Crash in TROOT::EndOfProcessCleanups if created a classic TRootBrowser in python session through c++

[will-cern]

Check duplicate issues.

• Checked for duplicates

Description

This a minimal reproducer (I think) of an issue I have been seeing with the xRooBrowser, which uses the classic TRootBrowser for its interface still. This reproducer uses just ROOT classes though, so I dont think its an xRooFit issue at its heart.
The issue seems to be that when the ListOfBrowsers is deleted, the TCanvasImp from the canvas in the browser gets deleted, but the pointer held by the TCanvas object doesn't get zeroed. Strangely, if one Deletes the Browsers list from the python prompt, the imp is zeroed (see reproducer below)
I am trying to figure out the best solution to this problem here, but thought I'd post the problem in case someone else can figure out the right solution.

Reproducer

import ROOT
ROOT.gInterpreter.ProcessLine("gEnv->SetValue(\"Browser.Name\", \"TRootBrowser\"); auto b = new TBrowser")
exit()

crashes inside TROOT::EndOfProcessCleanups() at the stage where it deletes the Canvases

But the following will work fine:

import ROOT
ROOT.gInterpreter.ProcessLine("gEnv->SetValue(\"Browser.Name\", \"TRootBrowser\"); auto b = new TBrowser")
ROOT.gROOT.GetListOfBrowsers().Delete()
exit()

ROOT version

master

Installation method

build from source

Operating system

MacOS (at least)

Additional context

No response

[vepadulano]

Dear @will-cern ,

Thanks for reaching out! While not discarding this issue, which will indeed need investigation, I still wanted to ask you a bit more details about your context. What is the motivation to launch a TBrowser from the interpreter like that, as opposed to just creating one from Python directly?

[will-cern]

This is more of a minimal reproducer than anything! The real-world case of launching of the TBrowser happens via some C++ code that is called from python through the pyROOT bindings. The examples that manifests it with xRooFit is:

import ROOT.Experimental.XRooFit as XRF
w = XRF.xRooNode()
w.Browse() # this creates a TBrowser
exit()

In the meantime, I made some progress understanding why the CanvasImp isn't getting zeroed if the browser is still open when python exits. Or rather, the opposite: closing the browser before python exits (including deleting the ListOfBrowsers as given in the reproducer) seems to call TRootCanvas::ReallyDelete(), which zeros the CanvasImp pointer. So one possible solution is to ensure that the call to fBrowsers.Delete() that happens in the end of process cleanup also ends up calling ReallyDelete on the canvases inside the browsers...