gnu/ci.go: Embeds build path, breaking reproducible builds

Something in gnu/ci.scm is embedding the build path when compiled into

gnu/ci.go, as can be seen:

https://tests.reproducible-builds.org/debian/rb-pkg/experimental/amd64/diffoscope-results/guix.html

./usr/lib/x86_64-linux-gnu/guile/3.0/site-ccache/gnu/ci.go

strings --all --bytes=8 {}

Offset 85, 15 lines modified Offset 85, 15 lines modified

...

92 /build/1st/guix-1.2.0~rc2/gnu/ci.scm 92 /build/2/guix-1.2.0~rc2/2nd/gnu/ci.scm

While guix builds of guix are typically built with a consistent build

path, it would be nice to fix this issue for other environments where

the build path may vary between builds.

My *wild* guess is it maybe has something to do with the use of

canonicalize-path:

(define (find-current-checkout arguments)

"Find the first checkout of ARGUMENTS that provided the current file.

Return #f if no such checkout is found."

(let ((current-root

(canonicalize-path

(string-append (dirname (current-filename)) "/.."))))

(find (lambda (argument)

(and=> (assq-ref argument 'file-name)

(lambda (name)

(string=? name current-root)))) arguments)))

Either directly or indirectly... does canonicalize-path resolve at build

time? run time? somewhere in-between? Or is this a red herring, and

something else entirely is responsible? I'll let a competent schemer

ponder this! :)

Thanks for all your help!

live well,

vagrant

Hi!

Vagrant Cascadian <vagrant@reproducible-builds.org> skribis:

‘canonicalize-path’ is called at run time, so that’s fine. However,

‘current-filename’ is a macro that captures the source file name at

build time, so it’s the likely culprit here.

I was going to go with something like:

… but I don’t think we can assume that the checkout is in the

‘%load-path’ when this code is executed. WDYT, Mathieu?

Looking at f71b0a0012d46bd30ead1a14ed58fd59647415e2, which introduced

this, there might be other options too.

Thanks,

Ludo’.

On 2020-11-26, Ludovic Courtès wrote:

This does resolve the reproducibility issues after some "quick" testing;

no idea if it breaks anything, though I didn't see any new test suite

failures (though my builds are patched to skip many tests)...

live well,

vagrant

Hey,

Cuirass happens to add checkouts to the %load-path just before loading

this file.

I tested your path, it works fine. I think it is acceptable as this (gnu

ci) interface needs a big rework anyway.

I can apply your patch if it's ok for you.

Thanks,

Mathieu

Hi,

Mathieu Othacehe <othacehe@gnu.org> skribis:

Is that systematic? Isn’t it only when ‘load_path_inputs’ is non-empty?

Sure if you’re confident you can go ahead. :-)

Thanks,

Ludo’.

On 2020-12-03, Ludovic Courtès wrote:

This looks to have been fixed some time ago in:

76bea3f8bcd951ded88dfb7f8cad5bc3e5a1701f ci: Remove hydra support.

live well,

vagrant