[llvm-dev] [RFC] Introduce the `!nocapture` metadata and "nocapture_use" operand bundle
Johannes Doerfert via llvm-dev
llvm-dev at lists.llvm.org
Thu Jan 21 08:17:30 PST 2021
Hi Artur,
I'm still having trouble reading your emails. As you see here
https://lists.llvm.org/pipermail/llvm-dev/2021-January/147963.html
there is no visual indication of the quoted text versus the new one. I
will try to answer you but I might simply overlook parts
of your last email. Apologies. Could you try to modify your email client
to indent the quoted text, or something?
On 1/20/21 4:07 PM, Artur Pilipenko wrote:
> On Jan 18, 2021, at 4:09 PM, Johannes Doerfert <johannesdoerfert at gmail.com<mailto:johannesdoerfert at gmail.com>> wrote:
>
> On 1/12/21 8:41 PM, Artur Pilipenko wrote:
> On Jan 11, 2021, at 9:44 PM, Johannes Doerfert<johannesdoerfert at gmail.com<mailto:johannesdoerfert at gmail.com>> wrote:
>
> The email formatting is somewhat broken. Might be my client.
> I try to answer inline, if I missed something or there is any other problem, let me know.
>
>
> On 1/11/21 6:39 PM, Artur Pilipenko wrote:
> On Jan 11, 2021, at 3:46 PM, Johannes Doerfert <johannesdoerfert at gmail.com<mailto:johannesdoerfert at gmail.com><mailto:johannesdoerfert at gmail.com>> wrote:
>
>
> On 1/11/21 5:13 PM, Artur Pilipenko wrote:
>
> On Jan 11, 2021, at 2:40 PM, Johannes Doerfert <johannesdoerfert at gmail.com<mailto:johannesdoerfert at gmail.com><mailto:johannesdoerfert at gmail.com>> wrote:
>
> Hi Artur,
>
> On 1/11/21 4:25 PM, Artur Pilipenko wrote:
> I'm a bit confused with nocapture_use. I guess you need this because without it BasicAA would assume that the pointer is not accessed by the call at all.
> Correct.
>
>
> So, as a workaround you introduce a use which implicitly reads and writes.
> Correct, for now. We could add "readonly"/"writeonly" etc. later on.
>
>
> But this might be a more general problem. For example:
>
> a = new ...
> store a, ptr, !nocapture
> a' = load ptr
> ; Now you have 2 pointers to the same object (a' and a ) which BasicAA considers as no aliasing.
> v1 = load a
> store 5, a'
> v2 = load a
>
> We would happily replace v2 with v1 even though the memory was clobbered by the store through a’.
> Right. But that is not strictly speaking a problem. You can build things with the annotation
> that are nonsensical, though, that is nothing new. Especially if you employ the annotations
> alone you might not find a good use case, seehttps://reviews.llvm.org/D93189#2485826 .
> My concern here is that we miscompile a program which is seemingly correct. None of the users
> of pointer a escape the pointer. So, I assume it should be correct to mark the store as !nocapture.
>
> It looks like you assume a more restrictive definition for !nocapture. The proposed lang ref says:
> "``!nocapture`` metadata on the instruction tells the optimizer that the pointer
> stored is not captured in the sense that all uses of the pointer are explicitly
> marked otherwise”
>
> a) What do you mean by "uses of the pointer” here? Is it uses of the pointer value stored by the
> Annotated instruction? Is it uses of the memory modified by the store?
>
> It is uses of the stored pointer. So if you never load the pointer from the location
> you stored it using `!nocapture`, there are no uses and "all uses" are explicitly
> marked. If you do load it, you should make sure the use is "explicitly marked otherwise"
> because you do not get a "dependence edge" from the `store %a %ptr !nocapture` to `%a' = load %ptr`.
> In your example, that explicit use is missing. So you load `ptr` but that instruction is
> not annotated with an explicit use of `a`. Now, this could actually be OK, depending on the use,
> but unlikely what you want.
>
> If we would have operand bundles on loads you could do: `%a' = load %ptr ["nocapture_use"(%a)]`,
> which would correspond to the intended use `call @f(%ptr) ["nocapture_use"(%a)]`. That way you would
> be able to communicate `%a` through memory (here `%ptr`) without causing it to be captured.
> (This assumes you ensured `%a'` is not captured.)
>
> I think we could require `!nocapture` to be used with "nocapture_use" but I resisted so far
> as it would be more complex.
> On the other hand, it would make it clear that usage of only one of them is, so far,
> discouraged since it can easily lead to unexpected results.
> So, basically every use of the value loaded from memory which was previously stored to as !nocapture
> needs to have an annotation indicating that this is an alias of the original pointer.
> Not necessarily an alias but a use of the original pointer. More below.
> In general case we might have a pointer which is not the original pointer but an alias for it.
>
> E.g. a select between the original pointer and some other value:
> a = new ...
> store a, ptr, !nocapture
> a' = load ptr
> s = select c, a', some other ptr
> v1 = load s
>
> Or a derived pointer based off the original pointer:
> a = new ...
> store a, ptr, !nocapture
> a' = load ptr
> gep = gep a', offset
> v1 = load get
> Do we need to annotate things like geps/bitcasts?
>
> What if the use is a phi or a select?
> a = new ...
> store a, ptr, !nocapture
> a' = load ptr
> s = select c, a', some other ptr ; do we annotate the select?
> v1 = load s ; or this load?
>
> It looks like currently we don’t have the means to annotate the uses of the loaded value. We might
> need to prohibit loads of !nocapture-stored values altogether (if this is a load in the same function as
> the nocapture store).
> Right, we could do that.
>
>
> b) Does the example above violate this statement somehow?
>
> So far, there is no violation per se possible. The semantics cannot be violated,
> as stated right now. Using the annotation changes the program semantic, if that change is not
> what you wanted, that is a problem but not a miscompile (IMHO).
> The way I’m looking at this is I have a program without !nocapture metadata and operand bundles
> and I want to derive it using some analysis.
> (We actually have an escape analysis capable of handling object graphs.
> https://llvm.org/devmtg/2020-09/slides/Pilipenko-Falcon-EA-LLVM-Dev-Mtg.pdf
> https://www.youtube.com/watch?v=WHiU2-h_kRM
> We are looking into ways we can communicate facts this analysis computes with the rest of the
> optimizer. But we were looking into noalias and alias.scope metadata for such purpose.)
> So, I don’t want to change the behavior using the metadata, I want to derive it and for that I need
> to understand the semantic it implies.
>
> BTW, in your motivating example, what are the benefits you expect from the nocapture property?
> You can mark the pointers in the caller nocapture, with all the benefits that might bring.
> For a more direct example, see this thread:https://lists.llvm.org/pipermail/cfe-dev/2020-December/067346.html
> Basically, we pass a bunch of pointers to an API. Since we need to pass an unknown number, we
> do it with an array of void* and a size. The runtime will only load the pointers and not cause
> them to escape, though it looks like they could. The same problem appears for memory latency hiding
> when you do an OpenMP target map, basically a bulk device memcpy. Since you do more than one, you
> pass a bunch of pointers via a void* array, which causes them to escape. This causes spurious aliases
> that we know cannot exist, though the runtime is not (supposed to be) linked statically and you need
> domain knowledge to deal with this. I split the OpenMP usage code from the patch but the initial version
> still had it:https://reviews.llvm.org/D93189?id=311479 I'll clean it up and put it on phab again soon.
> In general we want to have a way to express the results of some more powerful escape analysis in the IR
> so CaptureTracking/BasicAA can take advantage of this analysis. In your case this analysis is some domain
> specific knowledge, in our case it is a downstream analysis pass.
>
> We can easily express basic facts like capture/nocapture using attributes and metadata. Things get more
> difficult when we need to express aliasing properties. For example, your case when the call modifies an
> otherwise unescaped pointer. While the proposed solution works for the motivational example, I’m afraid
> it’s not really extensible for the general case.
>
> First, we would need to have the ability to attach operand bundles to instructions of any kind. Second, there
> are open questions, like how do we treat uses which are not a memory operations? How do we deal with
> aliases? These things are not addressed in the current proposal.
>
> The best idea I had so far regarding expressing the results of our EA in the IR was to use noalias and
> alias.scope metadata. Essentially every unescaped object can be assigned to a separate scope and every
> memory operation can be marked as noalias wrt the scopes we know it doesn’t touch.
>
> In fact, the scheme you propose where every user of a !nocapture pointer is annotated with the original
> pointer resembles the alias scopes and noalias metadata. The difference is alias scopes and noalias use
> Indirection through metadata while your proposal relies on operand bundles to directly link the original
> pointer.
>
> Do you think you can apply something like this to you problem?
>
> There is a fundamental difference with the proposal here and alias scopes that is not
> "indirection through metadata". `!nocapture` + `"nocapture_use"` are independent of the
> surrounding. This is crucial because it avoids the scaling problem that comes with the
> bidirectional approach of `noalias.scopes` + `!noalias` metadata. As part of the Fortran
> alias discussions and the AA monthly working group, we have seen how problematic large
> amounts of `noalias` metadata can be. You basically start listing edges in your
> alias/dependence graph at some point.
> It sounds like it's a matter of representation. Conceptually in both cases we are explicitly providing external
> aliasing facts. Note that with the proposed `!nocapture` + `"nocapture_use”` scheme every user of a `!nocapture`
> stored pointer needs to be annotated.
Yes. But not every potential aliasing access. So we are talking about
the number of users of the
memory it is stored in, which needs to be alias free or otherwise we
could not annotate it in the
first place. For the main motivational example the memory has only a
single use, a runtime call.
This can be a __kmpc_reduce (see PR48475),
__tgt_target_data_update_mapper ...
If we deduce the property the single uses in pthread_create, a
__kmpc_fork_call, ... become
interesting as well as cases where we have potentially multiple uses:
```
__attribute__((noinline)) void f1(struct S* s) { int* /* noescape */ p =
s->p; ... }
__attribute__((noinline)) void f2(struct S* s) { int* /* noescape */ p =
s->p; ... }
__attribute__((noinline)) void f3(struct S* s) { int* /* noescape */ p =
s->p; ... }
void g() { struct S s; int x; s.p = &x; f1(&s); f2(&s); f2(&s); /* x is
not captured here! */ ... }
```
The key difference is it doesn't depend on the surrounding, that is the
number of other pointers
floating around. It is linear in the uses of the memory stored in, which
is an OK price to
pay given that `noalias.scopes` + `!noalias` metadata can numerate edges
in a graph that grows
quadratic with the number of pointers.
Note that this *does not* manifest aliasing facts but capture facts,
which is arguably different and simpler.
>
> We are looking at alternative solutions but those
> are so far unrelated to he use case that is discussed here.
>
>
> Given an escape analysis you might be interested in the same use case. A pointer is stored but you can
> see from the uses of the stored-to location that it does not escape. As you pointed out, if those uses
> are non-calls we cannot attach "nocapture_use" for now, which is a problem. On the other hand, if they
> are in the same function, we could forward the pointer and eliminate the load, or the memory is already
> aliasing something and it is unclear how much we can deduce anyway.
> The unfortunately property of this approach is that we can’t simply compute the analysis and annotate the
> IR with the results. First we need to shape the IR in the way so we can attach the results of the analysis.
> This kind of shaping (store load forwarding to eliminate extra aliases) may require the results of this analysis
> itself (this is assuming we want to reuse some of the existing LLVM transforms to do store load forwarding).
>
> You seem to argue this approach is not a good fit for your analysis, which I agree. That seems
> to be not a problem on its own. Do you have an alternative proposal that would subsume this
> approach?
> To me the fact that you need to annonate every single user of a `!nocapture` stored pointer looks like the most
> fragile part of the proposal. It makes it hard to generalize this scheme. It makes it easy to misuse this scheme,
> both for the frontends and for the optimizer. Moreover, current wording of the proposal doesn't address these
> issues.
Hm, OK. So, you don't need to annotate each use, but the common usage
would expect you do.
For DSL usages you might actually not. Assuming you do annotate both
sides, which I can explain
better in the lang ref, this is no different from common techniques,
e.g., `!noalias`. I'm not
sure why this would be any harder. In fact, you don't need to combine
scopes, deal with the
issues of scope duplication, find all side-effect instructions to
annotate that are unrelated,
... so arguably simpler. Lastly, you don't annotate the users of the
memory it was stored in
with `!nocapture`, which is simply following def-use chains. One key
point here is that we might
not see the users of the pointer at all, but just the users of the
memory it was stored in.
In that case, we want to capture the "noescape" property we know from
domain knowledge.
>
> How far can we get with an approach which is conservatively correct without explicit `"nocapture_use”`
> annotations?
>
> Currently there are 2 places where BasicAA uses isNonEscapingLocalObject:
> * Aliasing between isNonEscapingLocalObject and isEscapeSource
> https://github.com/llvm/llvm-project/blob/main/llvm/lib/Analysis/BasicAliasAnalysis.cpp#L1563
> * ModRef for a isNonEscapingLocalObject and a call
> https://github.com/llvm/llvm-project/blob/main/llvm/lib/Analysis/BasicAliasAnalysis.cpp#L839
>
> If we were to support nocapture stores without explicit `"nocapture_use”` annotations we would need to
> differentiate between isNonEscapingLocalObject with StoreCaptures==true and StoreCaptures==false.
> These are two different properties and we would need to handle them separately.
>
> For the aliasing between isNonEscapingLocalObject and isEscapeSource we would need to change
> isEscapeSource for isNonEscapingLocalObject(StoreCaptures==false) case. Specifically, if StoreCaptures==false
> then isEscapeSource should not consider loads as escape sources.
>
> For the ModRef query around calls it's more compilcated. The easiest thing to do here is to not use
> isNonEscapingLocalObject(StoreCaptures==false). I don't know if this is too limiting for your example.
> But we can come up with some analysis here as well. What we need to know is the fact that a call doesn’t
> capture pointers reachable through its arguments. This would require a new argument attribute, something
> like nocapture_content.
I don't follow what this means. The way I understand this it seems more
complicated and brittle
than my proposal. What I suggested should work if you outline part of
the code, introduce int2ptr
+ ptr2int pairs in between the stores and use of the stored memory, etc.
I am always worried that
"implicit" schemes will break as soon as uses get disturbed and are hard
to track afterwards. That
is why I mark things explicitly:
The store would cause the `nocapture` property to be removed so we
say `!nocapture`.
This break the implicit dependence on the memory use calls which we
need to make explicit.
> My next comment is only somewhat related to `"nocapture_use”` issue. Your current proposal marks stores
> as `!nocapture`. But I assume that the store can be marked as nocapture because the underlying object is
> unescaped. Why don't we mark this fact instead? If we have this fact then any store to an unescaped memory
> can be treated as nocapture (assuming we are doing conservatively correct analysis for StoreCaptures==false).
> This will give us a useful fact about the underlying object as well (it can be used for things other than aliasing).
>
> There is a caveat to this suggestion. The proposed `!nocapture` metadata makes it possible to express
> flow-sensitive facts (a store into the same underlying object might be nocapture on one path, but capture on some
> other). We would give up this flow-sensitivity if we make the underlying object instead. On the other hand, global
> property instead of a flow-sensitive propery will make it easier for a conservatively correct analysis.
>
I'll try to capture some thoughts about your proposals:
1) Having `nocapture_memory` on the declaration of storage to indicate
store into it won't escape.
We could look to the pointer operand at the store side to find the
attribute, which is doable I guess.
We give up on field-sensitivity. We need a declaration to anchor the
attribute at, which inherently
comes with a scope. Inlining will become an issue then as the scope is
expanded; ref all the
`noalias.scope` intrinsic stuff we need to do to keep `noalias` scopes
alive.
2) Having `nocaptue_content` on the usage of a pointer in a call.
We give up on field-sensitivity again. This would not help you in the
non-call situations. We would
need to look from to store to all uses of the pointer stored into every
time we ask ourselves if the
memory escaped. This seems doable, albeit more involved for each query
and can be disturbed.
I think I could live with both approaches for the runtime library case,
some deduction cases will also
work. I am doubtful it is better though.
~ Johannes
>
> Partially related: Maybe you want to give a short presentation in our next AA call about your
> analysis and the ideas you have in this space:
>
> https://docs.google.com/document/d/1ybwEKDVtIbhIhK50qYtwKsL50K-NvB6LfuBsfepBZ9Y/edit?usp=sharing
> I’ll plan to join.
>
> Artur
>
>
> ~ Johannes
>
>
>
> Artur
> If we would want to use `!nocapture`
> more fine-grained we could try to come up with a way to tie it to a "nocapture_use", but that would
> certainly make it more complicated.
> ~ Johannes
>
>
> Artur
>
>
>
> Basically, what am I doing wrong that I get a miscompile on this example?
>
> You don't get the clobber because you did not explicitly mark the use of `%a` in `%a'`.
>
> WDYT?
>
> ~ Johannes
>
>
>
> Artur
>
> Note that we do not inline a call with an "unkown" operand bundle, so there is no fear we
> accidentally produce such a situation as you pointed out. A "proper" version of the example
> would be:
>
> ```
> a = new
> store a, ptr, !nocapture
> call foo(ptr, a) !nocapture_use(a)
>
> void foo(arg_ptr. arg_a) {
> a' = load arg_ptr
> v1 = load arg_a
> ...
> }
> ```
> which should be OK.
>
> Does that make sense?
>
> ~ Johannes
>
>
> Artur
>
> On Jan 7, 2021, at 4:20 PM, Johannes Doerfert via llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org><mailto:llvm-dev at lists.llvm.org>> wrote:
>
> TL;DR: A pointer stored in memory is not necessarily captured, let's add a way to express this.
>
> Phab:https://reviews.llvm.org/D93189
>
> --- Commit Message / Rational ---
>
> Runtime functions, as well as regular functions, might require a pointer
> to be passed in memory even though the memory is simply a means to pass
> (multiple) arguments. That is, the indirection through memory is only
> used on the call edge and not otherwise relevant. However, such pointers
> are currently assumed to escape as soon as they are stored in memory
> even if the callee only reloads them and use them in a "non-escaping" way.
> Generally, storing a pointer might not cause it to escape if all "uses of
> the memory" it is stored to all have the "nocapture" property.
>
> To allow optimizations in the presence of pointers stored to memory we
> introduce two new IR extensions. `!nocapture` metadata on stores and
> "nocapture_use" operand bundles for call(base) instructions. The former
> ensures that the store can be ignored for the purpose of escape
> analysis. The latter indicates that a call is using a pointer value
> but not capturing it. This is important as the call might still read
> or write the pointer and since the passing of the pointer through
> memory is not considered "capturing" with the "nocapture" metadata,
> we need to otherwise indicate the potential read/write.
>
> As an example use case where we can deduce `!nocapture` metadata,
> consider the following code:
>
> ```
> struct Payload {
> int *a;
> double *b;
> };
>
> int pthread_create(pthread_t *thread, const pthread_attr_t *attr,
> void *(*start_routine) (void *), void *arg);
>
> int use(double);
>
> void fn(void *v) {
> Payload *p = (Payload*)(v);
> // Load the pointers from the payload and then dereference them,
> // this will not capture the pointers.
> int *a = p->a;
> double *b = p->b;
> *a = use(*b);
> }
>
> void foo(int *a, double *b) {
> Payload p = {a, b};
> pthread_create(..., &fn, &p);
> }
> ```
>
> Given the usage of the payload struct in `fn` we can conclude neither
> `a` nor `b` in are captured in `foo`, however we could not express this
> fact "locally" before. That is, we can deduce and annotate it for the
> arguments `a` and `b` but only since there is no other use (later on).
> Similarly, if the callee would not be known, we were not able to
> describe the "nocapture" behavior of the API.
>
> A follow up patch will introduce `!nocapture` metadata to stores
> generated during OpenMP lowering. This will, among other things, fix
> PR48475. I generally expect us to find more APIs that could benefit from
> the annotation in addition to the deduction we can do if we see the callee.
>
> ---
>
> As always, feedback is welcome. Feel free to look at the phab patch as well.
>
> Thanks,
> Johannes
>
>
> --
> ──────────
> ∽ Johannes
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org><mailto:llvm-dev at lists.llvm.org>
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
More information about the llvm-dev
mailing list