C as intermediate language, signed integer overflow and -ftrapv

[Richard Biener]

On Fri, Jul 25, 2014 at 12:35 PM, Richard Biener
<richard.guenther@gmail.com> wrote:
> On Fri, Jul 25, 2014 at 10:43 AM, Thomas Mertes <thomas.mertes@gmx.at> wrote:
>> On Thu, Jul 24 at 10:36 PM, Richard Biener <richard.guenther@gmail.com> wrote:
>>> Fact is that if somebody is interested in
>>> -ftrapv he/she is welcome to contribute patches.  Especially testing
>>> coverage is poor.
>>
>> As I said I have test programs for integer overflow (not written
>> in C). Now I have converted one test program to C. This program
>> checks if an int64_t overflow raises SIGABRT or SIGILL. The name of
>> the program is chkovf64.c and I have uploaded it to
>>
>>   http://sourceforge.net/projects/seed7/files/
>>
>> It is licenced with GPL. You can use it to improve the testing
>> coverage of gcc. When I compile it with:
>>
>>   gcc -ftrapv chkovf64.c -o chkovf64
>>
>> it writes a lot of warnings about "integer overflow in expression".
>> Running chkovf64 shows that -ftrapv does not work correct.
>
> That's https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61893 - basically
> as soon as we can constant-fold we lose the trap.  Which is probably
> not important for practical purposes, but you have a point here.
>
> Note the slight complication with static initializers that _do_ have
> to simplify to something (well, maybe not with -ftrapv ...).  Joseph,
> I can easily make fold fail if you don't explicitely use a constant
> folding API (int_const_binop) and -ftrapv is set.  Is it valid to
> reject
>
> static int x = __INT_MAX__ + 1;
>
> with an initializer-not-constant error?
>
> Index: fold-const.c
> ===================================================================
> --- fold-const.c        (revision 212388)
> +++ fold-const.c        (working copy)
> @@ -1121,7 +1121,12 @@
>    STRIP_NOPS (arg2);
>
>    if (TREE_CODE (arg1) == INTEGER_CST)
> -    return int_const_binop (code, arg1, arg2);
> +    {
> +      tree res = int_const_binop (code, arg1, arg2);
> +      if (res && TYPE_OVERFLOW_TRAPS (TREE_TYPE (arg1)) && TREE_OVERFLOW (res))
> +        return NULL_TREE;
> +      return res;
> +    }
>
>    if (TREE_CODE (arg1) == REAL_CST)
>      {

Just checked and for

static int x = __INT_MAX__ + 1;
int main()
{
  return __INT_MAX__ + 1;
}

we still fold the initializer (and warn) but no longer warn about the overflow
in main() but generate a properly trapping call to __addvsi (at -O0, with
optimization CCP manages to constant fold this).

Richard.

>
>> It seems that gcc thinks that writing a warning is enough and
>> raising a signal at runtime is not necessary. For human
>> programmers this makes sense, since they read the warnings and
>> correct the code. But for generated C programs this is not the
>> right strategy. There are different needs when C is used as
>> intermediate language.
>>
>> Maybe all -ftrapv problems uncovered by chkovf64.c are because
>> of this. Unfortunately there are also other test cases where
>> a signal is not raised although a signed integer overflow occurred.
>> This happens in a much bigger program and until now I was not
>> able to create a simple test case from it.
>
> Yes, not all optimizations may be aware of -ftrapv.
>
> Richard.
>
>> I used clang version 3.4-1 to proof that chkovf64.c works correct.
>> When I compile it with:
>>
>>   clang -ftrapv chkovf64.c -o chkovf64
>>
>> and start chkovf64 afterwards it writes:
>>
>> Overflow checking of negation works correct.
>> Overflow checking of addition works correct.
>> Overflow checking of addition assignment works correct.
>> Overflow checking of subtraction works correct.
>> Overflow checking of subtraction assignment works correct.
>> Overflow checking of incr works correct.
>> Overflow checking of decr works correct.
>> Overflow checking of multiplication works correct.
>> Overflow checking of multiplication assignment works correct.
>>
>> Greetings Thomas Mertes
>>
>> --
>> Seed7 Homepage:  http://seed7.sourceforge.net
>> Seed7 - The extensible programming language: User defined statements
>> and operators, abstract data types, templates without special
>> syntax, OO with interfaces and multiple dispatch, statically typed,
>> interpreted or compiled, portable, runs under linux/unix/windows.