Align the runner structures on a memory bound.

That fixes #395 (closed) for me.

I don't know whether what we had was a bug in the first place but a posix_memalign there does not harm in any way.

changed the description

Didn't work for me. I also needed to redefine SWIFT_STRUCT_ALIGNMENT to 64, then all the various warnings disappeared.

Did the warnings on your machine say that the structs required 64 byte alignment?

It did but it also worked as such both with gcc 5.4.1 and gcc 6.3.0. I ran with --enable-debug CFLAGS="-fsanitize=undefined".

But I am happy to replace SWIFT_STRUCT_ALIGNMENT by SWIFT_CACHE_ALIGNMENT in the call to posix_memalign.

I see, yes that also does the trick for me as well, so we should use that.

Now for the next related issue. If I use an MPI job instead. Now we get:

cell.c:175:13: runtime error: member access within misaligned address 0x000002d70610 for type 'struct pcell', which requires 32 byte alignment
0x000002d70610: note: pointer points here
 00 00 00 00  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  00 00 00 00
              ^ 
.....

also:

cell.c:373:13: runtime error: store to misaligned address 0x00000444c090 for type 'struct gravity_tensors', which requires 32 byte alignment
0x00000444c090: note: pointer points here
 00 00 00 00  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  00 00 00 00
              ^ 
.....

added 1 commit

• e3ef36a3 - Also align the proxy-cells and the proxy-multipoles.

Compare with previous version

I am not quite sure why alignment would be required here either but there is again no reason not to do it.

added 1 commit

• 8f24ed26 - Add --enable-undefined-sanitizer as a configure option.

Compare with previous version

Thanks, that seems to have squashed all those warnings. I've added a new configure option --enable-undefined-sanitizer as we should probably expect to not raise these issues.

merged

mentioned in commit 733dbc51