why to (not) support "older" compiler versions

[Johannes Pfau via Digitalmars-d]

Am Tue, 3 Nov 2015 13:50:55 +0100
schrieb Iain Buclaw via Digitalmars-d <digitalmars-d at puremagic.com>:

> On 3 November 2015 at 12:57, drug via Digitalmars-d <
> digitalmars-d at puremagic.com> wrote:
> 
> > On 03.11.2015 14:11, Iain Buclaw via Digitalmars-d wrote:
> >  
> >> On 3 November 2015 at 11:35, Daniel Murphy via Digitalmars-d
> >> <digitalmars-d at puremagic.com <mailto:digitalmars-d at puremagic.com>>
> >> wrote:
> >>
> >>     On 3/11/2015 7:52 PM, drug wrote:
> >>
> >>         On 03.11.2015 11:22, Johannes Pfau wrote:
> >>
> >>
> >>             I guess it's to be compatible with the latest DMD, LDC
> >> and GDC. GDC
> >>             currently only provides the 2.066.1 frontend.
> >>
> >>         A bit offtopic - will the situation change with ddmd
> >> accepted? I mean
> >>         the situation with different frontend version in different
> >>         compilers.
> >>
> >>
> >>     While DDMD does not have any direct effect on our ability to
> >> keep the three compilers synced, some of the cleanup work that has
> >> been done does help.
> >>
> >>
> >> Whilst other clean-up work has destroyed years of stable
> >> compatibility between different 'ends'.  ;-)
> >>  
> > Hmm, I asked because I've heard that using ddmd would help with
> > keeping the compilers synced and we would have the same version of
> > frontend everywhere...
> >  
> 
> Well, how would that work?  :-)
> 
> What you've probably misheard is half of a phrase.  Moving to towards
> ddmd is not to be confused with moving towards a shared 'frontend'
> codebase, and is the first half of the correct sentence.  The second
> half is that even then, that has no guarantee of keeping things in
> sync without also integrating other 'ends' into the CI process.
> 
> This requires that we set-up an infrastructure where:
> 
> - New PRs are tested against all compilers before merging.  This not
> to be confused with our current set-up where all compilers build DMD.
> Specifically new changes upstream must:
>   1. Be able to apply the change cleanly in their local repositories
>   2. Build themselves without error.
> - We then need another process in place to keep each end in sync after
> changes upstream are applied.
> 
> It was hoped that moving towards ddmd would force a lot of the
> ABI-specific code to be moved into Target or Port (host) interfaces
> that are agnostic to the backend.  There are still many
> target-specific areas where this is not the case, and on top of that
> there are regressions in the host-specific interfaces.
> 
> In short, there will always be a heavy maintenance burden regardless
> of what language we're written in. :-)
> 
> Iain
> 

I'd like to see some statistics how many DMD pull requests are
frontend-only vs how many pull requests also touch the backend*.

A crazy idea:
Once gdc supports the latest frontend version we could theoretically
adjust the dmd pull request testing to also merge dmd pull requests
into the gdc frontend and test gdc with these frontend-only requests. We
would then only merge dmd pull requests that build for gdc as well. Then
we would need some hooks to also automatically pull these into gdc. Or
we could setup the frontend as a submodule.

The main problem is that even frontend-only changes will depend on
earlier backend changes. So we'd need to keep the compilers somehow in
sync. Every request touching the backend would have to be ported to
GDC before merging into dmd. This will keep GDC/DMD 100% in sync, but
it will also slow down DMD development.

So the interesting question now is what does the frontend/backend* pull
request ration look like? If only very few pull requests touch the
backend this approach could work. The situation for druntime is
similar, although I guess there are fewer compiler specific pull
requests for druntime. Phobos should be mostly compiler independent.


* DMD backend improvements are fine. The critical pull requests are
  requests which affect the explicit and implicit frontend/backend
  interface.