Hello,

I wonder if FPC optimize a code such as:

if a < b then
  c := d
else
  c := b;

as :

r := b + ((d - b) and -(a < b));

If so, do you check the type of CPU (because as I understand, some CPU will not execute as fast as other CPU's.

If not, then why, and how would you optimize such code ?

Thanks,

Ido

Op Wed, 6 Sep 2006, schreef ik:

The construction above would be fastest with conditional moves; in the above construction the optimal code generation would be:

mov c,d
cmp a,b
movge c,b

This would be much faster than any code generated for the construction you propose.

Free Pascal won't generate the optimal code, but it can already use conditional moves and I would expect that the compiler would use it in this situation.

Daniël

It does, if the reg. allocator starts correctly :)

This is not a generic optimization that will give real benefits. Your code also relies on the fact that (a<b) always returns 0 or 1. But this is not garantueed. A boolean is defined as 0=false and everything else is true.

Afaik only in case of reading a boolean everything 0 is true (for compatibility with other languages). In case of writing, Pascal (and FPC in particular) always stores a 1 for true and 0 for false, since boolean is defined as an enum with values (false, true)

Jonas

Op Thu, 7 Sep 2006, schreef Jonas Maebe:

Well, the last minute change in 2.0.4 by Florian was needed because constructors return non-1 true cases, i.e.:

if constructor_call then

... results actuall a pointer to the object, or nil.

Daniël

Even though it's a dirty construct by Borland, it does not contradict what I wrote.

Jonas

A constructor returns a qword/longbool which isn't defined as enumeration.