If someone here has followed this thread, ldc2 has big performance problems with the function "floor":

http://forum.dlang.org/thread/lo19l7$n2a$1@digitalmars.com

Bye,
bearophile

On 6/22/2014 1:49 PM, bearophile wrote:
> If someone here has followed this thread, ldc2 has big performance
> problems with the function "floor":
>
> http://forum.dlang.org/thread/lo19l7$n2a$1@digitalmars.com
>
> Bye,
> bearophile

Well, unless LDC is doing something somewhere that I can't find, Clang
is using llvm's floor intrinsic, while LDC is using the non-inline asm
version of floor in phobos. There is however a declaration for llvm's
floor intrinsic in ldc.intrinsics : llvm_floor, but, with my limited
knowledge of LDC's inner workings, I don't believe that is what is being
used when you call std.math : floor. core.stdc.math would be using the
standard C library's floor function, which would be using an intrinsic
for floor.

On 22 Jun 2014, at 21:33, Orvid King via digitalmars-d-ldc wrote:
> On 6/22/2014 1:49 PM, bearophile wrote:
>> If someone here has followed this thread, ldc2 has big performance
>> problems with the function "floor":
> Well, unless LDC is doing something somewhere that I can't find, […]

The problem is that we somehow ended up with Iain's pure, multi-bitwidth D implementation being the default.

We should make our std.math call llvm.floor.*, or at least the C standard library one by default. Anybody wants to cook up a pull request for ldc-developers/phobos?

Thanks,
David