[erlang-questions] R12.0B and Sparcs

[Bruce O'Neel]

Hi,

Thanks, but it's probably not worth spending hours on this.  I suspect that
the set of non-Ultra sparc users of erlang is quite small...

Anyway, one of the many error messages is:

/tmp//ccX10737.s:4437: Error: Architecture mismatch on "cas".
/tmp//ccX10737.s:4437:  (Requires v9|v9a|v9b; requested architecture is sparclite.)

The line in R12B-0 that fails is:

gcc -DUSE_THREADS  -D_THREAD_SAFE -D_REENTRANT  -g -O3 -I/home/edoneel/tmp/otp_src_R12B-0/erts/sparc-unknown-openbsd4.2    -fomit-frame-pointer -Wall -Wstrict-prototypes -Wmissing-prototypes -DHAVE_CONFIG_H  -I../include -I../include/sparc-unknown-openbsd4.2 -I../include/internal -I../include/internal/sparc-unknown-openbsd4.2 -c common/ethread.c -o obj/sparc-unknown-openbsd4.2/opt/r/ethread.o


As you pointed out, the code in erts/include/internal/sparc32 hasn't changed much,
but, the R11 code had:

#ifdef ETHR_TRY_INLINE_FUNCS

around the files.

When R11B-5 compiles the line that gets compiled is:

gcc -MM -DUSE_THREADS  -D_THREAD_SAFE -D_REENTRANT  -g -O3 -I/home/edoneel/tmp/otp_src_R11B-5/erts/sparc-unknown-openbsd4.2     -fomit-frame-pointer -Wall -Wstrict-prototypes -Wmissing-prototypes -DHAVE_CONFIG_H
-I../include -I../include/sparc-unknown-openbsd4.2 -I../include/internal -I../include/internal/sparc-unknown-openbsd4.2 common/ethread.c 

which looks quite similar.

In both cases the configure request was:

./configure --prefix=$HOME/local-`uname -m`

I did several runs with turning on and of threads, smp, kernel-poll, and hipe.  
None made this go away, though, I did not try all possible combinations, rather
I tried each one individually.  16 runs of configure would have been days...

Would you like some bits of config.log or config.h?  Also, the sparc
version of OpenBSD is special in that gcc is 2.95.3 rather than some sort
of gcc 3.x.

Still, my email message was more so that if I ever needed to solve this again
and didn't remember, google, with luck, would find it.  Also potentially it helps
the one other user of Sparc erlang :-) 

Thanks again!

cheers

bruce

----- Message d'origine -----
De: Mikael Pettersson <mikpe@REDACTED>
Date: Wed, 2 Jan 2008 14:26:26 +0100
Sujet: Re: [erlang-questions] R12.0B and Sparcs
À: "Bruce O'Neel" <erlang@REDACTED>
Cc: erlang-questions@REDACTED, Jon Olsson <jon@REDACTED>

>Bruce O'Neel writes:
> > Happy New Year,
> > 
> > Here at the home for ancient, disused, and discarded computers we noticed that 
> > while R11B-5 built and ran just fine on Sparcs, R12B-0 did not at least on 
> > OpenBSD and NetBSD.
>
>What exactly is the failure? Compile-time warning from the assembler
>on the V9 instructions? Runtime SIGILL?
>
> > No, not Ultras, the sparcs that Sun hasn't actually shipped for 10+ years, the old
> > V7 and V8 systems (ie SPARCstation 10s, 20s, 4s, and 5s etc).
> > 
> > It turns out that a small change to 
> > 
> > erts/include/internal/ethread.h 
> > 
> > allows them to work again.
> > 
> > 82a83,85
> > > /* BEO  for SPARC OpenBSD only.  Not SPARC64! */
> > > #  define ETHR_DISABLE_NATIVE_IMP
> > >
> > 
> > Basically the above define throws away the optimized assembly code in 
> > erts/include/internal/sparc32 which is V9 (ie Ultra) specific.
>
>The SPARC v9 (actually v8+) code referenced by ethread.h is
>essentially the same in R11B-5 and R12B-0.
>I suspect some configuration change is causing your problems.
>
>- Did you build R11B-5 with SMP support on SPARC v7/v8?
>  If so, was it autoenabled or did you enable it explicitly.
>- Does R12B-0 autoenable SMP on BSD/SPARC or do you enable
>  it explicitly?
>
>GCC's __sparcv9 macro is unfortunately useless here
>(it seems to be set if and only if 64-bit code is generated),
>so to disable the v8+/v9 code automatically we'd have to add
>a check in erts/configure.in and use that to #define a macro
>which we then check in ethread.h.
>
>/Mikael
>