Hi.

In the morning I was hunting valgrind "Invalid read of size 4" in my code and finally I found it. Example:

program str;
uses
  SysUtils;
var
  p: PChar;
begin
  GetMem(p, 1);
  p^ := #0;
  StrScan(p, 'x');
  FreeMem(p);
end.

==14379== Invalid read of size 4
==14379== at 0x80708F0: SYSUTILS_STRSCAN$PCHAR$CHAR$$PCHAR (strings.inc:439) ==14379== by 0x804828A: main (str.pas:9)
==14379== Address 0x416A494 is 4 bytes inside a block of size 5 alloc'd ==14379== at 0x401D38B: malloc (vg_replace_malloc.c:149) ==14379== by 0x8063398: CMEM_CGETMEM$LONGWORD$$POINTER (cmem.pp:50) ==14379== by 0x805AFA6: SYSTEM_GETMEM$POINTER$LONGWORD (heap.inc:235) ==14379== by 0x8048276: main (str.pas:7)

Problem is in assembler code for i386. rtl/i386/strings.inc. This routine reads always 4 bytes alligned by 4 regardless string length.

function strscan(p : pchar;c : char) : pchar;assembler; ...
// load new 4 bytes
movl (%edi),%edx <-here valgrind report Invalid read ...

Problem was discovered aleady here http://bugs.freepascal.org/view.php?id=6314

My assembler knowledge is unfortunately not so good yet to effectively repair it. Besides routine is very sophisticated. :)
But I think that solving this cosmetic error will save many developers hours during valgrind code checking. And preparing testcases fo fpc developers.

Petr

But I think that solving this cosmetic error will save many developers hours during valgrind code checking.

.. and slows down things a lot I fear.

It is a false positive by Valgrind. That code intentionally does not care about reading uninitialised memory from time to time, and treats that situation correctly. There is similar code in glibc, but Valgrind comes with standard suppression files for glibc which suppress these false positives. I'm not aware of any suppression files for FPC.

Jonas

I think that this is not false positive. We allocate 1 byte and read 4, report is right. But obviously effective solution do not exists.

What about to establish valgind suppressions file in fpc tree? Here is first record.
{
Assembler routine always reads 4 bytes alligned by 4 regardles str length Memcheck:Addr4
fun:SYSUTILS_STRSCAN$PCHAR$CHAR$$PCHAR
}

Petr

It is a false positive because users read Valgrind reports as lists of error locations in their program (which is indeed how the MemCheck backend of Valgrind is supposed to be used).

If you submit such a file, we can add it to svn.

Jonas

No problem. Here it is.

Petr

Thanks, committed (to the fpcbuild repository rather than to the fpc repository, because it is an installation-related file rather than a compilation-related file).

Jonas