[LLVM-C] Add more debug information accessors to GlobalObject and Instruction

[CodaFi]

Summary

Adds missing debug information accessors to GlobalObject. This puts the finishing touches on cloning debug info in the echo tests.

[whitequark]

How is a user of this API supposed to map the unsigned Kind values to something meaningful? We don't have an enum LLVMMetadataKind, do we?

include/llvm-c/Core.h

2150

Very small nit: Global. There's some code out there that generates a higher level API by parsing the C headers and inferring semantics from names, let's keep it easy to maintain.

[CodaFi]

In D51522#1220268, @whitequark wrote:

How is a user of this API supposed to map the unsigned Kind values to something meaningful? We don't have an enum LLVMMetadataKind, do we?

You’re supposed to be able to register your own kinds with LLVMContext::getMDKindID et al. I was planning to wrap those in a follow-up.

include/llvm-c/Core.h

2150

I wanted to distinguish these from their unnamedspaced counterparts that seem to work exclusively with the parallel APIs in llvm::Instruction. Is there some middle ground naming scheme here or should I remove the Global bits and add GlobalObject support to the Instruction one?

[whitequark]

include/llvm-c/Core.h

2150

Oh I just mean the argument name.

[CodaFi]

@whitequark Ping

[whitequark]

LGTM

include/llvm-c/Core.h

  * @see llvm::AllocaInst::setAlignment()
  * @see llvm::LoadInst::setAlignment()
  * @see llvm::StoreInst::setAlignment()
  * @see llvm::GlobalValue::setAlignment()
  */
 void LLVMSetAlignment(LLVMValueRef V, unsigned Bytes);
 
 /**
+ * Sets a metadata attachment, erasing the existing metadata attachment if
+ * it already exists for the given kind.
+ *
+ * @see llvm::GlobalObject::setMetadata()
+ */
+void LLVMGlobalSetMetadata(LLVMValueRef Global, unsigned Kind,
+                           LLVMMetadataRef MD);
+
+/**
+ * Erases a metadata attachment of the given kind if it exists.
+ *
+ * @see llvm::GlobalObject::eraseMetadata()
+ */
+void LLVMGlobalEraseMetadata(LLVMValueRef Global, unsigned Kind);
+
+/**
+ * Removes all metadata attachments from this value.
+ *
+ * @see llvm::GlobalObject::clearMetadata()
+ */
+void LLVMGlobalClearMetadata(LLVMValueRef Glob);
+
+/**
+ * Retrieves an array of metadata entries representing the metadata attached to
+ * this value. The caller is responsible for freeing this array by calling
+ * \c LLVMDisposeValueMetadataEntries.
+ *
+ * @see llvm::GlobalObject::getAllMetadata()
+ */
+LLVMValueMetadataEntry *LLVMGlobalCopyAllMetadata(LLVMValueRef Value,
+                                                  size_t *NumEntries);
+
+/**
+ * Destroys value metadata entries.
+ */
+void LLVMDisposeValueMetadataEntries(LLVMValueMetadataEntry *Entries);
+
+/**
+ * Returns the kind of a value metadata entry at a specific index.
+ */
+unsigned LLVMValueMetadataEntriesGetKind(LLVMValueMetadataEntry *Entries,
+                                         unsigned Index);
+
+/**
+ * Returns the underlying metadata node of a value metadata entry at a
+ * specific index.
+ */
+LLVMMetadataRef
+LLVMValueMetadataEntriesGetMetadata(LLVMValueMetadataEntry *Entries,
+                                    unsigned Index);
+
+/**
  * @}
  */
 
 /**
  * @defgroup LLVMCoreValueConstantGlobalVariable Global Variables
  *
  * This group contains functions that operate on global variable values.
  *

 LLVMValueRef LLVMGetMetadata(LLVMValueRef Val, unsigned KindID);
 
 /**
  * Set metadata associated with an instruction value.
  */
 void LLVMSetMetadata(LLVMValueRef Val, unsigned KindID, LLVMValueRef Node);
 
 /**
+ * Returns the metadata associated with an instruction value, but filters out
+ * all the debug locations.
+ *
+ * @see llvm::Instruction::getAllMetadataOtherThanDebugLoc()
+ */
+LLVMValueMetadataEntry *
+LLVMInstructionGetAllMetadataOtherThanDebugLoc(LLVMValueRef Instr,
+                                               size_t *NumEntries);
+
+/**
  * Obtain the basic block to which an instruction belongs.
  *
  * @see llvm::Instruction::getParent()
  */
 LLVMBasicBlockRef LLVMGetInstructionParent(LLVMValueRef Inst);
 
 /**
  * Obtain the instruction that occurs after the one specified.

include/llvm-c/Types.h

 /**
  * Represents an LLVM Named Metadata Node.
  *
  * This models llvm::NamedMDNode.
  */
 typedef struct LLVMOpaqueNamedMDNode *LLVMNamedMDNodeRef;
 
 /**
+ * Represents an entry in a Global Object's metadata attachments.
+ *
+ * This models std::pair<unsigned, MDNode *>
+ */
+typedef struct LLVMOpaqueValueMetadataEntry LLVMValueMetadataEntry;
+
+/**
  * Represents an LLVM basic block builder.
  *
  * This models llvm::IRBuilder.
  */
 typedef struct LLVMOpaqueBuilder *LLVMBuilderRef;
 
 /**
  * Represents an LLVM debug info builder.

lib/IR/Core.cpp

 }
 
 void LLVMSetMetadata(LLVMValueRef Inst, unsigned KindID, LLVMValueRef Val) {
   MDNode *N = Val ? extractMDNode(unwrap<MetadataAsValue>(Val)) : nullptr;
 
   unwrap<Instruction>(Inst)->setMetadata(KindID, N);
 }
 
+LLVMValueMetadataEntry *
+LLVMInstructionGetAllMetadataOtherThanDebugLoc(LLVMValueRef Value,
+                                               size_t *NumEntries) {
+  return llvm_getMetadata(NumEntries, [&Value](MetadataEntries &Entries) {
+    unwrap<Instruction>(Value)->getAllMetadata(Entries);
+  });
+}
+
 /*--.. Conversion functions ................................................--*/
 
 #define LLVM_DEFINE_VALUE_CAST(name)                                       \
   LLVMValueRef LLVMIsA##name(LLVMValueRef Val) {                           \
     return wrap(static_cast<Value*>(dyn_cast_or_null<name>(unwrap(Val)))); \
   }
 
 LLVM_FOR_EACH_VALUE_SUBCLASS(LLVM_DEFINE_VALUE_CAST)
else if (LoadInst *LI = dyn_cast<LoadInst>(P))
     LI->setAlignment(Bytes);
   else if (StoreInst *SI = dyn_cast<StoreInst>(P))
     SI->setAlignment(Bytes);
   else
     llvm_unreachable(
         "only GlobalValue, AllocaInst, LoadInst and StoreInst have alignment");
 }
 
+struct LLVMOpaqueValueMetadataEntry {
+  unsigned Kind;
+  LLVMMetadataRef Metadata;
+};
+
+using MetadataEntries = SmallVectorImpl<std::pair<unsigned, MDNode *>>;
+static LLVMValueMetadataEntry *
+llvm_getMetadata(size_t *NumEntries,
+                 llvm::function_ref<void(MetadataEntries &)> AccessMD) {
+  SmallVector<std::pair<unsigned, MDNode *>, 8> MVEs;
+  AccessMD(MVEs);
+
+  LLVMOpaqueValueMetadataEntry *Result =
+      static_cast<LLVMOpaqueValueMetadataEntry *>(
+        safe_malloc(MVEs.size() * sizeof(LLVMOpaqueValueMetadataEntry)));
+  for (unsigned i = 0; i < MVEs.size(); ++i) {
+    const auto &ModuleFlag = MVEs[i];
+    Result[i].Kind = ModuleFlag.first;
+    Result[i].Metadata = wrap(ModuleFlag.second);
+  }
+  *NumEntries = MVEs.size();
+  return Result;
+}
+
+LLVMValueMetadataEntry *LLVMGlobalCopyAllMetadata(LLVMValueRef Value,
+                                                  size_t *NumEntries) {
+  return llvm_getMetadata(NumEntries, [&Value](MetadataEntries &Entries) {
+    if (Instruction *Instr = dyn_cast<Instruction>(unwrap(Value))) {
+      Instr->getAllMetadata(Entries);
+    } else {
+      unwrap<GlobalObject>(Value)->getAllMetadata(Entries);
+    }
+  });
+}
+
+unsigned LLVMValueMetadataEntriesGetKind(LLVMValueMetadataEntry *Entries,
+                                         unsigned Index) {
+  LLVMOpaqueValueMetadataEntry MVE =
+      static_cast<LLVMOpaqueValueMetadataEntry>(Entries[Index]);
+  return MVE.Kind;
+}
+
+LLVMMetadataRef
+LLVMValueMetadataEntriesGetMetadata(LLVMValueMetadataEntry *Entries,
+                                    unsigned Index) {
+  LLVMOpaqueValueMetadataEntry MVE =
+      static_cast<LLVMOpaqueValueMetadataEntry>(Entries[Index]);
+  return MVE.Metadata;
+}
+
+void LLVMDisposeValueMetadataEntries(LLVMValueMetadataEntry *Entries) {
+  free(Entries);
+}
+
+void LLVMGlobalSetMetadata(LLVMValueRef Global, unsigned Kind,
+                           LLVMMetadataRef MD) {
+  unwrap<GlobalObject>(Global)->setMetadata(Kind, unwrapDI<MDNode>(MD));
+}
+
+void LLVMGlobalEraseMetadata(LLVMValueRef Global, unsigned Kind) {
+  unwrap<GlobalObject>(Global)->eraseMetadata(Kind);
+}
+
+void LLVMGlobalClearMetadata(LLVMValueRef Global) {
+  unwrap<GlobalObject>(Global)->clearMetadata();
+}
+
 /*--.. Operations on global variables ......................................--*/
 
 LLVMValueRef LLVMAddGlobal(LLVMModuleRef M, LLVMTypeRef Ty, const char *Name) {
   return wrap(new GlobalVariable(*unwrap(M), unwrap(Ty), false,
                                  GlobalValue::ExternalLinkage, nullptr, Name));
 }
 
 LLVMValueRef LLVMAddGlobalInAddressSpace(LLVMModuleRef M, LLVMTypeRef Ty,

test/Bindings/llvm-c/echo.ll

pad2:
   unreachable
 cleanup:
   %cp = cleanuppad within none []
   cleanupret from %cp unwind to caller
 exit:
   ret void
 }
 
-!llvm.module.flags = !{!0}
-!named = !{!1, !2, !3}
+define void @with_debuginfo() !dbg !4 {
+  ret void, !dbg !7
+}
+
+!llvm.dbg.cu = !{!0, !2}
+!llvm.module.flags = !{!3}
 
-!0 = !{i32 2, !"Debug Info Version", i32 3}
-!1 = distinct !{}
-!2 = distinct !{}
-!3 = !{!4}
-!4 = distinct !{}
+!0 = distinct !DICompileUnit(language: DW_LANG_C, file: !1, producer: "", isOptimized: true, runtimeVersion: 0, emissionKind: FullDebug)
+!1 = !DIFile(filename: "echo.ll", directory: "/llvm/test/Bindings/llvm-c/echo.ll")
+!2 = distinct !DICompileUnit(language: DW_LANG_C, file: !1, producer: "", isOptimized: true, runtimeVersion: 0, emissionKind: FullDebug)
+!3 = !{i32 2, !"Debug Info Version", i32 3}
+!4 = distinct !DISubprogram(name: "with_debuginfo", linkageName: "_with_debuginfo", scope: null, file: !1, line: 42, type: !5, isLocal: false, isDefinition: true, scopeLine: 1519, flags: DIFlagPrototyped, isOptimized: true, unit: !0, templateParams: !6, retainedNodes: !6)
+!5 = !DISubroutineType(types: !6)
+!6 = !{}
+!7 = !DILocation(line: 42, scope: !8, inlinedAt: !11)
+!8 = distinct !DILexicalBlock(scope: !9, file: !1, line: 42, column: 12)
+!9 = distinct !DISubprogram(name: "fake_inlined_block", linkageName: "_fake_inlined_block", scope: null, file: !1, line: 82, type: !5, isLocal: false, isDefinition: true, scopeLine: 82, flags: DIFlagPrototyped, isOptimized: true, unit: !2, templateParams: !6, retainedNodes: !6)
+!10 = distinct !DILocation(line: 84, scope: !8, inlinedAt: !11)
+!11 = !DILocation(line: 42, scope: !4)

tools/llvm-c-test/echo.cpp

 //
 // This command uses the C API to read a module and output an exact copy of it
 // as output. It is used to check that the resulting module matches the input
 // to validate that the C API can read and write modules properly.
 //
 //===----------------------------------------------------------------------===//
 
 #include "llvm-c-test.h"
+#include "llvm-c/DebugInfo.h"
 #include "llvm-c/Target.h"
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/Support/ErrorHandling.h"
 
 #include <stdio.h>
 #include <stdlib.h>
 
 using namespace llvm;
switch(Op) {
         break;
     }
 
     if (Dst == nullptr) {
       fprintf(stderr, "%d is not a supported opcode\n", Op);
       exit(-1);
     }
 
+    auto Ctx = LLVMGetModuleContext(M);
+    size_t NumMetadataEntries;
+    auto *AllMetadata =
+        LLVMInstructionGetAllMetadataOtherThanDebugLoc(Src,
+                                                       &NumMetadataEntries);
+    for (unsigned i = 0; i < NumMetadataEntries; ++i) {
+      unsigned Kind = LLVMValueMetadataEntriesGetKind(AllMetadata, i);
+      LLVMMetadataRef MD = LLVMValueMetadataEntriesGetMetadata(AllMetadata, i);
+      LLVMSetMetadata(Dst, Kind, LLVMMetadataAsValue(Ctx, MD));
+    }
+    LLVMDisposeValueMetadataEntries(AllMetadata);
+    LLVMSetInstDebugLocation(Builder, Dst);
+
     check_value_kind(Dst, LLVMInstructionValueKind);
     return VMap[Src] = Dst;
   }
 
   LLVMBasicBlockRef DeclareBB(LLVMBasicBlockRef Src) {
     // Check if this is something we already computed.
     {
       auto i = BBMap.find(Src);
while (true) {
     const char *Name = LLVMGetValueName2(Cur, &NameLen);
     LLVMValueRef G = LLVMGetNamedGlobal(M, Name);
     if (!G)
       report_fatal_error("GlobalVariable must have been declared already");
 
     if (auto I = LLVMGetInitializer(Cur))
       LLVMSetInitializer(G, clone_constant(I, M));
 
+    size_t NumMetadataEntries;
+    auto *AllMetadata = LLVMGlobalCopyAllMetadata(Cur, &NumMetadataEntries);
+    for (unsigned i = 0; i < NumMetadataEntries; ++i) {
+      unsigned Kind = LLVMValueMetadataEntriesGetKind(AllMetadata, i);
+      LLVMMetadataRef MD = LLVMValueMetadataEntriesGetMetadata(AllMetadata, i);
+      LLVMGlobalSetMetadata(G, Kind, MD);
+    }
+    LLVMDisposeValueMetadataEntries(AllMetadata);
+    
     LLVMSetGlobalConstant(G, LLVMIsGlobalConstant(Cur));
     LLVMSetThreadLocal(G, LLVMIsThreadLocal(Cur));
     LLVMSetExternallyInitialized(G, LLVMIsExternallyInitialized(Cur));
     LLVMSetLinkage(G, LLVMGetLinkage(Cur));
     LLVMSetSection(G, LLVMGetSection(Cur));
     LLVMSetVisibility(G, LLVMGetVisibility(Cur));
     LLVMSetUnnamedAddress(G, LLVMGetUnnamedAddress(Cur));
     LLVMSetAlignment(G, LLVMGetAlignment(Cur));
if (LLVMHasPersonalityFn(Cur)) {
       const char *FName = LLVMGetValueName2(LLVMGetPersonalityFn(Cur),
                                            &FNameLen);
       LLVMValueRef P = LLVMGetNamedFunction(M, FName);
       if (!P)
         report_fatal_error("Could not find personality function");
       LLVMSetPersonalityFn(Fun, P);
     }
 
+    size_t NumMetadataEntries;
+    auto *AllMetadata = LLVMGlobalCopyAllMetadata(Cur, &NumMetadataEntries);
+    for (unsigned i = 0; i < NumMetadataEntries; ++i) {
+      unsigned Kind = LLVMValueMetadataEntriesGetKind(AllMetadata, i);
+      LLVMMetadataRef MD = LLVMValueMetadataEntriesGetMetadata(AllMetadata, i);
+      LLVMGlobalSetMetadata(Fun, Kind, MD);
+    }
+    LLVMDisposeValueMetadataEntries(AllMetadata);
+
     FunCloner FC(Cur, Fun);
     FC.CloneBBs(Cur);
 
     Next = LLVMGetNextFunction(Cur);
     if (Next == nullptr) {
       if (Cur != End)
         report_fatal_error("Last function does not match End");
       break;