MPS issue job003683

TitleEpoch number can wrap around on 32-bit platforms
Assigned userGareth Rees
DescriptionThe epoch number is stored in an Epoch, which is a typedef for Size, and on 32-bit platforms (such as Windows) this is a 32-bit number. There are about 2^32 centiseconds in a year, so a long-running application that collects frequently can cause this number to wrap around on a timescale of years.
AnalysisWe could change Epoch so that it gets 64 bits even on 32-bit Windows (the type needs to be unsigned long long to get 64 bits in Visual Studio). The only places where an epoch number is stored is in the ArenaStruct [1] and in mps_ld_s [2] (where mps_word_t is used), so the memory consequences are not severe (an additional 4 bytes for each client structure containing an mps_ld_s).

When the epoch number overflows, what will actually go wrong?
* LDAge: fine.
* LDIsStaleAny: the assertion AVER(ld->_epoch <= arena->epoch) will go off. But if that is ignored, the function seems OK: the subtraction arena->epoch - ld->_epoch will wrap around correctly.
* LDMerge: two assertions will go off. And if those are ignored, it can pick the wrong epoch for the merged LD (it uses the test "from->_epoch < ld->_epoch", but this will no longer be reliable). Incorrect merges might happen less often if the test were changed to "arena->epoch - from->_epoch > arena->epoch - ld->_epoch".
How foundinspection
Evidence[1] <//>
[2] <//>
Created byGareth Rees
Created on2014-03-07 09:55:05
Last modified byGareth Rees
Last modified on2014-10-25 23:14:47
History2014-03-07 GDR Created.