Patch: rs6000-specific ICE

[Dale Johannesen]

On Friday, October 18, 2002, at 02:19  PM, Dale Johannesen wrote:

>
> On Friday, October 18, 2002, at 02:03  PM, David Edelsohn wrote:
>
>>>>>>> Dale Johannesen writes:
>>
>> Dale> It's true I couldn't make it fail without -ffast-math, but 
>> there is no
>> Dale> dependency on that in the pattern or the matching functions.  
>> I'll need
>> Dale> to look further to convince myself the problem inputs never 
>> arise
>> Dale> without -ffast-math.
>>
>> 	I assumed you knew that this ICE is PR target/7133. The failure
>> requires -ffast-math.
>
> The one in that bug report does, but that doesn't in itself prove the 
> crash
> can only occur with -ffast-math.  You may well be right, but I need to 
> convince
> myself.

Doh.  The new condition is simply testing whether the resulting
split RTL will be matchable later on; this is dependent only on the 
patterns,
not on -ffast-math.  It may be that without -ffast-math the new 
condition can
never fail (this seems to be your belief and I suspect you're right), 
but
it won't hurt anything.  So I don't think an adjustment is necessary.