Re: cmake 3.31.7 bootstrap illegal instruction on Sparc32 (qemu) ?

сб, 6 дек. 2025 г., 10:13 Andrew Randrianasulu <[email protected]>:

>
>
> сб, 6 дек. 2025 г., 07:02 Andrew Randrianasulu <[email protected]>:
>
>> On Sat, Dec 6, 2025 at 6:56 AM Andrew Randrianasulu
>> <[email protected]> wrote:
>> >
>> > On Sat, Dec 6, 2025 at 5:57 AM Andrew Randrianasulu
>> > <[email protected]> wrote:
>> > >
>> > > On Sat, Dec 6, 2025 at 4:35 AM Andrew Randrianasulu
>> > > <[email protected]> wrote:
>> > > >
>> > > > On Fri, Dec 5, 2025 at 4:06 PM Valery Ushakov <[email protected]>
>> wrote:
>> > > > >
>> > > > > On Fri, Dec 05, 2025 at 11:35:23 +0300, Andrew Randrianasulu
>> wrote:
>> > > > >
>> > > > > > So, I am still trying to build things from slightly older
>> pkgsrc tree,
>> > > > > > and Cmake 3.31.7 fails witch attached gzipped log, ending with
>> illegal
>> > > > > > instruction.
>> > > > > >
>> > > > > > Qemu line:
>> > > > > > qemu-system-sparc -hda sparc-netbsd-15g-compressed.qcow2 -m
>> 1536 -g
>> > > > > > 1024x768x24 -M SS-20 -smp 3 -accel tcg,thread=multi,tb-size=256
>> > > > > >
>> > > > > > qemu --version:
>> > > > > > qemu-system-sparc --version
>> > > > > > QEMU emulator version 9.1.50 (v9.1.0-1687-gf0cfd06786-dirty)
>> > > > > > Copyright (c) 2003-2024 Fabrice Bellard and the QEMU Project
>> developers
>> > > > > >
>> > > > > > Note that this configuration survived gnutls , libbluray,
>> > > > > > libtheora-1.2.0  builds
>> > > > > > so generally it works ?
>> > > >
>> > > > even installing pre-build pkg by Vom513 still gives illegal
>> instruction
>> > > >
>> > > > https://mastodon.online/@Andrew_R/115670749252048951
>> > > >
>> > > > _atomic_fetch_add_4 in /usr/pkg/lib/libatomic.so.3 ..
>> >
>> >
>> > 0xede93374 in __atomic_fetch_add_4 () from /usr/pkg/lib/libatomic.so.3
>> > (gdb) disassemble /s __atomic_fetch_add_4
>> > Dump of assembler code for function __atomic_fetch_add_4:
>> >    0xede9335c <+0>:       save  %sp, -96, %sp
>> >    0xede93360 <+4>:       mov  %i0, %g2
>> >    0xede93364 <+8>:       ld  [ %i0 ], %g1
>> >    0xede93368 <+12>:      unknown
>> >    0xede9336c <+16>:      mov  %g1, %i0
>> >    0xede93370 <+20>:      add  %g1, %i1, %g3
>> > => 0xede93374 <+24>:      casa  [ %g2 ] #ASI_P, %g1, %g3
>> >    0xede93378 <+28>:      unknown
>> >    0xede9337c <+32>:      cmp  %g3, %i0
>> >    0xede93380 <+36>:      unknown
>> >    0xede93384 <+40>:      mov  %g3, %g1
>> >    0xede93388 <+44>:      rett  %i7 + 8
>> >    0xede9338c <+48>:      nop
>> > End of assembler dump.
>>
>> ow, it was /sr
>>
>> (gdb) disassemble /sr  __atomic_fetch_add_4
>> Dump of assembler code for function __atomic_fetch_add_4:
>>    0xede9335c <+0>:       9d e3 bf a0     save  %sp, -96, %sp
>>    0xede93360 <+4>:       84 10 00 18     mov  %i0, %g2
>>    0xede93364 <+8>:       c2 06 00 00     ld  [ %i0 ], %g1
>>    0xede93368 <+12>:      81 43 e0 0f     unknown
>>    0xede9336c <+16>:      b0 10 00 01     mov  %g1, %i0
>>    0xede93370 <+20>:      86 00 40 19     add  %g1, %i1, %g3
>> => 0xede93374 <+24>:      c7 e0 90 01     casa  [ %g2 ] #ASI_P, %g1, %g3
>>    0xede93378 <+28>:      81 43 e0 0f     unknown
>>    0xede9337c <+32>:      80 a0 c0 18     cmp  %g3, %i0
>>    0xede93380 <+36>:      12 47 ff fa     unknown
>>    0xede93384 <+40>:      82 10 00 03     mov  %g3, %g1
>>    0xede93388 <+44>:      81 cf e0 08     rett  %i7 + 8
>>    0xede9338c <+48>:      01 00 00 00     nop
>> End of assembler dump.
>>
>> still hm.
>>
>
> https://gitlab.com/qemu-project/qemu/-/issues/1771
>
> but this affects qemu-sparc userspace emulator? (so Linux-on-linux)
>



strange, there was talk on Linux/glibc side of things about emulating those
CAS (compare and swap) instructions, but not sure if 32bit part was ever
written ... (considering Linux aggressively remove "obsoleted" arches.

https://sourceware.org/legacy-ml/libc-alpha/2016-11/msg00243.html

Not sure if NetBSD does something special on real hardware ... Because
there evidently it works.

probably 32bit host for qemu introduces this somehow, but I yet to verify
this ....


> Now, time to retest with -cpu LEON3 ...
>
>
>
>
>> Note, that this is on Generic.MP kernel from 10.1
>>
>> >
>> > Hm ...
>> >
>> > attempt at building libatomic inside this VM also failed (even without
>> > any -smp  or -cpu params)
>> >
>> >
>> > > >
>> > > > even with cpu set to
>> > > >
>> > > > qemu-system-sparc -hda sparc-netbsd-15g-compressed.qcow2 -m 1536 -g
>> > > > 1024x768x24 -M SS-20 -smp 3 -accel tcg,thread=multi,tb-size=256 -cpu
>> > > > TI-SuperSparc-60
>> > >
>> > > Strange, bug still exist on
>> > >
>> > > ./qemu/build/qemu-system-sparc --version
>> > > QEMU emulator version 10.1.92 (v10.2.0-rc2-31-g2257f52a97-dirty)
>> > > Copyright (c) 2003-2025 Fabrice Bellard and the QEMU Project
>> developers
>> > >
>> > > compiled for i586 on Slackware 15.0
>> > >
>> > > Next try: put VM to real netBSD amd64 partition, try qemu from there
>> > >
>> > >
>> > > >
>> > > >
>> > > > >
>> > > > > I had varying success with qemu sparc.  Incidentally, I was
>> trying to
>> > > > > build cmake too (that I needed to build some other project, not in
>> > > > > pkgsrc).
>> > > > >
>> > > > > cmake and/or modern c++ compilers are hogs... so I'm not surprised
>> > > > > that particular combination (builing cmake on qemu sparc) might
>> have
>> > > > > reached some darker corners where qemu is not that stable.
>> > > > >
>> > > > > That was an year ago or so, so my memory is hazy.  But IIRC, in
>> the
>> > > > > end it was faster to loan an ultra1 from a friend setup it and
>> build
>> > > > > cmake there.
>> > > > >
>> > > > > PS: My first attempt to build cmake was natively on a (diskless)
>> > > > > javastataion.  That gave me a new reference for "exercise in
>> futility"
>> > > > > (I think I gave up after couple of weeks or so).  That's when I
>> turned
>> > > > > to qemu.  For scale: I also needed to build cmake on macos/ppc
>> (for
>> > > > > the same project).  My mini g4 (1.25GHz cpu with 1G ram) took
>> about a
>> > > > > week to build cmake from macports from scratch.  That time was
>> mostly
>> > > > > spent builidng gcc10 and gcc14, though.  cmake proper was about a
>> day
>> > > > > only, I think.
>> > > > >
>> > > > > -uwe
>>
>