Re: www/firefox on earmv6hf

[Rin Okuyama <rokuyama.rk%gmail.com@localhost>]

Hi,

I am who added this assertion to libunwind.

As described in the comment, numbering schemes for VFP
registers are different for GCC and LLVM :(

I assumed mixing of two schemes cannot happen even if
objects compiled by the two compilers are linked...

It would be nice if you can obtain core, backtrace, and
crashed binary itself.

Thanks,
rin

On 2025/05/09 4:00, John Klos wrote:
> Hi,
>
> I'm trying to figure out why www/firefox fails on earmv6hf. However, I 
> have no idea why we'd get this assertion:
>
>      if (num >= REGNO_ARM32_S0 && num <= REGNO_ARM32_S31) {
>        /*
>         * XXX
>         * There are two numbering schemes for VFPv2 registers: s0-s31
>         * (used by GCC) and d0-d15 (used by LLVM). We won't support both
>         * schemes simultaneously in a same frame.
>         */
>        assert((flags & FLAGS_EXTENDED_VFPV2_REGNO) == 0);
>
> from:
>
>     Compiling jog v0.1.0 (/usr/pkgsrc/www/firefox/work/firefox-137.0.2/ 
> toolkit/components/glean/bindings/jog)
> assertion "(flags & FLAGS_EXTENDED_VFPV2_REGNO) == 0" failed: file "/ 
> usr/src/sys/lib/libunwind/Registers.hpp", line 448, function "void 
> _Unwind::Registers_arm32::copyFloatVectorRegister(int, uint64_t)"
> gmake[3]: *** [/usr/pkgsrc/www/firefox/work/firefox-137.0.2/config/ 
> makefiles/rust.mk:526: force-cargo-library-build] Abort trap (core dumped)
> gmake[3]: Leaving directory '/usr/pkgsrc/www/firefox/work/build/toolkit/ 
> library/rust'
>
> Does anyone have any suggestions?
>
> Thanks!
> John