Re: Andreas Seltenreich 2015-09-13 <87si6i1875.fsf@credativ.de>
> I managed disassemble RecordNewMultiXact from the core dump using a
> cross-binutils, and it reveals that the compiler[1] appears to have
> indeed generated a signed division here. I'm attaching a piece of C
> code that does the same computation as the assembly (I think), as well
> as the disassembly itself.
>
> Footnotes:
> [1] Sun C 5.12 SunOS_sparc Patch 148917-07 2013/10/18, 64-bit
Hi,
a short update here: the customer updated the compiler to a newer
version, is now compiling using -O2 instead of -O3, and the code
generated now looks sane, so this turned out to be a compiler issue.
(Though it's unclear if the upgrade fixed it, or the different -O
level.)
Thanks to all who provided feedback, it was very valuable in actually
tracking down the root of the issue.
Christoph