memset bugs.

[John Reiser]

On 11/25/2009 02:03 PM, Dave Jones wrote:
> On Wed, Nov 25, 2009 at 01:58:38PM -0500, Jakub Jelinek wrote:

>   >  glibc headers warn about this (when -D_FORTIFY_SOURCE=2), so a faster way
>   >  would be just grep through all packages' build.log files (preferrably on the
>   >  box where they are stored to avoid downloading it all).
>
> Can we make it fail the build instead of warning ?
> A zero sized memset is always a bug.

No, memset(,,0) is not always a bug.  A null region is not automatically a bug.
Here is one example:

     struct Foo {
         long x;
         char hole[8 - sizeof(long)];
     } foo;

     memset(&foo.hole, 0, sizeof(foo.hole));

On a LP64-bit machine such as x86_64, this is memset(&foo.hole, 0, 0),
which is *NOT* a bug.

Perhaps the best that can be expected is for the compiler to warn
if _builtin_memset has a third argument which is known to be a compile-time
constant zero.  But such a warning must be optional, for there are
legitimate use cases.  Also, if the second argument to _builtin_memset
is a compile-time constant which cannot be represented in one byte
(considering both signed and unsigned cases) then another optional warning
may be appropriate.

--