modular codegen of class template static member variables

Hi Richard,

(Lang, you’re here because I mentioned stumbling across this on Friday in ORC - this is the reduced test case (where ‘t’ is the NameMutex member and ‘nt’ is the Name member))

Working on getting all LLVM binaries linking successfully under modular codegen, I’ve hit something that seems it’ll need a bit more feature work (which I’m happy/planning to do myself - though always happy for help/advice/etc)…

The test case I have boils down to the following modular header:

struct trivial {};
struct nontrivial { nontrivial(); };
// namespace foo {
void sink(void *);
template struct bar {
static void baz() {
sink(&t);
sink(&nt);
}
static trivial t;

static nontrivial nt;
};
template trivial bar::t;
template nontrivial bar::nt;
//} // namespace foo
template struct bar;
// inline void use() { (void)bar::baz(); }

To build with modular codegen:

$ echo ‘module foo { header “foo.h” }’ > foo.cppmap
$ clang++ -cc1 -xc++ -emit-module -fmodules -w -fmodule-name=foo foo.cppmap -o foo.pcm -fmodules-codegen

So here are some interesting facts I know, some of which may be relevant, some of which may not:

  1. Code as written ends up with linkonce_odr definitions for t and nt

  2. Use use instead of the explicit instantiation and are both t and nt are only declarations

  3. Add the outer namespace foo and then t is emitted as a linkonce_odr definition and nt is emitted as a declaration
    That last one (which was the first result I got) really confuses me - any ideas why a namespace would change the behavior here?

In any case, all those mysteries/differences in behavior might be aside to actually fixing the behavior here, which is what this email is really about.

This is basically the same problem as inline variables, and maybe even would allow some support for static variables in headers too (not sure, will see).

Any ideas what the behavior should be here? Since there’s a desire not to run all global initializers if their specific submodule header isn’t included in the program (for iostream’s sake), how would this be done correctly under modular codegen?
My initial thought is potentially to defer the global initializers to the includers (that seems necessary to get the lazy/only-those-included behavior, right?) But that may not account for indirect inclusion? I guess that’s already handled somehow for the iostreams non-modular case, so maybe it works.

& then the modular object file would perhaps have the weak_odr definition of the global variable itself, but no global initializer - depending on any live codepaths that reference the global necessarily requiring the using TU to have caused the initializer to run? That seems vaguely concerning…

Is this making sense? Any good ideas? Pointers to where to start, etc?

Thanks,

  • Dave

Hi Richard,

(Lang, you're here because I mentioned stumbling across this on Friday in
ORC - this is the reduced test case (where 't' is the NameMutex member and
'nt' is the Name member))

Working on getting all LLVM binaries linking successfully under modular
codegen, I've hit something that seems it'll need a bit more feature work
(which I'm happy/planning to do myself - though always happy for
help/advice/etc)...

The test case I have boils down to the following modular header:

  struct trivial {};
  struct nontrivial { nontrivial(); };
  // namespace foo {
  void sink(void *);
  template <typename T> struct bar {
    static void baz() {
      sink(&t);
      sink(&nt);
    }
    static trivial t;
    static nontrivial nt;
  };
  template <typename T> trivial bar<T>::t;
  template <typename T> nontrivial bar<T>::nt;
  //} // namespace foo
  template struct bar<int>;
  // inline void use() { (void)bar<int>::baz(); }

To build with modular codegen:

  $ echo 'module foo { header "foo.h" }' > foo.cppmap
  $ clang++ -cc1 -xc++ -emit-module -fmodules -w -fmodule-name=foo
foo.cppmap -o foo.pcm -fmodules-codegen

So here are some interesting facts I know, some of which may be relevant,
some of which may not:

   1. Code as written ends up with linkonce_odr definitions for t and nt
   2. Use use instead of the explicit instantiation and are both t and nt are
   only declarations
   3. Add the outer namespace foo and then t is emitted as a linkonce_odr
   definition and nt is emitted as a declaration

That last one (which was the first result I got) really confuses me - any
ideas why a namespace would change the behavior here?

My first guess would be that something has registered a
ASTConsumer::HandleTopLevelDecl callback or similar, and they're assuming
that it gets called for every namespace-scope declaration.

In any case, all those mysteries/differences in behavior might be aside to
actually fixing the behavior here, which is what this email is really about.

This is basically the same problem as inline variables, and maybe even
would allow some support for static variables in headers too (not sure,
will see).

Any ideas what the behavior should be here? Since there's a desire not to
run all global initializers if their specific submodule header isn't
included in the program (for iostream's sake), how would this be done
correctly under modular codegen?
My initial thought is potentially to defer the global initializers to the
includers (that seems necessary to get the lazy/only-those-included
behavior, right?) But that may not account for indirect inclusion? I guess
that's already handled somehow for the iostreams non-modular case, so maybe
it works.

The explicit instantiation definition case is not especially interesting,
because by [temp.spec]/5.1, such things should never appear in modular
headers (because the header could only ever be used in one translation
unit).

So let's focus in on the "inline void use()" case. We need some kind of
mental model for what modular codegen means in order to figure out what
should happen. The way I'm thinking about modular codegen is roughly:

For each header for which we perform modular codegen, we act as if
* that header is a separate translation unit in the program (in *addition*
to being included into other places), and
* for that translation unit, we happen to emit definitions of inline
functions and class metadata, even if they're not otherwise used, and
* in other translation units, we don't need to emit those symbols as a
consequence.

Under that model, emitting the definition of use() should cause us to emit
linkonce_odr definitions of both t and nt into the modular codegen
translation unit. But it should not suppress the emission of linkonce_odr
definitions of t and nt in other translation units too.

However, we also want to *not* run initializers for modules that are not
actually used (eg, we don't want linking against the standard library to
run the iostreams initializer -- and thus link in the iostreams library --
if it's not used, such as for a freestanding / embedded compilation). For
modular codegen, this presumably also needs function sections, and section
GC enabled in the linker.

& then the modular object file would perhaps have the weak_odr definition

of the global variable itself, but no global initializer - depending on any
live codepaths that reference the global necessarily requiring the using TU
to have caused the initializer to run? That seems vaguely concerning...

It does. Mostly I think it works out: if another TU is relying on an inline
function definition to be provided by a modular codegen object, they must
have run the notional initializer for that module, which in turn would have
initialized those globals.

There's one case I'm concerned by: suppose the module is never actually
imported, and all the TUs actually include the header textually. Now,
suppose the inline function and global variables from the modular codegen
TU are selected at link time, and the other copies are all discarded, and
we cleverly put the per-variable global initializers in a COMDAT with the
variables, so they get discarded too. Now we're left with a reference to an
uninitialized global.

Perhaps we need to make a distinction between internal linkage globals with
dynamic initialization in headers (eg the iostreams initializer), which we
run in every user of the modular codegen header, and external linkage
globals with dynamic initialization in headers (eg, inline variables,
static data members of class templates, ...) which we run as part of
initializing the modular codegen translation unit itself. If we do make
that distinction, I worry that we'll lose some of the initialization order
guarantees, though.

Is this making sense? Any good ideas? Pointers to where to start, etc?

Do you reckon that’s worth looking into that? Or just some unimportant oddity?

Ah, good point.

One minor difference here is that the whole module is a translation unit, when it comes to the build details (which are somewhat relevant) - or at least a single object file. Which means used/unused behavior of linkers work at the module granularity, not the header. (eg: if you use a function from one header module object, you’ll pull in all the functions from the module, not just the one header (and all its initializers… ))

Hmm, not sure I followed this last bit. Which part /needs/ functions sections & GC sections?

Right, so we (Richard & I) talked about this over lunch & summarizing our (mostly Richard’s) thoughts on this to the mailing list for posterity:

The initialization order guarantees are that, if I recall correctly - a “happens after” thing. If there’s inline variable A and B, B happens after A at least. Now some other inline variable X might happen before B but after A (if A and X appear in some other translation unit) but importantly B can’t happen before A.

General goal/premise/proposed solution:
Treat modular codegen objects the same as a translation unit that includes the headers, /except/ for the internal linkage globals (eg: iostreams init).

This preserves ordering - it’s just as correct (except for the internal linkage globals) as if you had a separate source file that included all the modular headers & weren’t using modules to compile anything.

The only place it breaks down is the case where your modular codegen non-internal globals (class template static members, variable templates, inline variables, etc) use or depend on the internal linkage global initializer (eg: an inline variable with an initializer that prints to a stream, etc). We’re just going to accept this as broken, I think…

Is that right, Richard? I feel like I left this too long and didn’t remember/include quite all the nuance about ordering, etc. Feel free to add things.

Practically speaking, currently all the globals aren’t quite handled in modular codegen because a lot of this work is triggered specifically by the presence/handling of a module import. So I’ll need to refactor and reuse that code, that should make all the globals be handled - then specifically opt-out/skip over the internal linkage variables to get back to the right behavior there.

  • Dave

Hi Richard,

(Lang, you're here because I mentioned stumbling across this on Friday
in ORC - this is the reduced test case (where 't' is the NameMutex member
and 'nt' is the Name member))

Working on getting all LLVM binaries linking successfully under modular
codegen, I've hit something that seems it'll need a bit more feature work
(which I'm happy/planning to do myself - though always happy for
help/advice/etc)...

The test case I have boils down to the following modular header:

  struct trivial {};
  struct nontrivial { nontrivial(); };
  // namespace foo {
  void sink(void *);
  template <typename T> struct bar {
    static void baz() {
      sink(&t);
      sink(&nt);
    }
    static trivial t;
    static nontrivial nt;
  };
  template <typename T> trivial bar<T>::t;
  template <typename T> nontrivial bar<T>::nt;
  //} // namespace foo
  template struct bar<int>;
  // inline void use() { (void)bar<int>::baz(); }

To build with modular codegen:

  $ echo 'module foo { header "foo.h" }' > foo.cppmap
  $ clang++ -cc1 -xc++ -emit-module -fmodules -w -fmodule-name=foo
foo.cppmap -o foo.pcm -fmodules-codegen

So here are some interesting facts I know, some of which may be
relevant, some of which may not:

   1. Code as written ends up with linkonce_odr definitions for t and
   nt
   2. Use use instead of the explicit instantiation and are both t and
   nt are only declarations
   3. Add the outer namespace foo and then t is emitted as a
   linkonce_odr definition and nt is emitted as a declaration

That last one (which was the first result I got) really confuses me -
any ideas why a namespace would change the behavior here?

My first guess would be that something has registered a ASTConsumer::HandleTopLevelDecl
callback or similar, and they're assuming that it gets called for every
namespace-scope declaration.

Do you reckon that's worth looking into that? Or just some unimportant
oddity?

It would be useful to understand what's going on here, as I'd guess it has
impact in other cases too. But it's not something I'd prioritize.

In any case, all those mysteries/differences in behavior might be aside to

actually fixing the behavior here, which is what this email is really about.

This is basically the same problem as inline variables, and maybe even
would allow some support for static variables in headers too (not sure,
will see).

Any ideas what the behavior should be here? Since there's a desire not
to run all global initializers if their specific submodule header isn't
included in the program (for iostream's sake), how would this be done
correctly under modular codegen?
My initial thought is potentially to defer the global initializers to
the includers (that seems necessary to get the lazy/only-those-included
behavior, right?) But that may not account for indirect inclusion? I guess
that's already handled somehow for the iostreams non-modular case, so maybe
it works.

The explicit instantiation definition case is not especially interesting,
because by [temp.spec]/5.1, such things should never appear in modular
headers (because the header could only ever be used in one translation
unit).

Ah, good point.

So let's focus in on the "inline void use()" case. We need some kind of
mental model for what modular codegen means in order to figure out what
should happen. The way I'm thinking about modular codegen is roughly:

For each header for which we perform modular codegen, we act as if
* that header is a separate translation unit in the program (in
*addition* to being included into other places), and
* for that translation unit, we happen to emit definitions of inline
functions and class metadata, even if they're not otherwise used, and
* in other translation units, we don't need to emit those symbols as a
consequence.

One minor difference here is that the whole module is a translation unit,
when it comes to the build details (which are somewhat relevant) - or at
least a single object file. Which means used/unused behavior of linkers
work at the module granularity, not the header. (eg: if you use a function
from one header module object, you'll pull in all the functions from the
module, not just the one header (and all its initializers... ))

Under that model, emitting the definition of use() should cause us to
emit linkonce_odr definitions of both t and nt into the modular codegen
translation unit. But it should not suppress the emission of linkonce_odr
definitions of t and nt in other translation units too.

However, we also want to *not* run initializers for modules that are not
actually used (eg, we don't want linking against the standard library to
run the iostreams initializer -- and thus link in the iostreams library --
if it's not used, such as for a freestanding / embedded compilation). For
modular codegen, this presumably also needs function sections, and section
GC enabled in the linker.

Hmm, not sure I followed this last bit. Which part /needs/ functions
sections & GC sections?

I... am not sure exactly what case I had in mind there. But I don't think
this impacts the direction we ended up with in our lunchtime discussion.

& then the modular object file would perhaps have the weak_odr definition

of the global variable itself, but no global initializer - depending on any
live codepaths that reference the global necessarily requiring the using TU
to have caused the initializer to run? That seems vaguely concerning...

It does. Mostly I think it works out: if another TU is relying on an
inline function definition to be provided by a modular codegen object, they
must have run the notional initializer for that module, which in turn would
have initialized those globals.

There's one case I'm concerned by: suppose the module is never actually
imported, and all the TUs actually include the header textually. Now,
suppose the inline function and global variables from the modular codegen
TU are selected at link time, and the other copies are all discarded, and
we cleverly put the per-variable global initializers in a COMDAT with the
variables, so they get discarded too. Now we're left with a reference to an
uninitialized global.

Perhaps we need to make a distinction between internal linkage globals
with dynamic initialization in headers (eg the iostreams initializer),
which we run in every user of the modular codegen header, and external
linkage globals with dynamic initialization in headers (eg, inline
variables, static data members of class templates, ...) which we run as
part of initializing the modular codegen translation unit itself. If we do
make that distinction, I worry that we'll lose some of the initialization
order guarantees, though.

Right, so we (Richard & I) talked about this over lunch & summarizing our
(mostly Richard's) thoughts on this to the mailing list for posterity:

The initialization order guarantees are that, if I recall correctly - a
"happens after" thing. If there's inline variable A and B, B happens after
A at least. Now some other inline variable X might happen before B but
after A (if A and X appear in some other translation unit) but importantly
B can't happen before A.

I think the argument was this: Every global variable has either unordered
initialization (in which case we don't need to do anything special), or
ordered initialization (in which case we are required to initialize all
such variables in declaration order), or partially-ordered initialization
(whose order relative to other ordered or partially ordered initialization
must be at least somewhat preserved: specifically, if an inline variable A
is initialized before another variable B in every TU where both appear,
then A's initializer happens before B's:
[basic.start.dynamic]). In a modular header, there
can be no variables with ordered initialization and external linkage,
because that would be an ODR violation, as variables with ordered
initialization have strong linkage. So the only remaining cases are
partially-ordered initialization (inline variables) and internal linkage
globals.

For a modular header that appears in at least two distinct translation
units (which happens for any modular header that is imported), the relevant
ordering guarantees then reduce to: (1) internal linkage globals must be
initialized in declaration order, (2) inline variables must be initialized
in declaration order, and (3) the initialization of any inline variable
must happen before the initialization of any later-declared internal
linkage global. In particular, this permits us to either run the
initializers in declaration order, or to run all of the inline variable
initializers first followed by all of the internal-linkage variable
initializers.

General goal/premise/proposed solution:

Treat modular codegen objects the same as a translation unit that includes
the headers, /except/ for the internal linkage globals (eg: iostreams init).

This preserves ordering - it's just as correct (except for the internal
linkage globals) as if you had a separate source file that included all the
modular headers & weren't using modules to compile anything.

The only place it breaks down is the case where your modular codegen
non-internal globals (class template static members, variable templates,
inline variables, etc) use or depend on the internal linkage global
initializer (eg: an inline variable with an initializer that prints to a
stream, etc). We're just going to accept this as broken, I think...

Is that right, Richard? I feel like I left this too long and didn't
remember/include quite all the nuance about ordering, etc. Feel free to add
things.

I think it's actually not even broken. :slight_smile: That is, given two translation
units both containing this:

#include <iostream> // internal linkage variable initializes iostreams
inline auto &foo = std::cout << "hello world";

... there is no guarantee the iostream initializer runs before the 'foo'
variable's initializer. (The partial ordering rule doesn't apply.)

Ah, right - thanks for re-clarifying all the ordering guarantees. Took me a few times reading it over to get it.

So the partial ordering rules are that this:

inline T x;
inline T y;

is ordered (if they appear in that order in all TUs where both appear), and this is ordered:

inline T x;
static T y;

But this is not:

static T y;
inline T x;

I’m guessing the model this supports is “inline variables in headers, static variables after that/in the implementation file” - so it makes sense to require inline variables to be ordered WRT each other, when they always appear in a certain order (eg: two inline variables in the same header (or a header can depend on the inline variables of the other header it includes) - but not in unrelated headers (because they could be included in either order)). And inline variables come before static variables so you can include some headers, then in your implementation file you have some static variables that might depend on those inline variables.

So, as long as static variables in modular headers are not actually /used/ by anything (which would be ODR violations if they were) then it seems OK, if a bit sneaky, just not to emit them in the modular object (to support iostreams conditional init - but at the cost of breaking the “a modular object is just the same as including the header & pinning the inline functions” - now it specifically opts out of the static variables (which, admittedly, it already was doing in the implementation today))

What about those sneaky uses of static globals that aren’t ODR violations? I guess all of those require non-ODR uses, and are those all guaranteed in LLVM to actually not use the global but instead just inline the value? (eg: I think it’s acceptable to have a static global reference, or a const value that’s never ODR used (eg: could be used for non-type template parameters, static array bounds, etc)) fingers crossed

Sorry for the long-windedness, just talking it through to hopefully get it more clear in my head.

Thanks again!

  • Dave