Inline optimization

[Jonathan M Davis]

On Tuesday, February 26, 2019 7:07:40 PM MST Michelle Long via Digitalmars-d 
wrote:
> On Tuesday, 26 February 2019 at 14:30:58 UTC, Dennis wrote:
> > On Tuesday, 26 February 2019 at 02:43:36 UTC, Michelle Long
> >
> > wrote:
> >> Given a function such as a computation that would benefit from
> >> running at full speed, is it possible to optimize only that
> >> function rather than the full program?
> >>
> >> That is, compile with debug but optimize specific code:
> >>
> >> pragma(optimize, "O3")
> >> auto fastAdd(A,B)
> >
> > GCC has such a pragma, but as far as I know D compilers don't.
> > The closest thing I found was the optimization strategy
> > attribute in LDC:
> > https://wiki.dlang.org/LDC-specific_language_changes#.40.28ldc.attribute
> > s.optStrategy.28.22strategy.22.29.29
> >
> > But it seems you can only disable, not enable optimizations
> > with that.
> >
> > What you can do in any case is put the performance critical
> > code in a separate compilation unit (i.e. its own file) and
> > compile that with optimizations and link it with the rest of
> > the program compiled without optimizations. With a makefile
> > that shouldn't be hard, though for expressing that in Dub I'm
> > afraid you have to put it in its own sub package with an
> > optimization flag in the 'dflags' attribute.
>
> That's a lot of trouble...
>
> Surely it would be very simple for the compiler to do this... it
> has the flag inside and so when it is parsing a function it
> either runs an optimizer on it or not... given the command line
> options. But it could make this dynamic and modify it based on
> pragma.... seems like it would be much simpler to modify the
> compiler to do this?

As far as dmd goes, Walter seems to generally be against fine-grained
control of optimizations like this. It took a lot of arguing from folks to
even get pragma(inline, true), and it still doesn't do what most folks want,
since folks wanted to be able to force inlining, and it gives an error when
it can't inline so that you know that the function isn't being inlined, but
it doesn't actually force anything. The ldc or gdc devs may be more open to
fine-grained control of optimizations via pragmas and the like, but nothing
like that is likely to ever be standard.

- Jonathna M Davis