[llvm] [RemoveDIs] Add bitcode format (PR #83251)

via llvm-commits llvm-commits at lists.llvm.org
Wed Feb 28 03:42:56 PST 2024


llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-lto

Author: Orlando Cazalet-Hyams (OCHyams)

<details>
<summary>Changes</summary>

If --write-experimental-debuginfo-iterators-to-bitcode is true (default true) 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.

---

The bitcode format is converted back to the old debug mode (intrinsics) after loading as not all tools support RemoveDIs yet. This mirrors the approach taken for textual IR. We will eventually remove this back-and-forth conversion, but it's necessary to be able to incrementally update tools.

---

I experimented with a few different schemas and this seemed to work best. Using the unwrapped-value-reference trick used in FUNC_CODE_DEBUG_RECORD_VALUE_SIMPLE  (thanks @<!-- -->jmorse) doesn't have the same payoff for dbg.declares in an unoptimized bitcode file as it does for dbg.values in an optimized one (~1% savings over the intrinsic version instead of ~3%, for one test case, which is obviously only anecdotal, however this makes sense based on the nature of dbg.values/declares too; one dbg.declare is used per variable, which may be replaced with many dbg.values). 


---

Patch is 40.90 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/83251.diff


13 Files Affected:

- (modified) llvm/include/llvm/Bitcode/LLVMBitCodes.h (+11) 
- (modified) llvm/lib/Bitcode/Reader/BitcodeAnalyzer.cpp (+4) 
- (modified) llvm/lib/Bitcode/Reader/BitcodeReader.cpp (+86-3) 
- (modified) llvm/lib/Bitcode/Writer/BitcodeWriter.cpp (+100-18) 
- (modified) llvm/lib/Bitcode/Writer/BitcodeWriterPass.cpp (+10-10) 
- (modified) llvm/lib/Bitcode/Writer/ValueEnumerator.cpp (+100-49) 
- (modified) llvm/lib/IR/BasicBlock.cpp (+5) 
- (modified) llvm/lib/IR/Verifier.cpp (+11) 
- (modified) llvm/lib/Linker/IRMover.cpp (+20-3) 
- (modified) llvm/lib/Transforms/IPO/ThinLTOBitcodeWriter.cpp (+5-4) 
- (added) llvm/test/Bitcode/dbg-record-roundtrip.ll (+127) 
- (modified) llvm/tools/llvm-as/llvm-as.cpp (+7) 
- (modified) llvm/tools/verify-uselistorder/verify-uselistorder.cpp (+11) 


``````````diff
diff --git a/llvm/include/llvm/Bitcode/LLVMBitCodes.h b/llvm/include/llvm/Bitcode/LLVMBitCodes.h
index c6f0ddf29a6da8..528a6baa4870ee 100644
--- a/llvm/include/llvm/Bitcode/LLVMBitCodes.h
+++ b/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 {
diff --git a/llvm/lib/Bitcode/Reader/BitcodeAnalyzer.cpp b/llvm/lib/Bitcode/Reader/BitcodeAnalyzer.cpp
index 7005011980ebc9..81579a8469ddb0 100644
--- a/llvm/lib/Bitcode/Reader/BitcodeAnalyzer.cpp
+++ b/llvm/lib/Bitcode/Reader/BitcodeAnalyzer.cpp
@@ -270,6 +270,10 @@ 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)
     }
   case bitc::VALUE_SYMTAB_BLOCK_ID:
     switch (CodeID) {
diff --git a/llvm/lib/Bitcode/Reader/BitcodeReader.cpp b/llvm/lib/Bitcode/Reader/BitcodeReader.cpp
index 832907a3f53f5f..f223864a4a6605 100644
--- a/llvm/lib/Bitcode/Reader/BitcodeReader.cpp
+++ b/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 {
@@ -6357,6 +6354,81 @@ 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 record");
+      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 record");
+
+      // 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;
+        if (getValueTypePair(Record, Slot, NextValueNo, V, TyID, CurBB))
+          return error("Invalid record");
+        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)
@@ -6636,10 +6708,21 @@ 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);
 
diff --git a/llvm/lib/Bitcode/Writer/BitcodeWriter.cpp b/llvm/lib/Bitcode/Writer/BitcodeWriter.cpp
index 656f2a6ce870f5..8bd5913c7ea79e 100644
--- a/llvm/lib/Bitcode/Writer/BitcodeWriter.cpp
+++ b/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
@@ -3491,25 +3495,93 @@ 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)) {
@@ -3750,7 +3822,17 @@ 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();
 }
 
diff --git a/llvm/lib/Bitcode/Writer/BitcodeWriterPass.cpp b/llvm/lib/Bitcode/Writer/BitcodeWriterPass.cpp
index 93fb2a821dee74..de2396f31f6669 100644
--- a/llvm/lib/Bitcode/Writer/BitcodeWriterPass.cpp
+++ b/llvm/lib/Bitcode/Writer/BitcodeWriterPass.cpp
@@ -18,11 +18,12 @@
 #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 DPValue 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 =
@@ -30,7 +31,7 @@ PreservedAnalyses BitcodeWriterPass::run(Module &M, ModuleAnalysisManager &AM) {
                        : 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 DPValue 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;
     }
diff --git a/llvm/lib/Bitcode/Writer/ValueEnumerator.cpp b/llvm/lib/Bitcode/Writer/ValueEnumerator.cpp
index fccb2a606f7ed9..b2edb75eec8c6e 100644
--- a/llvm/lib/Bitcode/Writer/ValueEnumerator.cpp
+++ b/llvm/lib/Bitcode/Writer/ValueEnumerator.cpp
@@ -134,20 +134,28 @@ static OrderMap orderModule(const Module &M) {
     // Metadata used by instructions is decoded before the actual instructions,
     // so visit any constants used by it beforehand.
     for (const BasicBlock &BB : F)
-      for (const Instruction &I : BB)
-        for (const Value *V : I.operands()) {
-          if (const auto *MAV = dyn_cast<MetadataAsValue>(V)) {
-            if (const auto *VAM =
-                    dyn_cast<ValueAsMetadata>(MAV->getMetadata())) {
+      for (const Instruction &I : BB) {
+        auto OrderConstantFromMetadata = [&](Metadata *MD) {
+          if (const auto *VAM = dyn_cast<ValueAsMetadata>(MD)) {
+            orderConstantValue(VAM->getValue());
+          } else if (const auto *AL = dyn_cast<DIArgList>(MD)) {
+            for (const auto *VAM : AL->getArgs())
               orderConstantValue(VAM->getValue());
-            } else if (const auto *AL =
-                           dyn_cast<DIArgList>(MAV->getMetadata())) {
-              for (const auto *VAM : AL->getArgs())
-                orderConstantValue(VAM->getValue());
-            }
           }
+        };
+
+        for (DPValue &DPV : DPValue::filter(I.getDbgValueRange())) {
+          OrderConstantFromMetadata(DPV.getRawLocation());
+          if (DPV.isDbgAssign())
+            OrderConstantFromMetadata(DPV.getRawAddress());
         }
 
+        for (const Value *V : I.operands()) {
+          if (const auto *MAV = dyn_cast<MetadataAsValue>(V))
+            OrderConstantFromMetadata(MAV->getMetadata());
+        }
+      }
+
     for (const Argument &A : F.args())
       orderValue(&A, OM);
     for (const BasicBlock &BB : F)
@@ -261,33 +269,39 @@ static UseListOrderStack predictUseListOrder(const Module &M) {
   // constants in the last Function they're used in.  Module-level constants
   // have already been visited above.
   for (const Function &F : llvm::reverse(M)) {
+    auto PredictValueOrderFromMetadata = [&](Metadata *MD) {
+      if (const auto *VAM = dyn_cast<ValueAsMetadata>(MD)) {
+        predictValueUseListOrder(VAM->getValue(), &F, OM, Stack);
+      } else if (const auto *AL = dyn_cast<DIArgList>(MD)) {
+        for (const auto *VAM : AL->getArgs())
+          predictValueUseListOrder(VAM->getValue(), &F, OM, Stack);
+      }
+    };
     if (F.isDeclaration())
       continue;
     for (const BasicBlock &BB : F)
       predictValueUseListOrder(&BB, &F, OM, Stack);
     for (const Argument &A : F.args())
       predictValueUseListOrder(&A, &F, OM, Stack);
-    for (const BasicBlock &BB : F)
+    for (const BasicBlock &BB : F) {
       for (const Instruction &I : BB) {
+        for (DPValue &DPV : DPValue::filter(I.getDbgValueRange())) {
+          PredictValueOrderFromMetadata(DPV.getRawLocation());
+          if (DPV.isDbgAssign())
+            PredictValueOrderFromMetadata(DPV.getRawAddress());
+        }
         for (const Value *Op : I.operands()) {
           if (isa<Constant>(*Op) || isa<InlineAsm>(*Op)) // Visit GlobalValues.
             predictValueUseListOrder(Op, &F, OM, Stack);
-          if (const auto *MAV = dyn_cast<MetadataAsValue>(Op)) {
-            if (const auto *VAM =
-                    dyn_cast<ValueAsMetadata>(MAV->getMetadata())) {
-              predictValueUseListOrder(VAM->getValue(), &F, OM, Stack);
-            } else if (const auto *AL =
-                           dyn_cast<DIArgList>(MAV->getMetadata())) {
-              for (const auto *VAM : AL->getArgs())
-                predictValueUseListOrder(VAM->getValue(), &F, OM, Stack);
-            }
-          }
+          if (const auto *MAV = dyn_cast<MetadataAsValue>(Op))
+            PredictValueOrderFromMetadata(MAV->getMetadata());
         }
         if (auto *SVI = dyn_cast<ShuffleVectorInst>(&I))
           predictValueUseListOrder(SVI->getShuffleMaskForBitcode(), &F, OM,
                                    Stack);
         predictValueUseListOrder(&I, &F, OM, Stack);
       }
+    }
   }
 
   // Visit globals last, since the module-level use-list block will be seen
@@ -409,6 +423,41 @@ ValueEnumerator::ValueEnumerator(const Module &M,
 
     for (const BasicBlock &BB : F)
       for (const Instruction &I : BB) {
+        // Local metadata is enumerated during function-incorporation, but
+        // any ConstantAsMetadata arguments in a DIArgList should be examined
+        // now.
+        auto EnumerateNonLocalValuesFromMetadata = [&](Metadata *MD) {
+   ...
[truncated]

``````````

</details>


https://github.com/llvm/llvm-project/pull/83251


More information about the llvm-commits mailing list