On Friday, 5 March 2021 at 16:41:39 UTC, kinke wrote:
> On Friday, 5 March 2021 at 16:34:07 UTC, kinke wrote:
>> https://run.dlang.io/is/z0k6l9
>
> Small correction (IR generation was fine, compilation wasn't): that intrinsic requires a newer target shader model (-mcpu=sm_70) [and -betterC to avoid ModuleInfo etc.]. Then the asm can be generated and inspected too.

Yep.  The above works for me as well.  cuda_620 (sm_62 IIUC) is the highest currently allowed via the --mdcompute-targets interface.  I'll work on expanding that.

I'm not sure if the "i1" type can just plug in to gen_gccbuiltins.cpp as previously discussed but I'll look at that as well.

Thanks.