Building Sunfish

Has anyone succeeded in building Sunfish from source code?

I’ve had a bunch of issues with the Perl side of things refusing to run inside Make (ran those build commands and the CMHG ones manually), which were easy to work around. Now I’m having an issue where portmapper-calls.c fails to compile because there are conflicting definitions of EDOM, ERANGE, EILSEQ, EOVERFLOW and EFBIG.

I’ve also been unable to use the version of Subversion available through Packman to check the code out — it complains about the URL scheme for the server URL being unsupported (URL is http://svn.cp15.org/NFS/)

My goal is to get this building and see if I can figure out why it keeps crashing on my Pi.

Cheers,
Phil.

I’ve also been unable to use the version of Subversion available through Packman to check the code out

Try this one. It’s the previous version of Subversion (which at least works with my local server), patched for ARMv8 using PatchSWP.

If you get it built, please post how! I feel every person and their pet had their own variant on a build system.

Has anyone succeeded in building Sunfish from source code?

I just tried building the Sunfish front end with the GCCSDK autobuilder and it still built OK.

It’s only the Sunfish front end that has problems. I tracked it down to a bug in the RTK library the Sunfish front end uses. It’s in C++ so couldn’t work out how to fix it.

!MoonFish and the SunFish module seem to work fine on RISC OS 5.

It’s always the same stack trace.

Fatal signal received: Segmentation fault

Stack backtrace:

Running thread 0xf4280 (Main Thread)
  (   f6f30) pc:    d0a6c lr:    d0f18 sp:    f6f34  __write_backtrace()
  (   f6fa0) pc:    d0b8c lr:    afa04 sp:    f6fa4  __unixlib_raise_signal()
  (   f6fb0) pc:    af908 lr:    56964 sp:    fd8a8  __h_cback()

  Register dump at 000f6fb4:

    a1:    fd8a8 a2:        0 a3: fdfdfdfd a4: fdfdfdfd
    v1:    fd938 v2:   139c88 v3:        1 v4:    e4120
    v5:        0 v6:    ebe9f sl:    fd220 fp:    fd8e8
    ip:        0 sp:    fd8a8 lr:    56964 pc:    56968
    cpsr: 60000110

  00056954 : .0<9c>å : e59c3000 : LDR     R3,[R12,#0]
  00056958 : .. á : e1a0000d : MOV     R0,R13
  0005695c : .à á : e1a0e00f : MOV     R14,PC
  00056960 : .ð<93>å : e593f018 : LDR     PC,[R3,#24]
  00056964 : ...å : e5150004 : LDR     R0,[R5,#-4]
  00056968 : .  á : e1a0200d : MOV     R2,R13
  0005696c :  .<8d>â : e28d1020 : ADD     R1,R13,#&20        ; =" "
  00056970 : .0 á : e1a03006 : MOV     R3,R6
  00056974 : Éåÿë : ebffe5c9 : BL      &000500A0

  (   fd8e8) pc:    56788 lr:    5660c sp:    fd8ec  rtk::desktop::application::reformat(rtk::graphics::point const&, rtk::graphics::box const&)
  (   fd930) pc:    564c0 lr:     8438 sp:    fd934  rtk::desktop::application::run()
  (   f5fec) pc:     83f0 lr:    bc5d8 sp:    f5ff0  main()

Ahh, I’ve been trying to build it natively on a Raspberry Pi, not the GCCSDK builder. I have no idea what kind of environment the autobuilder even provides.

Seems like adding some debug to RTK might be the way to go.

Ahh, I’ve been trying to build it natively on a Raspberry Pi, not the GCCSDK builder.

I managed to get everything to build natively. I’ve created a zip file with everything needed to get it to build including the RTK library. I can make it available for download if you are still interested.

@Chris – very much so! Please let me know where it’s available, either here or by email (philpem at gmail). Cheers!

Chris, I saw your modification in the autobuilder script. tbqh the only reason I saw is it it refuses to download your patch because of an invalid certificate. Hopefully whatever is causing that passes..

refuses to download your patch because of an invalid certificate

Resolving <a href="http://www.cgransden.co.uk">www.cgransden.co.uk</a> (<a href="http://www.cgransden.co.uk">www.cgransden.co.uk</a>)... 107.170.139.70
Connecting to <a href="http://www.cgransden.co.uk">www.cgransden.co.uk</a> (<a href="http://www.cgransden.co.uk">www.cgransden.co.uk</a>)|107.170.139.70|:80... connected.
HTTP request sent, awaiting response... 302 Found
Location: <a href="https://www.cgransden.co.uk//sunfish.zip">https://www.cgransden.co.uk//sunfish.zip</a> [following]
--2020-04-26 06:56:24--  <a href="https://www.cgransden.co.uk//sunfish.zip">https://www.cgransden.co.uk//sunfish.zip</a>
Connecting to <a href="http://www.cgransden.co.uk">www.cgransden.co.uk</a> (<a href="http://www.cgransden.co.uk">www.cgransden.co.uk</a>)|107.170.139.70|:443... connected.
ERROR: no certificate subject alternative name matches
	requested host name ‘<a href="http://www.cgransden.co.uk’">www.cgransden.co.uk’</a>.
To connect to <a href="http://www.cgransden.co.uk">www.cgransden.co.uk</a> insecurely, use `--no-check-certificate'.
Package Sunfish: ***Failure***
Build for package "Sunfish" failed

The `—no-check-certificate’ suggestion given works.

In SetVars

PATCH_URL="http://www.cgransden.co.uk/${PATCH_ZIP} --no-check-certificate"

The build completes successfully but !Sunfish crashes as described.

refuses to download your patch because of an invalid certificate.

I’ve updated the setvars file so there is no longer a certificate error.

So that was the magic place to add —no-check-certificate.
I had a couple of attempts, but they were wrong.

!MoonFish and the SunFish module seem to work fine on RISC OS 5.

I get shedloads of ZeroPain errors from the Moonfish module (offset 00025344). It would be nice to stop them.

I can’t understand the makefiles – they look far too clever for their own good.

I’ve discovered that perl on Ubuntu will execute the required commands (although I don’t understand them, they populate the ‘auto’ directory, albeit with /c and /h suffixed filenames).

Has anyone looked at the ZeroPain errors that appear when a remote client reads a file served up by Moonfish?

The error is caused by the middle of these three instructions:

CMP     R0,#0
LDR     R12,[R0,#32]
CMNNE   R1,#1

Check if R0 is 0, then load from R0 + 32, then do something conditionally on the result of checking R0? What is this?

The ZeroPain error occurs because R0 is 0 at the time.

I can only see one assembly language file with any code in it, and the above code isn’t from it – it’s nowhere near it in the module. It has to be compiler-generated.

I get shedloads of ZeroPain errors from the Moonfish module

Does the module have function names embedded in it? If so you can dump the file in stronged and search for LDR R12,[R0,#32] scrolling back from that should reveal the function that failed.

What does it compile with? In AcornC/C++ if the CModule makefile is used you need

override C_NO_FNAMES    := 

before including CModule to embed function names in the code. The -ff flag disables function names.

Does the module have function names embedded in it?

Yes! I didn’t realise. It’s in state_getlockrange.

Two of the parameters are uint64_t. How are they passed in?

What does it compile with?

In theory it compiles with either Acorn C/C++ or GCC, but AFAIK no-one in recent history has managed to compile it. If you manage it, you might well make a lot of people happy.

Here are the first few lines of function state_checklocks:

  struct lock_stateid *current = file->lock_stateids;
  struct lock_owner *lock_owner = NULL;

  if ((file == NULL) || (stateid == STATEID_ANY)) return NFS_OK;

So it can be called with file == NULL, but it has already referenced file→lock_stateids.

There are some function labels in the file, but not all: the previous and the next functions’ labels are there, but state_checklocks’s label isn’t.

I’m having same problem with Perl from the as Phil in the first post.

I managed to get Sunfish to build a couple of months ago (with the intention of making it worked for untyped files, so my backups work properly), but then got distracted by other things, and updated a lot of software. I now can’t get it to build anymore.

AMU starts by running the Perl script which hangs (even left overnight), the same as if I run it from a TaskWindow. That’s using Perl 5.14.2 on my ARMx6 RO 5.27rc16, which otherwise works properly (at least for all my crusty Perl code), athough it wont run on a Pi 3B, so I suspect there is something wrong with it. I’ve tried 5.8.8 and 5.8.8-1 and they fail immediately with “UnixLib detected recursion of signal SIGSEGV”.

Any ideas?

Any ideas?

I used Perl on a Linux box to do the processing. It happened very quickly. My best suggestion is to do the same, copy the processed files over, and remove the Perl steps from the makefiles.

Looking at the source files and the processed results, I don’t understand why there was an indirect step in the build process. I’d have expected to just provide the processed files as source.

athough it wont run on a Pi 3B, so I suspect there is something wrong with it.

It’s built with an old UnixLib that contains the SWP instruction which causes it to hang. I can rebuild it with current UnixLib.

That would be great, and please remove the spurious !Perl.perl as !Perl.bin.perl is the copy of the interpreter which is actually run. When that is patched to removed SWPs, it does run on the Pi 3B and 3B+.

I’m not having any luck trying to build this on RISC OS.

1. Is it supposed to build with Norcroft or GCC? The makefile seems to have support for both, but neither get very far.
2. Is it supposed to use amu or make? amu builds the perl but then fails, make doesn’t build the Perl
3. Where does libiconv come from? All I can find is the standalone header and stubs from iconv module or the headers in UnixLib or GCC4
4. Is there a simple way of just building the modules and not the front end? It seems to want to build the front end regardless.

This was only supposed to be a mornings job, but its taken days to find the right Perl, and lots of buggering around an uncooperative makefile.
Some documentation on the build environment and location of dependencies would really help!

Right finally got the modules to build. I had to:-

1. Use amu to build the Perl, commented those lines and use GCC4’s make & Norcroft cc for the rest
2. Butchered the makefile to remove all traces of the front end
3. Put in the iconv header from iconv module, and changed it to use sys/errno.h to avoid redefinitions
4. Removed the conflicing READONLY DATA area attribute from base.s
5. Ignore the AS warnings about PLD not supported on the traget CPU

That wasn’t much fun.

Right finally got the modules to build.

Well done! That’s an important success to chalk up.

Once you’ve caught your breath, I hope you have equal success on the front ends.

Careful. Solving this might unravel the fabric of the universe.

Is it just me or are so many RO programs protected by obfuscated build procedures?

This is great news. No pressure, but I hope it leads to a build of Sunfish that is somewhat less explodey.