bug#46617: 28.0.50; nativecomp: native compile cache is not invalidated when file is re-byte compiled and changes

[Aaron Jensen <aaronjensen@gmail.com>]

On Sun, Mar 21, 2021 at 12:27 PM Andrea Corallo <akrl@sdf.org> wrote:
>
> Hi Aaron,
>
> I had to slightly modify your init.el reproducer as we have changed the
> semantic of `load' [1].

With this change, how do you specify that you want the eln to load if
it's there? If I understand correctly, this prevents loading the eln
with load unless it is specified explicitly? Or can you leave the
extension off when loading and that will prefer the eln, then elc,
then el?

> I think what happened is that in the original reproducer the second load
> of the .elc file still automatically loaded in place the corresponding
> .eln as the source file didn't changed.  The issue here is that in the
> compilation unit repro.elk is capturing some of the environment and this
> is not accounted in the hash computation of the .eln file.

The original repro simulated a restart without a restart. The real bug
(in straight.el) is that on start, it loads a byte compiled/native
compiled file that relies on outside environment. So, that may be
fixed now because I believe they load the elc explicitly, but that
means it'll never load the native compiled version. Ideally, they
would, but I don't know how to force loading byte compiled OR native
compiled in one command.


> The new reproducer (well it does not reproduce much :) ) works because
> now `load' called explicitly on a .elc file does not load automatically
> the corresponding .eln.  Alternatively to overcome the issue still using
> native compilation one has to force a new native compilation using
> `native-compile.  WDYT?

I think it sounds like straight.el will need to add detection for
native compilation, and any time they byte compile, also native
compile. I also think that this is extremely fragile and I wonder if
byte compiling should result in "invalidating" (deleting) the
previously corresponding eln. We essentially have a cache invalidation
that is relying purely on a cache key that is incomplete.

Thanks,

Aaron