On Wed, 23 Nov 2022 at 22:19, John Naylor <john.naylor@enterprisedb.com> wrote:
>
>
> On Wed, Nov 23, 2022 at 3:54 PM David Rowley <dgrowleyml@gmail.com> wrote:
> >
> > Going by [1], clang will actually use multiplication by 16 to
> > implement the former. gcc is better and shifts left by 4, so likely
> > won't improve things for gcc. It seems worth doing it this way for
> > anything that does not have HAVE__BUILTIN_OP_OVERFLOW anyway.
>
> FWIW, gcc 12.2 generates an imul on my system when compiling in situ.
I spent a bit more time trying to figure out why the compiler does
imul instead of bit shifting and it just seems to be down to a
combination of signed-ness plus the overflow check. See [1]. Neither
of the two compilers I tested could use bit shifting with a signed
type when overflow checking is done, which is what we're doing in the
new code.
In clang 15, multiplication is done in both smultiply16 and
umultiply16. These both check for overflow. The versions without the
overflow checks both use bit shifting. With GCC, only smultiply16 does
multiplication. The other 3 variants all use bit shifting.
David
[1] https://godbolt.org/z/EG9jKMjq5