Re: Removing cc0 the right way

[Jim Wilson <wilson at specifixinc dot com>]

Paul Woegerer wrote:
> When not using cc0 the following code gets sometimes generated:
>    compare instr
>    add instr
>    branch instr

When exactly is this generated?  This could be any number of problems 
depending on who is generating it.

If compare has an explicit set of the condition code register, and the 
branch has an explicit use, and add has an explicit clobber, then this 
really shouldn't be happening unless something is broken.

Are you using a pseudo-reg or hard-reg for the condition code register? 
 If you are using a pseudo, try using a hard reg instead.  Or vice versa.

> 1. Add clobber side-effect to all insns that effect the condition
> status - this of-course failed since the compiler sometimes generates
> pattern without this clobber.

The x86 port does this.  You might try looking at it for ideas.  x86 
however has lea instructions which can do adds without clobbering the 
condition code register that might be helping here.

> 2. Try to define compare and branch in the same define_expand and
> define a define_split for them using the condition "reload_completed".
> This solved the bug but I think it is still ugly because some
> instruction scheduling is done after reload is completed and it might
> still be dangerous.

The instruction scheduler won't move an instruction that clobbers a reg 
across an instruction that sets or uses the same reg.  So there should 
be no danger here.
--
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com