Re: PDTs, as you know few compilers support robust and reasonably full-featured implementations. Intel Fortran i.e., IFORT is nearly there but users have to beware, it too falls over when it comes to real-life problems. So I agree it’s still impractical to start “coding in anger” using PDTs in one’s day job, the “internal compiler errors (ICEs)” with PDTs alone will get one fired!
Given this, it is unfortunately still too early to fully evaluate performance, but you’re right: when compilers are struggling with implementing the feature itself, the most important aspect of compiler optimizations are essentially pending when it comes to PDTs.
Nonetheless, I suggest you take a look at these 2 links - I repost them since one of the links in my previous post doesn’t work anymore:
- An alternate method to support different floating-point kinds,
- An option to combine SoA approach with AoS class design.
With both of these cases, PDTs provide several benefits overall with program design and with little to no adverse effect on performance.
But I agree the performance aspect may end up varying significantly with quality of compiler implementations, thus YMMV may apply here too.
So I have to admit unfortunately: given the state of compilers, proceed with great caution when it comes to PDTs.