Re: [Emacs-diffs] master d826037 3/3: Remove the need for temacs.in

[Daniel Colascione]

>> The linker deciding to lay out objects
>> within sections in a different order will break the dump.
>
> But why would the linker do that?  It sounds like you're worrying that,
> even if
> we give the linker the same object files in the same order and ask it to
> link
> Emacs again, then the linker might generate a different executable, one
> that is
> incompatible with the previous one.

Who knows what linkers might do? Change is especially likely in an LTO
world  when different linker flags or the linker binary changes? I want to
future-proof the fingerprint mechanism by fingerprinting something as
close as possible to the actual binary we run.

> But if that's the case, the temacs.in
> solution has the same problem so we're no worse off than before. And if
> it's not
> the case, then what exactly is the failure scenario here? I'm not seeing
> any
> failure scenarios for the current approach that can't also be failure
> scenarios
> for the previous approach, even now that I know that pdumper cares about
> object
> order within a section.

It's likely that the linker to lay out objects in a section identically in
two different builds when the only difference between the builds is the
content of an array. If we're worried about the array being folded into
the code, we can make it volatile. If someone changes a linker flag that
changes object ordering, temacs.in will change. If the linker gets
upgraded and flips the order of two variables in .data, temacs.in will
change.

>>> The mechanism could also "break" with whole-program optimization that
>>> inlines the fingerprint array, or with other plausible future changes
>>> to
>>> linkers as they get smarter.
>>
>> The optimization you've described doesn't matter: as long as changing
>> the
>> *value* of the fingerprint array (not its length!) preserves Emacs
>> *object
>> layout* change in Emacs, we're good.
>
> The optimization I describe could migrate some or all of the fingerprint
> array's
> contents into the code, with the amount of migration depending on the
> contents
> of the fingerprint array, and with the migrated contents omitted from the
> fingerprint array. I don't see how object layout (in the sense that you
> describe) would be preserved in that scenario.

Even if the array is migrated into code, changes to linker configuration
will still change the temacs.in fingerprint.

The invariant here isn't that temacs.in has to be same as temacs, but
rather that if temacs_1 differs from temacs_2 in ways that pdumper cares
about, then temacs.in_1 must have hash different from temacs.in_2. My
objection is that the object-hashing approach can result in a situation in
which temacs_1 and temacs_2 differ in ways that pdumper cares about but
nevertheless have the same fingerprint.