Re: Speedup CSE by 5%

[Jeffrey A Law <law at redhat dot com>]

On Tue, 2005-01-25 at 21:03 -0500, Kazu Hirata wrote:
> Hi Jeff,
> 
> > Err, I don't think that patch is safe.  See the various calls to
> > validate_change from within fold_rtx.
> 
> The comment just above fold_rtx says
> 
>    INSN is the insn that we may be modifying.  If it is 0, make a copy
>    of X before modifying it.  */
> 
> If fold_rtx is called like fold_rtx (x, NULL_RTX), then all the
> recursive calls are of the form fold_rtx (x', NULL_RTX).
> 
> If fold_rtx is called like fold_rtx (x, insn), then all the recursive
> calls are of either one of the followoing forms
> 
>   fold_rtx (x', insn);
> 
>   fold_rtx (x', NULL_RTX);
> 
> In other words, as you go deeper into the recursion, the second
> argument to fold_rtx either stays nonzero or becomes zero, but it does
> not becomes nonzero if NULL_RTX is passed.
> 
> Furthermore, if the second argument is 0, we make a fair amount of
> effort to copy X before modiyfing it.  I don't know if we are doing a
> perfect job there, though.  I haven't looked at the code that
> carefully.
I understand all that.  However, if you look at the first call to
validate_change (inside an ASM_OPERANDS case) you'll see a case where
I think we can modify x without copying it first.

jeff