From mboxrd@z Thu Jan  1 00:00:00 1970
From: Richard Henderson <rth@cygnus.com>
To: Joern Rennecke <amylaar@cygnus.co.uk>
Cc: Jim Wilson <wilson@cygnus.com>, rth@cygnus.com, egcs@cygnus.com
Subject: Re: alpha failure on 920810-1
Date: Wed, 29 Apr 1998 22:57:00 -0000
Message-id: <19980429225753.08852@dot.cygnus.com>
References: <199804290430.VAA25220@rtl.cygnus.com> <199804292113.WAA32444@phal.cygnus.co.uk>
X-SW-Source: 1998-04/msg01176.html

On Wed, Apr 29, 1998 at 10:13:05PM +0100, Joern Rennecke wrote:
> If it tests only if all bits are zero, I think we should express this
> as a DImode equality test to zero:
> 
>         (set (reg:DF 1)
>           (eq:DF (reg:DF 2) (reg:DF 3)))
>         (set (reg:DF 4)
>           (if_then_else:DF (eq (subreg:DI (reg:DF 1) 0)
>                                (const_int 0))
>                            (reg:DF 5)
>                            (reg:DF 4)))

Unfortunately that is not the whole truth.  I read the fine print
last night and found that it handles -0.0 specially, so it is not
really a DImode compare.  Which is unfortunate, since I'd like to 
be able to do real DImode comparisons with them...

In conversation with Jim today, we decided that 

         (set (reg:DF 1) (eq:DF (reg:DF 2) (reg:DF 3)))
         (set (reg:DF 4)
           (if_then_else:DF (eq:CC (reg:DF 1) (const_int 0))
             (reg:DF 5) (reg:DF 4)))

is a reasonable compromise.  It turns out that combine already knows
that it can't do anything with CCmode compares, and with my previous
patch modified just a tad, that takes care of loop, jump, and cse.

I've bootstrapped succesfully with these changes, and am currently
running Spec95.  Patches will follow once no regressions are shown.


r~