[llvm-commits] [llvm] r61552 - in /llvm/trunk: lib/Transforms/IPO/FunctionAttrs.cpp test/Transforms/FunctionAttrs/2008-12-31-NoCapture.ll

[Nick Lewycky]

Duncan Sands wrote:
> Hi Nick,
> 
>> FunctionAttrs just went through and added the readnone/readonly 
>> attributes. Nobody's had the chance to eliminate dead calls yet.
> 
> it would be nice if SimplifyCFG interacted nicely with SCC passes:
> if you specify -simplifycfg -functionattrs it would be great if
> simplifycfg was run on each function in each SCC before functionattrs
> is.  Thus both passes would work their way up the callgraph.  Eg if
> f calls g, you would get:
>   simplifycfg g
>   functionattrs g
>   simplifycfg f
>   functionattrs f
> If g becomes readonly and is trivially dead in f, the simplifycfg
> run on f would eliminate the trivially dead use.  This would also be
> useful for pruneeh: it would no longer need to have its own mini
> version of simplifycfg, since if g becomes nounwind then simplifycfg
> would take care of turning invokes of g in f into calls.

I'm not too keen on the notion of 'opt -simplifycfg -functionattrs' 
behaving entirely differently from 'opt -simplifycfg | opt -functionattrs'.

Actually, that's one of the first things I check for when I encounter a 
case where two passes together produce the wrong result. It usually 
means someone isn't updating an analysis correctly.

Nick

>> I haven't redone the metrics with the new change. The answer is 
>> "probably not much" but I don't think it hurts to leave it in.
> 
> I agree.
> 
>>>> -        NumNoCapture++;
>>>> +        ++NumNoCapture;
>>>>     
>>> If this is an improvement, please change NumReadOnly++ etc.
> 
> I already changed these.
> 
> Ciao,
> 
> Duncan.
>