llvm dev - Apr 2016 - Lazily Loaded Modules and Linker::LinkOnlyNeeded

TL;DR - when linking from a lazily loaded module and using 
Linker::LinkOnlyNeeded, bodies of used functions aren't being copied 
during linking.

Previously on one of our products, we would lazily load our runtime 
module (around 9000 functions), and link some user module into this 
(which is in all practical use cases much smaller). Then, post linking, 
we have a pass that runs over the module and rips out all the 
un-materialized functions that weren't being used in the original user 
module.

I only just noticed that LinkModules has a flags parameter that can take 
a LinkOnlyNeeded flag, which made me wonder if I could reverse the link 
order (EG. link from the lazily loaded runtime module into the user 
module), set the LinkOnlyNeeded flag, and hey presto, we wouldn't need 
to have a cleanup pass that ran afterwards ripping out functions that 
weren't used.

So I tried it, and it failed. Basically any function that was still to 
be materialized wasn't getting its body copied over during linking.

The only line of code that differs when you set LinkOnlyNeeded is in 
LinkModules.cpp -> ModuleLinker::linkIfNeeded:

if (shouldLinkOnlyNeeded() && !(DGV && DGV->isDeclaration()))
     return false;


The isDeclaration() for functions has a call to isMaterializable().

Things I've tried:

  * If I don't pass LinkOnlyNeeded but still link from the lazily loaded
    runtime module into the user module, it works (albeit it is orders
    of magnitude slower like we'd expect).
  * If I don't lazily load the runtime module, it works (but again, much
    slower).
  * I tried doing the linking and then materializing the newly linked
    module, but the runtime functions were still missing their bodies
    (which implies the information was lost during linking).
  * If I hack the LinkModules.cpp code such that it checks if the DGV
    could be materialized, and if so materialize it, before checking for
    a declaration again, it works:

if (shouldLinkOnlyNeeded() && !(DGV && DGV->isDeclaration()))
{
     if (DGV && DGV->isMaterializable())
         DGV->materialize();

     if (!(DGV && DGV->isDeclaration()))
         return false;
}

Even with the extra cost of the hack above - this has resulted in a 2x 
speedup in our total link time.

So really here I am wondering - is this expected behaviour? A bug? And 
if so how best to go about fixing the issue would be some grand advice 
from people more in the know!

Cheers,
-Neil.

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160420/7b2229ad/attachment.html>

Hi Neil,
> On Apr 20, 2016, at 5:20 AM, Neil Henning via llvm-dev <llvm-dev at
lists.llvm.org> wrote:
> 
> TL;DR - when linking from a lazily loaded module and using
Linker::LinkOnlyNeeded, bodies of used functions aren't being copied during
linking.
> 
> Previously on one of our products, we would lazily load our runtime module
(around 9000 functions), and link some user module into this (which is in all
practical use cases much smaller).
It sounds reverse to what I would intuitively do (i.e. load the runtime into my
module).
> Then, post linking, we have a pass that runs over the module and rips out
all the un-materialized functions that weren't being used in the original
user module.
> 
> I only just noticed that LinkModules has a flags parameter that can take a
LinkOnlyNeeded flag, which made me wonder if I could reverse the link order (EG.
link from the lazily loaded runtime module into the user module), set the
LinkOnlyNeeded flag, and hey presto, we wouldn't need to have a cleanup pass
that ran afterwards ripping out functions that weren't used.
> 
> So I tried it, and it failed. Basically any function that was still to be
materialized wasn't getting its body copied over during linking.
> 
> The only line of code that differs when you set LinkOnlyNeeded is in
LinkModules.cpp -> ModuleLinker::linkIfNeeded:
> 
> if (shouldLinkOnlyNeeded() && !(DGV &&
DGV->isDeclaration()))
>     return false;
> 
> The isDeclaration() for functions has a call to isMaterializable().
> 
> Things I've tried:
> If I don't pass LinkOnlyNeeded but still link from the lazily loaded
runtime module into the user module, it works (albeit it is orders of magnitude
slower like we'd expect).
> If I don't lazily load the runtime module, it works (but again, much
slower).
> I tried doing the linking and then materializing the newly linked module,
but the runtime functions were still missing their bodies (which implies the
information was lost during linking).
> If I hack the LinkModules.cpp code such that it checks if the DGV could be
materialized, and if so materialize it, before checking for a declaration again,
it works:
> if (shouldLinkOnlyNeeded() && !(DGV &&
DGV->isDeclaration())) {
>     if (DGV && DGV->isMaterializable())
>         DGV->materialize();
> 
>     if (!(DGV && DGV->isDeclaration()))
>         return false;
> }
DGV is the GlobalValue in the *destination* Module, it is not clear to me how
materializing has an effect on the *source* Module.
I am probably missing something here...
> Even with the extra cost of the hack above - this has resulted in a 2x
speedup in our total link time.
> So really here I am wondering - is this expected behaviour? A bug? And if
so how best to go about fixing the issue would be some grand advice from people
more in the know!
> 
The linker was written before Module was lazy loaded I think. Many pieces in
LLVM assume things their working on are materialized.
On a side note (a bit off-topic), I wonder if `isDeclaration()` should return
false for materializable function?

CC Rafael, who knows this code better.

-- 
Mehdi


-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160420/e4a55fb0/attachment.html>

+cc Artem, who added the LinkOnlyNeeded flag.

On Wed, Apr 20, 2016 at 9:18 AM, Mehdi Amini via llvm-dev <
llvm-dev at lists.llvm.org> wrote:
> Hi Neil,
>
> On Apr 20, 2016, at 5:20 AM, Neil Henning via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
>
> TL;DR - when linking from a lazily loaded module and using
> Linker::LinkOnlyNeeded, bodies of used functions aren't being copied
during
> linking.
>
> Previously on one of our products, we would lazily load our runtime module
> (around 9000 functions), and link some user module into this (which is in
> all practical use cases much smaller).
>
>
> It sounds reverse to what I would intuitively do (i.e. load the runtime
> into my module).
>
> Then, post linking, we have a pass that runs over the module and rips out
> all the un-materialized functions that weren't being used in the
original
> user module.
>
> I only just noticed that LinkModules has a flags parameter that can take a
> LinkOnlyNeeded flag, which made me wonder if I could reverse the link order
> (EG. link from the lazily loaded runtime module into the user module), set
> the LinkOnlyNeeded flag, and hey presto, we wouldn't need to have a
cleanup
> pass that ran afterwards ripping out functions that weren't used.
>
> So I tried it, and it failed. Basically any function that was still to be
> materialized wasn't getting its body copied over during linking.
>
> The only line of code that differs when you set LinkOnlyNeeded is in
> LinkModules.cpp -> ModuleLinker::linkIfNeeded:
>
> if (shouldLinkOnlyNeeded() && !(DGV &&
DGV->isDeclaration()))
>     return false;
>
>
> The isDeclaration() for functions has a call to isMaterializable().
>
> Things I've tried:
>
>    - If I don't pass LinkOnlyNeeded but still link from the lazily
loaded
>    runtime module into the user module, it works (albeit it is orders of
>    magnitude slower like we'd expect).
>    - If I don't lazily load the runtime module, it works (but again,
much
>    slower).
>    - I tried doing the linking and then materializing the newly linked
>    module, but the runtime functions were still missing their bodies (which
>    implies the information was lost during linking).
>    - If I hack the LinkModules.cpp code such that it checks if the DGV
>    could be materialized, and if so materialize it, before checking for a
>    declaration again, it works:
>
> if (shouldLinkOnlyNeeded() && !(DGV &&
DGV->isDeclaration())) {
>     if (DGV && DGV->isMaterializable())
>         DGV->materialize();
>
>     if (!(DGV && DGV->isDeclaration()))
>         return false;
> }
>
>
> DGV is the GlobalValue in the *destination* Module, it is not clear to me
> how materializing has an effect on the *source* Module.
> I am probably missing something here...
>
I think the difference here is that the destination module is being lazy
loaded, whereas I typically see the source modules being lazily loaded. So
it sounds like the issue is that DGV has not *yet* been materialized in the
dest module, and therefore DGV->isDeclaration() is returning false, leading
the linkIfNeeded to return false, despite the fact that if we did
materialize DGV it would be a declaration and would decide to link in SGV.

Not sure that this usage mode of lazy loading has been tested before. As
Mehdi says, Rafael may have more insight.

Teresa
>
> Even with the extra cost of the hack above - this has resulted in a 2x
> speedup in our total link time.
>
> So really here I am wondering - is this expected behaviour? A bug? And if
> so how best to go about fixing the issue would be some grand advice from
> people more in the know!
>
> The linker was written before Module was lazy loaded I think. Many pieces
> in LLVM assume things their working on are materialized.
> On a side note (a bit off-topic), I wonder if `isDeclaration()` should
> return false for materializable function?
>
> CC Rafael, who knows this code better.
>
> --
> Mehdi
>
>
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
>

-- 
Teresa Johnson |  Software Engineer |  tejohnson at google.com |  408-460-2413
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160420/5b65e1a5/attachment.html>