Int division

[Roman Leshchinskiy]

Stefan O'Rear wrote:
> On Thu, Dec 13, 2007 at 11:17:10PM +1100, Roman Leshchinskiy wrote:
>> Hi everyone,
>>
>> here's a simple program:
>>
>> foo :: Int
>> foo = 8 `div` 1
>>
>> I was a bit surprised to see that GHC (both 6.8 and HEAD) generates the 
>> following Core for this:
>>
>> foo = case GHC.Base.divInt# 8 1 of wild21_afL
>>         { __DEFAULT -> GHC.Base.I# wild21_afL }
>>
>> It turns out that there are no rules for divInt#, quotInt#, remInt# and 
>> modInt# in GHC.Base and that no constant folding happens for divInt# and 
>> modInt# in PrelRules.lhs. In fact, the latter two don't even seem to be 
>> primops. This also seems to be the case for their Word# counterparts. I 
>> assume this is just an oversight but thought I'd ask before doing anything. 
>> Is it ok to add the primops and a couple of rules?
> 
> divInt# and modInt# are not primops because (i386) division corresponds
> to quotInt# and remInt#, so the other two absolutely need to be
> implemented in software.  But I'd call it an probably inliner bug if
> they show up in optimized Core output.

Hmm, I should have investigated further before posting. That's a very 
Intel-specific reason, though. On the Sparc, for instance, it wouldn't 
make much sense to implement divInt# in terms of quotInt#, I think (but 
then again, we don't have a Sparc NCG).

Anyway, I don't think it's an inliner bug because divInt# and friends 
aren't marked INLINE. If nobody objects, I'll go ahead and add the 
INLINE pragmas at least. We lose the opportunity to have rewrite rules 
for divInt# this way, though, which is a pity.

Roman