remove 2nd DVCcodec time regression

[Dale Johannesen]

On Thursday, March 21, 2002, at 06:54 PM, Richard Henderson wrote:
> On Thu, Mar 21, 2002 at 06:02:44PM -0800, Dale Johannesen wrote:
>> At least this time the FSF change was intentional, see thread:
>> http://gcc.gnu.org/ml/gcc-patches/2001-11/msg00680.html
>> I'm not sure this is a problem on ppc at all, and certainly
>> it should be safe if both regs are pseudos.
>
> Err.. should it?  In what way does this check for pseudos affect 
> match_dup?

My reasoning was that if an insn is legal with one pseudo, it should
still be legal with a different pseudo substituted; the mode checks already
there appear to be adequate.  If this is wrong I'd appreciate an 
explanation why.

> You're probably right about ppc, but I can't let you do this for mainline.

Still?

> Instead, move_movables should be taught to use validate_replace_rtx 
> instead
> of replace_regs, and to abort a combination or replacement if if fails.

I realize this patch doesn't do everything you'd want, since hard-reg
substitutions will often be OK, but I think it's an improvement.
(I hadn't actually intended this to be posted to gcc-patches
yet, but at this point it has bootstrapped and tested OK on Darwin.  I'll
fix up the comments and ChangeLog if approved.)