Re: RFA: Make compare_and_jump_seq check that it has a jump insn

[Paolo Bonzini <bonzini at gnu dot org>]

On 06/23/2010 07:16 PM, Jeff Law wrote:
> On 06/23/10 10:49, Eric Botcazou wrote:
> > > The problem is that compare_and_jump_seq was then using the
> > > JUMP_LABEL macro on an insn that was not a jump, and in the
> > > process corrupting a completely unrelated insn. (It took me a
> > > *long* time to track down exactly where this corruption was
> > > occurring...).
> >
> > Configuring with RTL checking (--enable-checking=yes,rtl) would
> > presumably have revealed the problem immediately.
>
> But isn't RTL checking still extremely expensive?

It is expensive indeed, but it is still worthwhile giving it a shot when
something looks like memory corruption.

I think it's absolutely not a problem for do_compare_rtx_and_jump to
return something not ending with a label, for example a jump if greater 
or equal could be realized like this on a machine that has UNGE but 
lacks GE:

     cc=compare(r1,r2)
     jump if unordered to L2
     jump if unge to L1
  L2:

I think compare_and_jump_seq is just trying to do some poor-man 
jump-to-jump optimization.  If the last insn is not a jump, it should 
work just as well to not adjust the label.

The problem is that in this case setting the REG_BR_PROB note will be 
not quite as easy as compare_and_jump_seq tries to put it.

Paolo