[llvm-commits] [PATCH] Aggressive FMA formation mode.

Roland Scheidegger sroland at vmware.com
Thu Jun 21 12:12:29 PDT 2012


On 06/21/2012 07:54 PM, Eli Friedman wrote:
> On Thu, Jun 21, 2012 at 11:50 AM, Roland Scheidegger<sroland at vmware.com>  wrote:
>> Isn't actually llvm IR desparately missing fma instruction?
>
> We have @llvm.fma.*; is that not sufficient?

Yeah I guess so (actually fmuladd is there as well if you don't care 
about rounding).
I think I'm missing though why it's an intrinisc, the docs say 
intrinsics are special features whose availability depend on the code 
generator, and I'm not sure it makes sense to restrict the availability 
of it? I'm also not sure if these intrinsics actually take vectors (well 
fma says so but fmuladd does not, neither do the half-float conversion 
functions, which would make them useless for my purpose), if it would be 
a normal ir function then it would be obvious it should work regardless 
(which as far as I can tell wouldn't actually really make that much of a 
difference in practice since not all code generators support everything 
of the ir neither anyway).

Roland



More information about the llvm-commits mailing list