[bug#36346] [PATCH] gnu: Allow building toolchain with non-default libc.

[Marius Bakke <mbakke@fastmail.com>]

[-- Attachment #1: Type: text/plain, Size: 2600 bytes --]

Carl Dong <contact@carldong.me> writes:

> Hi Marius!
>
>> It would be easier to digest this patch if it came with an actual user of this
>> change. Right now it complicates a very simple procedure for no apparent
>> reason. Can you elaborate a bit on the use case?
>
> Ah! This change is motivated by the work I've been doing in shifting the Bitcoin
> release process to a Guix-based one. The binaries we produce aim to be
> compatible with GLIBC_2_11, and we have glibc compat wrappers
> (https://github.com/bitcoin/bitcoin/blob/f373beebbcc0c7d1160e02dc638a00b3e6831d98/src/compat/glibc_compat.cpp)
> all the way up to 2.27 (since we need RISCV support). With Guix, I hope that we
> don't have to keep updating compat wrappers anymore, and pin our toolchain glibc
> version to a fixed one. See here for how I use this:
> https://github.com/bitcoin/bitcoin/blob/e8dd4da0b287e0fe252c99bb4a7cb26c2e947b71/contrib/guix/packages/gcc-bitcoin.scm#L91

I see, thanks for the links.

>> Guix excels at creating bespoke toolchains like these. It is easy to express
>> this change as a new 'make-gcc-toolchain-with-custom-libc' procedure. So I'm
>> not sure if it's worth changing 'make-gcc-toolchain', which serves a fairly
>> specific use case.
>>
>> I would expect any reasonably complex toolchain to need further tweaks, and we
>> cannot possibly support all such configuration inside 'make-gcc-toolchain'.
>>
>> It does sound useful to make these procedures more generally accessible
>> however. Perhaps 'make-gcc-toolchain' could be implemented in terms of a more
>> generic 'make-toolchain' interface?
>
> That all sound like promising solutions. My thought process comes from porting
> riscv64 to Guix, where I realized that I had to override the default gcc version
> (riscv64 requires gcc 7.1), glibc version (2.27), and kernel headers version
> (4.15). That makes me think that the sensible list of things to be overridable
> for a toolchain would be those three, in case of future architectures. I've
> submitted previous patches to cross-base.scm that added the ability to
> parameterize these three, and this patch was simply doing the same for
> gcc-toolchain.
>
> Anyway, please let me know which approach you'd prefer, and I'd be very happy to
> implement and change. :-)

I feel better about this patch now that I've seen its uses.  It would be
great if you could leave some comments at the top of the definition
about what the libc argument is for, and maybe even a usage example.

Otherwise it LGTM.  Let's hold it for a couple of days in case others
have additional suggestions.

Thanks!

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 487 bytes --]