Dear Freepascalers,

The multiplication in floating arthmetic units is usualy know as being slower than the addition. So, have I interest to code X+X, 2*X or X*2 when using reals (mainly double and extended) ?And with integers ?

Can the compiler optimise this operation choosing the fastest one - probably depending on the architecture x86, arm, aso.

Cheers, Raoul

forgot

X shl 1

normaly the fastest ...

Ched wrote

Dear Freepascalers, The multiplication in

floating arthmetic units is usualy know as being slower than the addition. So, have

I interest to code X+X, 2*X or X*2 when using

reals (mainly double and extended) ?And with integers ?

Can the

compiler optimise this operation choosing the fastest one - probably depending on the

architecture x86, arm, aso. Cheers, Raoul

But only for ints, not floating point values.

Regards,
Sven

Can't really comment on floats, but for integers the fasted to multiply by 2 should be a shift left.

Can the compiler optimise this operation choosing the fastest one - probably depending on the architecture x86, arm, aso.

It can if someone codes support for it (maybe some code generators already have such simplifications). I thought I've at least seen such code for integers (that exactly converts *2 to a shl), but I can't find it currently...

Regards,
Sven

That is right for positive integers. But for floating point numbers ?

Cheers, Raoul

Integer multiplication by a power of 2 constant is converted to a SHL in pass 1, see nadd.pas:2869.

Nikolay

Am 19.01.2014 22:50 schrieb "Nikolay Nikolov" <

already have such simplifications). I thought I've at least seen such code for integers (that exactly converts *2 to a shl), but I can't find it currently...

Integer multiplication by a power of 2 constant is converted to a SHL in

pass 1, see nadd.pas:2869.

Ah, there it is. Thanks. I must have been too tired already when I looked through the code ^^

Regards,
Sven

Sven Barth wrote

Can't really comment on floats, but for integers the fasted to multiply by 2 should be a shift left.

GCC automatically optimizes this. With the small "embedded" CPUs I usually use, it even converts more most multiplications and divisions with constants to a series of adds and shiftts.

-Michael

With the caveat that at least some Pentiums have a limited amount of shifting hardware (specifically, of the 2x ALUs in a processor only one has a barrel shifter on the output). That might have the effect that converting a long run of *2 to <<1 would have the effect of preventing work being routed to half of the ALUs.