Record[Slot++]));
+      DILocalVariable *Var =
+          cast<DILocalVariable>(getFnMetadataByID(Record[Slot++]));
+      DIExpression *Expr =
+          cast<DIExpression>(getFnMetadataByID(Record[Slot++]));
+
+      // Union field (3: LocationMetadata | Value).
+      Metadata *RawLocation = nullptr;
+      if (BitCode == bitc::FUNC_CODE_DEBUG_RECORD_VALUE_SIMPLE) {
+        Value *V = nullptr;
+        unsigned TyID = 0;
+        // We never expect to see a fwd reference value here because
+        // use-before-defs are encoded with the standard non-abbrev record
+        // type (they'd require encoding the type too, and they're rare). As a
+        // result, getValueTypePair only ever increments Slot by one here (once
+        // for the value, never twice for value and type).
+        unsigned SlotBefore = Slot;
+        if (getValueTypePair(Record, Slot, NextValueNo, V, TyID, CurBB))
+          return error("Invalid dbg record: invalid value");
+        (void)SlotBefore;
+        assert((SlotBefore == Slot - 1) && "unexpected fwd ref");
+        RawLocation = ValueAsMetadata::get(V);
+      } else {
+        RawLocation = getFnMetadataByID(Record[Slot++]);
+      }
+
+      DPValue *DPV = nullptr;
+      switch (BitCode) {
+      case bitc::FUNC_CODE_DEBUG_RECORD_VALUE:
+      case bitc::FUNC_CODE_DEBUG_RECORD_VALUE_SIMPLE:
+        DPV = new DPValue(RawLocation, Var, Expr, DIL,
+                          DPValue::LocationType::Value);
+        break;
+      case bitc::FUNC_CODE_DEBUG_RECORD_DECLARE:
+        DPV = new DPValue(RawLocation, Var, Expr, DIL,
+                          DPValue::LocationType::Declare);
+        break;
+      case bitc::FUNC_CODE_DEBUG_RECORD_ASSIGN: {
+        DIAssignID *ID = cast<DIAssignID>(getFnMetadataByID(Record[Slot++]));
+        DIExpression *AddrExpr =
+            cast<DIExpression>(getFnMetadataByID(Record[Slot++]));
+        Metadata *Addr = getFnMetadataByID(Record[Slot++]);
+        DPV = new DPValue(RawLocation, Var, Expr, ID, Addr, AddrExpr, DIL);
+        break;
+      }
+      default:
+        llvm_unreachable("Unknown DPValue bitcode");
+      }
+      Inst->getParent()->insertDPValueBefore(DPV, Inst->getIterator());
+      continue; // This isn't an instruction.
+    }
     case bitc::FUNC_CODE_INST_CALL: {
       // CALL: [paramattrs, cc, fmf, fnty, fnid, arg0, arg1...]
       if (Record.size() < 3)
@@ -6677,10 +6761,21 @@ Error BitcodeReader::materialize(GlobalValue *GV) {
   // Move the bit stream to the saved position of the deferred function body.
   if (Error JumpFailed = Stream.JumpToBit(DFII->second))
     return JumpFailed;
+
+  // Set the debug info mode to "new", forcing a mismatch between
+  // module and function debug modes. This is okay because we'll convert
+  // everything back to the old mode after parsing.
+  // FIXME: Remove this once all tools support RemoveDIs.
+  F->IsNewDbgInfoFormat = true;
+
   if (Error Err = parseFunctionBody(F))
     return Err;
   F->setIsMaterializable(false);
 
+  // Convert new debug info records into intrinsics.
+  // FIXME: Remove this once all tools support RemoveDIs.
+  F->convertFromNewDbgValues();
+
   if (StripDebugInfo)
     stripDebugInfo(*F);
 

llvm/lib/Bitcode/Writer/BitcodeWriter.cpp

@@ -99,6 +99,9 @@ namespace llvm {
 extern FunctionSummary::ForceSummaryHotnessType ForceSummaryEdgesCold;
 }
 
+extern bool WriteNewDbgInfoFormatToBitcode;
+extern llvm::cl::opt<bool> UseNewDbgInfoFormat;
+
 namespace {
 
 /// These are manifest constants used by the bitcode writer. They do not need to
@@ -128,6 +131,7 @@ enum {
   FUNCTION_INST_RET_VAL_ABBREV,
   FUNCTION_INST_UNREACHABLE_ABBREV,
   FUNCTION_INST_GEP_ABBREV,
+  FUNCTION_DEBUG_RECORD_VALUE_ABBREV,
 };
 
 /// Abstract class to manage the bitcode writing, subclassed for each bitcode
@@ -3512,25 +3516,95 @@ void ModuleBitcodeWriter::writeFunction(
       NeedsMetadataAttachment |= I.hasMetadataOtherThanDebugLoc();
 
       // If the instruction has a debug location, emit it.
-      DILocation *DL = I.getDebugLoc();
-      if (!DL)
-        continue;
-
-      if (DL == LastDL) {
-        // Just repeat the same debug loc as last time.
-        Stream.EmitRecord(bitc::FUNC_CODE_DEBUG_LOC_AGAIN, Vals);
-        continue;
+      if (DILocation *DL = I.getDebugLoc()) {
+        if (DL == LastDL) {
+          // Just repeat the same debug loc as last time.
+          Stream.EmitRecord(bitc::FUNC_CODE_DEBUG_LOC_AGAIN, Vals);
+        } else {
+          Vals.push_back(DL->getLine());
+          Vals.push_back(DL->getColumn());
+          Vals.push_back(VE.getMetadataOrNullID(DL->getScope()));
+          Vals.push_back(VE.getMetadataOrNullID(DL->getInlinedAt()));
+          Vals.push_back(DL->isImplicitCode());
+          Stream.EmitRecord(bitc::FUNC_CODE_DEBUG_LOC, Vals);
+          Vals.clear();
+          LastDL = DL;
+        }
       }
 
-      Vals.push_back(DL->getLine());
-      Vals.push_back(DL->getColumn());
-      Vals.push_back(VE.getMetadataOrNullID(DL->getScope()));
-      Vals.push_back(VE.getMetadataOrNullID(DL->getInlinedAt()));
-      Vals.push_back(DL->isImplicitCode());
-      Stream.EmitRecord(bitc::FUNC_CODE_DEBUG_LOC, Vals);
-      Vals.clear();
-
-      LastDL = DL;
+      // If the instruction has DPValues attached to it, emit them. Note that
+      // they come after the instruction so that it's easy to attach them again
+      // when reading the bitcode, even though conceptually the debug locations
+      // start "before" the instruction.
+      if (I.hasDbgValues() && WriteNewDbgInfoFormatToBitcode) {
+        /// Try to push the value only (unwrapped), otherwise push the
+        /// metadata wrapped value. Returns true if the value was pushed
+        /// without the ValueAsMetadata wrapper.
+        auto PushValueOrMetadata = [&Vals, InstID,
+                                    this](Metadata *RawLocation) {
+          assert(RawLocation && "RawLocation unexpectedly null in DPValue");
+          if (ValueAsMetadata *VAM = dyn_cast<ValueAsMetadata>(RawLocation)) {
+            SmallVector<unsigned, 2> ValAndType;
+            // If the value is a fwd-ref the type is also pushed. We don't
+            // want the type, so fwd-refs are kept wrapped (pushValueAndType
+            // returns false if the value is pushed without type).
+            if (!pushValueAndType(VAM->getValue(), InstID, ValAndType)) {
+              Vals.push_back(ValAndType[0]);
+              return true;
+            }
+          }
+          // The metadata is a DIArgList, or ValueAsMetadata wrapping a
+          // fwd-ref. Push the metadata ID.
+          Vals.push_back(VE.getMetadataID(RawLocation));
+          return false;
+        };
+
+        // Write out non-instruction debug information attached to this
+        // instruction. Write it after the instruction so that it's easy to
+        // re-attach to the instruction reading the records in.
+        for (DbgRecord &DR : I.DbgMarker->getDbgValueRange()) {
+          if (DPLabel *DPL = dyn_cast<DPLabel>(&DR)) {
+            Vals.push_back(VE.getMetadataID(&*DPL->getDebugLoc()));
+            Vals.push_back(VE.getMetadataID(DPL->getLabel()));
+            Stream.EmitRecord(bitc::FUNC_CODE_DEBUG_RECORD_LABEL, Vals);
+            Vals.clear();
+            continue;
+          }
+
+          // First 3 fields are common to all kinds:
+          //   DILocation, DILocalVariable, DIExpression
+          // dbg_value (FUNC_CODE_DEBUG_RECORD_VALUE)
+          //   ..., LocationMetadata
+          // dbg_value (FUNC_CODE_DEBUG_RECORD_VALUE_SIMPLE - abbrev'd)
+          //   ..., Value
+          // dbg_declare (FUNC_CODE_DEBUG_RECORD_DECLARE)
+          //   ..., LocationMetadata
+          // dbg_assign (FUNC_CODE_DEBUG_RECORD_ASSIGN)
+          //   ..., LocationMetadata, DIAssignID, DIExpression, LocationMetadata
+          DPValue &DPV = cast<DPValue>(DR);
+          Vals.push_back(VE.getMetadataID(&*DPV.getDebugLoc()));
+          Vals.push_back(VE.getMetadataID(DPV.getVariable()));
+          Vals.push_back(VE.getMetadataID(DPV.getExpression()));
+          if (DPV.isDbgValue()) {
+            if (PushValueOrMetadata(DPV.getRawLocation()))
+              Stream.EmitRecord(bitc::FUNC_CODE_DEBUG_RECORD_VALUE_SIMPLE, Vals,
+                                FUNCTION_DEBUG_RECORD_VALUE_ABBREV);
+            else
+              Stream.EmitRecord(bitc::FUNC_CODE_DEBUG_RECORD_VALUE, Vals);
+          } else if (DPV.isDbgDeclare()) {
+            Vals.push_back(VE.getMetadataID(DPV.getRawLocation()));
+            Stream.EmitRecord(bitc::FUNC_CODE_DEBUG_RECORD_DECLARE, Vals);
+          } else {
+            assert(DPV.isDbgAssign() && "Unexpected DbgRecord kind");
+            Vals.push_back(VE.getMetadataID(DPV.getRawLocation()));
+            Vals.push_back(VE.getMetadataID(DPV.getAssignID()));
+            Vals.push_back(VE.getMetadataID(DPV.getAddressExpression()));
+            Vals.push_back(VE.getMetadataID(DPV.getRawAddress()));
+            Stream.EmitRecord(bitc::FUNC_CODE_DEBUG_RECORD_ASSIGN, Vals);
+          }
+          Vals.clear();
+        }
+      }
     }
 
     if (BlockAddress *BA = BlockAddress::lookup(&BB)) {
@@ -3771,7 +3845,17 @@ void ModuleBitcodeWriter::writeBlockInfo() {
         FUNCTION_INST_GEP_ABBREV)
       llvm_unreachable("Unexpected abbrev ordering!");
   }
-
+  {
+    auto Abbv = std::make_shared<BitCodeAbbrev>();
+    Abbv->Add(BitCodeAbbrevOp(bitc::FUNC_CODE_DEBUG_RECORD_VALUE_SIMPLE));
+    Abbv->Add(BitCodeAbbrevOp(BitCodeAbbrevOp::VBR, 7)); // dbgloc
+    Abbv->Add(BitCodeAbbrevOp(BitCodeAbbrevOp::VBR, 7)); // var
+    Abbv->Add(BitCodeAbbrevOp(BitCodeAbbrevOp::VBR, 7)); // expr
+    Abbv->Add(BitCodeAbbrevOp(BitCodeAbbrevOp::VBR, 6)); // val
+    if (Stream.EmitBlockInfoAbbrev(bitc::FUNCTION_BLOCK_ID, Abbv) !=
+        FUNCTION_DEBUG_RECORD_VALUE_ABBREV)
+      llvm_unreachable("Unexpected abbrev ordering! 1");
+  }
   Stream.ExitBlock();
 }
 

llvm/lib/Bitcode/Writer/BitcodeWriterPass.cpp

@@ -18,19 +18,20 @@
 #include "llvm/Pass.h"
 using namespace llvm;
 
+extern bool WriteNewDbgInfoFormatToBitcode;
+
 PreservedAnalyses BitcodeWriterPass::run(Module &M, ModuleAnalysisManager &AM) {
-  // RemoveDIs: there's no bitcode representation of the DbgRecord debug-info,
-  // convert to dbg.values before writing out.
-  bool IsNewDbgInfoFormat = M.IsNewDbgInfoFormat;
-  if (IsNewDbgInfoFormat)
+  bool ConvertToOldDbgFormatForWrite =
+      M.IsNewDbgInfoFormat && !WriteNewDbgInfoFormatToBitcode;
+  if (ConvertToOldDbgFormatForWrite)
     M.convertFromNewDbgValues();
 
   const ModuleSummaryIndex *Index =
       EmitSummaryIndex ? &(AM.getResult<ModuleSummaryIndexAnalysis>(M))
                        : nullptr;
   WriteBitcodeToFile(M, OS, ShouldPreserveUseListOrder, Index, EmitModuleHash);
 
-  if (IsNewDbgInfoFormat)
+  if (ConvertToOldDbgFormatForWrite)
     M.convertToNewDbgValues();
 
   return PreservedAnalyses::all();
@@ -56,16 +57,15 @@ namespace {
     StringRef getPassName() const override { return "Bitcode Writer"; }
 
     bool runOnModule(Module &M) override {
-      // RemoveDIs: there's no bitcode representation of the DbgRecord
-      // debug-info, convert to dbg.values before writing out.
-      bool IsNewDbgInfoFormat = M.IsNewDbgInfoFormat;
-      if (IsNewDbgInfoFormat)
+      bool ConvertToOldDbgFormatForWrite =
+          M.IsNewDbgInfoFormat && !WriteNewDbgInfoFormatToBitcode;
+      if (ConvertToOldDbgFormatForWrite)
         M.convertFromNewDbgValues();
 
       WriteBitcodeToFile(M, OS, ShouldPreserveUseListOrder, /*Index=*/nullptr,
                          /*EmitModuleHash=*/false);
 
-      if (IsNewDbgInfoFormat)
+      if (ConvertToOldDbgFormatForWrite)
         M.convertToNewDbgValues();
       return false;
     }