On Sunday, 21 December 2014 at 09:34:33 UTC, bearophile wrote:
This sounds more ambitious than the hypothetical D3 language :-)
Feature complete should be ambitious! I am only asking for
something a little bit better than C++ ;-) I am not asking for
fringe features like "multiple alias this"...
Modular arithmetics is handy, when you want it.
But you always get it… you can just use the modulo-operator.
I don't understand enough some of your points, so you should
popularize better your ideas :-)
Which points? I assume point 1 was clear, more on point 2 and 3:
On point 2, AVX:
LLVM is getting AVX auto vectorization. AVX uses masks so that
you can vectorize loops with conditionals, when the mask goes to
zero then the instruction is turned into NOP. So with the proper
constructs you can get branch-free vectorized inner loops.
If D wants to stay competitive as a system level language, D need
to make sure that programs get efficient AVX auto vectorization
out-of-the-box.
I find this poster from the last LLVM meeting interesting:
http://llvm.org/devmtg/2014-10/Slides/Nis-AVX-512ArchPoster.pdf
On point 3, allocators:
If you want reasonable and fast ref counting without code bloat
you need to standardize how ref-counting is done. The simple
solution is to put the ref count on a negative offset and make
sure that all ref-countable class objects are allocated as
separate objects.
Then you need a fast allocator. That means you need a pool where
alloc/free is O(1) in the common case. By profiling allocation
patterns the compiler can generate code that makes objects
available when needed. Since most applications have uneven load
you can initialize objects to defaults and restructuring the free
memory in thread local/global heap when you are having cycles to
spare.
If the compiler controls allocations it can also reuse objects
when they are dead (by liveness analysis) rather than
freeing/allocating, put it on the stack etc…
