DMD performance issues with deeply-recursive templates

[Peter Alexander]

On Thursday, 25 July 2013 at 04:33:59 UTC, H. S. Teoh wrote:
> From what I can tell from a cursory glance, it's caused by the
> implementation of cartesianProduct using too many template 
> expansions
> (esp. in the variadic version of cartesianProduct), causing DMD 
> to run
> very slowly and eventually exhaust all available memory and 
> crash.

Actually, it's probably caused by the exponential size of 
template symbol names thanks to mangling of eponymous templates:

int[] a, b, c, d;
writeln(typeof(cartesianProduct(a, b)).mangleof.length);
writeln(typeof(cartesianProduct(a, b, c)).mangleof.length);
writeln(typeof(cartesianProduct(a, b, c, d)).mangleof.length);

534
4025
25003

I can't even get it to compile with 5 params, but it's increasing 
by a factor of 6 or 7 each time, so 5 will be hundred of KB, 6 
will be MBs and 7 will be 10s of MB.

This needs to be fixed. Simply compressing the names will not 
help, we need a better mangling scheme.