Dmitry Babokin

Probably we should just fix that. We have `sincos()` implementation for `float`, but not for `double` (which relies on system library availability, which is Linux-only).

So, the request is basically to introduce pow(float, int) in addition to pow(float, float). I'm pasting here the code snippet from the Godbolt, just in case: ``` float cubeFast(float num)...

This part of `alloy.py` functionality has rotten, as it was not used for quite some time. I think we should remove it (rebuilding LLVM as part of the testing). So...

This sounds like `export` keyword should be able to be used with struct declaration. > Alternatively, a way to never include them would also be helpful for my use case....

I can't speak about original intention. But I'm pretty sure that multiple headers are mostly to address the problem of different data types width. Try that: t.ispc: ```c struct S...

Do you need to access target-specific version of the function directly? The overhead of going through dispatch should be really small. The result of cpuid check is cached, so it's...

@xnorpx When you compile for a single target, there's not dispatch. Dispatch is generated only when multiple target are specified on the command line, i.e. `--target=avx2-i32x8,avx512skx-i32x16`. Note that in case...

@Lectem We are not allowing to expose varying types as part of extern function signatures (so we are, at least, able to generate dispatch functions), but we allow pointers to...

@Lectem I haven't realized that clang and gcc don't have auto-dispatch. I got used to ICC (Intel C/C++ compiler), which has it. Anyway, I'm glad that you like it, enjoy!...

I agree that exposing target-specific functions has its value. We see several use-cases, which would benefit from that. So we are going to redesign this part to satisfy all design...