Subject: Re: [boost] Interest in updated expression template library?
From: Zach Laine (whatwasthataddress_at_[hidden])
Date: 2017-01-11 09:48:17
On Wed, Jan 11, 2017 at 2:00 AM, Larry Evans <cppljevans_at_[hidden]>
wrote:
> On 01/10/2017 04:22 PM, Mathias Gaunard wrote:
>
>> On 10 January 2017 at 20:07, Zach Laine <whatwasthataddress_at_[hidden]>
>> wrote:
>>
>>
>>> I agree with all of these complaints. This is in fact why I wrote Yap.
>>> The compile times are very good, even for an obnoxious number of
>>> terminals
>>> (thousands), and there is no memory allocated at all -- or did you mean
>>> compiler memory usage? I haven't looked at that.
>>>
>>>
>> Compiler memory usage of course.
>> When a TU takes 4GB to compile or more, it leads to lots of problems, even
>> if RAM is cheap and you could put hundreds of gigabytes in your build
>> server.
>>
>
> Zach, w.r.t. compile-time benchmarking,
> Louis has a compiler benchmark library here:
>
> https://github.com/ldionne/metabench
Yes. It's quite nice.
> I did have a brief look at it; however, I didn't see an easy way
> to vary parameters of the benchmark, for example, the size of
> expressions, and compare the results (it uses embedded ruby, and
> I couldn't figure out from that how to do what I want).
>
> Instead, I resorted to gmake for loops, as shown here:
>
> https://github.com/cppljevans/spirit/blob/get_rhs/workbench/
> x3/rule_defns/Makefile#L97
>
> resulting in output shown here:
>
> https://github.com/cppljevans/spirit/blob/get_rhs/workbench/
> x3/rule_defns/bench.tmp
>
> Interestingly enough, the method that performed worse was the
> one (RULE2RHS_CTX_LIST) which stored the rule2rhs information
> in the context. In contrast, the GET_RHS_CRTP stored this
> information by overloading functions generated by macros.
> What's interesting is that, based on what Brook Milligan says
> in his message, proto uses a context as well, but your library
> uses a set of function overloads.
>
> With regard to the gmake method of comparing compiler times,
> I realize that's sorta kludgy, and, years ago, I used a series
> of python programs to do the equivalent. I you find the
> gmake-for-loop method unacceptable, I can try to find the
> python method instead.
>
I haven't done much compile time benchmarking for Yap. It's fast enough
that I haven't felt the need. If people start using Yap and noticing long
compile times, I'll certainly do my best to fix that.
Zach