[v2] c++/modules: Remember that header units have CMIs

On Tue, May 14, 2024 at 06:21:48PM -0400, Jason Merrill wrote:
> On 5/12/24 22:58, Nathaniel Shead wrote:
> > Bootstrapped and regtested on x86_64-pc-linux-gnu, OK for trunk?
> 
> OK.
> 

I realised as I was looking over this again that I might have spoken too
soon with the header unit example being supported. Doing the following:

  // a.H
  struct { int y; } s;
  decltype(s) f(decltype(s));  // { dg-error "used but never defined" }
  inline auto x = f({ 123 });
  
  // b.C 
  struct {} unrelated;
  import "a.H";
  decltype(s) f(decltype(s) x) {
    return { 456 + x.y };
  }

  // c.C
  import "linkage-3_a.H";
  int main() { auto a = x.y; }

Actually does fail to link, because in 'c.C' we call 'f(.anon_0)' but
the definition 'b.C' is f(.anon_1).

I don't think this is fixable, so I don't think this direction is
workable.

That said, I think that it might still be worth making header modules
satisfy 'module_has_cmi_p', since that is true to the name, and will be
useful in other places we currently use 'module_p ()': in which case we
could instead make all the callers in 'no_linkage_check' do
'module_maybe_has_cmi_p () && !header_module_p ()'; something like the
following, perhaps?

But I'm not too fussed about this overall if you think this will just
make things more complicated. Otherwise bootstrapped and regtested (so
far just modules.exp) on x86_64-pc-linux-gnu, OK for trunk if full
regtest passes?

-- >8 --

This appears to be an oversight in the definition of module_has_cmi_p.
This change will allow us to use the function directly in more places
that need to additional work only if generating a module CMI in the
future.

However, we do need to change callers of 'module_maybe_has_cmi_p'; in
particular header units, though having a CMI, do not provide a TU to
emit names into, and thus each importer will emit their own definitions
which may not match for no-linkage types.

gcc/cp/ChangeLog:

	* cp-tree.h (module_has_cmi_p): Also true for header units.
	* decl.cc (grokfndecl): Disallow no-linkage names in header
	units.
	* tree.cc (no_linkage_check): Likewise.

gcc/testsuite/ChangeLog:

	* g++.dg/modules/linkage-3.H: New test.

Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
---
 gcc/cp/cp-tree.h                         |  2 +-
 gcc/cp/decl.cc                           |  2 +-
 gcc/cp/tree.cc                           | 13 +++++++-----
 gcc/testsuite/g++.dg/modules/linkage-3.H | 25 ++++++++++++++++++++++++
 4 files changed, 35 insertions(+), 7 deletions(-)
 create mode 100644 gcc/testsuite/g++.dg/modules/linkage-3.H

On Fri, May 17, 2024 at 04:14:31PM +1000, Nathaniel Shead wrote:
> On Tue, May 14, 2024 at 06:21:48PM -0400, Jason Merrill wrote:
> > On 5/12/24 22:58, Nathaniel Shead wrote:
> > > Bootstrapped and regtested on x86_64-pc-linux-gnu, OK for trunk?
> > 
> > OK.
> > 
> 
> I realised as I was looking over this again that I might have spoken too
> soon with the header unit example being supported. Doing the following:
> 
>   // a.H
>   struct { int y; } s;
>   decltype(s) f(decltype(s));  // { dg-error "used but never defined" }
>   inline auto x = f({ 123 });
>   
>   // b.C 
>   struct {} unrelated;
>   import "a.H";
>   decltype(s) f(decltype(s) x) {
>     return { 456 + x.y };
>   }
> 
>   // c.C
>   import "linkage-3_a.H";
>   int main() { auto a = x.y; }
> 
> Actually does fail to link, because in 'c.C' we call 'f(.anon_0)' but
> the definition 'b.C' is f(.anon_1).
> 
> I don't think this is fixable, so I don't think this direction is
> workable.
> 
> That said, I think that it might still be worth making header modules
> satisfy 'module_has_cmi_p', since that is true to the name, and will be
> useful in other places we currently use 'module_p ()': in which case we
> could instead make all the callers in 'no_linkage_check' do
> 'module_maybe_has_cmi_p () && !header_module_p ()'; something like the
> following, perhaps?
> 
> But I'm not too fussed about this overall if you think this will just
> make things more complicated. Otherwise bootstrapped and regtested (so
> far just modules.exp) on x86_64-pc-linux-gnu, OK for trunk if full
> regtest passes?
> 
> -- >8 --
> 
> This appears to be an oversight in the definition of module_has_cmi_p.
> This change will allow us to use the function directly in more places
> that need to additional work only if generating a module CMI in the
> future.
> 
> However, we do need to change callers of 'module_maybe_has_cmi_p'; in
> particular header units, though having a CMI, do not provide a TU to
> emit names into, and thus each importer will emit their own definitions
> which may not match for no-linkage types.
> 
> gcc/cp/ChangeLog:
> 
> 	* cp-tree.h (module_has_cmi_p): Also true for header units.
> 	* decl.cc (grokfndecl): Disallow no-linkage names in header
> 	units.
> 	* tree.cc (no_linkage_check): Likewise.
> 
> gcc/testsuite/ChangeLog:
> 
> 	* g++.dg/modules/linkage-3.H: New test.
> 
> Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
> ---
>  gcc/cp/cp-tree.h                         |  2 +-
>  gcc/cp/decl.cc                           |  2 +-
>  gcc/cp/tree.cc                           | 13 +++++++-----
>  gcc/testsuite/g++.dg/modules/linkage-3.H | 25 ++++++++++++++++++++++++
>  4 files changed, 35 insertions(+), 7 deletions(-)
>  create mode 100644 gcc/testsuite/g++.dg/modules/linkage-3.H
> 
> diff --git a/gcc/cp/cp-tree.h b/gcc/cp/cp-tree.h
> index ba9e848c177..ac55b5579a1 100644
> --- a/gcc/cp/cp-tree.h
> +++ b/gcc/cp/cp-tree.h
> @@ -7381,7 +7381,7 @@ inline bool module_interface_p ()
>  inline bool module_partition_p ()
>  { return module_kind & MK_PARTITION; }
>  inline bool module_has_cmi_p ()
> -{ return module_kind & (MK_INTERFACE | MK_PARTITION); }
> +{ return module_kind & (MK_INTERFACE | MK_PARTITION | MK_HEADER); }
>  
>  inline bool module_purview_p ()
>  { return module_kind & MK_PURVIEW; }
> diff --git a/gcc/cp/decl.cc b/gcc/cp/decl.cc
> index 6fcab615d55..f89a7df30b7 100644
> --- a/gcc/cp/decl.cc
> +++ b/gcc/cp/decl.cc
> @@ -10802,7 +10802,7 @@ grokfndecl (tree ctype,
>         used by an importer.  We don't just use module_has_cmi_p here
>         because for entities in the GMF we don't yet know whether this
>         module will have a CMI, so we'll conservatively assume it might.  */
> -    publicp = module_maybe_has_cmi_p ();
> +    publicp = module_maybe_has_cmi_p () && !header_module_p ();
>  
>    if (publicp && cxx_dialect == cxx98)
>      {
> diff --git a/gcc/cp/tree.cc b/gcc/cp/tree.cc
> index 9d37d255d8d..00c50e3130d 100644
> --- a/gcc/cp/tree.cc
> +++ b/gcc/cp/tree.cc
> @@ -2974,9 +2974,9 @@ verify_stmt_tree (tree t)
>  
>  /* Check if the type T depends on a type with no linkage and if so,
>     return it.  If RELAXED_P then do not consider a class type declared
> -   within a vague-linkage function or in a module CMI to have no linkage,
> -   since it can still be accessed within a different TU.  Remember:
> -   no-linkage is not the same as internal-linkage.  */
> +   within a vague-linkage function or in a non-header module CMI to
> +   have no linkage, since it can still be accessed within a different TU.
> +   Remember: no-linkage is not the same as internal-linkage.  */
>  
>  tree
>  no_linkage_check (tree t, bool relaxed_p)
> @@ -3019,7 +3019,8 @@ no_linkage_check (tree t, bool relaxed_p)
>  	{
>  	  if (relaxed_p
>  	      && TREE_PUBLIC (CP_TYPE_CONTEXT (t))
> -	      && module_maybe_has_cmi_p ())
> +	      && module_maybe_has_cmi_p ()
> +	      && !header_module_p ())
>  	    /* This type could possibly be accessed outside this TU.  */
>  	    return NULL_TREE;
>  	  else
> @@ -3037,7 +3038,9 @@ no_linkage_check (tree t, bool relaxed_p)
>  	    {
>  	      if (relaxed_p
>  		  && (vague_linkage_p (r)
> -		      || (TREE_PUBLIC (r) && module_maybe_has_cmi_p ())))
> +		      || (TREE_PUBLIC (r)
> +			  && module_maybe_has_cmi_p ()
> +			  && !header_module_p ())))
>  		r = CP_DECL_CONTEXT (r);
>  	      else
>  		return t;
> diff --git a/gcc/testsuite/g++.dg/modules/linkage-3.H b/gcc/testsuite/g++.dg/modules/linkage-3.H
> new file mode 100644
> index 00000000000..a34ff084eaf
> --- /dev/null
> +++ b/gcc/testsuite/g++.dg/modules/linkage-3.H
> @@ -0,0 +1,25 @@
> +// { dg-additional-options "-fmodule-header" }
> +// { dg-module-cmi !{} }
> +
> +// Like linkage-1, but for header units.
> +
> +// External linkage definitions must be declared as 'inline' to satisfy
> +// [module.import] p6, so we don't need to care about voldemort types in
> +// function definitions.  However, we still care about anonymous types like
> +// this: because a header unit is not a TU, it's up to each importer to export
> +// the name declared here, and depending on what other anonymous types they
> +// declare they could give each declaration different mangled names.
> +// So we should still complain about this because in general it's not safe
> +// to assume that the declaration can be provided in another TU; this is OK
> +// to do by [module.import] p5.
> +
> +inline auto f() {
> +  struct A {};
> +  return A{};
> +}
> +decltype(f()) g();  // OK, vague linkage function
> +auto x = g();
> +
> +struct { int y; } s;
> +decltype(s) h();  // { dg-error "used but never defined" }
> +inline auto y = h();
> -- 
> 2.43.2
> 

Oops, I attached the wrong version of the test: it should be this one:

diff --git a/gcc/testsuite/g++.dg/modules/linkage-3.H b/gcc/testsuite/g++.dg/modules/linkage-3.H
new file mode 100644
index 00000000000..3e1b4bad057
--- /dev/null
+++ b/gcc/testsuite/g++.dg/modules/linkage-3.H
@@ -0,0 +1,26 @@
+// { dg-additional-options "-fmodule-header -Wno-error=c++20-extensions" }
+// { dg-module-cmi !{} }
+
+// Like linkage-1, but for header units.
+
+// External linkage definitions must be declared as 'inline' to satisfy
+// [module.import] p6, so we don't need to care about voldemort types in
+// function definitions.  However, we still care about anonymous types like
+// this: because a header unit is not a TU, it's up to each importer to export
+// the name declared here, and depending on what other anonymous types they
+// declare they could give each declaration different mangled names.
+// So we should still complain about this because in general it's not safe
+// to assume that the declaration can be provided in another TU; this is OK
+// to do by [module.import] p5.
+
+// OK, vague linkage function
+inline auto f() {
+  struct A {};
+  return A{};
+}
+decltype(f()) g();  // { dg-warning "used but not defined" "" { target c++17_down } }
+auto x = g();
+
+struct { int y; } s;
+decltype(s) h();  // { dg-error "used but never defined" }
+inline auto y = h();

On 5/17/24 02:14, Nathaniel Shead wrote:
> On Tue, May 14, 2024 at 06:21:48PM -0400, Jason Merrill wrote:
>> On 5/12/24 22:58, Nathaniel Shead wrote:
>>> Bootstrapped and regtested on x86_64-pc-linux-gnu, OK for trunk?
>>
>> OK.
>>
> 
> I realised as I was looking over this again that I might have spoken too
> soon with the header unit example being supported. Doing the following:
> 
>    // a.H
>    struct { int y; } s;
>    decltype(s) f(decltype(s));  // { dg-error "used but never defined" }
>    inline auto x = f({ 123 });
>    
>    // b.C
>    struct {} unrelated;
>    import "a.H";
>    decltype(s) f(decltype(s) x) {
>      return { 456 + x.y };
>    }
> 
>    // c.C
>    import "linkage-3_a.H";
>    int main() { auto a = x.y; }
> 
> Actually does fail to link, because in 'c.C' we call 'f(.anon_0)' but
> the definition 'b.C' is f(.anon_1).
> 
> I don't think this is fixable, so I don't think this direction is
> workable.

Since namespace-scope anonymous types are TU-local, we don't need to 
support that for proper modules, but it's not clear to me that we don't 
need to support it for header units.

OTOH, https://eel.is/c++draft/module#import-5.3 allows c.C to import a 
different header unit than b.C, in which case the type is different and 
x violates the odr.

> That said, I think that it might still be worth making header modules
> satisfy 'module_has_cmi_p', since that is true to the name, and will be
> useful in other places we currently use 'module_p ()': in which case we
> could instead make all the callers in 'no_linkage_check' do
> 'module_maybe_has_cmi_p () && !header_module_p ()'; something like the
> following, perhaps?

If we need that condition, it should be its own predicate rather than 
expecting callers to do that combined check.

But it's not clear to me how this is different from a type in the GMF of 
a named module, which is exactly the maybe_has_cmi case; there we could 
again see a different version of the type if another TU includes the header.

Jason