Disable polymorphisation

[davidtwco]

Fixes #74614.

This PR disables polymorphisation to fix the regression in #74614 after investigation into the issue makes it clear that the fix won't be trivial. I'll file an issue shortly to replace #74614 with the findings so far. #74636 has been filed to track the fix of the underlying regression.

r? @eddyb

[lcnr]

Can you add the following as a test?

fn test<T>() { std::mem::size_of::<T>(); }

pub fn foo<T>(_: T) -> &'static fn() {
    &(test::<T> as fn())
}

fn outer<T>() {
    foo(|| ());
}


fn main() {
    outer::<u8>();
}

[wesleywiser]

r=me with the test added. We shouldn't roll this up since it will swing perf results again.

@bors rollup=never

[wesleywiser]

@bors r+ p=1

Raising the priority because of nightly breakage

[bors]

📌 Commit 799d52e has been approved by wesleywiser

[bors]

⌛ Testing commit 799d52e with merge 16b5082861b7528b4f3d171bd6797961b906fdd5...

[Manishearth]

@bors retry yield

yielding to rollup

[Aaron1011]

@bors retry

[lqd]

@bors p=10 for nightly breakage

[bors]

⌛ Testing commit 799d52e with merge bbebe73...

[bors]

☀️ Test successful - checks-actions, checks-azure
Approved by: wesleywiser
Pushing bbebe73 to master...

[nnethercote]

#69749 was a perf loss, but disabling polymorphization was mostly neutral. @davidtwco: is that expected?

[Mark-Simulacrum]

@davidtwco, @wesleywiser -- could someone follow up on this being a performance-neutral PR? Thanks!

[davidtwco]

#69749 was a perf loss, but disabling polymorphization was mostly neutral. @davidtwco: is that expected?

@davidtwco, @wesleywiser -- could someone follow up on this being a performance-neutral PR? Thanks!

@njn @Mark-Simulacrum Apologies for the delay in looking into this - this was surprising but I think it makes sense; I did some local benchmarking of the PR which landed polymorphisation (with -Zpolymorphisation=off) and the merge before that.

I believe that the majority of the losses that we didn't win back are from cross-crate metadata/incr-comp changes - -Zpolymorphize=off makes unused_generic_params always return every-parameter-is-used, but that result is still serialized - there was nothing to serialize without polymorphisation having landed. Analysing perf results isn't something I've got much experience with, but as far as I can tell, this explains the inconsistency here.

[Mark-Simulacrum]

Thanks! Is that something we could plausibly "fix", or does that seem too hard? I guess we're already encoding them as bitsets. Maybe we can skip recording all-zero bitsets and decode them appropriately as well? It may also be worth switching to a lower minimum bound -- u64 implies 64 generics which I expect is way above the normal amount for most functions?

[davidtwco]

Thanks! Is that something we could plausibly "fix", or does that seem too hard?

I think both of these suggestions make sense, I've implemented them in #75155.