Problem with mingw32 DLL build

Folks, there is an issue pretty buried in the commits list that I suspect should have wider visibility.

See r262188 and subsequent discussion. To summarize: it appears that mingw32 was unable to correctly produce a static data member when instantiated as a base class. The “fix” is to then explicitly instantiate the base class separately from its use in a base class.

I think this is unacceptable in this case because these base classes are intended to be the primary way that users of LLVM will define new analysis passes. That means we’ll actually have to teach LLVM library users about this pattern, not just the LLVM developers. =/

So my questions are:

  1. Is there some more elegant way to fix this that doesn’t require every derived class to write an explicit instantiation definition of their base class? If so, then the rest of my questions are moot.

  2. If not, is this a problem with a native Windows 32-bit DLL build?

  3. If this is a problem with the native Windows 32-bit DLL build, should we back out of using this pattern of CRTP injection of the static data member and just deal with the significant (and error prone) boiler plate? It seems less error prone than the alternative.

  4. If this is not a problem with the native Windows 32-bit DLL, then I wonder how valuable it is to continue to support the conjunction of mingw32 and a DLL build? This seems to be a really high cost to carry.

First, I’d like to say it would be great if we could get away from relying on these globally unique pass IDs represented as addresses of globals. Long ago I tried to hack up a DLL build of LLVM and quickly discovered that these IDs are the biggest source of dllimported data, which has to be annotated with __declspec(dllimport/dllexport).

Here’s the issue as I understand it:

  • LLVM today supports using mingw64 compilers with -DBUILD_SHARED_LIBS=ON. This produces a DLL per LLVM library, just like Unix.
  • In lib/Analysis, each analysis inherits from a CRTP base class AnalysisBase.
  • AnalysisBase is very simple, it exists solely to provide the unique pass ID in the form of a static data member.
  • Because the ID is part of a template, it is only emitted when referenced. This is the same on Windows and Unix, no compiler bug here.
  • Previously, each analysis would manually provide it’s own ID, and the ID would be emitted exactly once in the TU providing the analysis.
  • The Windows loader is very simple, it does not merge symbols across DSOs as it does on Unix. As a result, the IDs are not unique, leading to runtime problems.
  • The auto-dll tools provided by mingw might not let you get this far, though, I haven’t actually seen the logs.

You introduced AnalysisBase to eliminate the boilerplate of declaring and defining the ID of each analysis, and discovered that it just means we have to put back a different kind of boilerplate in the form of explicit template instantiations.

I don’t think we can overcome the boilerplate here, if we want to support BUILD_SHARED_LIBS on Windows, regardless of compiler. It’s a limitation of DLLs, and a useful one because it avoids doing tons of useless symbol lookup at startup.

So, the options are:

  1. Go back to declaring IDs in each analysis. It’s tried and true and doesn’t have vague linkage like templates.
  2. Push forward with extern template declarations, which is a new kind of boilerplate that most users of LLVM probably don’t understand very well.
  3. Drop support for BUILD_SHARED_LIBS. This might not actually be viable if we want to try to support transforms and analysis provided by loadable plugins. I’m not sure what issues would arise here.

I guess I favor 1.

I don’t disagree with your analysis, but I’d like to get a build bot that tests DLL builds with a host toolchain other than mingw32 if this is in fact common to those configs (as no other bot broke that i saw)… But that’s a separate issue.

The base class isn’t just providing the static variable. It is also providing the accessor method. I really like using a method instead of raw access to the static data member. It is also providing the name of the pass. So what we’d likely end up with is still having the base class but having a friend declaration and private static data member in the derived class.

But before going there, I’d like to ask if there is another approach that would work: do you see any ways to effectively cause the static data member to be emitted reliably?

First, I'd like to say it would be great if we could get away from relying
on these globally unique pass IDs represented as addresses of globals. Long
ago I tried to hack up a DLL build of LLVM and quickly discovered that
these IDs are the biggest source of dllimported data, which has to be
annotated with __declspec(dllimport/dllexport).

Here's the issue as I understand it:
- LLVM today supports using mingw64 compilers with -DBUILD_SHARED_LIBS=ON.
This produces a DLL per LLVM library, just like Unix.
- In lib/Analysis, each analysis inherits from a CRTP base class
AnalysisBase.
- AnalysisBase is very simple, it exists solely to provide the unique pass
ID in the form of a static data member.
- Because the ID is part of a template, it is only emitted when
referenced. This is the same on Windows and Unix, no compiler bug here.
  - Previously, each analysis would manually provide it's own ID, and the
ID would be emitted exactly once in the TU providing the analysis.
- The Windows loader is very simple, it does not merge symbols across DSOs
as it does on Unix. As a result, the IDs are not unique, leading to runtime
problems.
  - The auto-dll tools provided by mingw might not let you get this far,
though, I haven't actually seen the logs.

You introduced AnalysisBase to eliminate the boilerplate of declaring and
defining the ID of each analysis, and discovered that it just means we have
to put back a different kind of boilerplate in the form of explicit
template instantiations.

I don't think we can overcome the boilerplate here, if we want to support
BUILD_SHARED_LIBS on Windows, regardless of compiler. It's a limitation of
DLLs, and a useful one because it avoids doing tons of useless symbol
lookup at startup.

So, the options are:
1. Go back to declaring IDs in each analysis. It's tried and true and
doesn't have vague linkage like templates.

+1

Looking at this from a documentation perspective, it seems the "tried and
true" method is easiest to document since we just say "declare this ID in
your analysis" and "declare your pass' name" (both with examples, of
course).
The current state is that we have to document "inherit from this CRTP base
class" and "oh, but if you want to manually specify this then here is how
to do it" (I forget where, but having the ability to override this was the
reason given for the PassManager not doing all this internally to itself,
which would be a great simplification I think (there would just be nothing
to document)).
Hence the current state is strictly more to document, and has more "magic"
which can cause problems for new users if they actually have to debug
anything.

Documentation is an absurdly effective way to make boilerplate a non-issue.

-- Sean Silva

I don't disagree with your analysis, but I'd like to get a build bot that
tests DLL builds with a host toolchain other than mingw32 if this is in
fact common to those configs (as no other bot broke that i saw)... But
that's a separate issue.

We don't support BUILD_SHARED_LIBS=ON with MSVC. It doesn't come with tools
that make it easy to support this configuration. CMake recently added some
support for building DLLs that export all symbols, but it's very new and
requires you to annotate all your exported global data, i.e. pass ids.

The base class isn't *just* providing the static variable. It is also

providing the accessor method. I really like using a method instead of raw
access to the static data member. It is also providing the name of the
pass. So what we'd likely end up with is still having the base class but
having a friend declaration and private static data member in the derived
class.

Sure, that'd work.

But before going there, I'd like to ask if there is another approach that
would work: do you see any ways to effectively cause the static data member
to be emitted reliably?

I'm sure we could come up with a way to get the ID emitted reliably at the
definition of the analysis, but we also need to ensure that the ID is
unique, which means we need something like the extern template declaration
in the header.

I think your idea of using CRTP to stamp out the pass name and ID accessor
is reasonable, and then we can use a normal C++ declaration/definition pair
in the derived class for the ID.

I’m about to switch over the the conclusion here since I’ve already gotten this wrong a few times. =[

However, I want to say for the record that I don’t really think we should keep supporting this kind of DLL builds given the limitations of DLLs when it comes to correctly implementing these aspects of C++. I suspect we should really move toward having an LLVM.dll (and a LTO.dll, etc) and only using static linking within that on Windows. (I know that actually producing an LLVM.dll is a huge challenge currently due to CMake and other issues.) Trying to support the fully shared build with the limitations of DLLs when it only even somewhat works on one platform (mingw32, as other things prevent it from working w/ MSVC from my understanding of conversations with Reid) seems like a losing battle and to have a high cost for very little gain for the community.