Repeated calls to TClass::GetClass/TClassEdit::ResolveTypedef hoard memory

[amete]

Hi everyone,

We came across this issue in ATLAS recently. There were similar reports in the past, such as ROOT-7254, but there is none that's still open as far as I could see.

Describe the bug

Repeatedly calling TClass::GetClass/TClassEdit::ResolveTypedef w/ an input such as std::vector<int>::value_size seems to create memory hoarding.

To Reproduce

Either of these python scripts should reproduce the issue:

import ROOT

while True:
    ROOT.TClass.GetClass("std::vector<int>::value_size")

import ROOT

while True:
    ROOT.TClassEdit.ResolveTypedef("std::vector<int>::value_size")

Setup

Tested with both ROOT versions 6.24/00 and 6.24/06 in the ATLAS production environment.

Please let us know, many thanks.

Best,
Serhan

[eguiraud]

At a quick glance it looks like TClass::GetClass calls the interpreter and jitting allocates a little bit every time, notoriously never giving that memory back until the end of the application. I can't tell whether TClass::GetClass could avoid any and all jitting on subsequent calls, but it seems reasonable to think so. @Axel-Naumann @pcanal ?

[eguiraud]

In case it helps here's the output of massif for the first of the two reproducers: https://cernbox.cern.ch/index.php/s/aACGp2JdPzAQUfW (it should be readable with ms_print, the most relevant frame is the very last one)

[akmorley]

I was wondering if there was a time scale for this issue being fixed as it impacts one of the calibration workflows in ATLAS?