Re: Building Gforth for Apple's new Arm-based Macs

[Anton Ertl]

On Wed, Jan 20, 2021 at 10:07:10AM -0500, Craig Treleaven wrote:
> 
> > On Jan 19, 2021, at 6:08 PM, Anton Ertl <anton@mips.complang.tuwien.ac.at> 
> > wrote:
> > 
> > On Tue, Jan 19, 2021 at 02:42:10PM -0500, Craig Treleaven wrote:
> >> Hi:
> >> 
> >> I contribute a little to the MacPorts project.  MacPorts packages many, 
> >> many open source software titles for use on Macintosh.  I noticed that 
> >> Gforth failed to build on our new buildbot running on an M1 machine(aka 
> >> Apple’s 64 bit ARM SOC).  The configure phase reported the following error:
> >> 
> >> checking build system type... Invalid configuration 
> >> `arm64-apple-darwin20.1.0': machine `arm64-apple' not recognized
> >> configure: error: /bin/sh ./config.sub arm64-apple-darwin20.1.0 failed
> > 
> > config.sub apparently does not know this architecture under the name
> > arm64, but instead under the name aarch64.
> > 
> >> For a number of years, we’ve invoked configure thusly on Intel 64 
> >> bit-based Macs:
> >> 
> >> ./configure --prefix=/opt/local --build=x86_64-apple-darwin14.5.0 
> >> 
> >> I browsed through configure.ac but I have only a rudimentary knowledge of 
> >> auto tools and couldn’t determine if the “-build” flag was even necessary;
> > 
> > If you are not cross-building, you don't need --build.  configure then
> > looks at the current machine and guesses the machine from that (with
> > config.guess).
> 
> Many thanks for explaining so clearly.
> 
> So I’ve tried several different combinations…sorry for the length of this 
> email.
> 
> I’m running these on an older MacBook Pro (Core i7) which identifies thusly:
> 
> $ uname -m
> x86_64
> $ uname -rs
> Darwin 14.5.0
> 
> Running config.guess gives:
> 
> $ ./config.guess
> i386-apple-darwin14.5.0
> 
> If I omit the build flag, configure says (in part):
> 
> ./configure --prefix=/opt/local
> ...
> checking build system type... i386-apple-darwin14.5.0
> checking host system type... i386-apple-darwin14.5.0

Ok, config.guess will only produce one architecture on a
multi-architecture machine, and you have to set up the other variants
manually.

> Check for arch/386//gforth.ld (false)

Don't worry about gforth.ld.

> But it fails testing:
> 
> --- ./test/coretest.out       2013-10-11 17:31:28.000000000 -0400
> ***************
> *** 1,10 ****
> ! redefined {  
> ! 
> ! Floating point exception: 8.
> ! INCORRECT RESULT: { -1 S>D 1 FM/MOD -> 0 -1 }
> ! INCORRECT RESULT: { -2 S>D 1 FM/MOD -> 0 -2 }
> ! INCORRECT RESULT: { 1 S>D -1 FM/MOD -> 0 -1 }
> ! INCORRECT RESULT: { 2 S>D -1 FM/MOD -> 0 -2 }
> ! INCORRECT RESULT: {  7 S>D -3 FM/MOD -> -2 -3 }
> ! INCORRECT RESULT: { -7 S>D  3 FM/MOD -> 2 -3 }
> ! INCORRECT RESULT: { -7 S>D -3 FM/MOD -> -1 2 }
> --- 1,17 ----
> ! redefined {  redefined GDX  YOU SHOULD SEE THE STANDARD GRAPHIC CHARACTERS:
> !  !"#$%&'()*+,-./0123456789:;<=>?@
> ! ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`
> ! abcdefghijklmnopqrstuvwxyz{|}~
> ! YOU SHOULD SEE 0-9 SEPARATED BY A SPACE:
> ! 0 1 2 3 4 5 6 7 8 9 
> ! YOU SHOULD SEE 0-9 (WITH NO SPACES):
> ! 0123456789
> ! YOU SHOULD SEE A-G SEPARATED BY A SPACE:
> ! A B C D E F G 
> ! YOU SHOULD SEE 0-5 SEPARATED BY TWO SPACES:
> ! 0  1  2  3  4  5  
> ! YOU SHOULD SEE TWO SEPARATE LINES:
> ! LINE 1
> ! LINE 2
> ! you should see this first. 
> ! you should see this later. 
> make[3]: *** [checkone] Error 1

That should not happen.  One theory is that you actually compile for
AMD64, but due to the i386 setting get some 386-specific setup (in
particular the multiplication and division instructions), and these
then don't work.  If that is the case, you may have to tell the
compiler to use -m32 (at least if it's gcc; clang usually understands
the same options, but I am not sure).  Or alternatively tell configure
to configure for amd64:

> If I set the build flag to x86_64 on my machine, Gforth configures and builds 
> OK and passes all tests.  Configure says:
> 
> ./configure --prefix=/opt/local --build=x86_64 
> ...
> checking build system type... x86_64-pc-none
> checking host system type... x86_64-pc-none
> Check for arch/amd64//gforth.ld (false)
> 
> 
> Unless you have a better solution, it appears we need to keep adding the 
> build flag to configure.  

An alternative may be to get a more recent config.guess (from the
autotools).

> Re the arm-based macs, configure does not seem to like aarch64:
> 
> ./configure --prefix=/opt/local --build=aarch64 
> …
> checking build system type... Invalid configuration `aarch64': machine 
> `aarch64' not recognized
> configure: error: /bin/sh ./config.sub aarch64 failed

I found that you use gforth 0.7.3, which has a fairly old config.guess
and config.sub, which don't know aarch64 yet.  One option is to
replace them with newer versions.  An alternative is to tell configure
that you use a machine that config.sub knows and that gforth-0.7.3 has no
special support for, e.g.:

configure --build=vax-apple-darwin14.5.0

Alternatively, you could also try a recent Gforth development snapshot
from

https://www.complang.tuwien.ac.at/forth/gforth/Snapshots/current/

These have config.sub that know about aarch64, and they have special
support for aarch64.

> If I pass just “-build=arm”, configure at least completes without error.

It then builds a version that tries to use dynamic native code
generation without having the necessary cache consistency support;
using --build=vax should disable dynamic native code generation.
Alternatively, Gforth development snapshots know how to deal with
aarch64.

- anton