An issue with new PM's requirements on call graph changes

Project Infrastructure LLVM Dev List Archives
1

I have hit a fairly isolated practical issue deploying the new PM, but it does point to a latent theoretical issues as well. I see various ways to address it, but want feedback from others before moving forward.

The issue is that we can introduce out-of-thin-air calls to known library functions (SimplifyLibCalls, etc). These can be introduced in function passes (InstCombine in particular) and that seems highly desirable.

These all look like one of these cases:
1a) Introducing a new call to an LLVM intrinsic
1b) Replacing an existing call with a call to an LLVM intrinsic
2a) Introducing a new call to a declared library function (but not defined)
2b) Replacing an existing call with a call to a declared library function
3a) Introducing a new call to a defined library function
3b) Replacing an existing call with a call to a defined library function

Both #1 and #2 are easy to handle in reality. Intrinsics and declared functions don’t impact the PM’s call graph because there is no need to order the walk over them. But #3 is a real issue.

The only case I have found that actually hits #3 at all hits #3b when building FORTIFY code with the new pass manager because after inlining we do a lot of (really nice) optimizations on library calls to remove unnecessary FORTIFY checks. But this is in theory a problem when LTO-ing with libc. More likely it could be a problem when LTO-ing with a vector math library.

So what do we do?

My initial idea: find all defined library functions in the module, and every time we create a ref edge to one of them, synthesize a ref edge to all of them. This should completely solve #3b above. But it doesn’t really address #3a at all.

Is that OK? It would be very convenient to say that if we want to introduce truly novel and new calls to library functions, we should have an LLVM intrinsic to model those routines.

But we actually have an example (I think) of #3a, introducing a call to a library function out of the blue: memset_pattern. =/

The only way I see to reasonably handle #3a is to have every function implicitly contain a reference edge to every defined library function in the module. This is, needless to say, amazingly wasteful. Hence my email. How important is this?

If we need to correctly handle this, I think I would probably implement this by actually changing the iteration of reference edges in the graph to just implicitly walk the list of defined library functions so that we didn’t burn any space on this. But it will make iteration of reference edges slower and add a reasonable amount of complexity. So I’d like to hear some other opinions before going down either of these roads.

Thanks,
-Chandler

2

Hi Chandler,

I have hit a fairly isolated practical issue deploying the new PM, but it
does point to a latent theoretical issues as well. I see various ways to
address it, but want feedback from others before moving forward.

The issue is that we can introduce out-of-thin-air calls to known library
functions (`SimplifyLibCalls`, etc). These can be introduced in function
passes (`InstCombine` in particular) and that seems highly desirable.

These all look like one of these cases:
1a) Introducing a new call to an LLVM intrinsic
1b) Replacing an existing call with a call to an LLVM intrinsic
2a) Introducing a new call to a declared library function (but not defined)
2b) Replacing an existing call with a call to a declared library function
3a) Introducing a new call to a defined library function
3b) Replacing an existing call with a call to a defined library function

An alternate way to model 3b is to say that if a call to X can be
replaced by a call to Y, X needs to have a ref edge to Y (even if X is
a declaration). The list of these "hidden" ref edges will be provided
by TargetLibraryInfo (or something close).

3a seems logically equivalent to an outliner pass[0] followed by
function commoning to me, so I think we should ask ourselves what the
behavior of an outliner pass needs to be in the new RefSCC world.
Perhaps we're okay not doing 3a from function passes (just like we
would not allow outlining from function passes)? This would involve
refactoring SimplifyLibCalls though, to create a variant that can only
be invoked from a CGSCC pass.

[0]: Assuming we //want// to apply the PM's post-order traversal rule
-- the rule is less beneficial for outlining though, since the
outlined function will probably be simplified already because we
carved it out of a function that has been simplified.

Thanks!
-- Sanjoy

3

I have hit a fairly isolated practical issue deploying the new PM, but it
does point to a latent theoretical issues as well. I see various ways to
address it, but want feedback from others before moving forward.

The issue is that we can introduce out-of-thin-air calls to known library
functions (`SimplifyLibCalls`, etc). These can be introduced in function
passes (`InstCombine` in particular) and that seems highly desirable.

These all look like one of these cases:
1a) Introducing a new call to an LLVM intrinsic
1b) Replacing an existing call with a call to an LLVM intrinsic
2a) Introducing a new call to a declared library function (but not defined)
2b) Replacing an existing call with a call to a declared library function
3a) Introducing a new call to a defined library function
3b) Replacing an existing call with a call to a defined library function

Both #1 and #2 are easy to handle in reality. Intrinsics and declared
functions don't impact the PM's call graph because there is no need to
order the walk over them. But #3 is a real issue.

The only case I have found that actually hits #3 at all hits #3b when
building FORTIFY code with the new pass manager because after inlining we
do a lot of (really nice) optimizations on library calls to remove
unnecessary FORTIFY checks. But this is in *theory* a problem when LTO-ing
with libc. More likely it could be a problem when LTO-ing with a vector
math library.

So what do we do?

My initial idea: find all *defined* library functions in the module, and
every time we create a ref edge to one of them, synthesize a ref edge to
all of them. This should completely solve #3b above. But it doesn't really
address #3a at all.

Is that OK? It would be very convenient to say that if we want to
introduce truly novel and new calls to library functions, we should have an
LLVM intrinsic to model those routines.

That seems reasonable, though it would in theory block inlining of, say,
memset_pattern if we have a definition available when LTO'ing but that
seems acceptable (and we already live with similar restrictions with other
memory intrinsics anyway and I don't hear anyone complaining).

But we actually have an example (I think) of #3a, introducing a call to a
library function out of the blue: memset_pattern. =/

The only way I see to reasonably handle #3a is to have *every* function
implicitly contain a reference edge to every defined library function in
the module. This is, needless to say, amazingly wasteful. Hence my email.
How important is this?

This seems like the simplest and most natural solution overall. Can you get
some numbers on how wasteful it actually is? (and for the alternative
approach)

Even if it ends up not being that wasteful and worth doing just for
simplicity's sake, I do like the idea of "out-of-thin-air calls must be
intrinsics".

-- Sean Silva

4

Out of curiosity, is this the only example we have? In the context of , I was thinking about whether it might make sense to have intrinsic form of memset_pattern anyways. If we did that, could we disallow such cases in practice? On the other hand, it does seem less than desirable to prevent inlining of such cases when we do in fact have the implementation linked in.

5

I have hit a fairly isolated practical issue deploying the new PM, but it does point to a latent theoretical issues as well. I see various ways to address it, but want feedback from others before moving forward.

The issue is that we can introduce out-of-thin-air calls to known library functions (`SimplifyLibCalls`, etc). These can be introduced in function passes (`InstCombine` in particular) and that seems highly desirable.

These all look like one of these cases:
1a) Introducing a new call to an LLVM intrinsic
1b) Replacing an existing call with a call to an LLVM intrinsic
2a) Introducing a new call to a declared library function (but not defined)
2b) Replacing an existing call with a call to a declared library function
3a) Introducing a new call to a defined library function
3b) Replacing an existing call with a call to a defined library function

Both #1 and #2 are easy to handle in reality. Intrinsics and declared functions don't impact the PM's call graph because there is no need to order the walk over them. But #3 is a real issue.

The only case I have found that actually hits #3 at all hits #3b when building FORTIFY code with the new pass manager because after inlining we do a lot of (really nice) optimizations on library calls to remove unnecessary FORTIFY checks. But this is in *theory* a problem when LTO-ing with libc. More likely it could be a problem when LTO-ing with a vector math library.

This latter case concerns me most. When the vectorizer creates the vectorized version of a loop, that's new code (the original code for the loop stays in place as a fall back). Further, the vectorizer can (today) create calls to vector math library functions, and a setup where we LTO with the definitions of those functions is certainly possible (and desirable). As a result, this issue does not seem all that theoretical to me.

Moreover, once we have support for OpenMP simd functions, we'll end up in exactly this situation on a regular basis, and we can't have intrinsics for all of the possible user-defined functions (unless the intrinsic just takes a function pointer and we clean it up afterwards somehow). In short, I think we do need to correctly handle this situation.

FWIW, I can also see this situation come up in other instrumentation cases as well. There are plenty of cases where it is useful to LTO with a runtime library.

Thanks again,
Hal

6

I have hit a fairly isolated practical issue deploying the new PM, but it
does point to a latent theoretical issues as well. I see various ways to
address it, but want feedback from others before moving forward.

The issue is that we can introduce out-of-thin-air calls to known library
functions (`SimplifyLibCalls`, etc). These can be introduced in function
passes (`InstCombine` in particular) and that seems highly desirable.

These all look like one of these cases:
1a) Introducing a new call to an LLVM intrinsic
1b) Replacing an existing call with a call to an LLVM intrinsic
2a) Introducing a new call to a declared library function (but not
defined)
2b) Replacing an existing call with a call to a declared library function
3a) Introducing a new call to a defined library function
3b) Replacing an existing call with a call to a defined library function

Both #1 and #2 are easy to handle in reality. Intrinsics and declared
functions don't impact the PM's call graph because there is no need to
order the walk over them. But #3 is a real issue.

The only case I have found that actually hits #3 at all hits #3b when
building FORTIFY code with the new pass manager because after inlining we
do a lot of (really nice) optimizations on library calls to remove
unnecessary FORTIFY checks. But this is in *theory* a problem when LTO-ing
with libc. More likely it could be a problem when LTO-ing with a vector
math library.

This latter case concerns me most. When the vectorizer creates the
vectorized version of a loop, that's new code (the original code for the
loop stays in place as a fall back). Further, the vectorizer can (today)
create calls to vector math library functions, and a setup where we LTO
with the definitions of those functions is certainly possible (and
desirable). As a result, this issue does not seem all that theoretical to
me.

Moreover, once we have support for OpenMP simd functions, we'll end up in
exactly this situation on a regular basis, and we can't have intrinsics for
all of the possible user-defined functions

Can you clarify how OpenMP simd would require introducing out-of-thin-air
references to an open-ended set of user-defined functions?

(unless the intrinsic just takes a function pointer and we clean it up
afterwards somehow).

Note that simply referencing a function pointer out-of-thin-air would still
run afoul of the same issue. It would constitute a ref edge and have
similar implications as a direct call, at least as far as the fundamental
problem here is concerned (guaranteeing bottom-up iteration order). So an
intrinsic taking a function pointer wouldn't really circumvent the issue
(if I understand correctly what you're saying).

-- Sean Silva

7

Because OpenMP simd functions (i.e. the “declare simd” functionality) allows the user to specify that vectorized versions of a given scalar function are to be made available, meaning generated, or are externally available. For example, let’s say we have this: #pragma omp declare simd notinbranch float min (float a, float b) { Return a < b ? a : b; } void minner (float *a, float b, float c) { #pragma omp parallel for simd for (i=0; i<N; i++) c[i] = min(a[i], b[i], c[i]); } And, assume for a moment that min() was not inlined before vectorization. The pragma says that we’ll generate some vectorized version of the min function taking vector arguments (), and then the vectorizer, when generating the vectorized loop body, will insert a call to said vectorized min function at the appropriate place. There are already patches floating around phabricator to implement this (although I’m not sure of their status), and it certainly is an important functionality. () Intel, at least, has a well-defined ABI for these functions: Good point. Thanks again, Hal