For similar languages / compilers in this area:

1. ISPC [1], the Intel® Implicit SPMD Program Compiler also compiles SIMD programs with branches and other control flow efficiently using predication/masking etc.

2. Futhark [2] compiles nice-looking functional programs into efficent parallel GPU/CPU code.

[1]: https://github.com/ispc/ispc [2]: https://futhark-lang.org/

What about real performance? Does parallelization via SIMD instructions work well? What about supporting older CPUs with fewer SIMD instructions?

Does it have some loophole to allow loops? Or it just allows linear execution?

Can it read or write external memory (allocated within other language like C++)?