On Wednesday, 7 December 2016 at 18:50:20 UTC, Stefan Koch wrote:
On Wednesday, 7 December 2016 at 18:00:40 UTC, Stefan Koch
wrote:
I have an update about ctfe performance.
I wondered for quite a while why newCTFE had a 5 millisecond
overhead, when compared the old interpreter.
Since on my charts it had comparable or better stats.
I finally figured it out.
Because the interpreter is supposed to be ctfe-able itself it
uses a gc-allocated stack.
And that in turn call envokes the gc.
The mark phase of the garbage-collector matches up exactly
with the extra time taken.
I have no idea, why this happens though since the GC is
supposed to be disabled.
It turns out that almost classes in dmd that are em-placed.
Whereas by bytecode visitor uses the regular new, because it is
fairly sizable. The gc is much more visible.
Although that does still not explain why GC.disable seems to
have no effect.
After a few more iterations the observed slow down seems to
disappear.
I am suspicious.
