Dmitry Babokin

Try `--math-lib=system` to see if this has an effect - it should be slow, but may help you isolate the problem. "default" library depends on ISA quite heavily and it's...

This is definitely a smarter way to do the shuffle when the indices are non-static. That implementation that we use is pretty naive (but very generic). The logic is currently...

This tricks are implemented by LLVM backend (codegen), ISPC can handle it, but preferably it should be done in LLVM. I suggest verifying that LLVM doesn't do that for C/C++...

I would experiment with `SimplifyCFG` and DCE to see if they are able to solve the problem and if they trigger in the right place in the opt pipeline.

It's this `TODO`: https://github.com/ispc/ispc/blob/ea4617c61c7eb9682b7846d95f42d262b48ac0e8/builtins/target-avx512spr-x32.ll#L536 Here's this code in `-x16` version: https://github.com/ispc/ispc/blob/ea4617c61c7eb9682b7846d95f42d262b48ac0e8/builtins/target-avx512spr-x16.ll#L22 `rcp_fast_*` should map to pure instructions/intrinsics without extra refining steps.

Not a call to action, but just to add broader context here. The logic of dispatch module exists in ISPC compiler in 3 different incarnations: - dispatch.ll - library code...

> Given that it looks like the ncurses dependency is obsolete (#2855). Thanks for explaining this. If we can drop this dependency, we should probably do that. But we need...

> Because I don't want to rebuild the previous versions. Up to you, but I vote for removing it once and for all versions - so we get better testing,...

Works for me (macOS, llvm-17 built by alloy.py).

> So, I will reduce this PR to change related to CMake files only, just to have option to build LLVM/ISPC with `libc++` in future. I'm ok with current set...