On Wed, 2 Jun 2021, Jason Merrill wrote:
> On 6/2/21 4:56 PM, Patrick Palka wrote:
> > On Wed, 2 Jun 2021, Patrick Palka wrote:
> >
> > > On Wed, 2 Jun 2021, Jason Merrill wrote:
> > >
> > > > On 6/2/21 2:39 PM, Patrick Palka wrote:
> > > > > Here, the dependent template name in the return type of f() resolves
> > > > > to
> > > > > an alias of int& after substitution, and we end up complaining about
> > > > > qualifying this reference type with 'const' from
> > > > > cp_build_qualified_type
> > > > > rather than just silently dropping the qualification as per
> > > > > [dcl.ref]/1.
> > > >
> > > > Hmm, the patch looks fine, but why does the TYPE_DECL test fail for the
> > > > alias?
> > >
> > > Ah, I hadn't considered investigating that. It seems make_typename_type
> > > always returns a _TYPE instead of a TYPE_DECL when resolving a dependent
> > > name that's a template-id, regardless of the tf_keep_type_decl flag.
> > > This can be easily fixed like so, and this change alone is sufficient to
> > > fix the PR (no changes to qualttp20.C needed). Note that this change
> > > should only have an effect when tf_keep_type_decl is passed to
> > > make_typename_type, and the only such caller is the TYPENAME_TYPE case
> > > of tsubst in question, so this change seems pretty safe.
> > >
> > > The downside is that we don't get the __restrict__-dropping
> > > "improvement" as exhibited by qualttp20.C that the original patch
> > > provides, so this other approach is more conservative in that sense.
> > >
> > > So shall we go with the original patch, or something like the following?
> > > (If we go with the original patch, it just occurred to me that we could
> > > remove tf_keep_type_decl altogether.) Testing in progress.
> >
> > For sake of concreteness, here's the full alternative patch for
> > consideration (modulo ChangeLog):
>
> This seems better. I think the only non-type return from
> lookup_template_class is error_mark_node; does it work to check that
> specifically rather than !TYPE_P?
Indeed, checking for error_mark_node instead works nicely. Does the
following look OK? Bootstrapped and regtested on x86_64-pc-linux-gnu.
-- >8 --
Subject: [PATCH] c++: cv-qualified dependent name of alias tmpl [PR100592]
Here, the dependent template name in the return type of f() resolves to
an alias of int& after substitution, and we end up complaining about
qualifying this reference type with 'const' from cp_build_qualified_type
rather than just silently dropping the qualification as per [dcl.ref]/1.
The problem is ultimately that make_typename_type ignores the
tf_keep_type_decl flag when the dependent name is a template-id. This
in turn causes the TYPE_DECL check within tsubst <case TYPENAME_TYPE>
to fail, and so we end up not passing tf_ignore_bad_quals to
cp_build_qualified_type. This patch fixes this by making
make_typename_type respect the tf_keep_type_decl flag even in the case
of a dependent template-id name.
Bootstrapped and regtested on x86_64-pc-linux-gnu, does this look OK for
trunk?
PR c++/100592
gcc/cp/ChangeLog:
* decl.c (make_typename_type): After dispatching to
lookup_template_class, adjust the result to its TYPE_NAME
and then consider the tf_keep_type_decl flag.
gcc/testsuite/ChangeLog:
* g++.dg/cpp0x/alias-decl-71.C: New test.
* g++.dg/template/qualttp20.C: Remove dg-error and augment.
---
gcc/cp/decl.c | 13 +++++++++----
gcc/testsuite/g++.dg/cpp0x/alias-decl-71.C | 13 +++++++++++++
2 files changed, 22 insertions(+), 4 deletions(-)
create mode 100644 gcc/testsuite/g++.dg/cpp0x/alias-decl-71.C
diff --git a/gcc/cp/decl.c b/gcc/cp/decl.c
index fb21a3a1ae8..a3687dbb0dd 100644
--- a/gcc/cp/decl.c
+++ b/gcc/cp/decl.c
@@ -4136,10 +4136,15 @@ make_typename_type (tree context, tree name, enum
tag_types tag_type,
return error_mark_node;
if (want_template)
- return lookup_template_class (t, TREE_OPERAND (fullname, 1),
- NULL_TREE, context,
- /*entering_scope=*/0,
- complain | tf_user);
+ {
+ t = lookup_template_class (t, TREE_OPERAND (fullname, 1),
+ NULL_TREE, context,
+ /*entering_scope=*/0,
+ complain | tf_user);
+ if (t == error_mark_node)
+ return error_mark_node;
+ t = TYPE_NAME (t);
+ }
if (DECL_ARTIFICIAL (t) || !(complain & tf_keep_type_decl))
t = TREE_TYPE (t);
diff --git a/gcc/testsuite/g++.dg/cpp0x/alias-decl-71.C
b/gcc/testsuite/g++.dg/cpp0x/alias-decl-71.C
new file mode 100644
index 00000000000..6a61f93a0b0
--- /dev/null
+++ b/gcc/testsuite/g++.dg/cpp0x/alias-decl-71.C
@@ -0,0 +1,13 @@
+// PR c++/100592
+// { dg-do compile { target c++11 } }
+
+template<bool>
+struct meta {
+ template<class> using if_c = int&;
+};
+
+template<bool B>
+typename meta<B>::template if_c<void> const f();
+
+using type = decltype(f<true>());
+using type = int&;
--
2.32.0.rc2
