Compile time performance for metaprogramming is somewhat inconsistent

[maik klein via Digitalmars-d]

On Thursday, 3 March 2016 at 11:40:29 UTC, John Colvin wrote:
> On Thursday, 3 March 2016 at 02:03:01 UTC, maik klein wrote:
>> Consider the following code
>>
>> void main()
>> {
>>     import std.stdio;
>>     import std.range: iota, join;
>>     import std.algorithm.iteration: map;
>>     import std.conv: to;
>>     import std.meta: aliasSeqOf, staticMap, AliasSeq;
>>     enum types = "AliasSeq!(" ~ iota(0,10000).map!(i => 
>> to!string(i)).join(",") ~ ")";
>>     alias t = AliasSeq! (mixin(types));
>>     //alias t1 = aliasSeqOf!(iota(0, 10000));
>> }
>>
>> 't' compiles on my machine in ~3.5 seconds while 't1' needs ~1 
>> minute to compile. It seems that mixins are just way more 
>> performant than template instantiations. Any ideas why? What 
>> causes the slowdown and what can I improve?
>
> What happens if you add a few extra branches to 
> std.meta.aliasSeqOf, e.g. 
> https://github.com/D-Programming-Language/phobos/commit/5d2cdf103bd697b8ff1a939c204dd2ed0eec0b59
>
> Only a linear improvement but maybe worth a try?

I have tried the same thing in general and stuff like this is 
always a huge improvement.

In this case it goes down from ~60 seconds to ~3.8 seconds. I 
have done the same thing with my compile time map function, which 
gave me a drastic improvement.

I think recursion is just really bad in general for compile time 
stuff, for example your version

alias t1 = aliasSeqOf!(iota(0, 10000));

compiles in 3.8 seconds and uses roughly 600mb of ram while

alias t1 = aliasSeqOf!(iota(0, 20000));

compiles in 10.2 seconds and uses 1.9gb ram.

The mixin version is always the fastest but it also consumes way 
more memory and explodes before 20k elements