On Sunday, 11 December 2016 at 17:04:24 UTC, safety0ff wrote:
On Sunday, 11 December 2016 at 16:26:29 UTC, Ethan Watson wrote:
At the very least, I now have an idea of which parts of the
compiler I'm taxing and can attempt to write around that. But
I'm also tempted to go in and optimise those parts of the
compiler.
Have a look at this issue:
https://issues.dlang.org/show_bug.cgi?id=16513
In a nutshell:
Dmd puts template instances in an AA using a terrible hash
function.
When looking for a match it does an expensive comparison for
each collision.
The hash function is easily fixed but it does not obviate the
expensive comparison, so it trades hashing time for comparison
time. Therefore the performance impact of fixing it is unclear.
Martin Nowak is suggesting that using the mangled name is the
way forward.
That means you have to compute the mangled name which is crazy
expensive.
And you can't cache the parent part of mangle because it all
freshly generated by the template.
It seems like I fail to express the problem properly so let me
try again.
AliasSeq's that are append to look like this
: AliasSeq!(AliasSeq!(AliasSeq(! ....))))
An aliasSeq that is "appended to" n times will produce (n^2) with
((n-1)^2) sub instances in them all the way till n is 0.
When you want to compare thier paramterter times you need to go
through all of them and recursively call findTemplateInstance.
