[ntuple] create streamer infos from TClass names

Get the streamer info records of class fields from their meta (TClass)
normalized name. Previously, the names came from the descriptor, which
has them already RNTuple normalized.

Author: jblomer

tree/ntuple/inc/ROOT/RPageStorage.hxx

@@ -521,7 +521,7 @@ public:
 
    /// Updates the descriptor and calls InitImpl() that handles the backend-specific details (file, DAOS, etc.)
    void InitImpl(RNTupleModel &model) final;
-   void UpdateSchema(const ROOT::Internal::RNTupleModelChangeset &changeset, ROOT::NTupleSize_t firstEntry) final;
+   void UpdateSchema(const ROOT::Internal::RNTupleModelChangeset &changeset, ROOT::NTupleSize_t firstEntry) override;
    void UpdateExtraTypeInfo(const ROOT::RExtraTypeInfoDescriptor &extraTypeInfo) final;
 
    /// Initialize sink based on an existing descriptor and fill into the descriptor builder, optionally copying over

tree/ntuple/inc/ROOT/RPageStorageFile.hxx

@@ -65,6 +65,10 @@ private:
    std::unique_ptr<ROOT::Internal::RNTupleFileWriter> fWriter;
    /// Number of bytes committed to storage in the current cluster
    std::uint64_t fNBytesCurrentCluster = 0;
+   /// On UpdateSchema(), the new class fields register the corresponding streamer info here so that the
+   /// streamer info records in the file can be properly updated on dataset commit
+   ROOT::Internal::RNTupleSerializer::StreamerInfoMap_t fInfosOfClassFields;
+
    RPageSinkFile(std::string_view ntupleName, const ROOT::RNTupleWriteOptions &options);
 
    /// We pass bytesPacked so that TFile::ls() reports a reasonable value for the compression ratio of the corresponding
@@ -98,6 +102,8 @@ public:
    RPageSinkFile(RPageSinkFile &&) = default;
    RPageSinkFile &operator=(RPageSinkFile &&) = default;
    ~RPageSinkFile() override;
+
+   void UpdateSchema(const ROOT::Internal::RNTupleModelChangeset &changeset, ROOT::NTupleSize_t firstEntry) final;
 }; // class RPageSinkFile
 
 // clang-format off

tree/ntuple/src/RPageStorageFile.cxx

@@ -30,6 +30,7 @@
 #include <RVersion.h>
 #include <TDirectory.h>
 #include <TError.h>
+#include <TVirtualStreamerInfo.h>
 
 #include <algorithm>
 #include <cstdio>
@@ -85,6 +86,37 @@ void ROOT::Internal::RPageSinkFile::InitImpl(unsigned char *serializedHeader, st
    fWriter->WriteNTupleHeader(zipBuffer.get(), szZipHeader, length);
 }
 
+void ROOT::Internal::RPageSinkFile::UpdateSchema(const ROOT::Internal::RNTupleModelChangeset &changeset,
+                                                 ROOT::NTupleSize_t firstEntry)
+{
+   RPagePersistentSink::UpdateSchema(changeset, firstEntry);
+
+   auto fnAddStreamerInfo = [this](const ROOT::RFieldBase *field) {
+      auto clField = dynamic_cast<const RClassField *>(field);
+      if (!clField)
+         return;
+
+      const auto clName = clField->GetClass()->GetName();
+      auto cl = TClass::GetClass(clName);
+      if (!cl) {
+         throw RException(R__FAIL(std::string("cannot get TClass for ") + clName));
+      }
+      auto streamerInfo = cl->GetStreamerInfo(clField->GetTypeVersion());
+      if (!streamerInfo) {
+         throw RException(R__FAIL(std::string("cannot get streamerInfo for ") + clName + " [" +
+                                  std::to_string(clField->GetTypeVersion()) + "]"));
+      }
+      fInfosOfClassFields[streamerInfo->GetNumber()] = streamerInfo;
+   };
+
+   for (const auto field : changeset.fAddedFields) {
+      fnAddStreamerInfo(field);
+      for (const auto &subField : *field) {
+         fnAddStreamerInfo(&subField);
+      }
+   }
+}
+
 inline ROOT::RNTupleLocator
 ROOT::Internal::RPageSinkFile::WriteSealedPage(const RPageStorage::RSealedPage &sealedPage, std::size_t bytesPacked)
 {
@@ -244,7 +276,18 @@ ROOT::Internal::RPageSinkFile::CommitClusterGroupImpl(unsigned char *serializedP
 
 void ROOT::Internal::RPageSinkFile::CommitDatasetImpl(unsigned char *serializedFooter, std::uint32_t length)
 {
-   fWriter->UpdateStreamerInfos(fDescriptorBuilder.BuildStreamerInfos());
+   // Add the streamer info records from streamer fields: because of runtime polymorphism we may need to add additional
+   // types not covered by the type names of the class fields
+   for (const auto &extraTypeInfo : fDescriptorBuilder.GetDescriptor().GetExtraTypeInfoIterable()) {
+      if (extraTypeInfo.GetContentId() != EExtraTypeInfoIds::kStreamerInfo)
+         continue;
+      // Ideally, we would avoid deserializing the streamer info records of the streamer fields that we just serialized.
+      // However, this happens only once at the end of writing and only when streamer fields are used, so the
+      // preference here is for code simplicity.
+      fInfosOfClassFields.merge(RNTupleSerializer::DeserializeStreamerInfos(extraTypeInfo.GetContent()).Unwrap());
+   }
+   fWriter->UpdateStreamerInfos(fInfosOfClassFields);
+
    auto bufFooterZip = MakeUninitArray<unsigned char>(length);
    auto szFooterZip =
       RNTupleCompressor::Zip(serializedFooter, length, GetWriteOptions().GetCompression(), bufFooterZip.get());

tree/ntuple/test/ntuple_emulated.cxx

@@ -32,6 +32,12 @@ TEST(RNTupleEmulated, EmulatedFields_Simple)
       auto model = RNTupleModel::Create();
       model->AddField(RFieldBase::Create("f", "Outer_Simple").Unwrap());
 
+      // TStreamerInfo::Build will report a warning for interpreted classes (but only for members).
+      // See also https://github.com/root-project/root/issues/9371
+      ROOT::TestSupport::CheckDiagsRAII diagRAII;
+      diagRAII.optionalDiag(kWarning, "TStreamerInfo::Build", "has no streamer or dictionary",
+                            /*matchFullMessage=*/false);
+
       auto writer = RNTupleWriter::Recreate(std::move(model), "ntpl", fileGuard.GetPath());
       writer->Fill();
 
@@ -42,13 +48,6 @@ TEST(RNTupleEmulated, EmulatedFields_Simple)
       ProcessLine("ptrOuter->fInner.fInt2 = 82;");
       ProcessLine("ptrOuter->fInt1 = 93;");
       writer->Fill();
-
-      // TStreamerInfo::Build will report a warning for interpreted classes (but only for members).
-      // See also https://github.com/root-project/root/issues/9371
-      ROOT::TestSupport::CheckDiagsRAII diagRAII;
-      diagRAII.optionalDiag(kWarning, "TStreamerInfo::Build", "has no streamer or dictionary",
-                            /*matchFullMessage=*/false);
-      writer.reset();
    });
 
    auto reader = RNTupleReader::Open("ntpl", fileGuard.GetPath());
@@ -145,6 +144,12 @@ TEST(RNTupleEmulated, EmulatedFields_Vecs)
       auto model = RNTupleModel::Create();
       model->AddField(RFieldBase::Create("outers", "std::vector<Outer_Vecs>").Unwrap());
 
+      // TStreamerInfo::Build will report a warning for interpreted classes (but only for members).
+      // See also https://github.com/root-project/root/issues/9371
+      ROOT::TestSupport::CheckDiagsRAII diagRAII;
+      diagRAII.optionalDiag(kWarning, "TStreamerInfo::Build", "has no streamer or dictionary",
+                            /*matchFullMessage=*/false);
+
       auto writer = RNTupleWriter::Recreate(std::move(model), "ntpl", fileGuard.GetPath());
       writer->Fill();
 
@@ -155,13 +160,6 @@ TEST(RNTupleEmulated, EmulatedFields_Vecs)
       ProcessLine("(*ptrOuters)[0].fInners.push_back(Inner_Vecs{42.f});");
       ProcessLine("(*ptrOuters)[0].fInner.fFlt = 84.f;");
       writer->Fill();
-
-      // TStreamerInfo::Build will report a warning for interpreted classes (but only for members).
-      // See also https://github.com/root-project/root/issues/9371
-      ROOT::TestSupport::CheckDiagsRAII diagRAII;
-      diagRAII.optionalDiag(kWarning, "TStreamerInfo::Build", "has no streamer or dictionary",
-                            /*matchFullMessage=*/false);
-      writer.reset();
    });
 
    auto reader = RNTupleReader::Open("ntpl", fileGuard.GetPath());
@@ -318,15 +316,14 @@ TEST(RNTupleEmulated, EmulatedFields_EmptyStruct)
       auto model = RNTupleModel::Create();
       model->AddField(RFieldBase::Create("f", "Outer_EmptyStruct").Unwrap());
 
-      auto writer = RNTupleWriter::Recreate(std::move(model), "ntpl", fileGuard.GetPath());
-      writer->Fill();
-
       // TStreamerInfo::Build will report a warning for interpreted classes (but only for members).
       // See also https://github.com/root-project/root/issues/9371
       ROOT::TestSupport::CheckDiagsRAII diagRAII;
       diagRAII.optionalDiag(kWarning, "TStreamerInfo::Build", "has no streamer or dictionary",
                             /*matchFullMessage=*/false);
-      writer.reset();
+
+      auto writer = RNTupleWriter::Recreate(std::move(model), "ntpl", fileGuard.GetPath());
+      writer->Fill();
    });
 
    auto reader = RNTupleReader::Open("ntpl", fileGuard.GetPath());