On 2017/09/05 10:41, Stuart Henderson wrote:
> ----- Forwarded message from Stuart Henderson <st...@openbsd.org> -----
>
> From: Stuart Henderson <st...@openbsd.org>
> Date: Mon, 4 Sep 2017 13:27:56 -0600 (MDT)
> To: ports-chan...@openbsd.org
> Subject: CVS: cvs.openbsd.org: ports
>
> CVSROOT: /cvs
> Module name: ports
> Changes by: st...@cvs.openbsd.org 2017/09/04 13:27:56
>
> Modified files:
> graphics/ffmpeg: Makefile
>
> Log message:
> Use ports clang to build ffmpeg on amd64. The AAC encoder hangs on amd64 when
> compiled with clang 4.0.x (either 4.0.0 in base or the older ports 4.0.1).
> Based on a diff from Brad.
>
> Add a comment explaining this and note that there is another issue where the
> H264 decoder segfaults in ff_deblock_v_luma_8_sse2 on i386, but this isn't
> helped by the newer ports clang.
>
>
> ----- End forwarded message -----
>
> So this works around the recently-reported hangs seen on amd64 but we
> still have a segfault issue on i386 (I've only seen this so far when
> decoding H264, though given the function other things may use it too).
>
> I've tried a few things:
>
> - clang (ports or base): segfault
>
> - ports gcc: works, but needs -march=i686 to build, otherwise build fails
> "libpostproc/postprocess_template.c:3184:5: error: 'asm' operand has
> impossible constraints"
>
> - CC=/usr/bin/gcc: works, no extra flags
>
> - I haven't tried a --disable-inline-asm build yet, I suspect that
> this would also work but likely kill performance.
>
> Any thoughts on how to proceed? As far as getting things ship-shape
> for release goes, setting CC=/usr/bin/gcc for i386 seems the least
> worst option to me.
>
> Brad points out we had runtime problems in the past with clang-built
> FFmpeg on i386 (ports/graphics/ffmpeg/Makefile r1.119, r1.120) though
> I can't find the ports@ posts referred to in the commit log any more.
>
So in the absence of better ideas, this is what I'm suggesting.
It's a fairly dirty "fix" but this problem breaks much of the common
use of ffmpeg and its dependent packages on i386 (and the alternative
of -march=i686 will mean SIGILL on older machines).
Any comments, objections or OKs?
Index: Makefile
===================================================================
RCS file: /cvs/ports/graphics/ffmpeg/Makefile,v
retrieving revision 1.164
diff -u -p -r1.164 Makefile
--- Makefile 4 Sep 2017 19:27:56 -0000 1.164
+++ Makefile 6 Sep 2017 14:24:47 -0000
@@ -5,7 +5,7 @@ COMMENT= audio/video converter and strea
V= 20170825
DISTNAME= ffmpeg-git-${V}
PKGNAME= ffmpeg-${V}
-REVISION= 0
+REVISION= 1
CATEGORIES= graphics multimedia
MASTER_SITES= http://comstyle.com/source/
EXTRACT_SUFX= .tar.xz
@@ -35,10 +35,15 @@ WANTLIB= SDL X11 Xext Xv ass bz2 c crypt
# AAC encoder hangs on amd64 when built with clang 4.0.0 or 4.0.1 but
# succeeds with ports clang 5.0.0rc4.
-# XXX h264 decoder segfaults in ff_deblock_v_luma_8_sse2 on i386
COMPILER= base-gcc ports-clang base-clang
COMPILER_LANGS= c
MODCLANG_ARCHS= amd64
+# ff_deblock_v_luma_8_sse2 segfaults on i386/clang (e.g. in H264 decode)
+# works with ports gcc, but requires -march=i686 to build
+# works with base gcc
+.if ${MACHINE_ARCH} == "i386"
+CC= /usr/bin/gcc
+.endif
BUILD_DEPENDS= textproc/texi2html
.if ${MACHINE_ARCH} == "amd64" || ${MACHINE_ARCH} == "i386"
