Re: [MLton-user] MLton on ARM Mac?

[Matthew F. <mat...@gm...>]

On Tue, Dec 29, 2020 at 2:32 PM Chris Cannam
<ca...@al...> wrote:
>
> On Tue, 29 Dec 2020, at 18:35, Matthew Fluet wrote:
> > Try running
> > `./bin/regression -chunk-tail-call false`, which will force the C and
> > LLVM codegens to use a non-tail-calling trampoline.
>
> Yes! that works. Log here:
> https://all-day-breakfast.com/m/mlton-logs/mlton-bootstrap0-regression-notailcall.log
>
> (should I be concerned about the discrepancies in the "real" test?)

No, low-order bits in the floating point operations often differ.
(The IEEE-754 only species correct rounding for the basic math
operations; the trig and other operations often differ between
platforms in the low order bits.)  Eventually, we should create a
`real.arm64-darwin.ok` file with the output from that platform.

> Running the same compiler with -chunk-tail-call false on a real-world program suggests that this is going to work there as well - I will rebuild with the control-flags change to make sure.
>
> What would we be losing by this? Was there a performance or other advantage to the tail-call trampoline?

There should be some performance wins.  See
https://github.com/MLton/mlton/pull/304, though those results don't
specifically call out the effect of `-chunk-tail-call`.

> First quick check has the new native build running maybe 20-30% faster than translated amd64 code - which doesn't seem like much but is plausible on this platform, as the Rosetta2 Intel recompiler thing is really remarkably effective.

Nice to know.

-- 
To unsubscribe from this group and stop receiving emails from it, send an email to mlt...@ml....