[Bug c/45884] New: Incorrect removal of check for "less than zero" after adding value to (signed) long

[anders_jagd at yahoo dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45884

   Summary: Incorrect removal of check for "less than zero" after
adding value to (signed) long
   Product: gcc
   Version: 4.4.3
Status: UNCONFIRMED
  Severity: normal
  Priority: P3
 Component: c
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: anders_j...@yahoo.com


Created attachment 21959
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=21959
Source (test.c), Makefile, Binaries, test.i, gcc -dumpspecs etc. See README.txt

When compiled with -O2, the below (val < 0) check is, under some conditions
(see attachment), optimized away:

...
long val10, val;
char *s;
...
if((*s >= '0') && (*s <= '9')) {
val10 = val * 10; // Skip overflow check, not the issue we are showing
val   = val10 + (*s - '0');

/** OPTIMIZED AWAY IN CASE OF -O2 **/
if(val < 0) {
/* Overflow */
return -1;
}
...
}

Build on ubuntu (2.6.32-25-generic)
gcc version 4.4.3-4ubuntu5
machine i486-linux-gnu

Attaching complete test case with:

   Makefile
   Source (test.c)
   Binaries
   Result from compiling with -save-temps (test.i)
   System configuration, gcc -dumpspecs, etc.

[Bug c/45884] Incorrect removal of check for "less than zero" after adding value to (signed) long

[anders_jagd at yahoo dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45884

--- Comment #2 from Anders Jagd  2010-10-04 
18:12:16 UTC ---
(In reply to comment #1)
> Signed integer overflow is undefined so what GCC is doing is correct.  Use
> -fwarpv if you want it to be defined to wrapping.

I acknowledge that ISO/IEC 9899 defines integer overflow to be undefined. What
GCC doing is thus "not incorrect". However, would this maybe be a bit too
aggressive optimization at -O2 ?