Data corruption computing character tables

[fingolfin]

See gap-system/gap#3114

This leads to an unexpected break loop when using the Julia GC instead of GASMAN:
g:= AtlasGroup( "U3(8).3^2" );; Irr( g );;

[fingolfin]

A good way to get into a break loop just before the error happens is to apply this patch:

diff --git a/lib/ctblgrp.gi b/lib/ctblgrp.gi
index 0c45458005..21430b574b 100644
--- a/lib/ctblgrp.gi
+++ b/lib/ctblgrp.gi
@@ -1768,7 +1768,11 @@ StandardClassMatrixColumn := function(D,M,r,t)
 
     T:=DoubleCentralizerOrbit(D,r,t);
     Info(InfoCharacterTable,2,Length(T[1])," instead of ",D.classiz[r]);
-
+if t = 37 and Length(T[1]) = 1700 and D.classiz[r] = 96768 then
+    Print("HANDLE_OBJ(T) = ", HANDLE_OBJ(T), " = 0x", HexStringInt(HANDLE_OBJ(T)), "\n");
+    Print("HANDLE_OBJ(T[2]) = ", HANDLE_OBJ(T[2]), " = 0x", HexStringInt(HANDLE_OBJ(T[2])), "\n");
+    Error("Breakpoint");
+fi;
     if IsDxLargeGroup(D.group) then
       # if r and t are unique,the conjugation test can be weak (i.e. up to
       # galois automorphisms)

From that break prompt, I can trigger the issue just by printing T[2]; it seems that this triggers a GC, and apparently the master pointer for T[2] is freed, and then reused. To be clear: The content of the plist T of length 2 stays unchanged, but the master pointer stored in its second slot now points to another bag. So HANDLE_OBJ(T[2]) prints the same value, but the object it refers to has changed -- usually to a string.

Using watchpoints in lldb, I traced when this change happens, and it was in .... NewBag! Specifically, in SET_PTR_BAG(bag, DATA(header));. This is why I say it looks as if the master pointer object was freed and its memory reused.

I tried to prevent this, by assigning T[2] to a gobal variable from that break loop before resuming, but that did not stop the object from being collected.

[fingolfin]

Resolved by gap-system/gap#3199