About LoopDeletion and infinite loops ... again! (RFC?)

Hello!

I read a bunch of discussions about the matter on this very mailing-list that are relatively recent or relatively old and I couldn’t find much agreement on the matter, so … here again :smiley:

LoopDeletion and infinite loops …

Currently LoopDeletion bails if non-detectable trip count loops are encountered and that’s fine, there are languages where infinite loops without side effects cannot be removed.

If I read the C++ spec right though N4527 (sec 1.10) point 27 says:

27 The implementation may assume that any thread will eventually do one of the following:

  1. 27.1 — terminate,

    27.2 — make a call to a library I/O function

    (27.3)— read or modify a volatile object, or

    (27.4)— perform a synchronization operation or an atomic operation.

    [Note: This is intended to allow compiler transformations such as removal of empty loops, even when

termination cannot be proven. — end note ]

So, the spec seems to be explicitly calling out that removing infinite loops is ok unless they do any of the things mentioned there (at least that is my interpretation, is that correct?).

If that is the case we could be missing out for languages that have such a behavior (and in particular in C++).

I was wondering how it would be viewed the possibility of adding a flag to loop deletion that allows the removal of loops with loop counts that are not SCEVable.

Thanks,
Marcello

You probably have seen it, but, for reference
https://reviews.llvm.org/D38336 (which goes in the exact opposite direction).

I don't necessarily fancy the idea of having a flag, instead, maybe,
the frontend could emit enough information to tell the optimizer
whether it is safe to remove the loop?

Thanks,

I see the usecase for mixed language compilation (that’s probably why you fancy something like the side-effect thing instead right?)

BTW if the other proposal passes can we basically assume that if a loop doesn’t have the sideeffect intrinsic in it is then removable?

That patch seems to suggest that in its current state llvm is mostly broken for languages that consider all infinite loops as unremovable ... (but I didn’t read all the discussion)

Marcello

I see the usecase for mixed language compilation (that’s probably why you fancy something like the side-effect thing instead right?)

It's also about not having parameterized semantics for the IR. I'd certainly find that undesirable. We could make it part of datalayout, or similar, but that has problems with mixed-language compilation. It also has problems with C, where we generally get to assume that loops terminate, but loops with a constant controlling condition (at the source level) can be infinite. Whether or not something is a constant condition at the source level is something that we need the frontend to mark only for specific loops.

BTW if the other proposal passes can we basically assume that if a loop doesn’t have the sideeffect intrinsic in it is then removable?

If it does not have the sideeffect intrinsic, or some other actual side effect (or atomic, etc.), then yes.

That patch seems to suggest that in its current state llvm is mostly broken for languages that consider all infinite loops as unremovable ... (but I didn’t read all the discussion)

As I understand it, it is broken in places, but we also don't take advantage of assuming that loops terminate as much as we could (assuming we fully commit to that assumption).

  -Hal

I see the usecase for mixed language compilation (that’s probably why you fancy something like the side-effect thing instead right?)

That could be a reason, but it wasn't my main motivation. Basically
I'm worried about having per-passes specific flags for non-debug
purposes.
Also, whether it's valid to perform some optimizations is something
that the source language rules should dictate, not something the user
should control. Exposing such flag is dangerous as people could just
add `-mllvm -remove-my-loops` with any frontend and then complain the
semantic of their code is not preserving. In some sense they are
shooting themselves in the foot, but I'm under the impression that
flags shouldn't really change the semantic (although some do, e.g.
-ffast-math, which many regret).

BTW if the other proposal passes can we basically assume that if a loop doesn’t have the sideeffect intrinsic in it is then removable?

The patch as far as I can tell blocks some optimizations in presence
of the intrinsic. The intrinsic itself doesn't really lower to
anything, as far as I can tell. I'm not sure the semantic has settled
down (and I still need to digest the whole proposal).

That patch seems to suggest that in its current state llvm is mostly broken for languages that consider all infinite loops as unremovable ... (but I didn’t read all the discussion)

I think at least Rust (and maybe Java? but don't quote me on that) are broken.
https://github.com/rust-lang/rust/issues/42009
The original bugzilla for this issue has 3 digits which says a lot
about how old the problem is :slight_smile:
https://bugs.llvm.org//show_bug.cgi?id=965

Thanks,

I see the usecase for mixed language compilation (that’s probably why you fancy something like the side-effect thing instead right?)

It’s also about not having parameterized semantics for the IR. I’d certainly find that undesirable. We could make it part of datalayout, or similar, but that has problems with mixed-language compilation. It also has problems with C, where we generally get to assume that loops terminate, but loops with a constant controlling condition (at the source level) can be infinite. Whether or not something is a constant condition at the source level is something that we need the frontend to mark only for specific loops.

BTW if the other proposal passes can we basically assume that if a loop doesn’t have the sideeffect intrinsic in it is then removable?

If it does not have the sideeffect intrinsic, or some other actual side effect (or atomic, etc.), then yes.

Ok, great, that seems to match with my original understanding then!

That patch seems to suggest that in its current state llvm is mostly broken for languages that consider all infinite loops as unremovable … (but I didn’t read all the discussion)

As I understand it, it is broken in places, but we also don’t take advantage of assuming that loops terminate as much as we could (assuming we fully commit to that assumption).

Yeah, that’s what I meant with “mostly”, but probably it would have been better “sometimes” :stuck_out_tongue:

Thanks!
Marcello

I see the usecase for mixed language compilation (that’s probably why you fancy something like the side-effect thing instead right?)

That could be a reason, but it wasn't my main motivation. Basically
I'm worried about having per-passes specific flags for non-debug
purposes.
Also, whether it's valid to perform some optimizations is something
that the source language rules should dictate, not something the user
should control. Exposing such flag is dangerous as people could just
add `-mllvm -remove-my-loops` with any frontend and then complain the
semantic of their code is not preserving. In some sense they are
shooting themselves in the foot, but I'm under the impression that
flags shouldn't really change the semantic (although some do, e.g.
-ffast-math, which many regret).

Yeah, I totally understand your point and is valid. Considering the work that is being done in the other patch it totally makes sense to wait for that to be done and then “fix” LoopDeletion to allow
non-sideeffected infinite loops to be removed … unless some other issues arise with that.

Thanks,
Marcello