From mjh at edg.com  Mon Jun  4 15:36:13 2012
From: mjh at edg.com (Mike Herrick)
Date: Mon, 4 Jun 2012 11:36:13 -0400
Subject: [cxx-abi-dev] Substitutions for nested template parameters
Message-ID: <8A5D71C3-2542-429F-8B98-5D41C03C4E3B@edg.com>

A customer recently brought this to our attention.  In some cases, EDG and GNU generate different mangled names than clang in cases with nested template parameters in local types.  For example:

$ cat test.cpp
 template<typename T> T foo(T t) { return t; }
 template<typename T> struct X {
   template<typename U> void f(U) {
     struct L {} l;
     foo(l);
   }
 };
 int main() {
   X<int> x;
   x.f(0);
 }
$ eccp --c++11 -tused -c test.cpp && nm test.o | grep foo
0000000000000018 W _Z3fooIZN1XIiE1fIiEEvT_E1LET_S5_
$ g++470 -std=c++11 -c test.cpp && nm test.o | grep foo
0000000000000000 W _Z3fooIZN1XIiE1fIiEEvT_E1LET_S5_
$ clang -std=c++11 -c test.cpp && nm test.o | grep foo
0000000000000030 t _Z3fooIZN1XIiE1fIiEEvT_E1LES3_S3_

Internally, the template parameter for "f" and "foo" are different (because of the nesting level), so we (and apparently GNU) don't use the substitution (hence multiple occurrences of "T_" in the mangled names).

If the nesting level of the template argument is changed, EDG, GNU, and clang all use a substitution for the second occurrence of the template parameter:

$ cat test.cpp
 template<typename T> T foo(T t) { return t; }
 struct X {
   template<typename U> void f(U) {
     struct L {} l;
     foo(l);
   }
 };
 int main() {
   X x;
   x.f(0);
 }
$ eccp --c++11 -tused -c test.cpp && nm test.o | grep foo
0000000000000018 W _Z3fooIZN1X1fIiEEvT_E1LES2_S2_
$ g++470 -std=c++11 -c test.cpp && nm test.o | grep foo
0000000000000000 W _Z3fooIZN1X1fIiEEvT_E1LES2_S2_
$ clang -std=c++11 -c test.cpp && nm test.o | grep foo
0000000000000030 t _Z3fooIZN1X1fIiEEvT_E1LES2_S2_

I don't see anything in the IA-64 ABI documentation that says that template parameter nesting level has any bearing on mangling, so I think clang's behavior is correct.  I plan to change EDG's behavior to match that of clang, but thought I'd see if there are any other opinions first.

Mike Herrick
Edison Design Group