On Tue, 20 Jul 2021 15:09:04 +0000
Charlene Wendling <[email protected]> wrote:
> Hi,
>
> Long time ago, gkoehler@ found out the guile2 32 bit big endian
> bootstrap files were leading to segfaults during the build [0] on
> powerpc.
>
> Not using a bootstrap increases the building time by a factor of ~ 6.
> Currently, guile2 build time is 25 hours in the powerpc bulk. To give you
> an idea of what this represents on this arch, it's one hour more than
> x11/qt5/qtbase, or twice the time needed to build devel/boost.
>
> The people at Macports saw gkoehler@'s mail, and managed to build
> bootstraps [1].
>
> I took their recipe and adapted to our port system, so we can provide
> a valid 32-bit big endian bootstrap that actually works, based on
> what we do with lang/gcc/8, with the ability for anyone to
> produce a bootstrap.
>
> This reduces the build time from 19 to 3 hours on my Mac Mini G4! [2]
One issue: the port is extracting your g2strap-2.2.7-32be.tgz twice,
in both WRKDIR and WRKSRC. If you can fix that (maybe with
EXTRACT_ONLY), the diff will be ok gkoehler@
You don't bump REVISION, and I don't see a reason to bump.
We're building the same guile2, but building faster.
There is still no 64-bit-big-endian bootstrap, so powerpc64 (and
probably mips64 and sparc64) still go the slow way. I have not tried
to make a 64-bit-big-endian bootstrap. My build times for guile2 are
about
5 h 45 m on macppc Cube G4 (450 MHz) with your g2strap
4 h 50 m on powerpc64 (2700 to 3800 MHz)
I have avoided building guile2 for a long time, because the build was
too slow. Now that we have a bootstrap, I ran "make test". 2 tests
fail on all archs because OpenBSD's getrlimit(2) doesn't have
RLIMIT_AS, but macppc has a 3rd failure in the ffi (foreign function
interface to call C from Guile):
bad return from expression `(f-s64-)': expected -2000000000; got -873842568
bad return from expression `(f-u64-)': expected 4000000000; got 3421124728
bad return from expression `(f-s64-u8 10)': expected -1999999990; got
-4294967294
...
FAIL: test-ffi
guile2 uses devel/libffi so this might be some libffi problem. I have
not yet studied issue; I know that lang/ruby uses libffi and wonder if
I can reproduce the problem by writing ruby code.
--George
> This has no impact on !{macppc,hppa}. I know it's quite a big change
> given the limited scope of that diff. But actually guile moved to 3.x,
> and 2.2.x is now considered legacy [3], so it's very unlikely we
> see an update for this version anyway. I'm hoping they did not
> reproduce the mistake for 3.x.
>
> Comments and feedback are welcome,
>
> Charlène.
>
>
> [0] https://marc.info/?l=openbsd-ports&m=157611545827058&w=2
> [1] https://trac.macports.org/ticket/54124#comment:78
> [2] https://bin.charlenew.xyz/guile2-2.2.7.log
> [3] https://www.gnu.org/software/guile/download/#releases
>
>
> Index: Makefile
> ===================================================================
> RCS file: /cvs/ports/lang/guile2/Makefile,v
> retrieving revision 1.27
> diff -u -p -u -p -r1.27 Makefile
> --- Makefile 3 Aug 2020 07:27:46 -0000 1.27
> +++ Makefile 20 Jul 2021 14:53:41 -0000
> @@ -1,7 +1,8 @@
> # $OpenBSD: Makefile,v 1.27 2020/08/03 07:27:46 jasper Exp $
>
> -# When updating, check that x11/gnome/aisleriot MODGNOME_CPPFLAGS references
> the
> -# proper guile2 includes directory
> +# When updating, check that x11/gnome/aisleriot MODGNOME_CPPFLAGS references
> +# the proper guile2 includes directory. Also, 32-bit BE arches bootstraps may
> +# need to be rebuilt.
>
> COMMENT= GNU's Ubiquitous Intelligent Language for Extension
> # '
> @@ -25,6 +26,9 @@ WANTLIB += c curses ffi gc gmp iconv int
> WANTLIB += unistring
>
> MASTER_SITES= ${MASTER_SITE_GNU:=guile/}
> +MASTER_SITES0= https://charlenew.xyz/distfiles/
> +DISTFILES= ${DISTNAME}${EXTRACT_SUFX}
> +SUPDISTFILES= ${BOOTSTRAP_32BE}:0
>
> # internal compiler error: in unify_nodes, at tree-ssa-structalias.c:1231
> COMPILER = base-clang ports-gcc
> @@ -49,9 +53,22 @@ CONFIGURE_ARGS= --program-suffix=${V}
> # warning: format '%ji' expects type 'intmax_t', but argument 4 has type
> 'scm_t_intmax'
> CONFIGURE_ARGS += --disable-error-on-warning
>
> -# powerpc: Prevent "Segmentation fault (core dumped)" during build.
> -post-patch:
> - mv ${WRKSRC}/prebuilt/32-bit-big-endian{,-broken}
> +.include <bsd.port.arch.mk>
> +
> +# The prebuilt bootstrap makes the build segfault on 32-bit BE arches.
> +# Provide ours instead.
> +BOOTSTRAP_32BE= g2strap-2.2.7-32be.tgz
> +.if !${PROPERTIES:Mlp64} && ${PROPERTIES:Mbe}
> +BOOTSTRAP_TARGET= 32-bit-big-endian
> +BOOTSTRAP= ${BOOTSTRAP_32BE}
> +DISTFILES+= ${SUPDISTFILES}
> +.endif
> +
> +post-extract:
> +.if "${BOOTSTRAP_TARGET}" != ""
> + mv ${WRKSRC}/prebuilt/${BOOTSTRAP_TARGET}{,-broken}
> + cd ${WRKSRC} && tar xzf ${FULLDISTDIR}/${BOOTSTRAP}
> +.endif
>
> post-install:
> install -d ${PREFIX}/share/guile/site/${V}/
> @@ -59,3 +76,20 @@ post-install:
> mv ${PREFIX}/share/aclocal/guile{,${V}}.m4
>
> .include <bsd.port.mk>
> +
> +# The below bits only matter on 32-bit BE archs for now.
> +
> +G2PKG= ${PACKAGE_REPOSITORY}/${MACHINE_ARCH}/all/${FULLPKGNAME}.tgz
> +
> +.PHONY: g2strap
> +
> +g2strap: ${G2PKG}
> +.if "${BOOTSTRAP_TARGET}" != ""
> + mkdir -p ${WRKBUILD}/g2strap/prebuilt/${BOOTSTRAP_TARGET}
> + cp -a ${WRKBUILD}/bootstrap/{Makefile.*,ice-9,language,srfi,system} \
> + ${WRKBUILD}/g2strap/prebuilt/${BOOTSTRAP_TARGET}
> + tar -C ${WRKBUILD}/g2strap -czf ${FULLDISTDIR}/${BOOTSTRAP} .
> + @echo "The bootstrap can be found at: " ${FULLDISTDIR}/${BOOTSTRAP}
> +.else
> + @echo "You don't need to build a bootstrap on ${MACHINE_ARCH}."
> +.endif
> Index: distinfo
> ===================================================================
> RCS file: /cvs/ports/lang/guile2/distinfo,v
> retrieving revision 1.9
> diff -u -p -u -p -r1.9 distinfo
> --- distinfo 3 Aug 2020 07:27:46 -0000 1.9
> +++ distinfo 20 Jul 2021 14:53:41 -0000
> @@ -1,2 +1,4 @@
> +SHA256 (g2strap-2.2.7-32be.tgz) =
> MZmNCiGNR+GcVb73Bk7qnVGOn86juxflnEDEGzDjkZo=
> SHA256 (guile-2.2.7.tar.gz) = RLTF+74lfM3r6hhCAhLJs+kMPIalSSDYVUA5/GdpoAc=
> +SIZE (g2strap-2.2.7-32be.tgz) = 4057585
> SIZE (guile-2.2.7.tar.gz) = 18129833
--
George Koehler <[email protected]>
