Hello,
In the example of the patch, we crash because convert_nontype_argument
calls maybe_constant_value with the PTRMEM_CST for &B::i (in cxx0x
mode) which ends up trying to poke at the offset of i, but struct B is
not yet laid out.
My understanding is that using a pointer to non-static member as an
argument of the template A (in this context) is invalid anyway, so we
could avoid trying to see if it's a constant value altogether, similar
to what's done in the 'else if' right after the offending 'if'.
Thus the patch below, bootstrapped and tested on
x86_64-unknown-linux-gnu against trunk.
gcc/cp/
PR c++/51476
* pt.c (convert_nontype_argument): Don't call maybe_constant_value
for PTRMEM_CST nodes.
gcc/testsuite/
PR c++/51476
* cpp0x/ptrmem-cst-arg1.C: New test.
---
gcc/cp/pt.c | 3 ++-
gcc/testsuite/g++.dg/cpp0x/ptrmem-cst-arg1.C | 9 +++++++++
2 files changed, 11 insertions(+), 1 deletions(-)
create mode 100644 gcc/testsuite/g++.dg/cpp0x/ptrmem-cst-arg1.C
diff --git a/gcc/cp/pt.c b/gcc/cp/pt.c
index bb5aa0c..8c9f97b 100644
--- a/gcc/cp/pt.c
+++ b/gcc/cp/pt.c
@@ -5720,7 +5720,8 @@ convert_nontype_argument (tree type, tree expr,
tsubst_flags_t complain)
to a null value, but otherwise still need to be of a specific form. */
if (cxx_dialect >= cxx0x)
{
- if (INTEGRAL_OR_ENUMERATION_TYPE_P (type))
+ if (INTEGRAL_OR_ENUMERATION_TYPE_P (type)
+ && TREE_CODE (expr) != PTRMEM_CST)
expr = maybe_constant_value (expr);
else if (TYPE_PTR_P (type)
|| (TYPE_PTR_TO_MEMBER_P (type)
diff --git a/gcc/testsuite/g++.dg/cpp0x/ptrmem-cst-arg1.C
b/gcc/testsuite/g++.dg/cpp0x/ptrmem-cst-arg1.C
new file mode 100644
index 0000000..b6c81d5
--- /dev/null
+++ b/gcc/testsuite/g++.dg/cpp0x/ptrmem-cst-arg1.C
@@ -0,0 +1,9 @@
+// Origin PR c++/51476
+// { dg-options "-std=c++11" }
+
+template<int> struct A {};
+struct B
+{
+ int i;
+ A<&B::i> a; // { dg-error "could not convert template argument" }
+};
--
1.7.6.4
--
Dodji
