On Wed, Feb 4, 2015 at 11:08 PM, John McCall <[email protected]> wrote:
> On Feb 4, 2015, at 9:36 PM, Richard Smith <[email protected]> wrote: > On Wed, Feb 4, 2015 at 6:38 PM, John McCall <[email protected]> wrote: > >> On Jun 5, 2014, at 4:17 PM, Richard Smith <[email protected]> wrote: >> > John: any chance we could get the ABI document updated with these? ( >> http://sourcerytools.com/pipermail/cxx-abi-dev/2012-January/000024.html) >> >> After much delay, added. We don’t seem to get this right, though, at >> least not when the destination type isn’t dependent: >> >> template <class T, class U> T fst(T, U); >> struct A { >> int x[3]; >> }; >> template <class T> decltype(fst(A{1,2},T())) foo(T t) {} >> >> int main() { >> foo(1); >> } >> >> We produce: >> _Z3fooIiEDTcl3fstcv1AililLi1ELi2EEEcvT__EEES1_ >> It should be: >> _Z3fooIiEDTcl3fsttl1ALi1ELi2EcvT__EEES1_ >> > > There are quite a few bugs conspiring to give that result :( Our AST is > also poorly-suited to this mangling, because the braces are not considered > to be part of the functional cast itself; they're part of its subexpression. > > If you parenthesize the argument to A: >> template <class T> decltype(fst(A({1,2}),T())) foo(T t) {} >> We produce: >> _Z3fooIiEDTcl3fstcv1AcvS0_ililLi1ELi2EEEcvT__EEES1_ >> It should be: >> _Z3fooIiEDTcl3fstcv1AliLi1ELi2EcvT__EEES1_ >> > > Somewhat related, we also get this wrong: > > struct X { X(int); }; > int f(X); > template<typename T> void f(decltype(f(0), T())) { f(0); } > void g() { f<int>(0); } > > ... because we explicitly mangle the implicit conversion from int to X. I > see > > _Z1fIiEvDTcmcl1fLi0EEcvT__EE from EDG > _Z1fIiEvDTcmclL_Z1f1XELi0EEcvT__EE from GCC > _Z1fIiEvDTcmclL_Z1f1XEcvS0_cvS0_Li0EEcvT__EE from Clang > > > Ugh, that’s awful. > > I think GCC and Clang are right to use the resolved name L_Z1f1XE rather > than the unresolved name 1f here, and GCC's mangling is right overall. Do > you agree? > > > As an aside: if we have a fully-resolved call in an > instantiation-dependent expression, should we really be putting any used > default arguments into the mangling? > > > I feel like both of these points need to be asked on the cxx-abi-dev. I > definitely don’t think we should be mangling default arguments, but I’m not > sure that resolving ‘f’ here is really consistent with the general dictate > to follow the syntactic tree. > > All of the above fixed in r228274. I'm not really very happy with our AST > representation here; we've overloaded CXXConstructExpr to mean too many > different syntactic things that it's hard to reconstruct the right mangling. > > > The rule used to be that a “bare" CXXConstructExpr — neither a specific > subclass nor the implementation of a cast — was always implicit, and that > there were subclasses which provided additional syntactic information. I > think it would make sense to have a dedicated subclass for the truly > implicit case as well. The implicit case is always a constructor > conversion or copy-construction, right? > Right. The oddball cases are CXXConstructExpr-used-for-list-initialization: f({1}, {2}) (which clearly isn't a CXXFunctionalCastExpr but probably shouldn't be just a CXXConstructExpr either) and CXXConstructExpr-used-for-direct-intiialization: T var(1, 2); // #1 T var{1, 2}; // #2 According to the rules we use in -ast-print, the parens belong to the initialization of the variable, not to the CXXConstructExpr, so that we can support int var(1); with no additional AST nodes beyond the IntegerLiteral expression, but the braces in #2 usually belong to the CXXConstructExpr. Except when T has a constructor that takes std::initializer_list<int>, when they don't, because the braces belong to the construction of the underlying array of int. *sigh*
_______________________________________________ cfe-commits mailing list [email protected] http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
