Re: toolchain/60904: compiler dumps core cross-building from MacOS on Apple Silicon

["GOVIND PANDEY via gnats" <gnats-admin%NetBSD.org@localhost>]

The following reply was made to PR toolchain/60904; it has been noted by GNATS.

From: GOVIND PANDEY <govindup63%gmail.com@localhost>
To: gnats-bugs%netbsd.org@localhost
Cc: 
Subject: Re: toolchain/60904: compiler dumps core cross-building from MacOS on
 Apple Silicon
Date: Sun, 18 Oct 2026 20:53:12 -0400

 --00000000000051f966065e26f0c2
 Content-Type: text/plain; charset="UTF-8"
 
 Hi,
 
 I can confirm both the problem and the fix on the netbsd-11 branch, in
 case it helps with the pullup.
 
 Host: Darwin 27.0.0 arm64, Apple M4 (MacBook Air, 10 cores)
 Tree: netbsd-11 from the git mirror, 36cf5108d (2026-10-14)
 Build: ./build.sh -U -u -m amd64 -j10 tools
 
 With tools/gmp/Makefile at 1.6 (so assembly enabled on arm64, GMP 6.2.1
 with ten mpn/arm64/*.asm files that use x18), I took the resulting
 x86_64--netbsd-gcc and compiled a 300-line C file of long double
 constant expressions 600 times, 24 in parallel. 62 of them died:
 
 44 internal compiler error: Segmentation fault: 11
 27 internal compiler error: Bus error: 10
 6 internal compiler error: Abort trap: 6
 1 internal compiler error: Trace/BPT trap: 5
 
 (some runs printed two signatures, hence the total over 62.)
 
 I then applied the 1.6 -> 1.7 change to tools/gmp/Makefile, which
 applies cleanly with patch -p1 since the branch copy is identical to
 HEAD at 1.6, removed the gmp, mpfr, mpc and gcc objects and rebuilt
 tools. GMP's config.status now shows --disable-assembly and nothing
 under mpn/ is built from .asm. The same stress run, done twice:
 
 600 compiles, 24 parallel, 0 failures
 600 compiles, 24 parallel, 0 failures
 
 I haven't tried netbsd-10, mostly because I'm short on disk for a
 second tree.
 
 Thanks,
 Govind Pandey
 
 --00000000000051f966065e26f0c2
 Content-Type: text/html; charset="UTF-8"
 Content-Transfer-Encoding: quoted-printable
 
 <html><body><div dir=3D"auto">Hi,<br><br>I can confirm both the problem and=
  the fix on the netbsd-11 branch, in<br>case it helps with the pullup.<br><=
 br>Host: Darwin 27.0.0 arm64, Apple M4 (MacBook Air, 10 cores)<br>Tree: net=
 bsd-11 from the git mirror, 36cf5108d (2026-10-14)<br>Build: ./build.sh -U =
 -u -m amd64 -j10 tools<br><br>With tools/gmp/Makefile at 1.6 (so assembly e=
 nabled on arm64, GMP 6.2.1<br>with ten mpn/arm64/*.asm files that use x18),=
  I took the resulting<br>x86_64--netbsd-gcc and compiled a 300-line C file =
 of long double<br>constant expressions 600 times, 24 in parallel. 62 of the=
 m died:<br><br>  44 internal compiler error: Segmentation fault: 11<br>  27=
  internal compiler error: Bus error: 10<br>   6 internal compiler error: Ab=
 ort trap: 6<br>   1 internal compiler error: Trace/BPT trap: 5<br><br>(some=
  runs printed two signatures, hence the total over 62.)<br><br>I then appli=
 ed the 1.6 -&gt; 1.7 change to tools/gmp/Makefile, which<br>applies cleanly=
  with patch -p1 since the branch copy is identical to<br>HEAD at 1.6, remov=
 ed the gmp, mpfr, mpc and gcc objects and rebuilt<br>tools. GMP&#39;s confi=
 g.status now shows --disable-assembly and nothing<br>under mpn/ is built fr=
 om .asm. The same stress run, done twice:<br><br>  600 compiles, 24 paralle=
 l, 0 failures<br>  600 compiles, 24 parallel, 0 failures<br><br>I haven&#39=
 ;t tried netbsd-10, mostly because I&#39;m short on disk for a<br>second tr=
 ee.<br><br>Thanks,<br>Govind Pandey<br></div></body></html>
 
 --00000000000051f966065e26f0c2--