Re: [Valgrind-users] --partial-loads-ok and aligned SSE loads

[Julian S. <js...@ac...>]

> Background: I am using Valgrind on optimized SSE4.2 output from the
> Intel C Compiler.  I am seeing massive amounts of "Invalid read of
> size 8" false positives.  I believe these false positives would
> disappear if "--partial-loads-ok=yes" worked as documented.  The
> problem is that it does not work for 16-byte SSE loads; Valgrind's
> behavior appears to be to treat such loads as a pair of 8-byte loads.
> 
> Note that GCC has started to emit similar code
> (https://bugs.kde.org/show_bug.cgi?id=294285), and I suspect it will
> only become more common as compilers get smarter.

For whatever it's worth ..

The documentation never claimed that Memcheck will work sanely for
ultra optimised code.  In fact it recommends restricting yourself
to gcc -g -O1, which works fairly well in practice as a tradeoff
between performance and debuggability.

That said, I did do some work to make --partial-loads-ok=yes
work correctly for 16 byte SSE loads.  But it never got 
committed, and is still sitting in some tree around here somewhere.
The main problem, as you surmise, is that such loads get
treated as two 8 byte loads, and so there's no easy way for
--partial-loads-ok=yes to do the right thing.  I think the hack
I did makes it possible for mc_main.c to know whether an 8 byte
load is really "itself", or whether it's half of a 16 byte load
(and in that case, which half).

I don't really remember though.  I'd have to look over the diff
again.

J