This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: Obscure crashes due to gcc 4.9 -O2 => -fisolate-erroneous-paths-dereference
- From: Richard Biener <richard dot guenther at gmail dot com>
- To: Jeff Law <law at redhat dot com>
- Cc: Martin Sebor <msebor at redhat dot com>, Jakub Jelinek <jakub at redhat dot com>, Florian Weimer <fweimer at redhat dot com>, Sandra Loosemore <sandra at codesourcery dot com>, Jeff Prothero <jprother at altera dot com>, GCC Development <gcc at gcc dot gnu dot org>
- Date: Wed, 4 Mar 2015 13:36:21 +0100
- Subject: Re: Obscure crashes due to gcc 4.9 -O2 => -fisolate-erroneous-paths-dereference
- Authentication-results: sourceware.org; auth=none
- References: <pdf61azt48b dot fsf at sj-interactive3 dot altera dot com> <20150218192943 dot GR1746 at tucnak dot redhat dot com> <54E64DFF dot 8030100 at codesourcery dot com> <54E71534 dot 8070805 at redhat dot com> <20150220121014 dot GY1746 at tucnak dot redhat dot com> <54E76865 dot 4010305 at redhat dot com> <54F61214 dot 7030407 at redhat dot com> <54F645DB dot 1080903 at redhat dot com>
On Wed, Mar 4, 2015 at 12:38 AM, Jeff Law <law@redhat.com> wrote:
> On 03/03/15 12:57, Martin Sebor wrote:
>>
>>
>> As a data point(*) it might be interesting to note that GCC itself
>> relies on memcpy providing stronger guarantees than the C standard
>> requires it to by emitting calls to the function for large structure
>> self-assignments (which are strictly conforming, as discussed in bug
>> 65029).
>
> Right. I actually spent quite a bit of time struggling with this a while
> back in a different context. The only case I could come up with where GCC
> would generate an overlapping memcpy was self assignment, but even that was
> bad and while we ultimately punted, I've always considered it a wart.
?
struct A { int large[100]; };
void foo (struct A *x, struct A *y)
{
*x = *y;
}
call it as foo (&a, &a); (on x86 you need -mstringop-strategy=libcall,
even at -O0, to emit a memcpy call)
The self-assignment doesn't have to be visible to the compiler - so
to fix this we'd have to assume pointer equality everywhere and
either emit a conditional call to memcpy or always emit a call to
memmove.
Richard.
>
> [*] IMO, one in favor of tightening up the memcpy specification
>>
>> to require implementations to provide the expected semantics.
>
> That works for me :-)
>
> The things done in glibc's memcpy are a bit on the absurd side and the pain
> caused by the changes over time is almost impossible to overstate. If the
> Austin group tightens memcpy to require fewer surprises I think most
> developers would ultimately be happy with the result -- a few would complain
> about the performance impacts for specific workloads, but I suspect they'd
> be in the minority.
>
>
> jeff
>