Re: 9.0 RC1/Clang / illegal instruction (Signal 4) in gengtype while building cc_tools on i586.

On Sun, Oct 23, 2011 at 11:43:30AM +0300, Kostik Belousov wrote:
> On Sun, Oct 23, 2011 at 10:24:12AM +0200, Roman Divacky wrote:
> > > Program received signal SIGILL, Illegal instruction.
> > > 0x08048b24 in do_typedef (s=0x80532bf "CUMULATIVE_ARGS", pos=0x805e1a4)
> > >     at /usr/src/gnu/usr.bin/cc/cc_tools/../../../../contrib/gcc/gengtype.c:103
> > > 103     {
> > > 
> > > (gdb) disas 0x08048b24
> > > Dump of assembler code for function do_typedef:
> > > 0x08048b10 <do_typedef+0>:      push   %ebp
> > > 0x08048b11 <do_typedef+1>:      mov    %esp,%ebp
> > > 0x08048b13 <do_typedef+3>:      push   %ebx
> > > 0x08048b14 <do_typedef+4>:      push   %edi
> > > 0x08048b15 <do_typedef+5>:      push   %esi
> > > 0x08048b16 <do_typedef+6>:      sub    $0xc,%esp
> > > 0x08048b19 <do_typedef+9>:      mov    $0x805e1d4,%edi
> > > 0x08048b1e <do_typedef+14>:     mov    0x10(%ebp),%esi
> > > 0x08048b21 <do_typedef+17>:     mov    0x8(%ebp),%ebx
> > > 0x08048b24 <do_typedef+20>:     nopw   %cs:0x0(%eax,%eax,1)
> > 
> > LLVM attempts to use an optimal nop sequence when writing N-byte nop,
> > by using these nop instructions
> > 
> >   static const uint8_t Nops[10][10] = {
> >     // nop
> >     {0x90},
> >     // xchg %ax,%ax
> >     {0x66, 0x90},
> >     // nopl (%[re]ax)
> >     {0x0f, 0x1f, 0x00},
> >     // nopl 0(%[re]ax)
> >     {0x0f, 0x1f, 0x40, 0x00},
> >     // nopl 0(%[re]ax,%[re]ax,1)
> >     {0x0f, 0x1f, 0x44, 0x00, 0x00},
> >     // nopw 0(%[re]ax,%[re]ax,1)
> >     {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
> >     // nopl 0L(%[re]ax)
> >     {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
> >     // nopl 0L(%[re]ax,%[re]ax,1)
> >     {0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
> >     // nopw 0L(%[re]ax,%[re]ax,1)
> >     {0x66, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
> >     // nopw %cs:0L(%[re]ax,%[re]ax,1)
> >     {0x66, 0x2e, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
> >   };
> > 
> > There's no checking for a supported CPU, is it so that AMD geode doesnt support any of these?
> > Any other cpu that doesnt support these? If this is CPU dependant, I suggest to open a bug
> > report upstream as it's a bug.
> 
> Long nops are supported only on specific CPUs. Unconditional use of them
> is a plain bug, like unconditional use of cmovXX.

Yes, it's a bug, I filed http://llvm.org/bugs/show_bug.cgi?id=11212 upstream.
Patric, as a temporary workaround please add "-no-integrated-as" to your
CFLAGS, that will make clang use gnu as instead of its own intergrated
assembler, thus avoiding this problem.

Thanks for the great analysis!

roman