Re: pgsql: Generational memory allocator

On 2017-11-25 12:00:15 -0800, Andres Freund wrote:
> Hi,
> 
> On 2017-11-25 14:50:41 -0500, Tom Lane wrote:
> > I wrote:
> > > Tomas Vondra <tv@fuzzy.cz> writes:
> > >> BTW I also see these failures in hstore:
> >
> > >> ==15168== Source and destination overlap in memcpy(0x5d0fed0, 0x5d0fed0, 40)
> > >> ==15168==    at 0x4C2E00C: memcpy@@GLIBC_2.14 (vg_replace_strmem.c:1018)
> > >> ==15168==    by 0x15419A06: hstoreUniquePairs (hstore_io.c:343)
> > >> ==15168==    by 0x15419EE4: hstore_in (hstore_io.c:416)
> >
> > > Huh ...
> >
> > I tried to duplicate this on my RHEL6 workstation, and failed to,
> > even though adding an assertion easily proves that the hstore
> > regression test does exercise the case.  So apparently the answer
> > as to why skink isn't reporting this is just "not all versions of
> > valgrind check it".
> 
> I suspect that the issue rather is that the compiler will sometimes
> replace the memcpy() with an in-line member-by-member version. That'll
> not be visible as a memcpy to valgrind.

That's indeed the case. Here's the disassembly from skink, albeit for
v10, because those objects were currently present:

disassemble /s hstoreUniquePairs
...
342                             res++;  0x00000000000005c2 <+174>:   add    $0x28,%rbx

343                             memcpy(res, ptr, sizeof(Pairs));  0x00000000000005c6 <+178>:   mov    (%r12),%rax
0x00000000000005ca<+182>:   mov    0x8(%r12),%rdx  0x00000000000005cf <+187>:   mov    %rax,(%rbx)  0x00000000000005d2
<+190>:  mov    %rdx,0x8(%rbx)  0x00000000000005d6 <+194>:   mov    0x10(%r12),%rax  0x00000000000005db <+199>:   mov
0x18(%r12),%rdx  0x00000000000005e0 <+204>:   mov    %rax,0x10(%rbx)  0x00000000000005e4 <+208>:   mov
%rdx,0x18(%rbx) 0x00000000000005e8 <+212>:   mov    0x20(%r12),%rax  0x00000000000005ed <+217>:   mov
%rax,0x20(%rbx)

344                     }
345
...

Greetings,

Andres Freund