bad user, or optimization bug with worked-around lvalue casts ?

[Andrew Haley]

Michael Haubenwallner writes:
 > Hi,
 > 
 > after switching to gcc-4.1.1 from gcc-3.4.5, I'm facing a problem with
 > this code searching for non-blank in a string (extracted test-case):
 > 
 > #include <stdio.h>
 > int check(const void *pvBuf)
 > {
 >     for(; *((char*)pvBuf) == ' '; ++ *((char**)&pvBuf));

This is wrong.  Try this:

for(; *((char*)pvBuf) == ' '; pvBuf = (char*)pvbuf + 1);

 >     return *((char*)pvBuf);
 > }
 > int main(void)
 > {
 >         printf("%d\n", check(" x"));
 >         return 0;
 > }
 > 
 > When building without optimization, it works as expected.
 > When building with -O2 or higher, the loop hangs.
 > This seems to be platform independant, as it occurs on x86-linux as well
 > as powerpc-aix, hppa-hpux, ia64-hpux, x86-solaris and sparc-solaris.
 > 
 > Now the question is:
 > Is this a user-bug only, not looking at the output of "-Wall" (because
 > this code builds and works successfully for 5 years now),
 > or is it a gcc-bug too, producing working code without optimization
 > and non-working code with optimization level >= 2 ?
 > 
 > And: can this code be changed without using an intermediate variable ?

This is an aliasing bug.  Please see The C standard, Section 6.3.2.3,
Pointers.

Alternatively, you may compile your code with -fno-strict-aliasing.

Andrew.