Discussion:
Little question about the
(too old to reply)
J. Gareth Moreton
2018-08-10 15:43:41 UTC
Permalink
Thanks Sven,
Sorry if I'm bringing up obvious points.  I've set "pure" to be
interface-only - it's a little tricky because if you now specify it in the
main source file, it says that it can't be used in the implementation
section, even though, really, the main file doesn't have such section. 
Still, that can be dealt with.
This will require a lot of testing once I start submitting patches!

Gareth aka. Kit

On Fri 10/08/18 06:45 , "Sven Barth" ***@googlemail.com sent:
Gareth sent me the following two mails in private, but they were ment
Thanks Sven. Normally I would agree with
"pure" only belonging in the interface
section, but it causes problems when you
try to put a pure function inside the main
program block, as the entire thing is
considered equivalent to implementation.
Also, inline is allowed in the
implementation section, and the two follow
similar rules in regards to their calls
being modified.
Just to clarify, "pure" doesn't change
anything in regards to the parameter types
or how a call with variable arguments is
handled. The raw signature shouldn't
change. It's an optimisation hint. At
least it is in a perfect world!
I asked the question because I stumbled across something interesting. 
I've been using the following set of functions to see how the compiler
handles things that are a bit out of order (using 'inline' as a
program PureTest;
function PureMax(const a: Double; const b: Double): Double; forward;
procedure TestFunc;
begin
  WriteLn(PureMax(2.0, 3.0));
end;
function PureMax(const a: Double; const b: Double): Double; inline;
begin
  if (a > b) then
    Result := a
  else
    Result := b;
end;
begin
  TestFunc;
end.
****
Turns out, to my surprise, after analysing the nodes and the
disassembly, that "PureMax" is not inlined inside the TestFunc
routine.  I haven't tested units yet, but would a similar situation
occur if a function is not defined as inline in the "interface"
section and only the "implementation" section (unless the caller
appears after said function in the source file)?
Either way, I'm leaning towards making the new "pure" directive an
interface-only directive as Sven suggested, as that will make things a
lot easier if a pure function is used as part of a constant
definition.  After all, "pure" would have a bit more of a notable
effect than "inline" because it dictates where the function can and
cannot be used.
I hadn't thought that I need to explain that, but apparantly you haven't
yet reached that far in your research of the compiler, so:
- of course I only mean that it's an interface-only modifier if there
*is* an interface section; routines that are - inside a unit - only
declared inside the implementation section are obviously exempt from
this (thus also for the program file)
- again I wrote that I don't mean that "pure" modifies the parameters,
but it modifies the meta data (most importantly checksum) of the
interface section when you add a modifier flag inside the implementation
section. The compiler first handles the interface sections of units and
might then compile interface sections or even implementation sections of
other units depending on the dependencies between the units. If a flag
only appears in the implementation section than this messes up this
whole thing as the other units would need to be recompiled. This is
already a mess with "inline" so we don't need to do that with "pure" as
well, especially as we don't need to be Delphi compatible here.
- yes, it is correct that the compiler's capability to inline depends on
the order of the routine definitions. The compiler generates the code
routine by routine so if the routine's body has not yet been encountered
then it can't be inlined. (When using generic specializations it
triggeres the generation of the specialized routine bodies before
generating the routine to avoid "inline" not working for generics) For
"pure" we can even go so far as to say that a routine body *must* be
available to be able to call it inside another "pure" routine.
Everything else would be an error.

I hope this clears things up a bit.

Regards,
Sven
Sven Barth via fpc-devel
2018-08-10 18:51:47 UTC
Permalink
Post by J. Gareth Moreton
Thanks Sven,
Sorry if I'm bringing up obvious points. I've set "pure" to be
interface-only - it's a little tricky because if you now specify it in the
main source file, it says that it can't be used in the implementation
section, even though, really, the main file doesn't have such section.
Still, that can be dealt with.
You shouldn't declare it as interface only, but you should disallow the
situation that it is not set for the forward def in the interface, but for
the one in the implementation section. Other combinations are valid.
I think the function to look for is add_proc_definition (can't check
currently).
Also the defs should contain methods to check whether they had been
declared in the interface or implementation section so you can use those.

Regards,
Sven
J. Gareth Moreton
2018-08-11 00:56:46 UTC
Permalink
Sorry for not quite understanding what you mean.  Are you meaning that if
I detect something like this in the main file...

****
program SomeProg;

function PureFunc(A: Integer): Integer; forward;

procedure TestFunc;
begin  { Call PureFunc }
end;

function PureFunc(A: Integer): Integer; pure;
begin
  { etc. }
end;
****

... an error should be thrown?  But if something like this appears:

****
unit SomeUnit;

interface

function PureFunc(A: Integer): Integer;

implementation

procedure TestFunc;
begin  { Call PureFunc }
end;

function PureFunc(A: Integer): Integer; pure;
begin
  { etc. }
end;
****

... all is well?  Is it only problematic if a function is defined as
"forward" and not "pure" on the same line?

Gareth aka. Kit

On Fri 10/08/18 19:51 , "Sven Barth" ***@googlemail.com sent:
J. Gareth Moreton schrieb am Fr., 10. Aug. 2018, 18:44:
Thanks Sven,
Sorry if I'm bringing up obvious points.  I've set "pure" to be
interface-only - it's a little tricky because if you now specify it in the
main source file, it says that it can't be used in the implementation
section, even though, really, the main file doesn't have such section. 
Still, that can be dealt with.
You shouldn't declare it as interface only, but you should disallow the
situation that it is not set for the forward def in the interface, but for
the one in the implementation section. Other combinations are valid.  I
think the function to look for is add_proc_definition (can't check
currently). Also the defs should contain methods to check whether they had
been declared in the interface or implementation section so you can use
those. 
Regards, Sven 

Loading...