https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108845

--- Comment #4 from Qirun Zhang <qrzhang at gatech dot edu> ---
(In reply to Qirun Zhang from comment #3)
> (In reply to Jakub Jelinek from comment #2)
> > I'm not convinced it is a good idea.
> > Sure, in the above case it is obvious it will never trigger, but if we say
> > use ranger to decide if the operation can or can't overflow, then VRP is in
> > many cases based on assumptions which only hold for valid code, but
> > sanitizers actually want to diagnose invalid code.
> 
> 
> Thanks!
> 
> Here is another (similar) example.  Earlier versions of GCC will not inject
> UBSAN_CHECK_ADD. However, the latest version of GCC will.
> 
> the code example:
> ======
> void main() {
>   int a = 0;
>   for (; a != 2; a++)
>     ;
> }
> ======
> 
> Compile with "gcc-11 -fsanitize=signed-integer-overflow -O3 
> -fdump-tree-optimized", we got no UBSAN checks:
> 
> ======
> void main ()
> {
>   int a;
> 
>   <bb 2> [local count: 118111600]:
> 
>   <bb 3> [local count: 955630225]:
>   # a_6 = PHI <1(3), 0(2)>
>   a_3 = a_6 + 1;
>   if (a_3 != 2)
>     goto <bb 3>; [87.64%]
>   else
>     goto <bb 4>; [12.36%]
> 
>   <bb 4> [local count: 118111600]:
>   return;
> 
> }
> ======
> 
> Compile with "gcc-trunk -fsanitize=signed-integer-overflow -O3 
> -fdump-tree-optimized", we got one:
> 
> ======
> void main ()
> {
>   int a;
> 
>   <bb 2> [local count: 118111600]:
> 
>   <bb 3> [local count: 955630225]:
>   # a_5 = PHI <a_3(3), 0(2)>
>   a_3 = .UBSAN_CHECK_ADD (a_5, 1);
>   if (a_3 != 2)
>     goto <bb 3>; [89.00%]
>   else
>     goto <bb 4>; [11.00%]
> 
>   <bb 4> [local count: 118111600]:
>   return;
> 
> }
> ======
> 
> $ gcc-trunk -v
> gcc version 13.0.1 20230218 (experimental) [master r13-6132-g32b5875c911]
> (GCC)

I did a bisect. For the testcase provided in comment #3, the behavior was
introduced in the following commit:


commit 502ffb1f389011b28ee51815242c7397790802d5
Author: Andrew MacLeod <amacl...@redhat.com>
Date:   Tue Nov 2 21:26:44 2021 -0400

    Switch vrp2 to ranger.

    This patch flips the default for the VRP2 pass to execute ranger vrp rather
    than the assert_expr version of VRP.

            * params.opt (param_vrp2_mode): Make ranger the default for VRP2.

Reply via email to