llvm dev - Sep 2020 - [RFC] Introducing the maynotprogress IR attribute

On 9/5/20 12:40 AM, Atmn Patel wrote:
> On Sat, Sep 5, 2020 at 1:07 AM Johannes Doerfert
> <johannesdoerfert at gmail.com> wrote:
>>
>> On 9/4/20 7:39 PM, Hal Finkel via llvm-dev wrote:
>>   >
>>   > On 9/4/20 6:31 PM, Atmn Patel via llvm-dev wrote:
>>   >> Hi All,
>>   >>
>>   >> We’ve prepared a new function attribute `maynotprogress` and
loop
>>   >> metadata `llvm.loop.mustprogress` in order to better
formalize the way
>>   >> LLVM deals with infinite loops without observable
side-effects. This
>>   >> is deeply related to the implicit forward progress
requirements in the
>>   >> IR.
>>   >>
>>   >> Background:
>>   >> There has been a demonstrated need for clarity within the
forward
>>   >> progress requirements in LLVM IR. This discussion started
with the
>>   >> longstanding bug [1], followed by subsequent discussion in
[2,3].
>>   >> TL;DR is that LLVM IR needed a clear way to deal with the
fact that
>>   >> C/C++ deems certain infinite side-effect free loops as UB
while other
>>   >> loops, and other languages, have well defined infinite
side-effect
>>   >> free loops.
>>   >>
>>   >> The proposed mechanism was to add an intrinsic:
`llvm.sideeffect()`
>>   >> that should be inserted by the frontend into loops to
indicate that
>>   >> this loop should be treated as having side-effects, and
should not be
>>   >> optimized out. This was implemented in [4], and long story
short, it’s
>>   >> been an imperfect solution for other language frontends in
various
>>   >> ways.
>>   >
>>   >
>>   > Can you make the long story not quite so short? What kinds of
problems
>>   > have cropped up with this solution?
>>
>> You mean, in addition to the conceptual problem of introducing an
>> arbitrary side-effect to prevent transformations from happening?
>> (And yes, we should revisit `llvm.assume` next.)
>>
>> So why do this at all:
>> 1) We settle the discussion and finally define what semantic LLVM-IR
>>      has. Basically revisiting [1] and [5] with all the reasoning
there.
>>      IMHO, it is in itself a good idea to write "something"
about forward
>>      progress down. Either non-attributed IR has it or not, but limbo
is
>>      bad. For example, we are fairly conservative right now but still
>>      don't promise not to assume forward progress, the worst
situation.
>>      In fact, we have optimizations that assume forward progress and
some
>>      that don't, ..
>> 2a) Assuming we always had an implicit forward progress guarantee:
>>       We avoid generic pessimizations for languages that require other
>>       semantics for the IR: Rust, C with constant loop expressions, ...
>>       see [1]. A function that might have an endless loop but which
does
>>       not write memory does not write memory. I mean, we still want to
>>       perform transformations even if there is a path in which such a
>>       function is called.
>> 2b) Assuming we do not have an implicit forward progress guarantee:
>>       We can delete loops that do not have side effects even if the
trip
>>       count is unknown. I mean, we could have done that in the other
case
>>       (2a) but we didn't in loop deletion. Though, we do remove a
call if
>>       the function only contained such a loop, ...
>>
>> Why this way:
>>    - The discussions before concluded IR does have a forward process
>>      guarantee [1,5]. So we don't want to pessimize existing code.
>>      That said, we want to exploit the property, e.g. in LoopDeletion,
>>      during the deduction of `willreturn` (and thereby for
>>      `isGuaranteedToTransferExecutionToSuccessor` and everything that
>>      transitively uses it), ...
>>    - Function attributes are the most fine-grained level, except
>>      instructions, to provide sticky semantics. The instruction level
has
>>      however only coarse grained effects, that is why we used
>>      `llvm.sideeffect` so far.
>>    - Loop metadata is reasonably sticky in the few cases we might need
it
>>      for optimizations, the key is dropping it doesn't compromise
>>      correctness.
> Another reason why the intrinsic is an imperfect solution is that the
> `llvm.sideeffect()` intrinsic would in theory need to be emitted by
> the frontend many times - potentially once for each loop and perhaps
> one for each function too, depending on how the frontend language
> views forward-progress.
>
> The Rust Team implemented the intrinsic into the Rust compiler, and
> found that there are significant compile time regressions (5-30%) [7]
> (which is not entirely attributable to rustc needing to emit more
> instructions, see [8] for some indication of this and further Rust
> dialogue) because these intrinsics end up being ubiquitous and LLVM
> isn't able to effectively coalesce redundant occurrences.

Thanks, Atmn, Johannes. I definitely agree that putting an artificial 
side effect in the loop is a conservative solution, and that we can do 
better. It's important that the rationale is in the RFC, and now it is.

There is another thing that I would like for us to document: do infinite 
loops have any relationship to thread synchronization? As I recall, this 
issue came up in the past when we discussed the deletion, or not, of 
infinite loops. If a thread writes a value to memory, and then enters an 
infinite loop, is that value eventually visible to other threads? My 
understanding is that some code depends on this property. If this is 
something that people would like to see guaranteed, I suspect that we 
may need specific handling (if nothing else, so we don't sink stores 
past possibly-infinite loops).

  -Hal

>
>>   >>   Even C/C++ has loops with and loops without this
requirement,
>>   >> though we could not distinguish so far.
>>   >
>>   >
>>   > What kinds of loop could we not distinguish? Can you please
provide an
>>   > example?
>>
>> IIRC, this loop is UB and cannot be reached:
>>
>> ```
>> int x = ((int)y) + 1;
>> while (x < y);
>> ```
>>
>> while this loop is not UB and can be reached:
>>
>> ```
>> while (1);
>> ```
>>
>> Though, I'm not a C language lawyer.
>>
>>
>>   >> In addition, there has been ambiguity regarding the
forward-progress
>>   >> requirements in IR, as pointed out in [5].
>>   >>
>>   >> The introduction of the `maynotprogress` IR function
attribute and the
>>   >> `llvm.loop.mustprogress` loop metadata tries to resolve this
>>   >> situation. The changes proposed are:
>>   >> - Document that IR Functions are by default expected to make
>>   >> forward-progress (as defined in the C++ Spec [6]).
>>   >> - Implement a `maynotprogress` IR function attribute (not
droppable)
>>   >> to indicate that this function is exempt from the above
requirement.
>>   >> This will allow frontends to disable IR optimizations that
would
>>   >> otherwise optimize away their infinite loops without
side-effects.
>>   >> - Implement `llvm.loop.mustprogress` as a loop metadata to
notify to
>>   >> the LLVM optimization passes such as LoopDeletion that even
if this
>>   >> function is permitted to not make progress, this loop is
still
>>   >> required to make progress because it is not one of the
infinite
>>   >> side-effect free loops permitted by the frontend language
specs. Note
>>   >> that loop metadata can be dropped, so even if this metadata
is
>>   >> dropped, we would not optimize away loops that we don’t
optimize now
>>   >> and we wouldn’t preserve loops that we don’t preserve now.
>>   >
>>   >
>>   > I'm a bit worried about the following: It seems like you
want to
>>   > handle C functions that have loops that might be infinite (i.e.,
those
>>   > with constant controlling expressions) by adding the
maynotprogress
>>   > attribute to the containing function, and then this attribute to
all
>>   > of the other loops. Is that correct?
>>
>> That was the idea (for C/C++), yes.
>>
>>
>>   > Also, it seems semantically incorrect to inline functions with
the
>>   > attribute into functions without the attribute unless you add
the
>>   > attribute to the caller. Is that correct?
>>
>> Yes. We actually talked about this and simply forgot to make the
>> attribute "sticky", same as for example
SpeculativeLoadHardeningAttr.
>> Patch is underway.
>>
>>
>>   > If those are true, then we can end up with cases where, with
functions
>>   > from the same C source file, we're either disallowing
inlining or
>>   > pessimizing the optimization of all loops in the caller.
>>
>> Technically true, assuming we do not use the loop metadata or it is
>> dropped. However, that is still an improvement to the status quo. Right
>> now, either we have a forward progress guarantee and the C loop that
>> should not be deleted might be deleted, or we don't have such a
>> guarantee and no loop is deleted regardless of the presence of one that
>> should not be deleted.
>>
>> Note that it is not only about "empty" loops. Assuming finite
loops
>> helps for loops that do something too. Take:
>>
>> ```
>> unsigned count = 0;
>> for (void *p = s; p != e; p += 4)
>>     ++count;
>> ```
>>
>> which is just fine optimized but if we dropped the `inbounds` from the
>> GEP, at some point, we end up with a loop instead of a closed form
>> expression:
>>     https://clang.godbolt.org/z/6dYTYe
>>
>>
>>   > Unless, when inlining, we add the attribute to the caller and
also add
>>   > this metadata to all other loops?
>>
>> For "maximal optimization opportunity" yes. For correctness,
we don't
>> need the metadata at all. Atmn is putting a patch up now to make the
>> attribute sticky, it's a one liner + tests, and then we can look at
the
>> metadata in a follow up.
>>
>>
>>   > In any case, please explain the intended behavior of the
attribute and
>>   > the metadata upon inlining.
>>
>> The attribute will be attached to the caller upon inlining as this is
>> the only way to keep semantics correct. Metadata can be added to the
>> loops of the caller if the caller did not have the attribute, but that
>> is an optimization and not required. The patch for the first part will
>> be part of this change set.
>>
>>
>>   >> The current implementations are in:
>>   >> - Changes to the LoopDeletion Pass:
https://reviews.llvm.org/D86844
>>   >> - Changes to the Clang Frontend:
https://reviews.llvm.org/D86841
>>   >> - Changes to LangRef: https://reviews.llvm.org/D86233
>>   >> - Changed to IR: https://reviews.llvm.org/D85393
>>   >>
>>   >> The changes preserve what was previously accepted as the
“default
>>   >> behavior” [5]. That is, you get forward progress assumption
in case a
>>   >> function is not marked with the `maynotprogress` attribute.
Here the
>>   >> default behavior is that LLVM IR functions are required to
make
>>   >> forward-progress and are assumed to make forward progress.
These
>>   >> attributes are aimed at helping frontends write correct code
as per
>>   >> their language specs, and in addition, optimize out some
dead loops
>>   >> that we weren’t able to optimize out before but could’ve.
>>   >>
>>   >> Feedback welcome.
>>   >>
>>   >> (The name of the function attribute and the loop metadata
are still
>>   >> under discussion, and we’re definitely open to changing
them.)
>>   >
>>   >
>>   > Some additional thoughts on the bikeshedding:
>>   >
>>   > I'm not in love with this name. might_not_progress would be
better. We
>>   > could choose something more colloquial, like may_inf_loop. Or
more
>>   > formal, like no_forward_progress_guarantee. I like this because
it
>>   > seems the most technically accurate and isn't likely to be
misread.
>>   > It's long, however, and even though we have attribute lists,
it will
>>   > still appear a lot.
>>   >
>>   > I think that I like nfpg the best (for "no forward-progress
>>   > guarantee"). Why nfpg? It's technically accurate and
short. Short is
>>   > useful because, for some languages (any maybe even for some IR
from
>>   > C), the attribute is going to appear on nearly every function. I
know
>>   > that we have attribute lists, but even so. no_fpg is good too.
>>
>> I'm given up on the name. To be honest, if you don't read the
langref
>> you don't know what these things do half the time, at least not in
>> detail. Let's just number them, this would be `attribute70` I think
;)
>>
>> Cheers,
>>     Johannes
>>
>>
>>   > Thanks again,
>>   >
>>   > Hal
>>   >
>>   >
>>   >>
>>   >> Atmn and Johannes
>>   >>
>>   >> [1] https://bugs.llvm.org/show_bug.cgi?id=965
>>   >> [2]
https://lists.llvm.org/pipermail/llvm-dev/2015-July/088103.html
>>   >> [3]
https://lists.llvm.org/pipermail/llvm-dev/2017-October/118558.html
>>   >> [4] https://reviews.llvm.org/D38336
>>   >> [5] https://reviews.llvm.org/D65718
>>   >> [6] https://eel.is/c++draft/intro.progress
>>   >> _______________________________________________
>>   >> LLVM Developers mailing list
>>   >> llvm-dev at lists.llvm.org
>>   >> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>   >
>>
> Atmn
>
> [7]
https://github.com/rust-lang/compiler-team/issues/177#issuecomment-578549329
> [8] https://blog.rust-lang.org/inside-rust/2020/03/19/terminating-rust.html
-- 
Hal Finkel
Lead, Compiler Technology and Programming Languages
Leadership Computing Facility
Argonne National Laboratory

On 9/8/20 9:08 PM, Hal Finkel wrote:
 > There is another thing that I would like for us to document: do
 > infinite loops have any relationship to thread synchronization? As I
 > recall, this issue came up in the past when we discussed the deletion,
 > or not, of infinite loops. If a thread writes a value to memory, and
 > then enters an infinite loop, is that value eventually visible to
 > other threads? My understanding is that some code depends on this
 > property. If this is something that people would like to see
 > guaranteed,

I doubt we guarantee that now, though we might not actively exploit it.
I also think we should not treat termination as synchronization, at
least not in the case where progress is required. So, in C (and arguably
we might want to do this also in C++), we would say that:
```
global_signal = 1;
while (1);
```
is well defined and we will not remove the write.

FWIW, I would agree that we should write this down though.


 > I suspect that we may need specific handling (if nothing
 > else, so we don't sink stores past possibly-infinite loops).

If (possibly-)infinite loops are well defined, e.g., in a function with
the `mayprogress` attribute, we certainly cannot move any effect past
them. If they are not, they are UB if they loop infinitely and otherwise
finite and therefore a no-op. So we can either delete them, by assuming
they are finite, or, if we know they are not, actually eliminate
anything that is known to transfer execution to the loop [0].

[0] https://reviews.llvm.org/D87149

~ Johannes

 >  -Hal

Hi Nicolai and Hal,

I was wondering if your present concerns regarding the directions of
the proposed attributes and semantics of the current direction had
been addressed, so I thought I'd send over a friendly ping. Have they?

Atmn

On Wed, Sep 9, 2020 at 1:08 AM Johannes Doerfert
<johannesdoerfert at gmail.com> wrote:
>
>
> On 9/8/20 9:08 PM, Hal Finkel wrote:
>  > There is another thing that I would like for us to document: do
>  > infinite loops have any relationship to thread synchronization? As I
>  > recall, this issue came up in the past when we discussed the
deletion,
>  > or not, of infinite loops. If a thread writes a value to memory, and
>  > then enters an infinite loop, is that value eventually visible to
>  > other threads? My understanding is that some code depends on this
>  > property. If this is something that people would like to see
>  > guaranteed,
>
> I doubt we guarantee that now, though we might not actively exploit it.
> I also think we should not treat termination as synchronization, at
> least not in the case where progress is required. So, in C (and arguably
> we might want to do this also in C++), we would say that:
> ```
> global_signal = 1;
> while (1);
> ```
> is well defined and we will not remove the write.
>
> FWIW, I would agree that we should write this down though.
>
>
>  > I suspect that we may need specific handling (if nothing
>  > else, so we don't sink stores past possibly-infinite loops).
>
> If (possibly-)infinite loops are well defined, e.g., in a function with
> the `mayprogress` attribute, we certainly cannot move any effect past
> them. If they are not, they are UB if they loop infinitely and otherwise
> finite and therefore a no-op. So we can either delete them, by assuming
> they are finite, or, if we know they are not, actually eliminate
> anything that is known to transfer execution to the loop [0].
>
> [0] https://reviews.llvm.org/D87149
>
> ~ Johannes
>
>  >  -Hal
>