On Monday, 25 June 2018 at 05:14:31 UTC, Jonathan M Davis wrote:
On Monday, June 25, 2018 05:03:26 Mr.Bingo via
Digitalmars-d-learn wrote:
The compiler should be easily able to figure out that foo(3)
can be precomputed(ctfe'ed) and do so. It can already do this,
as you say, by forcing enum on it. Why can't the compiler
figure it out directly?
The big problem with that is that determining whether the
calculation can be done at compile time or not means solving
the halting problem. In general, the only feasible way to do it
would be to attempt it for every function call and then give up
when something didn't work during CTFE, which would balloon
compilation times and likely cause the compiler to run out of
memory on a regular basis given how it currently manages memory
and how CTFE tends to use a lot of memory.
It was decided ages ago that the best approach to the problem
was to use CTFE only when CTFE was actually required. If an
expression is used in a context where its value must be known
at compile time, then it's evaluated at compile time;
otherwise, it isn't. The compiler never attempts CTFE as an
optimization or because it thinks that it might be possible to
evaluate the value at compile time. As things stand, it should
be pretty trivial to be able to look at an expression and
determine whether it's evaluated at compile time or not based
on how it's used.
If you're looking to have the compiler figure out when to do
CTFE based on the fact that an expression could theoretically
be evaluated at compile time, or because you want the compiler
to optimize using CTFE, then you're going to be disappointed,
because that's never how CTFE has worked, and I'd be _very_
surprised if it ever worked any differently.
- Jonathan M Davis
The docs say that CTFE is used only when explicit, I was under
the impression that it would attempt to optimize functions if
they could be computed at compile time. The halting problem has
nothing to do with this. The ctfe engine already complains when
one recurses to deep, it is not difficult to have a time out
function that cancels the computation within some user definable
time limit... and since fail can simply fall through and use the
rtfe, it is not a big deal.
The problem then, if D can't arbitrarily use ctfe, means that
there should be a way to force ctfe optionally!
This means that the compiler will force ctfe if the input values
are known, just like normal but if they are not known then it
just treats the call as non-ctfe.
so, instead of
enum x = foo(y); // invalid or valid depending on if y is known
at CT
we have
cast(enum)foo(y)
which, hypothetically of course, would attempt to precompute
foo(y) as in the first case but if it is not precomputable then
just calls it at compile time.
So, the above code becomes
static if (ctfeable(y))
enum x = foo(y);
return x; // compile time version ("precomputed")
else
return foo(y) // runtime version
The semantic above is defined for something like cast(enum)(might
be confusing but anything could be used that works better).
hence
auto x = cast(enum)foo(y);
will result in x being an enum if y is an enum(since chaining can
then occur), else a runtime variable with the return of foo
calculated at runtime.
So, this has nothing to do with the halting problem. I'm not
asking for the compiler to do the impossible, I'm asking for a
notation that combines to different syntaxes in to a composite
pattern so one can benefit from the other. Don't make it harder
than it seems, it's a pretty easy concept.
It might even be possible to do this in a template:
import std.stdio;
auto IsCTFE()
{
if (__ctfe) return true;
return false;
}
template AutoEnum(alias s)
{
static if (IsCTFE)
{
pragma(msg, "CTFE!");
alias AutoEnum = s;
}
else
{
pragma(msg, "RTFE!");
alias AutoEnum = s;
}
}
double foo(int x) { return 3.42322*x; }
void main()
{
int x = 3;
writeln(AutoEnum!(foo(x)));
}
So, the problem is that if we replace x with 3 the above code is
executed at compile time(a precomputation).
But the way it is it won't allow it to be computed even at
runtime! There is no reason that the compiler can't just replace
the code with `writeln(foo(x));` for RTFE... yet it errors
*RATHER THAN* default to RTFE!
The simple fix is to allow for the alternative to not try to ctfe
and just compute the value at runtime.