Linker issue

Project Infrastructure LLVM Dev List Archives
1

Some background: We have an issue with in loop values being correctly marked uniform but the out of loop uses can be non-uniform. Currently the out of loop users are not marked as divergent because the in loop value is uniform inside the loop. We have gotten around this problem for the moment by applying LCSSA which inserts a PHI in the loop exit for the in loop uniform value that allows the divergent info to be passed onto isel.

The LCSSA is being inserted into XXXDAGToDAGISel class but this causes a pass scheduling conflict with StackProtector. So what we’ve done is try to preserve StackProtector in LCSSA, the issue is that the linker fails in Debug build (but not Release build).

StackProtector lies in CodeGen while LCSSA lies in Tranforms/Utils.

Matt had mentioned to me that you can’t refer to a preserved pass if it’s not on the same library and that it might make sense to move the transform IDs into a separate library. We need a way to mark transforms/analyses as preserved without depending on linking the transform itself.

Thanks,

Ryan

2

Hi all,

re-upping and renaming this thread to get some attention, as we'd like
some feedback on a change that affects the future design direction.

The concrete problem: We would like to mark the LCSSA pass as
preserving the StackProtector pass in order to be able to properly
express all dependencies we have. Unfortunately, those passes are
defined in different modules that do not (and should not) depend on
each other, which causes linking to fail.

Our proposal is to introduce a collection of cross-module pass IDs in
the core module (lib/IR) to work around such issues. This is done by
the RFC patch at https://reviews.llvm.org/D62802.

The only potential alternative I can see would be to generate the IDs
using some kind of template magic and then rely on (runtime) linker
magic to common the generated symbols. That feels like a bit too much
magic for a fundamentally simple problem to me, hence the more
low-tech proposal in the linked Phabricator review.

Please let us know your thoughts!

Thanks,
Nicolai

3

Any comments at all on this? Chandler perhaps?

I've since dug a bit further, and it seems like the template-based
solution wouldn't work anyway because DLL loading on Windows can't do
the required commoning. So the general approach taken in
https://reviews.llvm.org/D62802 seems to be the only technically
viable path forward, though it would still be good to get an outside
look at the concrete details.

Thanks,
Nicolai

4

Any comments at all on this? Chandler perhaps?

I've since dug a bit further, and it seems like the template-based
solution wouldn't work anyway because DLL loading on Windows can't do
the required commoning. So the general approach taken in
⚙ D62802 [RFC][AMDGPU] Uniform values being used outside loop marked non-divergent seems to be the only technically
viable path forward, though it would still be good to get an outside
look at the concrete details.

Thanks,
Nicolai

Hi all,

re-upping and renaming this thread to get some attention, as we'd like
some feedback on a change that affects the future design direction.

The concrete problem: We would like to mark the LCSSA pass as
preserving the StackProtector pass in order to be able to properly
express all dependencies we have. Unfortunately, those passes are
defined in different modules that do not (and should not) depend on
each other, which causes linking to fail.

You say "should not", but I'm not sure that's clear. If LCSSA has
knowledge of the StackProtector, then there is a legitimate dependence.
Either we express that directly, or alternatively, we abstract the
relevant properties to create an abstract dependence. For example, we
have the setPreservesCFG/CFGOnly property pair. In general, I think that
we should have more of these abstract properties. That having been said,
if the orthogonality is very specific, then a specific dependence may be
appropriate.

-Hal

5

New week, new attempt at a ping...?

6

Huh, Gmail failed at threading, causing me to miss this response. Anyway:

> Any comments at all on this? Chandler perhaps?
>
> I've since dug a bit further, and it seems like the template-based
> solution wouldn't work anyway because DLL loading on Windows can't do
> the required commoning. So the general approach taken in
> ⚙ D62802 [RFC][AMDGPU] Uniform values being used outside loop marked non-divergent seems to be the only technically
> viable path forward, though it would still be good to get an outside
> look at the concrete details.
>
> Thanks,
> Nicolai
>
>> Hi all,
>>
>> re-upping and renaming this thread to get some attention, as we'd like
>> some feedback on a change that affects the future design direction.
>>
>> The concrete problem: We would like to mark the LCSSA pass as
>> preserving the StackProtector pass in order to be able to properly
>> express all dependencies we have. Unfortunately, those passes are
>> defined in different modules that do not (and should not) depend on
>> each other, which causes linking to fail.

You say "should not", but I'm not sure that's clear. If LCSSA has
knowledge of the StackProtector, then there is a legitimate dependence.
Either we express that directly, or alternatively, we abstract the
relevant properties to create an abstract dependence. For example, we
have the setPreservesCFG/CFGOnly property pair. In general, I think that
we should have more of these abstract properties. That having been said,
if the orthogonality is very specific, then a specific dependence may be
appropriate.

There is mutual knowledge between the passes, yes. However:

1. Does that really mean that there is a _dependence_? And what kind
of dependence are we talking about anyway?
2. Even if there is a dependence, why should a linker be the correct
way of expressing it?

More to the point, what are you really trying to achieve here?

To the "abstract property" example, I don't think the analogy with the
CFG quite holds because that is about analyses, unlike our situation.
In our use case, there are two passes modifying the IR to establish
what one might think of as different "normal forms", and we want to
express the fact that those are compatible and can both be established
simultaneously. So they are conceptually quite different things.

Note that this directly interacts with the linking question. If you
have two passes each establishing a kind of normal form, and they are
compatible with each other, then beyond the minimal change I've put on
Phabricator which is the bare minimum to make our use case work, you
quite likely want to say that they preserve each other in many cases,
in which case the linker requirement would cause a cyclical dependency
that breaks everything.

As an additional food for thought, I've also started to wonder if we'd
even be able to express this kind of dependency in the non-legacy pass
manager. Is this something we _want_ to be able to express in the pass
manager?

Cheers,
Nicolai