[tree-ssa] RFC: Never gimplify ASM_EXPRs

[Diego Novillo <dnovillo at redhat dot com>]

Jason,

Andrew Haley found widespread breakage in libjava due to a
miscompilation of boehm-gc.


--------------------------------------------------------------------------------
        From: Andrew Haley <aph@redhat.com>
        Date: Fri, 16 May 2003 18:39:01 +0100
        
        Consider this code.
        
        void fubar (void *x)
        {
          __asm__ __volatile__ ("       prefetch        %0": : "m"(*(char *)(x)));
        }
        
        With gcc 3.3 you get
        
        fubar:
        #APP
                       prefetch        (%rdi)
        #NO_APP
                ret
        
        No problem there, everything is fine.
        
        With tree-ssa you get
        
        fubar:
                movzbl  (%rdi), %edx
                movb    %dl, -1(%rsp)
        #APP
                       prefetch        -1(%rsp)
        #NO_APP
                ret
        
        and this crashes the garbage collector.  Because of this, no gcj
        programs even start.
        
        I can fix the problem with the simple patch appended, but it might be
        better to fix the compiler bug.  Assuming, of course, that we agree it
        really is a bug.
        
        Andrew.
--------------------------------------------------------------------------------

The root of the problem is that the gimplifier is converting that
ASM_EXPR into:

fubar (void *x)
{
  char *x.1;
  char T.2;

  x.1 = (char *)x;
  T.2 = *x.1;
  __asm__ __volatile__("prefetch %0"::"m" T.2);
}

I think this is a real bug and neither the gimplifier nor the optimizers
should ever attempt to look into ASM_EXPRs.  The same way we never even
try to do anything with MD builtins.

You could argue that we may be able to deal with ASMs, but I don't think
we should.  ASM_EXPRs and MD builtins are very low-level and we can't
pretend to understand them at the tree level.  I'm sure this bug will be
very popular when trying to compile things like kernels.

I'm testing a patch to completely remove processing of ASM_EXPRs from
the gimplifier and tree optimizers.  We are going to be marking
ASM_EXPRs with TREE_NOT_GIMPLE.

Ideas to fix the problem some other way?


Diego.