[InstrProf] Add vtables with type metadata into symtab (#81051)

The indirect-call-promotion pass will look up the vtable to find out
the virtual function [1],
and add vtable-derived information in icall
candidate [2] for cost-benefit analysis.

[1] https://github.com/llvm/llvm-project/pull/81442/files#diff-a95d1ac8a0da69713fcb3346135d4b219f0a73920318d2549495620ea215191bR395-R416
[2] https://github.com/llvm/llvm-project/pull/81442/files#diff-a95d1ac8a0da69713fcb3346135d4b219f0a73920318d2549495620ea215191bR195-R199

Author: mingmingl-llvm

llvm/include/llvm/ProfileData/InstrProf.h

@@ -17,6 +17,7 @@
 
 #include "llvm/ADT/ArrayRef.h"
 #include "llvm/ADT/BitmaskEnum.h"
+#include "llvm/ADT/DenseMap.h"
 #include "llvm/ADT/IntervalMap.h"
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/StringRef.h"
@@ -470,6 +471,12 @@ class InstrProfSymtab {
   // A map from MD5 keys to function define. We only populate this map
   // when build the Symtab from a Module.
   std::vector<std::pair<uint64_t, Function *>> MD5FuncMap;
+  // A map from MD5 to the global variable. This map is only populated when
+  // building the symtab from a module. Use separate container instances for
+  // `MD5FuncMap` and `MD5VTableMap`.
+  // TODO: Unify the container type and the lambda function 'mapName' inside
+  // add{Func,VTable}WithName.
+  DenseMap<uint64_t, GlobalVariable *> MD5VTableMap;
   // A map from function runtime address to function name MD5 hash.
   // This map is only populated and used by raw instr profile reader.
   AddrHashMap AddrToMD5Map;
@@ -488,12 +495,18 @@ class InstrProfSymtab {
 
   // Add the function into the symbol table, by creating the following
   // map entries:
-  // name-set = {PGOFuncName} + {getCanonicalName(PGOFuncName)} if the canonical
-  // name is different from pgo name
+  // name-set = {PGOFuncName} union {getCanonicalName(PGOFuncName)}
   // - In MD5NameMap: <MD5Hash(name), name> for name in name-set
   // - In MD5FuncMap: <MD5Hash(name), &F> for name in name-set
   Error addFuncWithName(Function &F, StringRef PGOFuncName);
 
+  // Add the vtable into the symbol table, by creating the following
+  // map entries:
+  // name-set = {PGOName} union {getCanonicalName(PGOName)}
+  // - In MD5NameMap:  <MD5Hash(name), name> for name in name-set
+  // - In MD5VTableMap: <MD5Hash(name), name> for name in name-set
+  Error addVTableWithName(GlobalVariable &V, StringRef PGOVTableName);
+
   // If the symtab is created by a series of calls to \c addFuncName, \c
   // finalizeSymtab needs to be called before looking up function names.
   // This is required because the underlying map is a vector (for space
@@ -555,6 +568,7 @@ class InstrProfSymtab {
   Error create(const FuncNameIterRange &FuncIterRange,
                const VTableNameIterRange &VTableIterRange);
 
+  // Map the MD5 of the symbol name to the name.
   Error addSymbolName(StringRef SymbolName) {
     if (SymbolName.empty())
       return make_error<InstrProfError>(instrprof_error::malformed,
@@ -630,6 +644,10 @@ class InstrProfSymtab {
   /// Return function from the name's md5 hash. Return nullptr if not found.
   inline Function *getFunction(uint64_t FuncMD5Hash);
 
+  /// Return the global variable corresponding to md5 hash. Return nullptr if
+  /// not found.
+  inline GlobalVariable *getGlobalVariable(uint64_t MD5Hash);
+
   /// Return the name section data.
   inline StringRef getNameData() const { return Data; }
 
@@ -709,6 +727,12 @@ Function* InstrProfSymtab::getFunction(uint64_t FuncMD5Hash) {
   return nullptr;
 }
 
+GlobalVariable *InstrProfSymtab::getGlobalVariable(uint64_t MD5Hash) {
+  if (auto Iter = MD5VTableMap.find(MD5Hash); Iter != MD5VTableMap.end())
+    return Iter->second;
+  return nullptr;
+}
+
 // To store the sums of profile count values, or the percentage of
 // the sums of the total count values.
 struct CountSumOrPercent {

llvm/lib/ProfileData/InstrProf.cpp

@@ -476,11 +476,43 @@ Error InstrProfSymtab::create(Module &M, bool InLTO) {
       return E;
   }
 
+  SmallVector<MDNode *, 2> Types;
+  for (GlobalVariable &G : M.globals()) {
+    if (!G.hasName() || !G.hasMetadata(LLVMContext::MD_type))
+      continue;
+    if (Error E = addVTableWithName(
+            G, getIRPGOObjectName(G, InLTO, /* PGONameMetadata */ nullptr)))
+      return E;
+  }
+
   Sorted = false;
   finalizeSymtab();
   return Error::success();
 }
 
+Error InstrProfSymtab::addVTableWithName(GlobalVariable &VTable,
+                                         StringRef VTablePGOName) {
+  auto mapName = [&](StringRef Name) -> Error {
+    if (Error E = addSymbolName(Name))
+      return E;
+
+    bool Inserted = true;
+    std::tie(std::ignore, Inserted) =
+        MD5VTableMap.try_emplace(GlobalValue::getGUID(Name), &VTable);
+    if (!Inserted)
+      LLVM_DEBUG(dbgs() << "GUID conflict within one module");
+    return Error::success();
+  };
+  if (Error E = mapName(VTablePGOName))
+    return E;
+
+  StringRef CanonicalName = getCanonicalName(VTablePGOName);
+  if (CanonicalName != VTablePGOName)
+    return mapName(CanonicalName);
+
+  return Error::success();
+}
+
 /// \c NameStrings is a string composed of one of more possibly encoded
 /// sub-strings. The substrings are separated by 0 or more zero bytes. This
 /// method decodes the string and calls `NameCallback` for each substring.

llvm/unittests/ProfileData/InstrProfTest.cpp

@@ -6,6 +6,8 @@
 //
 //===----------------------------------------------------------------------===//
 
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/IR/DerivedTypes.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/LLVMContext.h"
@@ -1730,6 +1732,34 @@ TEST(SymtabTest, instr_prof_symtab_module_test) {
   Function::Create(FTy, Function::WeakODRLinkage, "Wblah", M.get());
   Function::Create(FTy, Function::WeakODRLinkage, "Wbar", M.get());
 
+  // [ptr, ptr, ptr]
+  ArrayType *VTableArrayType = ArrayType::get(
+      PointerType::get(Ctx, M->getDataLayout().getDefaultGlobalsAddressSpace()),
+      3);
+  Constant *Int32TyNull =
+      llvm::ConstantExpr::getNullValue(PointerType::getUnqual(Ctx));
+  SmallVector<llvm::Type *, 1> tys = {VTableArrayType};
+  StructType *VTableType = llvm::StructType::get(Ctx, tys);
+
+  // Create two vtables in the module, one with external linkage and the other
+  // with local linkage.
+  for (auto [Name, Linkage] :
+       {std::pair{"ExternalGV", GlobalValue::ExternalLinkage},
+        {"LocalGV", GlobalValue::InternalLinkage}}) {
+    llvm::Twine FuncName(Name, StringRef("VFunc"));
+    Function *VFunc = Function::Create(FTy, Linkage, FuncName, M.get());
+    GlobalVariable *GV = new llvm::GlobalVariable(
+        *M, VTableType, /* isConstant= */ true, Linkage,
+        llvm::ConstantStruct::get(
+            VTableType,
+            {llvm::ConstantArray::get(VTableArrayType,
+                                      {Int32TyNull, Int32TyNull, VFunc})}),
+        Name);
+    // Add type metadata for the test data, since vtables with type metadata
+    // are added to symtab.
+    GV->addTypeMetadata(16, MDString::get(Ctx, Name));
+  }
+
   InstrProfSymtab ProfSymtab;
   EXPECT_THAT_ERROR(ProfSymtab.create(*M), Succeeded());
 
@@ -1751,6 +1781,23 @@ TEST(SymtabTest, instr_prof_symtab_module_test) {
     EXPECT_EQ(PGOName, PGOFuncName);
     EXPECT_THAT(PGOFuncName.str(), EndsWith(Funcs[I].str()));
   }
+
+  StringRef VTables[] = {"ExternalGV", "LocalGV"};
+  for (auto [VTableName, PGOName] : {std::pair{"ExternalGV", "ExternalGV"},
+                                     {"LocalGV", "MyModule.cpp;LocalGV"}}) {
+    GlobalVariable *GV =
+        M->getGlobalVariable(VTableName, /* AllowInternal=*/true);
+
+    // Test that ProfSymtab returns the expected name given a hash.
+    std::string IRPGOName = getPGOName(*GV);
+    EXPECT_STREQ(IRPGOName.c_str(), PGOName);
+    uint64_t GUID = IndexedInstrProf::ComputeHash(IRPGOName);
+    EXPECT_EQ(IRPGOName, ProfSymtab.getFuncOrVarName(GUID));
+    EXPECT_EQ(VTableName, getParsedIRPGOName(IRPGOName).second);
+
+    // Test that ProfSymtab returns the expected global variable
+    EXPECT_EQ(GV, ProfSymtab.getGlobalVariable(GUID));
+  }
 }
 
 // Testing symtab serialization and creator/deserialization interface