[SOLVED] Re: [Qemu-devel] Overoptimization

[Alexander E. Patrakov]

Herbert Poetzl wrote:
> On Fri, May 28, 2004 at 04:58:33PM +0600, Alexander E. Patrakov wrote:
>
> > Herbert Poetzl wrote:
> >
> > > On Fri, May 28, 2004 at 07:19:35AM +0600, Alexander E. Patrakov wrote:
> > >
> > >
> > > > The CFLAGS in Makefile.target are bad for gcc 3.3.2. This is not a my 
> > > > problem with my gcc, because it passes the "make check" testsuite except 
> > > > some known errors in libstdc++, which cannot affect qemu.
> > > >
> > > > The victim is the tcp_input function in slirp. When compiling with the 
> > > > decault CFLAGS, for some reason all packets are dropped as having a 
> > > > wrong TCP checksum. I tried debugging this by placing some debug printfs 
> > > > in tcp_input.c file, but the problem disappeared after I inserted them. 
> > > > I concluded that it is very strange and compiled this tcp_input.c file 
> > > > without my debug statements, but with -O1 instead of -O2. The problem 
> > > > went away.
> > >
> > >
> > > gcc isn't supposed to change the semantics of
> > > a program based on the optimization level.
> > >
> > > nevertheless several cases are known where this
> > > happened, and not seldom the bug was in gcc
> > >
> > > testing the same opts with 3.3.3 or 3.4 will
> > > show if gcc is to blame or not.
> >
> > Done (with gcc 3.4.0). I used a truly minimal configuration of qemu in 
> > both cases to make sure that this is not an SDL/X/whatever else problem:
> >
> > ./configure --prefix=/usr --target-list=i386-softmmu --enable-slirp 
> > --disable-sdl
> >
> > GCC 3.4.0, however, required one more file to be compiled with -O1 
> > (because of some register allocation error), and there were some "struct 
> > timezone" errors that were trivial to fix by including <time.h> where 
> > appropriate.
> >
> > The result is the same. If I compile tcp_input.c with -O2, all tcp 
> > packets are rejected in user-net mode because of wrong tcp checksum. If 
> > tcp_input.c is compiled with -O1, qemu works.
>
>
> okay, sounds like some 'broken' struct, which might
> get misaligned or maybe a problem with the binutils ...

HJL binutils 2.15.90.0.3, there are 5 known failures during "make 
check", all are TLS-related and probably don't affect qemu.

> I'd suggest trying -Os and -O9 just to give some hints
> to the developers, also the detailed error messages
> of the specific code with -Wall might shed some light
> on that ...

In fact I just had to read the gcc info page and remove optimizations 
one by one. Result, with gcc 3.3.2:

-Wall -O2 -g -fomit-frame-pointer (defaults) = unsafe,
-Wall -O2 -g -fomit-frame-pointer -fno-strict-aliasing = safe.

So it looks like slirp code breaks strict aliasing rules assumed by -O2. 
Since the problem has been identified precisely, it has to be fixed in 
qemu CVS.

--
Alexander E. Patrakov