Which pass should be propagating memory copies

Project Infrastructure LLVM Dev List Archives
1

Keno,
Perhaps you can view the problem to be the memcpys themselves,
We humans can look at the memcpys and see loads and stores
but to almost all optimizer passes they aren’t what it is looking for,
They instead see function calls which they mostly don’t touch,

If these memcpys were inlined into plain old loads and stores
The redundant loads and stores should be deleted by existing opts

A question I have for you is, because this looks like “copy-in-copy-out” argument semantics,
Which to me looks more like Ada than C, what was the source language ?

Peter Lawrence.

2

No, I very much want the memcpys there. With individual stores I'd give up
hope that the optimizer can figure out what's going on here, esp. if it
gets beyond a few bytes, but I with memcpys it does seem doable. As for
which frontend produced this, we're considering adding language semantics
that would produce lots of code like this to julia, so we're looking into
getting the optimizer to fold the extra copies away.

Keno

3

Well, mostly I want to hoist the store to the stack and transform it into a store to the heap. After that the memcpys are essentially trivially dead, so instcombine or dse will delete them for me. If the memcpys were made of individual stores instead, there’d have to be some sort of exponential search somewhere in the compiler to figure that out. For the extreme case consider [100000000 x double]. The same optimization can apply here, but if it tried to do 100000000 stores instead, I wouldn’t expect the compiler to really figure that out. What I meant was that I think the memcpys are the correct representation of this from the frontend, it’s just that I’d like more optimization to happen here.

4

Argh, I’m sorry. I realize the original IR has the most unfortunate typo. What I meant was
that the intermediate store should have happened on the stack, i.e.

%dataptr = getelementptr inbounds [4 x double], [4 x double] *%stack, i32 0, i64 %idx

I’m sorry for any confusion that caused.

5

Keno,
Hmmm, seems like you are saying “copy-in-copy-out” argument semantics are required by the language,
Otherwise you would be using “by-reference" argument semantics,
And that CICO is easiest for you to represent with memcpy.

Usually there are some very subtle issues with CICO and the memory model,
Typically the original object isn’t supposed to be modified until the function returns,
IE multiple stores, especially of different values, to a field in the original object should not be visible, only the final store,
This is clearly “observable" in multithreaded programs, but can also be observable in a single threaded program
If the same object is visible from within the called function for example as a global variable
Which would be seen to have its internal values change multiple times, even though the
Intent of the language using CICO is to try to ensure all-at-once state changes (at least for single-threaded programs)

My advice, if the above applies to you, is to add a new pass to the compiler that figures out if
The transformation from memcpy to explicit multiple load/store is actually legal (won’t produce intermediate
State changes before the exit of the function which would violate the strict CICO calling convention),
And also profitable (I don’t view the code explosion of [1000000 x double] as profitable!),
Or if the transformation from “CICO" to pure “by-reference” is both legal, and profitable.

(Also, don’t forget to check what the language spec says about this function passing the object,
Or parts of it, to other functions before or after making modifications)

My advice regarding teaching GVN about memcpy is not to. It would be one thing if the memcpy
Were copying in/out a single variable, in that case the memcpy can and should be viewed as a load / store pair,
But in your case it isn’t being used that way, it is being used to copy multiple values, and the only
Logical thing that GVN could do is expand those out to multiple individual loads and stores. GVN should not
Be doing this, instead your new pass (that first checks to see if it is legal w.r.t. calling convention) is
The place to do this, or if should convert to pure “by-reference” if legal, which also shouldn’t be done in GVN.

—Peter Lawrence.

6

Hi Peter,

thank you for concern and advice. Since we both write the compiler and design the language, we are not particularly
bound by any pre-existing spec. The concerns about multi-threaded data races are very relevant of course
and we’re well aware of the implications. In the particular case where this comes up, language semantics

generally guarantee that this is unobservable both in single-threaded and multi-threaded contexts (though
we generally do allow the user to shoot themselves in the foot if they want to, the primary concern here
is not really observability, but what the programmer expects from the semantics of the language). For what
it’s worth, this isn’t exactly CICO. Our calling convention is generally by reference. However, we do have
notions of semantic immutability, which is where this particular pattern arises (in cases where a new immutable
gets created by taking an existing field and modifying it in one place only). Because of these semantic
guarantees, we know that there’s no aliasing of the kind that would be problematic (and expose this
information to LLVM through the various AA mechanisms). Now, similar issues of course arise with mutable
memory locations as well. However, in such cases the data race would be explicitly present in the source
program, so we don’t have a problem with the compiler making this optimization. FWIW, our multi-threading
programming model is in the early stages, and we’re considering various language level constraints
on concurrent data modification to mostly disallow that situation unless explicitly opted in to by the user,
but that’s a bit off.

email in the first place). It would of course be very possible for us to write our own pass that pattern matches
this and performs the transformation that we want. However, we generally tend to prefer working with the
community to put the optimizations in the best possible place such that others may automatically take advantage.
It sounds like the community consensus is that GVN should be able to do this kind of optimization (and thanks
to Daniel for providing some guidance on implementation!). If people feel strongly that memcpyopt (or a new pass)
with appropriate pattern matching would be a better place, I’d be happy to go that way as well of course.

Keno

7

Keno,
I suspect that if you, Daniel B., and I were to have an in person meeting this would take 5~10 minutes
For everyone to (in terms D.B. will appreciate) “converge to a fixed point” (:-)! of understanding. Meanwhile we
are stuck with limited email bandwidth, but a meet-up at the next llvm Bay Area social might be a good idea.

Whether GVN is an appropriate place for this opt is going to hinge on the precise details of the calling
Convention, which you are still being vague about. You are simultaneously saying

  1. the memcpy are necessary
  2. the memcpy can and should be opt away
    But these two statements are mutually exclusive, so you need to be more precise about the CC.

Why is this important?, because even if Daniel B. can enhance GVN to “look through” the memcpys
And optimize the loads from the local-stack-copy into loads through the original pointer argument,
And optimize the stores into the local-stack-copy into stores through the original pointer argument,
There is still the issue of deleting the memcpys themselves, which is the actual performance problem.

But the rules of the C/C++ programming language aren’t typically going to allow these deletions,
For example if the original pointer argument is passed to another function, or the address of any
Or all of the local stack copy are passed to another function, or simply calling any function because
It could by the C/C++ rules modify the original data, requiring the memcpys to be preserved.

Also the same logical arguments apply to the loads and stores that Daniel B. thinks he can optimize
In GVN, it depends on where they occur relative to calls to other functions within this function.

The only thing that allows the deletion of the memcpys is intimate knowledge of the Julia-specific
Calling convention. Again similar conclusions apply to even the loads and stores.

And that, IMHO, is inappropriate to include in GVN, which is otherwise a purely C/C++ optimizer,
So a separate Julia calling convention pass is indicated.

PS, Don’t be intimidated by writing an IR-to-IR pass, I’ve already written one, they are easy.
Yours will be particularly easy (after verifying the transform is legal) as it is just a “replace-all-
Uses-of” which already exists, deleting the memcpys, and finally deleting the stack object.

Peter Lawrence.

8

Hi Peter,