[m-rev.] [b.schmidt at ugrad.unimelb.edu.au: PPC Darwin linking fix] (fwd)

[Julien Fischer]

On Mon, 18 Dec 2006, Julien Fischer wrote:

> ----- Forwarded message from Ben Schmidt <b.schmidt at ugrad.unimelb.edu.au> 
> -----
>
> Date: Thu, 14 Dec 2006 16:52:38 +1100
> From: Ben Schmidt <b.schmidt at ugrad.unimelb.edu.au>
> To: Zoltan Somogyi <zs at csse.unimelb.edu.au>
> Subject: PPC Darwin linking fix
>
> Proposed diff and log message attached.
>
> Unfortunately, I presume that I can't include all the files required for
> the complete fix, as they are part of GCC and Darwin (as documented in
> the fix here). If someone knows more about this, or how to enforce a
> link order to work around the problem (as documented), let me know.
>
> Ben.
>
>
> Index: README.MacOS
> =================================================================== RCS
> file: /home/mercury/mercury1/repository/mercury/README.MacOS,v
> retrieving revision 1.9 diff -u -r1.9 README.MacOS --- README.MacOS
> 31 Jan 2006 13:31:01 -0000	1.9 +++ README.MacOS	11 Dec 2006
> 23:51:00 -0000 @@ -60,9 +60,12 @@  *building* such distributions!) The
> version of tar that comes with Mac OS X  10.3.3 doesn't have this
> problem.   -The linker on OS X may also have problems linking large
> executables in debug -grades.  This affects all versions of ld that we
> have tested and there is -currently no known workaround. +On PowerPC
> Darwin, large executables (requiring branches > +/- 32 MB) will +cause
> complaints about relocation displacements being too large in crt1.o and
> +crt2.o. To solve this problem, versions of crt1.o (Darwin module
> 'Csu') and +crt2.o (part of gcc) need to be compiled with gcc options
> -mlongcall and +-mlong-branch (modify the Makefiles) and replace the
> system-provided versions +of those files.
> -----------------------------------  Known bugs and unsupported
> features Index: compiler/compile_target_code.m
> =================================================================== RCS
> file:
> /home/mercury/mercury1/repository/mercury/compiler/compile_target_code.m,v
> retrieving revision 1.92 diff -u -r1.92 compile_target_code.m ---
> compiler/compile_target_code.m	22 Aug 2006 05:03:40 -0000
> 1.92 +++ compiler/compile_target_code.m	11 Dec 2006 23:51:03 -0000 @@
> -715,6 +715,29 @@          AppleGCCRegWorkaroundOpt = ""      ),   +
> % On PPC Darwin, large executables need to be compiled with -mlongcall
> +    % (and in GCC  +/- 32 MB. The linker will choose  +    % whether
> to ignore or use the glue code the compiler generates, +    % however,
> so there should be no real speed penalty, only a size +    % penalty,
> to always enabling this. Versions of crt1.o (Darwin  +    % module
> 'Csu') and crt2.o (part of gcc) also need to be compiled  +    % with
> this option(s) (modify the Makefiles) and replace the +    %
> system-provided ones for this to fully work, or a link order enforced +
>   % so that the jumps from system initialisation code to the program +
>   % (probably just to the entry point) are small. + +    ( +
> string.prefix(FullArch, "powerpc-apple-darwin") +    -> +
> AppleGCCLongCallOpt = "-mlongcall -mlong-branch" +    ; +
> AppleGCCLongCallOpt = "" +    ), +      % Be careful with the order
> here!  Some options override others,      % e.g. CFLAGS_FOR_REGS must
> come after OptimizeOpt so that      % it can override
> -fomit-frame-pointer with -fno-omit-frame-pointer. @@ -744,6 +767,7 @@
>        InlineAllocOpt, " ",           AnsiOpt, " ",
> AppleGCCRegWorkaroundOpt, " ",  +        AppleGCCLongCallOpt, " ",
>     WarningOpt, " ",           CFLAGS,           " -c ", C_File, " ",
> Index: scripts/mgnuc.in
> =================================================================== RCS
> file: /home/mercury/mercury1/repository/mercury/scripts/mgnuc.in,v
> retrieving revision 1.116 diff -u -r1.116 mgnuc.in --- scripts/mgnuc.in
> 15 Aug 2006 04:19:39 -0000	1.116 +++ scripts/mgnuc.in	11 Dec
> 2006 23:51:30 -0000 @@ -548,6 +548,25 @@      esac  esac   +# On PPC
> Darwin, large executables need to be compiled with -mlongcall +# (and
> in GCC  +/- 32 MB. The linker will choose  +# whether to ignore or use
> the glue code the compiler generates, +# however, so there should be no
> real speed penalty, only a size +# penalty, to always enabling this.
> Versions of crt1.o (Darwin  +# module 'Csu') and crt2.o (part of gcc)
> also need to be compiled  +# with this option(s) (modify the Makefiles)
> and replace the +# system-provided ones for this to fully work, or a
> link order enforced +# so that the jumps from system initialisation
> code to the program +# (probably just to the entry point) are small. +
> +case $FULLARCH in powerpc*apple*darwin*) +    ARCH_OPTS="$ARCH_OPTS
> -mlongcall -mlong-branch" ;; +esac +  case $assemble in true)      case
> $verbose in true)          echo $AS $AS_OPTS "$@" ;;
>
>
>
>
> Estimated hours taken: 2 + many (20ish) looking at GCC, Darwin, PPC,
> etc. Branches:
> Add compiler flags relevant to PPC Darwin (Mac OS X).
> README.MacOS: 	Updated documentation; read it for more info.
> compiler/compile_target_code.m: scripts/mgnuc.in: 	Add compiler
> flags relevant to PPC Darwin (Mac OS X) when 	appropriate.

I suspect most users are not going to bother to replace bits of the
OS just in order to use mdb.  I suggest that we just document the
workaround and leave it at that, so something like this:

Index: README.MacOS
===================================================================
RCS file: /home/mercury/mercury1/repository/mercury/README.MacOS,v
retrieving revision 1.10
diff -u -r1.10 README.MacOS
--- README.MacOS	7 Nov 2006 05:00:36 -0000	1.10
+++ README.MacOS	18 Dec 2006 06:43:50 -0000
@@ -57,9 +57,18 @@
  *building* such distributions!) The version of tar that comes with Mac OS X
  10.3.3 doesn't have this problem.

-The linker on OS X may also have problems linking large executables in debug
-grades.  This affects all versions of ld that we have tested and there is
-currently no known workaround.
+The linker on Mac OS X may also have problems linking large executables
+in debug grades.  It complains about relocation displacements being too
+large in crt1.o and crt2.o.  The only known problem workaround for this
+problem is to replace the system provided versions of those files with
+versions that are compiled the gcc options `-mlongcall' and `-mlong-branch'.
+crt1.o is in Darwin module 'Csu' and crt2.o is part of gcc.  C code generated
+by the Mercury compiler will also need to be compiled with the above options.
+This can be achieved by adding:
+
+	EXTRA_CFLAGS=-mlongcall -mlong-branch
+
+to your Mercury.options or Mmakefile.


Actually, I don't think replacing the system provided versions of crt1.o
and crt2.o should be necessary either.  It should be sufficient to just
build the new ones, put them in same directory in which the executable
will be built and compile with -nostartfiles (I haven't tested that though.)

Julien.
--------------------------------------------------------------------------
mercury-reviews mailing list
Post messages to:       mercury-reviews at csse.unimelb.edu.au
Administrative Queries: owner-mercury-reviews at csse.unimelb.edu.au
Subscriptions:          mercury-reviews-request at csse.unimelb.edu.au
--------------------------------------------------------------------------