[cling] Memory hogging when checking if type is an enum

[etejedor]

Initially reported here:

https://root-forum.cern.ch/t/pyroot-and-std-vector-dramatic-ram-usage/49711

when a user was getting a vector<int> branch from a tree in a loop in PyROOT.

After a bit of digging, it seems the culprit is a call that is made from cppyy to TInterpreter::ClassInfo_IsEnum. The following is a C++ reproducer (tested with current master):

float getMem() {
        auto info = ProcInfo_t();
        gSystem->GetProcInfo(&info);
        float mem = (float)info.fMemResident;
        return mem*1e-3;
}

void test()
{
    for (int i=0; i < 100000; ++i) {
        gInterpreter->ClassInfo_IsEnum("vector<int>");
        if (i%1000 == 0)
            printf("RSS at iteration %d is %f\n", i, getMem());
    }
}

When running the macro above, it can be seen how the memory usage increases after each iteration.

[eguiraud]

Related (or maybe the exact same issue): https://github.com/root-project/root/issues/9029

[pcanal]

I don't know if it is an option at that place in the code but calling a combination of TClass::GetClass(name_of_type); and TEnum::GetEnum(name_of_type); should be more memory efficient.

[aescalante]

Dear experts, @etejedor,

I think I am facing a similar memory usage problem in PyROOT (ROOT 6.24/06).

Do you know if this issue has been fixed? If yes, in which release?

Thanks a lot! Cheers, Alberto

[etejedor]

Dear Alberto,

I'm afraid this is not yet fixed, but @Axel-Naumann can confirm.

[Axel-Naumann]

@etejedor https://github.com/root-project/root/issues/10454#issuecomment-1110193565 suggests a change in PyROOT. Can you confirm whether that helps?

[etejedor]

So iiuc the suggestion is to replace this line by something like return TEnum::GetEnum(tn_short.c_str());? Where should TClass::GetClass be used here and for what purpose?

[Axel-Naumann]

Something along the lines of

    if (TClass::GetClass(tn_short.c_str()) return false;
    return gInterpreter->ClassInfo_IsEnum(tn_short.c_str());

maybe?

[etejedor]

Thanks @Axel-Naumann , so TClass::GetClass is used to discard classes faster instead of just going directly for TEnum::GetEnum? (I assume the second line in your code snippet meant to use TEnum::GetEnum and not the current ClassInfo_IsEnum, which seems to be behind the memory hogging).

[Axel-Naumann]

Right - TEnum::GetEnum() is doing these things already, unlike ClassInfo_IsEnum(). So just using TEnum::GetEnum() instead of ClassInfo_IsEnum() would be enough!

[etejedor]

TEnum::GetEnum() does not seem to suffer from the same issue. https://github.com/root-project/root/pull/11412 replaces the use of ClassInfo_IsEnum() by TEnum::GetEnum() in clingwrapper so there is no memory hogging in PyROOT for this case.

[etejedor]

#11412 revealed a test failure caused by the switch to TEnum::GetEnum. This is a reproducer in C++, extracted from the test that fails:

enum EFruit {kApple=78, kBanana=29, kCitrus=34};

void repro()
{
    auto type = "std::vector<EFruit>::value_type";
    cout << gInterpreter->ClassInfo_IsEnum(type) << endl;
    cout << TEnum::GetEnum(type) << endl;
}

which prints

1
0

The different answer between the two calls is what makes the test fail.

[pcanal]

Another instance of the issue can be seen at: https://root-forum.cern.ch/t/possible-memory-leakage/54519

[pcanal]

in https://github.com/root-project/root/issues/10454#issuecomment-1257779330, the typedef is not being resolved, i.e.

enum EFruit {kApple=78, kBanana=29, kCitrus=34};

void repro()
{
    auto type = "std::vector<EFruit>::value_type";
    cout << gInterpreter->ClassInfo_IsEnum(type) << endl;
    cout << TEnum::GetEnum(type) << endl;
    auto resolved = TClassEdit::ResolveTypedef (type, kTRUE);
    cout << TEnum::GetEnum(resolved.c_str()) << endl;
}

prints

1
0
0x2048750