[llvm-commits] [llvm-gcc-4.2] r55796 - /llvm-gcc-4.2/trunk/gcc/llvm-convert.cpp

[Dale Johannesen]

On Sep 5, 2008, at 2:09 PMPDT, Duncan Sands wrote:

>> This would be correct, but is an overly conservative approach.
>
> Yes, but for the moment there is nothing better.
>
>> IMO worrying about the rounding mode when we don't model it at all is
>> premature (I don't believe anyone has complained about the behavior  
>> of
>> sqrt, which has all the problems you note above).  But I can see  
>> where
>> some would prefer a conservatively correct implementation.
>
> Those crazy numerics guys for example!

Yes, unless they want limited precision answers instead:)
Perhaps we should have a -fi-am-a-numerics-pedant flag, there are  
about 3 people who care about this working correctly.

>> I guess what I can do is mark these intrinsics (including sqrt) as  
>> readonly
>> and leave the tests in llvm-gcc alone.  Does that work for you?  (If
>> it causes performance problems I may need to back that out, but I
>> don't anticipate this.)
>
> Good plan!
>
> Ciao,
>
> Duncan.