Kai Henningsen wrote:
On Sat, 11 Oct 2008 17:43:33 -0400
John Fine <johnsfine@xxxxxxxxxxx> wrote:
(not on my main topic, but I'd like to
know that too).
How about this?
typedef unsigned int uint;
inline void foo0(uint k, uint n, double const& m, double* s, uint const*
c, double const* v) { #pragma ivdep
while (++k < n) s[ c[k] ] -= m * v[k]; }
template <class Ts, class Tc, class Tv>
inline void foo(uint k, uint n, double const& m, Ts s, Tc c, Tv v)
{ foo0(k, n, m, s, c, v); }
Thanks. Maybe I should have stuck to my main topic. I'm looking at
many interacting issues at once. The code I posted is simplified (in
the template declaration, not in the asm code) from what I'm actually
working on and that caused a few of my side issues to loose something in
translation.
In cases similar to the current one, I keep hoping there would be some
way to select the right template specialization based on such ability to
cast. But I understand that it doesn't work that way. The fall back is
to at least check the correctness of the cast after this template is
selected, which your suggestion does and my code failed to do (I found
and corrected that this morning).
In fact, in the current example, I was wrong about what cast is even
legal for some cases. Actually it was (&v[0]) that could be cast to
(double const*). v itself might not be possible to cast to (double
const*). That would also be easy to cover in your form. Using (&v[0])
trusts that the caller won't use any container with noncontiguous
storage. That is safe in the project I'm working on, but not a valid
approach for more general tools.
