[gdal-dev] Requiring numpy for the Python bindings

[Greg Troxel]

Thomas Knudsen <knudsen.thomas at gmail.com> writes:

> Den tirs. 5. dec. 2023 kl. 02.15 skrev Greg Troxel via gdal-dev <
> gdal-dev at lists.osgeo.org>:
>>  Building the
>> rust compiler requires the *immediately preceding* compiler, which seems
>> unprecented, and I don't understand how people can think that's ok
>
> I don't think this is true - it's merely the common case:

I am guessing you have not struggled with packaging  the rust compiler,
and are just using binaries provided by other people, which is the
Linuxy approach.

> Over at https://rustc-dev-guide.rust-lang.org/building/bootstrapping.html
> the description is
>
>> The stage0 compiler is usually the current beta rustc compiler and its
> associated
>> dynamic libraries, which x.py will download for you.
>> (You can also configure x.py to use something else.)
>
> which makes it hardly any different than any compiler bootstrapping process.

That page *totally omits* discussion of how far back the bootstrap
compiler can be.  That is not about the concept of bootstrapping, but
about whether the project has rules that the compiler codebase must only
use language features from some earlier times.

In theory, yes, any particular version of any particular self-hosted
compiler needs at least some previous version.  But there is a vast
practical difference between "needs some not super ancient version" and
"needs the version that was released a few months ago".  The rust
community appears to think it's ok to require the bootstrap compiler
that is really recent.

Based on hearing from and helping the people working on packaging rust
in pkgsrc, it seems quite clear that the previous release is almost
always required.  Compiling bootstrap kits is a source of considerable
difficulty, and if e.g. rust 1.60 could be used to build 1.61 through
1.74, they'd be doing it that way.

In contrast, I think people are building gcc 12 with gcc 7 without
issues.  It's not really "5 versions old" so much as "gcc from 5 years
ago can build new gcc".  With gcc, it's probably even better than that,
as they seem to get it that bootstrapping sequences are painful.  I
basically don't hear complaints about gcc bootstrapping, only discussion
from the guix/mes community about getting to full gcc from a binary seed
small enough to be hand reviewed (Thompson's trusting trust attack and
all that).

go is another example.  go 1.4 can be built from C, go 1.18 can be built
with 1.4, and later versions are buildable with 1.18.

To disprove my statement, please install a rust compiler released in
2020 and use it to build 1.74, without the system under the hood
fetching a 1.73 binary somebody else built.